the impact of depression on outcomes ......iv acknowledgements when embarking on a phd dissertation,...

THE IMPACT OF DEPRESSION ON OUTCOMES FOLLOWING

ACUTE MYOCARDIAL INFARCTION.

by

Paul Andrew Kurdyak

A thesis submitted in conformity with the requirements for the degree of

Doctor of Philosophy in Clinical Epidemiology

Graduate Department of Health Policy, Management, and Evaluation in the

University of Toronto

© Copyright by Paul Andrew Kurdyak 2009

ii

ABSTRACT

Paul Andrew Kurdyak The Impact of Depression on Outcomes Following Acute Myocardial Infarction Doctor of Philosophy, 2009 Department of Health Policy, Management, and Evaluation (HPME) University of Toronto This thesis uses observational study design methods to explore the relationship

between depression and various outcomes following acute myocardial infarction (AMI).

There are three main studies. First, the relationship between depression and mortality

following AMI was measured. The main finding was that the factor determining the

increased mortality rate in depressed patients is reduced cardiac functional status. The

main implication was that efforts to address increased mortality in depressed patients

with cardiovascular illnesses should focus on processes that impact cardiac functional

status. Second, the impact of depression on service consumption following AMI was

examined. Depressive symptoms were associated with a 24% (Adjusted RR:1.24; 95%

CI:1.19-1.30, P<0.001), 9% (Adjusted RR:1.09; 95% CI:1.02-1.16, P=0.007) and 43%

(Adjusted RR: 1.43; 95% CI:1.34-1.52, P<0.001) increase in total, cardiac, and non-

cardiac hospitalization days post-AMI respectively, after adjusting for baseline patient

and hospital characteristics. Depressive-associated increases in cardiac health service

consumption were significantly more pronounced among patients of lower than higher

cardiac risk severity. The disproportionately higher cardiac health service consumption

among lower-risk AMI depressive patients may suggest that health seeking behaviors are

mediated by psychosocial factors more so than by objective measures of cardiovascular

risk or necessity. Third, methodological issues related to missing data were explored. A

systematic review of three psychiatric journals revealed that a small minority of studies

iii

(5.8%) addressed the impact of missing data in a meaningful way. An example using real

data demonstrated the potential bias introduced by missing data and different ways to

address this bias. The paper concludes with recommendations for both reporting and

analyzing studies with substantial amounts of missing data.

Overall, the studies add to the literature exploring the relationship between

depression and outcomes following acute myocardial infarction. Future studies measuring

the relationship between depression and mortality will need to factor the mediating

relationship between depression and cardiac functional status. The increased health

service utilization associated with depression will need to be replicated in other illness

models. Together, the studies add to the existing conceptual framework for measuring

relationships between depression and outcomes in patients with cardiovascular illnesses.

iv

Acknowledgements

When embarking on a PhD dissertation, and during the process, I often felt that I was on a solo journey. At the end of the road, I can now, with great insight, state that this feeling was entirely delusional. There are a great many people who have been directly and indirectly along for the journey and from whom I have received valuable input. First and foremost, my supervisor and committee members. Dr. Paula Goering has been exceptionally kind and supportive – she has been very generous and her generosity of spirit was felt well before I initiated my PhD. Dr. William Gnam has always been willing to provide advice related to both research and career. He is most responsible for steering me towards a PhD in the first place. Finally, Dr. David Alter has been an unbelievably gracious mentor and colleague. I greatly admire his intellectual curiosity and creativity and am extremely thankful for his willingness to allow me access to his amazing SESAMI database. I have a great many colleagues at CAMH who have been very accommodating while I have worked on my thesis. In particular, I am very appreciative of Dr. Peter Voore and Linda Mohri; they have always been supportive of my need for time to devote to my PhD. I have been fortunate to receive financial support from a number of sources during my PhD. The Canadian Institutes for Health Research (CIHR), the Ontario Mental Health Foundation, the CAMH Foundation, and the Canadian Psychiatric Research Foundation have all provided financial support that has allowed me to protect time to devote to research. The Clinical Epidemiology Program in Health Policy, Management, and Evaluation has been very helpful. Drs. Gillian Hawker and Ahmed Bayoumi have both been very capable Program Directors during my tenure as a PhD student. The administrative staff in Clinical Epidemiology have provided much needed assistance over the years. There are a number of colleagues who have been available to hear me out over the years. One longstanding friend, Damon Scales, deserves particular mention as he has been such a steady and friendly presence. Rob Quinn and Ghirish Kulkarni provided much needed levity for the hours spent in the basement at ICES. Drs. David Juurlink and Muhammad Mamdani have also been wonderful colleagues and mentors at ICES. All previously mentioned individuals have provided instrumental support. However, the individuals I most value and need to acknowledge most earnestly are my family. I have had the good fortune of watching a toddler (Robert) and a baby (Laura) grow into beautiful children during my PhD. I have also benefited hugely from the love and support of Patricia, my wonderful wife and partner. More than anything else, my family provides me with much-needed perspective and grounding. They always have been and always will be the most important thing to me.

v

Table of Contents

ABSTRACT....................................................................................................................... ii

Acknowledgements .......................................................................................................... iv

Chapter 1: Introduction to Depression and AMI .......................................................... 1

Depression and Medical Comorbidity ................................................................................ 1

Measurement of depression with comorbid medical illness ............................................... 2

Depression and health care consumption............................................................................ 3

Depression Following Acute Myocardial Infarction .......................................................... 3

Depression and Health Service Consumption Following AMI .......................................... 5

Depression and AMI – a conceptual framework ................................................................ 6

Data Sources ....................................................................................................................... 7

The impact of missing data ................................................................................................. 8

Thesis Questions ............................................................................................................... 10

Chapter 2: Depression, mortality, and cardiac functional status following AMI..... 12

ABSTRACT...................................................................................................................... 12

Introduction....................................................................................................................... 14

Methods............................................................................................................................. 15

Results............................................................................................................................... 21

Discussion......................................................................................................................... 24

Chapter 3: Depressive symptoms, health service use, and prognosis following AMI

........................................................................................................................................... 38

Abstract ............................................................................................................................. 38

vi

Introduction....................................................................................................................... 40

Methods............................................................................................................................. 41

Results............................................................................................................................... 48

Discussion......................................................................................................................... 52

Chapter 4: Missing data: scope and impact on outcomes. .......................................... 67

Abstract ............................................................................................................................. 67

Introduction....................................................................................................................... 68

Missing Data – The Problem ............................................................................................ 69

Review of the Literature ................................................................................................... 75

A Demonstration Using Real Data ................................................................................... 77

Discussion......................................................................................................................... 80

Chapter 5: Conclusions .................................................................................................. 88

Summary of Research ....................................................................................................... 88

Limitations ........................................................................................................................ 95

Directions for Future Research ......................................................................................... 97

Implications....................................................................................................................... 99

Appendix 1: Overview of SESAMI Depression Measures ........................................ 103

References...................................................................................................................... 112

Table of Figures and Tables

Table 2.1: Comparison of baseline characteristics between depressed, non-depressed, and

non-respondents to one-month follow-up survey. ............................................................ 28

Table 2.2: Baseline characteristics across three depression severity categories: minimal,

moderate, and severe......................................................................................................... 29

Table 2.3: Multivariate predictors of mortality................................................................. 31

Table 2.4: Sensitivity analyses – multivariate predictors of mortality for three different

depression measures.......................................................................................................... 32

Figure 2.1: Self-reported cardiac functional status value in three depression severity

categories. ......................................................................................................................... 34

Figure 2.2: Proportion of patients in each depression category who died within 2 years of

AMI discharge. ................................................................................................................. 35

Figure 2.3: Mortality hazard ratio for severe relative to minimal depression category with

sequential adjustment........................................................................................................ 36

Figure 2.4: Mortality hazard ratio for severe relative to minimal depression category with

sequential adjustment (VO2max last).................................................................................. 37

Table 3.1: Baseline characteristics of depressed vs. non-depressed AMI patients........... 57

Table 3.2: Health service consumption in depressed and non-depressed post-AMI

patients. ............................................................................................................................. 59

Table 3.3: Multivariate service utilization rate for depression stratified across cardiac

illness severity................................................................................................................... 60

Table 3.4: Multivariate health service consumption rates for depression measure using

only 3 replacement items. ................................................................................................. 62

viii

Figure 3.1: The adjusted relative rate of service consumption attributable to depression.63

Figure 3.2: The adjusted relative rate of service consumption attributable to depression –

multiple imputation results. .............................................................................................. 65

Figure 4.1: Flow chart of literature review of psychiatric journals for missing data........ 85

Figure 4.2: Test for Missing at Random (MAR). ............................................................. 86

Figure 4.3: Mortality risk attributable to depression in three different analyses. ............. 87

Appendix Table 1: BCDRS* items (included and missing) and replacement items from

SESAMI survey. ............................................................................................................. 106

Appendix Table 2: Measures of agreement for SESAMI depression measures............. 108

Appendix Table 3: Baseline characteristics for three SESAMI depression measures.... 109

1

Chapter 1: Introduction to Depression and AMI

The purpose of this chapter is to:

1) Highlight the importance of depression after acute myocardial infarction (AMI)

2) Describe the role depression might play in health service consumption following

AMI

3) Describe issues related to missing depression data

4) Introduce the primary thesis questions

Depression and Medical Comorbidity

Comorbidity exists when any two medical or psychiatric conditions co-occur.

The prevalence of depression is substantially higher in populations with medical illnesses

than in populations without medical illnesses 1, 2. The high rate of depression in persons

with medical illnesses is understandable given the burden imposed by chronic medical

illnesses. However, the burden of depression is increasingly being considered substantial

in its own right, with disability related to depression projected to be second only to

ischemic heart disease in developed countries by the year 2020 3.

The ways that depression interacts with medical illnesses are complex. In the

Medical Outcomes Study 4, Wells et al. demonstrated that depressed patients perceived

their general health and social and vocational functioning as more impaired than patients

with one of seven other chronic medical conditions. Furthermore, the disability was

additive when depression was comorbid with other chronic medical conditions 4. These

findings have been replicated in a prospective cohort study of primary care patients 5.

2

Thus, depression is a chronic, disabling illness that is quite common and, when co-

morbid with chronic medical conditions, cumulatively disabling.

Measurement of depression with comorbid medical illness

A diagnosis of depression is made after certain criteria are met for a specific

period of time and if a level of distress or functional impairment has occurred6.

Measuring depression when it co-occurs with a comorbid medical condition is a

challenge. Given the symptomatic burden of chronic medical conditions, certain criteria

(e.g., reduced appetite, sleep disturbance, reduced energy) could be ascribed to either

depression or a medical condition. If a somatic symptom that is a diagnostic criterion for

major depressive disorder is due to a medical condition, there is the potential for a bias

towards increased depression prevalence because the depression threshold is reached due

to somatic symptoms potentially unrelated to depression.

Adding to this potential bias is the finding that persons with both depression and

chronic medical conditions tend to amplify physical symptoms related to a medical

condition. A recent systematic review suggests that patients with both depression and

medical conditions compared with patients with medical conditions alone reported

significantly higher numbers of medical symptoms even after adjusting for severity of

illness 7. This phenomenon would likely create problems both for generating a

prevalence rate for depression (given the reliance on reporting of physical symptoms to

make a depression diagnosis) as well as an accurate ascertainment of medical illness

severity where objective measures of severity do not exist.

3

Depression and health care consumption

Several studies have shown significantly higher health care costs for depressed

patients in primary care settings compared to non-depressed patients 8-10. The increased

service utilization persists after adjustment for chronic medical conditions 8, 9. The

increased costs also occur in any health utilization category measured, including primary

care visits, medical specialty visits, lab tests, pharmacy costs, inpatient medical costs, and

mental health visits 11. Other studies have shown that approximately 35% of “high

utilizers” of health care services in primary care settings have recurrent major depression

12. Hospitalizations are the most costly services provided in a health care system.

Depression, after controlling for medical illness severity, increases both length of stay

and likelihood for readmission 13, 14.

Depression Following Acute Myocardial Infarction

Depression is highly prevalent following AMI, occurring in approximately 20%

of AMI survivors 15, 16 compared with the approximately 3% prevalence in a community-

based sample 17. Furthermore, depression after AMI tends to persist when assessed up to

four months following AMI 16, 18.

Numerous studies have suggested that depression independently increases the risk

of mortality following AMI 16, 19-24. However, the relationship between depression and

mortality following AMI has not been consistent, with numerous other studies suggesting

no increased risk of mortality related to depression 15, 18, 25-29. The studies differ widely in

4

terms of the way depression is measured, patient population, timing of depression

measurement, and risk adjustment. Authors have cited the timing of depression

measurement following AMI25 and confounding by non-cardiac comorbidity, somatic

symptoms or cardiac illness severity23, 27 as potential reasons for discrepant results.

Depression measurement with comorbid medical illness has already been

described. Following AMI, patients are likely to experience some of the somatic

symptoms of depression whether depressed or not. Additionally, these symptoms are

more likely to occur immediately following an acute event like AMI than in the weeks

following. Many of the studies that have found a large positive association between

depression and mortality following AMI have measured depression within 7 days of

incident AMI22, 30-34. One study found that depression measured during hospitalization

for AMI predicted mortality at 4 months but was not a predictor of mortality after 8 years

of follow-up 29. There is a high likelihood that somatic symptoms could be erroneously

attributed to depression rather than natural phenomena immediately following AMI,

especially since depression rating scales are unable to make a distinction between the

potential causes of somatic symptoms. Additionally, a diagnosis of depression requires a

two week duration of symptoms6. Patients following AMI may experience immediate

distress following AMI that does not persist to become a depressive episode.

The variation in the relationship between depression and mortality following AMI

suggests that the relationship may be quite complex. Biological mechanisms that

increase risk of cardiac events may be more prevalent in depressed patients following

AMI such as increased platelet aggregation35-37 and decreased heart rate variability38-40.

Depression is twice as common in women compared to men and women have a higher

5

likelihood of mortality following AMI41. In addition, individuals with depression are

more likely to have cardiac risk factors such as smoking and diabetes and less likely to

adhere to recommendations and medications that can alter risk of mortality following

AMI 42-44. Furthermore, cardiac functional status, a measure of peak oxygen capacity,

has been shown to be a robust predictor of mortality following AMI 45, 46 and to be lower

in patients with depression 47. Thus, the relationship between depression and mortality

following AMI could be related to depression per se or to factors that are known to both

correlate with depression and to be associated with an increased risk of mortality

following AMI.

It is not possible to use rigorous methodologies such as randomized controlled

trials to study the relationship between depression and mortality following AMI because

depression does not occur randomly and is associated with a number of factors described

above that are also related to mortality following AMI. In other words, depressed and

non-depressed individuals differ in systematic ways, and many of these differences are

related to the likelihood to experience an AMI and to have differential survival following

AMI21. Ascertainment of any relationship between depression and mortality following

AMI requires careful observational study design, elucidation of potential mediating

factors, and risk adjustment for potential confounding factors.

Depression and Health Service Consumption Following AMI

Individuals with depression use 50% more health care services than individuals

without depression 48, 49. Such increases in depression-related health service consumption

are associated with high health care costs 50 and have prompted targeted interventions to

try to address the needs of depressed people while simultaneously reducing health service

6

consumption51. Similar high rates of health service consumption have been observed in

patients with cardiovascular illnesses generally 52, 53 and following AMI specifically54.

The reason for elevated rates of health service consumption in patients with

depression following AMI is unclear. On the one hand, the elevated health service

consumption might reflect greater clinical severity 21 as studies showing a significantly

positive association between depression and mortality following AMI would suggest 16, 18,

22-24. However, others have argued that the relationship between depression and health

service consumption may reflect depression-related physical symptom amplification and

health seeking behaviours 48. If the increased health service consumption reflects factors

other than increased illness severity, then the increased health service consumption is a

potential source of unnecessary health care costs. The novel contribution of this thesis is

to assess the relationship between depression, cardiovascular illness severity and health

service consumption following AMI.

Depression and AMI – a conceptual framework

The conceptual framework for this thesis is adapted from earlier investigators

who conducted a review of the literature on the relationship between depression and AMI

55 and is illustrated in Figure 1.1. This framework outlines a causal pathway for factors

that are involved in the relationship between depression and AMI. Cardiac risk factors

and patient demographics are important determinants of both AMI and the subsequent

development of depression. Once depression has occurred, cardiac illness severity and

functional status are important predictors of mortality following AMI. In addition,

receiving revascularization procedures and the presence of medical comorbidities

7

influence mortality following AMI. These three factors (cardiac risk factors, cardiac

functional status, and revascularization procedures) also influence health service

consumption following AMI. Perceived general health and, possibly, provider effects

may also influence the likelihood of depression-related health service variation following

AMI.

The relationship between depression and outcomes such as mortality or health

service consumption following AMI is complex. Investigating these relationships

requires data that describe and measure the various factors outlined in the conceptual

model (Figure 1.1).

Data Sources

The primary data source for this study is the Socioeconomic Status and Acute

Myocardial Infarction (SESAMI) cohort, a prospective cohort of AMI survivors. The

SESAMI study group is led by Dr. David Alter. The study involved survivors of AMI in

53 hospitals across Ontario from December 1, 1999 to February 28, 2004. Data were

collected during the index AMI hospitalization, from chart abstraction of index AMI

hospitalization, and from a follow-up phone survey administered one month post-

discharge from the index hospitalization. In addition, administrative health data were

linked to the data collected from the baseline and follow-up survey and chart abstraction

to permit prospective and retrospective follow-up for mortality and health service

consumption measures. The sample and data collection processes have been described in

detail elsewhere56.

The SESAMI study is an ideal source of data because 1) it included measures for

all the factors outlined in Figure 1.1; 2) when compared to previous studies assessing the

8

relationship between depression and AMI, the sample size of the SESAMI study is quite

large and adequately powered; and 3) unlike prior studies, the measures for medical

comorbidity, cardiac illness severity, and cardiac functional status have been validated,

allowing for robust assessment of the influence of these factors on the relationship

between depression and the two outcomes.

The depression measures were included in the one-month follow-up survey. Of

the 2829 patients with successfully abstracted index AMI admissions and administrative

data linkage, 888 subjects did not complete the one-month follow-up survey, for a survey

response rate of 69%. This response rate is typical for population-based survey data

collection.

While the response rate is typical, it is unlikely that the 31% of AMI patients who

did not respond to the follow-up survey can be ignored. Sicker patients may be less

likely to participate in surveys than less acutely ill patients. If non-response is biased

towards sicker patients, then mortality rates would be falsely lowered by non-response.

The impact of missing data

Research studies with large sample sizes involving psychiatric diagnoses are

prone to missing data because psychiatric diagnoses are generated by completion of

psychiatric rating scales administered in a survey format. Invariably, there is less than

100% participation rate in surveys. The missing data are unlikely to be missing

randomly; that is, there are likely systematic differences between respondents and non-

respondents. If respondents differ systematically compared to non-respondents and the

variables determining such systematic differences are correlated with the outcome

9

measures, then the missing data can confer bias to estimates generated from analyses in

which missing data are ignored.

Typically, data collected from surveys are analyzed by removing any subject with

missing values for any variable included in a multivariate model. This method of

analyzing data is called a “complete case analysis” 57. In the SESAMI study, 31% of

patients did not respond to the phone survey that had depression measures. Thus, the

depression status is unknown for 31% of the sample. Any complete case analysis

including depression status as a variable would automatically remove those 31% of

subjects with no depression status value. This is problematic for two reasons: 1)

removing these subjects may result in biased estimates from analyses; and 2) by

removing the subjects, there is no way to estimate the effect of bias on the results from

the analysis.

Studies assessing the relationship between depression and mortality following

AMI are similar to other psychiatric studies in that the depression status (or any mental

illness diagnoses) is generated from survey instruments that generate missing data. These

studies have routinely ignored the impact of missing data by conducting complete case

analyses when there is reason to believe ignoring such data may confer bias on estimates

generated by complete case analyses in this instance. The SESAMI study data are useful

for estimating the impact of bias due to missing data because of the richness of data that

were collected at baseline, through chart abstraction and from linked administrative

sources. In other words, there are a large number of variables with complete values for

individuals with missing depression values; these complete variables in subjects with

missing depression values can be used to analyze the impact of missing data.

10

Thesis Questions

1) What is the relationship between depression and mortality following AMI and what is

the relative contribution of the physical health status of depressed patients in the

relationship? SESAMI data, a registry of AMI patients followed prospectively via follow-

up phone survey and linked administrative health, will be analyzed. (Chapter 2).

2) What is the impact of depression on health service consumption following AMI? This

part of the thesis also uses SESAMI data to determine the magnitude by which depression

increases health service consumption following AMI, as well as the degree to which any

increased depression-related health service consumption reflects underlying cardiac

illness severity (Chapter 3).

3) What impact do missing data have on the analysis of depression-related outcomes

following AMI? In this chapter, missing data mechanisms are reviewed, as well as

different ways to analyze data with missing values. The three psychiatric journals with

the highest impact factors are reviewed over a two year period to determine how missing

data are handled currently. Finally, SESAMI data are re-analyzed using modern missing

data techniques to illustrate the impact of missing data on estimates generated from

different types of analyses (Chapter 4).

Figure 1.1: Conceptual framework for thesis.

Cardiac Risk Factors •Hypertension •Diabetes •Smoking •Elevated Serum Lipids

Demographics •Age •Sex •SES

AMI Depression

• Cardiac Functional Status

• Cardiac illness severity

• Medical comorbidity

Mortality

Health Service Consumption

Processes of Care

12

Chapter 2: Depression, mortality, and cardiac functional

status following AMI


1) Evaluate the relationship between depression and mortality following AMI using

SESAMI data

2) Determine, using sequential multivariate modeling, the contribution of specific

variables on the relationship between depression and mortality

3) Discuss the limitations of the SESAMI data for assessing the relationship between

depression and mortality following AMI

4) Describe the implications for the role cardiac functional status plays in the

relationship between depression and mortality following AMI

ABSTRACT

Background: The association between depression and mortality following acute

myocardial infarction (AMI) has been observed in a large number of studies. The cause

of increased post-AMI mortality associated with depression remains poorly elucidated.

The objective of this study was to examine the extent to which self-reported cardiac

functional status accounted for depression-mortality associations following AMI.

Methods and Results: Using a prospective cohort design, the authors obtained self-

reported measures of depression and developed profiles of the patients' prehospitalization

13

cardiac risks, comorbid conditions and drugs and revascularization procedures during or

following index AMI hospitalization. To create these profiles, the patients' self-reports

were retrospectively linked to no less than 12 years' worth of previous hospitalization

data. Mortality rates 2 years after acute MI were examined with and without sequential

risk adjustment for age, sex, income, cardiovascular risk, comorbid conditions, selected

process-of-care factors, and cardiac functional status measured as VO2peak. Depression

was strongly correlated with 2-year mortality rate (crude hazard ratio (HR) of severe vs.

minimal depression category, 2.48 [95% CI 1.20 - 5.15]; P=0.01). However after

sequential adjustment for age, sex, income and VO2peak, the effect of depression was

greatly attenuated (adjusted HR for severe vs. minimal depression category, 1.35 [95%

CI 0.63 – 2.87]; P=0.44). Cardiac risk factors and non-cardiac comorbidities had

negligible explanatory effect.

Conclusions: The main factor determining the increased mortality rate in depressed

patients is self-reported cardiac functional status. Efforts to address increased mortality

in depressed patients with cardiovascular illnesses should focus on processes that impact

cardiac functional status.

14

Introduction

For many decades and across several studies, depression has been associated with an

increase in mortality following acute myocardial infarction (AMI) 16, 19-24 (AMI).

However, there are a large number of studies that do not support a positive association

between depression and mortality following AMI 18, 26, 28. The reasons for discrepant

results across different studies are unknown, although there is substantial variation in

how different studies are conducted, including the use of different types of depression

rating instruments28, 55.

Studies measuring the association between depression and mortality following AMI rely

on self-report measures or questionnaires to derive a diagnosis of depression. Depression

is comprised of both psychological and physical health attributes 6. While the two

attributes are not necessarily mutually exclusive, one may hypothesize that the

depression-mortality association that has been observed following AMI is mediated more

by the self-reported impact on physical health status and functional capacity than on

mental health consequences. This is further corroborated by recent evidence that

depression interventions following AMI have no impact on survival 58. No study has

examined the extent to which the association between self-reported depression and

mortality is explained by self-reported health status. If the association between

depression and mortality is explained by a second, independent self-reported functional

status measure, then future interventions should concentrate on improvements in physical

health status more so than on mental health to have an impact on survival following AMI.

15

Accordingly, the objective of this study was to examine the extent to which self-reported

cardiac functional status accounted for depression-mortality associations following AMI.

The Duke Activity Status Index (DASI) served as an ideal instrument for this objective

because it has been among the most widely validated self-reported measures of cardiac

functional status 59-65, and is a predictor of mortality 66. Moreoever, there are no items in

the DASI that directly reflect mood, anxiety, or any other psychosocial factors that

comprise the psychological attributes of depression.

Methods

Data Source and Study Sample

This study is a sub-study of the Socio-Economic and Acute Myocardial Infarction Study

(SESAMI) study, a prospective observational study of patients who were hospitalized for

AMI throughout Ontario, Canada67. Data came from four sources: 1) a baseline survey;

2) chart abstraction from index AMI; 3) a one-month follow-up phone survey; and 4)

linked health administrative data. From 2829 patients with successfully abstracted index

AMI admissions and linked health administrative data, 888 subjects did not complete the

one-month follow-up survey either because of death prior to the survey (N=73) or a

refusal to participate (N=815). A total of 1941 patients remained available for analysis.

The SESAMI cohort has been described in detail elsewhere67. The Sunnybrook Health

Sciences Centre Review Committee approved this study and all subjects gave informed

consent to participate.

16

Depression Measures

The SESAMI survey consisted of psychometric questions that explored various domains

related to depression, including low mood, loss of interest, sleep disturbance, reduced

appetite/wt loss, agitation/slowing, low self-esteem/guilt, suicidal thoughts, reduced

energy level, loss of concentration. The depression measures used in the SESAMI cohort

study are described in Appendix 1. The depression measure used in this study is a scale

containing 9 of the 12 items from the Brief Carroll Depression Rating Scale (BCDRS) 68.

The surveys were telephone administered by standardized trained health care personnel

(nurses). For the purposes of this study, we were interested in determining whether the

association between depression and mortality followed a gradient based on severity of

depression. As such, the study sample was stratified across three depression categories:

minimal depression symptoms (BCDRS score of 0 to 3), moderate depression symptoms

(BCDRS score of 4 to 6), and severe depression symptoms (BCDRS score of 7 to 9).

These categories were defined prior to data analysis.

Self-reported Cardiac Functional Status

The DASI is a self-report measure that includes items that measure exertion required for

various physical activities 63. It has been validated in a number of different patient

populations59-62, 64, 65. However, it has not been validated in depressed cardiac patients.

17

The DASI is weakly correlated with the 9-item depression scale (r = 0.30). Furthermore,

the DASI and the 9-item depression scale categories were less correlated in the full,

multivariate model (severe category, r = 0.25; modest category, r = 0.26), suggesting that

collinearity was not an issue between the two variables.

Demographic Factors

Demographic factors such as age, sex, and income tertile were acquired from the baseline

survey.

Cardiovascular risk factors and prognostic index (Global Registry of Acute

Coronary Events index)

Cardiovascular risk factors that included diabetes, hypertension, hyperlipidemia, and

current or former cigarette use. Smoking history was ascertained by questionnaire. Major

cardiovascular risk factors were determined by reviewing diagnostic fields in

computerized hospital discharge abstracts from April 1, 1988 to the index hospitalization

date using all primary and secondary ICD-9 and ICD-10 (where applicable) discharge

codes. Agreement between self-reported risk factors and chart audit ranged baseline 73%

(hyperlipidemia) and 95% (diabetes) 67.

Our measure of cardiovascular risk-severity was the Global Registry of Acute Coronary

Events (GRACE) prognostic index. The GRACE prognostic index provides a score that

reflects the probability of dying within 6 months post-AMI based on age, development

(or history) of heart failure, peripheral vascular disease, systolic blood pressure, Killip

class, initial serum creatinine concentration, elevated initial cardiac markers, cardiac

18

arrest on admission, and ST segment deviation69. The GRACE prognostic index has been

validated among SESAMI patients70.

Noncardiovascular Comorbid Conditions

Noncardiovascular risk factors consisted of all co-morbid diseases that were captured

through primary and secondary diagnostic fields of hospital discharge abstracts (Canadian

Institute for Health Information) from 1 April 1988 to the presenting hospitalization. Only

pre-existing conditions were included. Non-cardiovascular conditions were categorized as

cancer and as diseases of the central nervous system, endocrine system, hematology

system, musculoskeletal system, respiratory system, gastrointestinal system, and

genitourinary system. We categorized diabetes, secondary hypertension, and

hyperlipidemia as cardiovascular risk factors, not as diseases of the endocrine system71.

Processes of care

All of our outcomes could be influenced by whether or not a patient has received

revascularization by angioplasty or coronary bypass surgery. All revascularization

procedures occurring within 30 days of discharge from the index AMI hospitalization

were recorded within the Ontario Health Insurance Program database. In addition, we

examined the influence of cardiovascular medication (Beta-blockers, statins, ACE

inhibitors, ASA, and nitrates) prescribed at discharge from index admission as acquired

from chart abstraction.

19

Outcome

Our primary outcome was two-year all-cause mortality. Mortality data were obtained

from both hospital discharge abstracts and the Ontario Registered Persons Data Base.

Statistical Analysis

Differences in baseline characteristics across the three depression categories were

compared using the Mantel-Haenszel test for trend for categorical data, and analysis of

variance for continuous data. Cox proportional hazard models estimated hazard ratios of

both mortality and recurrent AMI. We examined depression as a 3-level ordinal variable

in which we compared the severe and moderate depression categories to the minimal

depression category (reference category). Changes in the depression-mortality hazard

ratio were examined after sequential adjustment for 1) demographic factors (age, sex,

income), 2) cardiac functional status (Duke Activity Status Index; DASI), 3) cardiac

illness severity (Global Registry of Acute Coronary Events (GRACE) score) and cardiac

risk factors (smoking history, hyperlipidemia, hypertension, and diabetes), 4) non-cardiac

medical comorbidities, and 6) drugs at discharge and revascularization procedures

(coronary artery bypass graft (CABG) surgery and percutaneous trans-thoracic coronary

angiography (PTCA)) occurring after the index AMI. CABG and PTCA were included

in the regression models as time-dependent variables to account for the fact that

likelihood of surviving would change after receiving revascularization procedures. For

the Cox proportional hazard models, subjects were censored at death. Violations of the

proportionality assumption in all proportional hazard model specifications were tested.

20

All analyses were performed using SAS statistical software, version 9.1 (SAS Institute,

Cary, NC).

Sensitivity Analyses

We repeated the sequentially adjusted multivariate analyses of the relationship between

depression severity categories with two other depression scales. One is an imputed 12-

item scale with items taken from the BCDS (Appendix 1). The other is the Duke

Psychological Well-being scale, a rating scale developed for the Global Utilization of

Streptokinase and Tissue Plasminogen Activator for Occluded Coronary Arteries

(GUSTO) trial72. The 12-item depression rating scale was categorized into minimal (0-

4), moderate (5-8) and severe (9-12) depression categories. The Duke Psychological

Well-being scale was similarly categorized into minimal (31-40), moderate (21-30) and

severe (10-20) depression categories. We also used multiple imputations to impute

depression measures for the 888 missing depression values due to non-response57 to test

whether our outcomes were affected by systematic differences between survey

responders and non-responders.

The mortality multivariate analyses were repeated using recurrent AMI as an outcome.

Recurrent AMI was defined from administrative health data using the most-responsible

diagnosis field of ICD-9 and ICD-10(AMI: ICD-9 410, 412, 4141; ICD-10 I21, I22, and

I23) 73. We used traditional Cox proportional hazard survival analysis and competing

21

risk survival analysis. We also estimated hazard ratios attributable to depression for both

recurrent AMI and mortality using competing risk survival analysis. The advantage of

competing risk survival analysis is that it does not assume that recurrent AMI is

independent of mortality. It permits informative censoring to account for the fact that if a

subject dies before having a recurrent AMI, the probability of having a recurrent AMI is

zero. Cox proportional hazard survival analysis uses non-informative censoring in such

circumstances.

Results

Baseline Characteristics

The 888 patients who did not respond to the one month follow-up survey were similar to

the 1941 subjects in terms of gender, cardiac risk factors, likelihood of receiving

revascularization procedures within 30 days of discharge, and the likelihood of receiving

ACE inhibitors, Beta blockers, and nitrates upon discharge. However, the sample of 888

survey non-respondents were older (mean (SD) age of respondents 66.3 (13.7) years vs.

62.4 (12.8) years; P<0.001), more likely to have 3 or more non-cardiac medical

comorbidities (55.7% vs. 47.2%; P < 0.001), less likely to receive statins upon discharge

(48.7% vs. 54.9%; P = 0.002), and had a higher GRACE score (predictive of 6-month

mortality)(4.1 vs. 3.2; P < 0.001) than the 1941 survey respondents who consented and

completed the 1 month post-myocardial infarction evaluation (Table 2.1).

22

Among the 1941 patients with complete data, patients with increasing number of

depression symptoms were younger (P = 0.02), more likely to be women (P<0.001), and

had lower household incomes (P<0.001) (Table 2.2). A higher proportion of the most

severely depressed patients also had more cardiac risk factors (diabetes – P = 0.004;

hypercholesterolemia – P = 0.04; and smoking history – P = 0.03) as well as a higher

number of non-cardiac medical comorbidities (P<0.001)(Table 2.2). Patients in the upper

two categories of depression symptoms were significantly more likely (P = 0.01) to have

been prescribed nitrates on discharge from index AMI hospitalization (Table 2.2).

Depression and cardiac functional status

The mean (SD) DASI score in our sample was 17.4 (5.1) , which correlates with a

VO2peak of 17.1 ml/kg/min, or a 23.6 minute mile. As the severity of depression

increased, the DASI score decreased, with the mean (SD) DASI score in the most

severely depressed category of 7.49 (6.25) as compared to 20.1 (11.8) among non-

depressed individuals (P<0.001)(Figure 2.1).

The Impact of Sequential Risk Adjustment on Depression-Mortality Associations

Crude two-year mortality rates were inversely correlated with depression severity (14.0%

(severe) vs. 6.4 (moderate) and 5.8 (no depression)), although the difference did not

achieve statistical significance (P = 0.07)(Figure 2.2). Relative to the minimal depression

symptom category, the most severe depression category was highly predictive of

increased 2-year mortality rates after adjusting for age, sex, and socioeconomic status

23

(HR = 2.91 [95% CI 1.40-6.07]; P<0.001) (Fig. 2.3). Additional sequential adjustments

for cardiac risk factors, prognostic severity (GRACE index), and non-cardiac medical

comorbidities, attenuated the relationship between severe depression and mortality;

however, the association remained statistically significant (HR = 2.12 [95% CI 0.99 –

4.52]; P = 0.05) (Fig. 2.3). After adjusting for self-reported cardiac functional status (the

DASI), the magnitude of association between depression and mortality was attenuated by

an additional 40% and was no longer statistically significant (HR = 1.37 [95% CI 0.63 –

2.97]; P = 0.43)(Figure 2.3). Moreover, the explanatory impact of self-reported cardiac

functional capacity was not a function of ordering. For example, when re-ordering the

sequential adjustment so that the DASI was included prior to cardiac risk factors,

prognostic severity (GRACE index), and non-cardiac medical comorbidities, the

relationship between the most severe depression category and 2-year mortality rates was

non-significant (HR – 1.35 [95% CI 0.63 – 2.97]; P=0.44)(Figure 2.4). The results of the

full model, including parameter estimates for all covariates included in the model, are

shown in Table 2.3.


Several sensitivity analyses were conducted. A reanalysis using two alternative

depression measurements did not meaningfully alter our results (Table 2.4). The impact

of cardiac functional status on the relationship between depression and recurrent AMI

was similar to the impact on the relationship between depression and mortality.

Repeating the analysis with recurrent AMI as an outcome using competing risk analysis

also did not substantially change the outcome (adjusted mortality AMI HR in the most

24

severe depression category relative to the minimal depression category – 1.09 [95% CI

0.65-1.59]; P = 0.94). Finally, multiple imputation to impute missing values for

depression for non-respondents did not alter the hazard ratio for mortality in the most

severe depression category relative to the minimal depression category (HR – 1.00 [95%

CI 0.47-2.11]; P = 0.99).

Discussion

In this study, the association between depression and mortality following AMI was

explained predominantly by self-reported cardiac functional status. Non-cardiovascular

medical conditions, pre-existing cardiac risk factors, and prognostic illness severity all

attenuated the relationship between depression symptoms and 2-year mortality, but

exerted only modest explanatory effects when compared to the importance of cardiac

functional status.

While there have been many hypotheses proposed, the actual explanatory mechanisms

linking depression with cardiovascular prognosis remain unknown. To our knowledge,

ours is the first study to have demonstrated the importance of self-reported cardiac

functional status as a key mediator for depression’s association with mortality among

AMI populations. Indeed, the importance of functional capacity as a determinant of

survival is not new. For example, Myers et al demonstrated a dose-response relationship

between exercise fitness and survival among 6213 VA patients74. Functional capacity has

25

also shown to be among the strongest predictors of long-term survival among women 75.

Accordingly, its role in mediating depression-associated outcomes is not surprising.

In contrast, cardiac risk factors, infarction risk-severity using the GRACE index, a

validated prognostic risk score, and non-cardiovascular comorbidity exerted only modest

explanatory effects on depression-mortality associations following AMI. While

adjustment for cardiac risk-factors and GRACE did attenuate the relationship between

severe depression and mortality by 28% over that explained by sociodemographic factors

alone, the effect was markedly smaller than that observed after adjustment for the DASI.

Moreover, neither non-cardiovascular comorbidity, cardiac risk-factors, or the GRACE

risk-index explained the association between depression and mortality once the DASI

was included in multivariate models, suggesting that any variation in mortality risk

attributable to cardiac risk factors, infarct risk-severity, and non-cardiovascular comorbid

illnesses could in themselves be mediated through cardiac functional status as a final

common pathway.

Our findings have important population health implications. Available evidence suggests

that the efficacy associated with psychosocial behavioral interventions in post-AMI

populations has been mixed15, 76, with a more recent meta-analysis indicating no impact

on survival58. However, most studies that have examined the effects of psychosocial

behavioral interventions have not routinely incorporated exercise components within the

intervention. Indeed, no study has specifically evaluated the survival impact of exercise-

interventions among depressed AMI populations. We hypothesize that improving

26

exercise capacity among patients with post-AMI depression, particularly among those

with moderate baseline functional capacity, may improve survival. Such an intervention

might be particularly effective since depressed individuals are less likely to participate in

cardiac rehabilitation programs42, 77. Therapeutic life-style and disease-management

interventions, which include exercise as one component, are associated with decreases in

avoidable hospitalizations among populations with depression78. Accordingly, the

inclusion of exercise into self-care and/or psychosocial behavioral interventions is logical

and preliminary evidence would indicate that it would be successful 79. However, beyond

traditional on-site cardiac rehabilitation, no such programs in Canada exist.

Our study has several limitations. First, the DASI is a self-reported measure of cardiac

functional status63. While it has been validated in several different clinical populations,

it has not been validated in depressed patients. Accordingly, it is not possible to postulate

with certainty that DASI scores in depressed patients are a valid reflection of cardiac

functional status. However, there is no compelling explanation besides reduced cardiac

functional status that could explain the explanatory impact of the DASI score in

mediating the relationship between depression and mortality following AMI. 63, 64.

Second, 31% of the sample did not respond to the one-month follow-up survey.

Nonetheless, sensitivity analyses using multiple imputation to impute depression values

for non-respondents did not meaningfully change our results. Third, depression was

measured at one-month following index AMI. This may confer a survival bias to the

sample. However, the depression-mortality gradients observed in our study are consistent

with what others have reported elsewhere 22. Fourth, this study assumes that self-

27

reported cardiac functional status is a confounding variable in relation to depression. In

reality, while our outcomes were gathered prospectively, our predictor variable data were

gathered cross-sectionally; therefore, it is not possible to make claims about temporal

causality for the relationship between the DASI and depression. However, the

relationship between depression and reduced cardiac functional status has merit from a

theoretical perspective given the loading of cardiac risk factors and inability to adhere to

lifestyle and treatment recommendations observed in depressed patients, all of which

would lead to reduced cardiac functional status. Finally, while our study did adjust for

various processes of care, the inclusion of such factors were limited to treatment factors

during, or soon following the index AMI hospitalization. It is possible that our results

might have changed, had we adjusted for treatment factors longitudinally throughout the

entire follow-up duration.

In conclusion, the repeated observation of a depression-mortality gradient in patients

following AMI has led to speculations about possible causes for the gradient. Our

findings suggest that cardiac functional status may serve as a key causal pathway

determinant explaining the association between depression and mortality following AMI.

The role of cardiac functional status in mediating the relationship between depression and

mortality is novel and has not been demonstrated before. Efforts aimed at reducing

mortality outcomes among depressed AMI patients should focus on the evaluation of

interventions that improve cardiac functional status among such patient populations.

28

Table 2.1: Comparison of baseline characteristics between depressed, non-

depressed, and non-respondents to one-month follow-up survey.

Depressed Non-

Depressed Missing P Value

N (%) 494 (17.5) 1447 (51.1) 888 (31.4) Age (mean, SD) 61.4 (13.1) 63.5 (12.7) 66 (14.3) 0.01

Male Sex (%) 62.8 73.1 67.0 <0.001

Income, Can $ (%)

Low (<30,000) 32.5 23.7 39.8 <0.001

Intermediate (30,000-59,999) 34.4 34.5 31.3

High (>59,999) 33.1 60.0 24.6

Coronary Risk Factors (%)

Diabetes 29.0 21.6 26.8 <0.001

Hypercholesterolemia 44.7 38.9 39.8 0.07

Hypertension 49.0 46.4 49.5 0.3

Smoking History 42.9 38.5 41.4 0.15

Medical Comorbidities (%)

0 5.7 8.5 5.4 <0.001

1 16.4 22.7 17.6

2 26.5 23.1 21.3

3 or greater 51.4 45.8 55.7

Processes of Care (%)

PTCA 8.1 8.2 7.2 0.67

CABG 11.9 10.1 10.0 0.46

Drugs at Discharge (%)

Beta Blocker 68.4 70.3 66.7 0.19

ACE Inhibitor 63.2 62.0 60.5 0.60

Statin 54.7 55.0 48.7 0.009

Nitrate 36.6 30.7 33.3 0.04

ASA 70.2 73.7 67.6 0.09

GRACE Score (mean, SD) 112 (30.3) 114 (28.5) 119 (31.2) 0.07 Legend: Depressed status is based on the 9-item depression scale. Abbreviations: PTCA – percutaneous, transluminal coronary angiography; CABG – coronary artery bypass graft; GRACE – Global Registry of Acute Coronary Events

29

Table 2.2: Baseline characteristics across three depression severity categories: minimal, moderate, and severe.

Depression Symptom Categories

Variables 0-3 4 to 6 7 to 9 P Value N = 1134 n = 697 n = 110

Age -- % 19 to 49 14.6 19.9 19.1 0.001 50 to 64 35.7 34.4 40.0 65 to 74 30.0 23.7 16.4 > 74 19.8 22.0 24.6

Sex -- % male 73.2 68.2 56.4 0.001 Income, Canadian $ -- %

Low (<$30,000) 22.1 29.7 41.0 <0.001 Intermediate ($30,000 - $59,999) 35.2 33.3 34.0 High (>$59,999) 42.7 37.0 25.0

Coronary Risk Factors -- % Diabetes 20.6 27.0 30.9 0.001 Hypercholesterolemia 38.8 42.2 45.5 0.19 Hypertension 45.8 48.8 50.0 0.17 Smoking 37.8 42.0 42.7 0.07

Non-cardiac comorbidities -- % None 14.6 10.3 10.0 <0.001 1 31.5 28.6 22.7 2 28.0 28.4 25.5 3 or more 25.9 32.7 41.8

Non-cardiac comorbidities – types -- % Respiratory conditions 7.8 10.0 13.6 0.02 Cancer 2.3 2.2 0.9 0.46 Central Nervous System Conditions 0.5 0.9 -- 0.4

30

Genitourinary illnesses Conditions 6.4 7.2 6.4 0.64 Hematological Conditions 0.8 -- -- 0.32 Musculoskeletal Conditions 4.1 3.9 8.2 0.28 Endocrinological Conditions 2.0 2.0 0.9 0.60 GRACE Prognostic Index; mean (SD) 113.4 (28.3) 112.1 (30.1) 116.1 (31.7) 0.53 Revascularization Procedures -- %

PTCA 8.4 7.9 -- 0.31 CABG 9.7 12.4 12.3 0.08

Drugs at Discharge -- % Beta Blocker 69.9 70.2 63.2 0.63 Statin 54.8 55.1 56.1 0.84 ACE Inhibitors 62.2 63.2 56.1 0.87 Nitrates 30.3 36 36.8 0.01 ASA 74.5 69.2 70.2 0.02

Legend: Depression categories based on 12-item Brief Carroll Depression Scale scores (maximum score = 12). Abbreviations: GRACE – Global Registry of Acute Coronary Events; PTCA – percutaneous, transluminal coronary angiography; CABG – coronary artery bypass graft. Cells with a – instead of a value indicate that the cell size was less than 5 subjects.

31

Table 2.3: Multivariate predictors of mortality.

Hazard Ratio 95% CI P Value 12-Item Depression Category

2 vs. 1 0.73 0.45-1.16 0.18 3 vs. 1 1.38 0.63-3.03 0.42

Age 1.02 0.99-1.05 0.21 Sex 1.20 0.76-1.88 0.44 Income Category

2 vs. 1 1.11 0.70-1.76 0.64 3 vs. 1 0.98 0.58-1.66 0.93

Cardiac Risk Factors Diabetes 1.51 0.96-2.38 0.08 Hypercholesterolemia 0.44 0.27-0.72 0.001 Hypertension 0.77 0.47-1.26 0.29 Smoking 0.70 0.43-1.13 0.15

GRACE Prognostic Index 1.02 1.01-1.03 <0.001 DASI 0.94 0.91-0.97 <0.001

Non-cardiac comorbidities 1.80 1.30-2.51 <0.001 Procedures

PTCA 0.94 0.56-1.59 0.82 CABG 0.27 0.11-0.69 0.01

Drugs at Discharge Beta Blockers 1.25 0.76-2.04 0.37 ACE Inhibitors 1.01 0.64-1.59 0.98 Statins 0.91 0.59-1.42 0.69 ASA 0.83 0.51-1.36 0.46 Nitrates 1.08 0.70-1.67 0.73

Legend: Abbreviations: GRACE – Global Registry of Acute Coronary Events; PTCA – percutaneous, transluminal coronary angiography; CABG – coronary artery bypass graft.

32

Table 2.4: Sensitivity analyses – multivariate predictors of mortality for three different depression measures.

BCDS 9-Item Depression Scale Duke Depression Scale Hazard Ratio (95% CI) P Value Hazard Ratio (95% CI) P Value Hazard Ratio (95% CI) P Value Depression Category

2 vs. 1 0.76 (0.49-1.18) 0.23 0.93 (0.60-1.43) 0.73 1.11 (0.73-1.67) 0.63 3 vs. 1 1.33 (0.61-2.88) 0.48 1.15 (0.66-1.99) 0.62 0.81 (0.41-1.60) 0.54

Age 1.01 (0.98-1.04) 0.40 1.01 (0.99-1.04) 0.36 1.01 (0.98-1.04) 0.42 Sex 1.28 (0.83-1.96) 0.27 1.29 (0.84-2.00) 0.24 1.27 (0.82-1.94) 0.28 Income Category

2 vs. 1 1.11 (0.70-1.76) 0.64 1.14 (0.72-1.81) 0.58 1.11 (0.70-1.76) 0.66 3 vs. 1 0.98 (0.58-1.66) 0.93 0.97 (0.57-1.65) 0.91 0.97 (0.57-1.65) 0.91

Cardiac Risk Factors Diabetes 1.65 (1.08-2.54) 0.02 1.67 (1.09-2.56) 0.02 1.68 (1.10-2.58) 0.02 Hypercholesterolemia 0.46 (0.29-0.73) 0.001 0.47 (0.29-0.75) 0.001 0.48 (0.30-0.76) 0.002 Hypertension 0.86 (0.54-1.37) 0.53 0.88 (0.55-1.41) 0.59 0.87 (0.55-0.40) 0.57 Smoking 0.79 (0.50-1.25) 0.31 0.80 (0.50-1.27) 0.34 0.79 (0.50-1.26) 0.32 GRACE Prognostic Index 1.02 (1.01-1.03) <0.001 1.02 (1.01-1.03) <0.001 1.02 (1.01-1.03) <0.001 VO2 Peak 0.86 (0.80-0.92) <0.001 0.86 (0.80-0.93) <0.001 0.86 (0.80-0.92) <0.001

Non-cardiac comorbidities 1.62 (1.19-2.22) 0.002 1.60 (1.17-2.19) 0.003 1.59 (1.16-2.18) 0.003 Procedures

PTCA 0.66 (0.30-1.44) 0.29 0.67 (0.30-1.46) 0.31 0.65 (0.30-1.43) 0.28 CABG 0.32 (0.14-0.74) 0.01 0.32 (0.14-0.74) 0.008 0.31 (0.14-0.73) 0.007

Drugs at Discharge Beta Blockers 1.07 (0.69-1.67) 0.77 1.06 (0.68-1.66) 0.79 1.05 (0.67-1.65) 0.83 ACE Inhibitors 0.89 (0.58-1.36) 0.59 0.86 (0.56-1.31) 0.48 0.88 (0.57-1.34) 0.55 Statins 0.90 (0.59-1.37) 0.62 0.88 (0.58-1.34) 0.55 0.88 (0.58-1.34) 0.55 Nitrates 1.09 (0.72-1.65) 0.68 1.07 (0.71-1.62) 0.74 1.10 (0.72-1.66) 0.67

33

Legend: Abbreviations: BCDS – Brief Carroll Depression Scale; Abbreviations: GRACE – Global Registry of Acute Coronary Events; PTCA – percutaneous, transluminal coronary angiography; CABG – coronary artery bypass graft.

34

Figure 2.1: Self-reported cardiac functional status value in three depression severity categories.

DASI Across Depression Categories

0

10

20

30

40

Mild Moderate Severe

12-Item Depression Categories

DA

SI

Sco

re

35

Figure 2.2: Proportion of patients in each depression category who died within 2 years of AMI discharge.

0

2

4

6

8

10

12

14

0 to 3 4 to 6 7 to 9

9-Item Depression Scale Categories

Figure 2.3: Mortality hazard ratio for severe relative to minimal depression

category with sequential adjustment.

Legend: Hazard ratio reflects risk of mortality for most severe depression category relative to least severe. Abbreviation: GRACE – Global Registry of Acute Coronary Events.

37

Figure 2.4: Mortality hazard ratio for severe relative to minimal depression

category with sequential adjustment (VO2max last).

Legend: VO2max was included before cardiac risk factors, GRACE prognostic index, and other variables. Hazard ratio reflects risk of mortality for most severe depression category relative to least severe. Abbreviation: GRACE – Global Registry of Acute Coronary Events.

38

Chapter 3: Depressive symptoms, health service use,

and prognosis following AMI


1) Determine the relationship between depression and health service consumption

following AMI

2) Demonstrate that increased health service consumption following AMI related to

depression is not a function of illness severity

3) Determine health system implications and costs associated with depression-

related health service consumption following AMI

Abstract

Background: The use of cardiovascular health services is greater among patients with

depressive symptoms than among patients without. However, the extent to which such

associations between depressive symptoms and health service utilization are attributable

to variations in comorbidity and prognostic disease severity is unknown. This paper

explores the relationship between depressive symptoms, health service cardiovascular

consumption, and prognosis following acute myocardial infarction (AMI).

Methods: The study design was a prospective cohort study with follow-up telephone

interviews of 1,941 patients 30 days following AMI discharged from 53 hospitals across

Ontario, Canada between December 1999 and February, 2003. Outcome measures were

post-discharge use of cardiac and non-cardiac health care services. The service

utilization outcomes were adjusted for age, sex, income, comorbidity, two validated

39

measures of prognosis (cardiac functional status and risk adjustment severity index),

cardiac procedures (CABG or PTCA) and drugs prescribed at discharge.

Results: Depressive symptoms were associated with a 24% (Adjusted RR:1.24; 95%

CI:1.19-1.30, P<0.001), 9% (Adjusted RR:1.09; 95% CI:1.02-1.16, P=0.007) and 43%

(Adjusted RR: 1.43; 95% CI:1.34-1.52, P<0.001) increase in total, cardiac, and non-

cardiac hospitalization days post-AMI respectively, after adjusting for baseline patient

and hospital characteristics. Depressive-associated increases in cardiac health service

consumption were significantly more pronounced among patients with lower than higher

cardiac risk severity. Depressive symptoms were not associated with increased mortality

after adjusting for baseline patient characteristics.

Conclusions: Depressive symptoms are associated with significantly higher cardiac and

non-cardiac health service consumption following AMI despite adjustments for

comorbidity and prognostic severity. The disproportionately higher cardiac health service

consumption among lower-risk AMI depressive patients may suggest that health-seeking

behaviors are mediated by psychosocial factors more so than by objective measures of

cardiovascular risk or necessity.

40

Introduction

Available evidence has demonstrated that health service consumption is higher among

patients with depression than those without. In ambulatory care settings, depression is

associated with a 50% increase in general medical service use, even after adjusting for

age, sex and chronic medical comorbidities49,50. Similar depression-related service

consumption patterns have been described among cohorts with cardiovascular

disease.5452.

Some authors advocate that increased health service consumption among cardiac-specific

patients with depression is appropriate and concordant with their underlying

cardiovascular prognosis21. However, others contend that depressed patients seek more

health care services regardless of illness severity. Evidence from utilization patterns of

depressed patients in primary care settings suggest that depressed patients use more

health care services than non-depressed patients regardless of medical illness severity 48.

Few studies have quantified the relationship between depressive symptoms, illness

severity, and health service consumption.

Accordingly, the objective of our study was to evaluate the impact of depressive

symptoms on health service consumption and cardiovascular prognosis following AMI.

AMI serves as a useful test case because the natural history of the disease has been well-

described69 and validated measures of cardiac prognostic risk severity exist80. Despite

these factors, unexplained variations in health service consumption exist across AMI

41

patient populations8167; similar health service consumption variations may be explained

by psychosocial factors. Because Canada's federal–provincial Medicare plan covers

medically necessary services based on need rather than affordability82, differences in

health service consumption are more likely to reflect differences in health service

behaviors than in a system where services are provided based on an ability to pay 54. We

hypothesized that health service consumption following AMI would be increased among

patients with depressive symptoms as compared to those without and would be

independent of comorbidity and cardiac illness severity.

Methods

Health system context

Canada's universal health insurance system provides comprehensive coverage for

most medical and hospital services without user fees at point of service. Under such

provisions, patients are entitled to equitable access to health care services based on

medical need, regardless of age, financial status, or financial circumstances82.


This study is a sub-study of the Socio-Economic and Acute Myocardial Infarction Study

(SESAMI) study, a prospective observational study of patients who were hospitalized for

AMI throughout Ontario, Canada67. Patients were included if they were English-

speaking and if 2 of 3 AMI criteria were met: presence of symptoms, abnormal

electrocardiographic findings, or elevated serum levels of cardiac enzymes. Patients

42

were excluded if they were younger than 19 years of age or older than 101 years of age,

lacking a valid health card number issued by the province of Ontario, and those who were

transferred to the recruiting hospital. Eligibility for the study required completion of a

self-administered baseline survey; patients who died within 24 hours of admission, who

had very severe illness, who had language barriers, or who underwent early discharge or

transfer were therefore ineligible. Data came from four sources: 1) a baseline survey; 2)

chart abstraction from index AMI; 3) a one-month follow-up phone survey; and 4) linked

health administrative data. From 2829 patients with successfully abstracted index AMI

admissions and linked health administrative data, 888 (31%) subjects did not complete

the one-month follow-up survey either because of death prior to the survey (N=73; 2%)

or a refusal to participate (N=815; 29%). A total of 1941 (69%) patients remained

available for analysis. The SESAMI cohort has been described in detail elsewhere67.

The Sunnybrook Health Sciences Centre Review Committee approved this study and all

subjects gave informed consent to participate.

Depressive Symptom Measures

The SESAMI survey consisted of psychometric questions that explored various domains

related to depression, including low mood, loss of interest, sleep disturbance, reduced

appetite/wt loss, agitation/slowing, low self-esteem/guilt, suicidal thoughts, reduced

energy level, loss of concentration (Table 1). The SESAMI surveys were telephone

administered by standardized trained health care personnel (nurses).

43

Nine of twelve questions were directly abstracted from the Brief Carroll Depression

Rating Scale (BCDRS). These nine questions served as our primary determinant of

depressive symptoms. The original 12-item BCDRS is a depression rating scale which

has been validated among hospitalized mentally ill populations, and has a sensitivity of

92% and a specificity of 89% using a cut-off scale of 6 68. These depression rating scales

are described in Appendix 1.

While a score of 6 or more of the original 12-item BCDRS has been used to define

depression, affirmative responses to 5 or more of the 9 administered BCDRS items were

used to define patients as having “depressive symptoms” for the purposes of our study. A

score of 5 rather than 6 was used because it corresponded most closely with the median

score of 6, and because five or more depressive constructs are required to fulfill DSM-IV

criteria for depression, and hence is concordant with the diagnostic criterion threshold

defined in the DSM-IV 6.

Demographic Factors Demographic factors such as age, sex, and income tertile were acquired from the baseline

survey and were based on patient self-report.

Cardiovascular Risk Severity

Two validated measures of cardiovascular risk-severity were used in this study:

The first measure was the Global Registry of Acute Coronary Events (GRACE)

prognostic index. The GRACE prognostic index provides a score that reflects the

44

probability of dying within 6 months post-AMI based on age, development (or history) of

heart failure, peripheral vascular disease, systolic blood pressure, Killip class, initial

serum creatinine concentration, elevated initial cardiac markers, cardiac arrest on

admission, and ST segment deviation80. The GRACE prognostic index has been validated

among SESAMI patients70. In addition to the GRACE prognostic index, we examined

other cardiovascular risk factors, which included diabetes, hypertension, hyperlipidemia,

and current or former cigarette use. Smoking history was ascertained by questionnaire.

Major cardiovascular risk factors were determined by reviewing diagnostic fields in

computerized hospital discharge abstracts from April 1, 1988 to the index hospitalization

date using all primary and secondary ICD-9 and ICD-10 (where applicable) discharge

codes. Agreement between self-reported risk factors and chart audit range was between

73% (hyperlipidemia) and 95% (diabetes) 67.

The second measure was the Duke Activity Status Index (DASI), administered to all

participants. The DASI is a self-report measure validated to determine functional

capacity which correlates well with peak oxygen uptake63, one of the most important

single predictors of long-term cardiovascular survival across populations83.

Preexisting Noncardiovascular Comorbid Conditions

Noncardiovascular risk factors consisted of all co-morbid diseases that were captured

through primary and secondary diagnostic fields of hospital discharge abstracts (Canadian

Institute for Health Information) from 1 April 1988 to the presenting hospitalization. This

45

method has been used and demonstrated to increase the prevalence of chronic conditions,

which historically are known to be under-coded using single data sources alone84 56.

Furthermore, supplementing clinical data sources with longer retrospective ascertainment

of comorbidities using administrative data has been shown to improve accuracy 85. Non-

cardiovascular conditions were categorized as cancer and as diseases of the central

nervous system, endocrine system, hematology system, musculoskeletal system,

respiratory system, gastrointestinal system, and genitourinary system. We categorized

diabetes, secondary hypertension, and hyperlipidemia as cardiovascular risk factors, not

as diseases of the endocrine system71. Available evidence has demonstrated that the

number of non-cardiac comorbidities have independent prognostic significance in

patients with cardiovascular illness86. For the purposes of our study, non-cardiac

comorbidities were analyzed as a count variable (0, 1, 2 or 3 or more). However, a re-

analysis of our data in which we incorporated both a count variable and indicator

variables for the type of non-cardiac comorbidity did not alter our results.

Early peri-infarction procedure and medication use

All of our outcomes could be influenced by whether or not a patient has received

revascularization by angioplasty or coronary bypass surgery. All revascularization

procedures occurring within 30 days of discharge from the index AMI hospitalization

were recorded within the Ontario Health Insurance Program database. In addition, we

examined the influence of cardiovascular medication (Beta-blockers, statins, ACE

inhibitors, and nitrates) prescribed at discharge from index admission as acquired from

46

chart abstraction.

Outcomes

Our primary outcomes included cumulative outpatient physician and emergency room

visits, the number and duration of recurrent hospitalizations, as well as total

hospitalizations and ambulatory physician visit over the 18 months following index AMI

hospitalization discharge. These outcomes were chosen because they reflect health care

consumption measures that can be accurately measured in the linked administrative

health data sets. Secondary outcomes included mortality, recurrent AMI, as well as time

to first re-admission for cardiac-specific readmissions as general measures of prognosis.

We examined recurrent AMI hospitalizations separately from other cause-specific

admissions because recurrent AMI represents a deleterious prognostic outcome indicator.

Moreover, unlike our other hospitalizations which are based on more discretionary

criteria, recurrent AMI admissions are based on more standardized, objective clinical,

laboratory, and ECG parameters. However, a re-analysis in which recurrent AMI was

categorized together with the other health service consumption variables did not alter our

results. Recurrent AMI was defined using the most-responsible diagnosis field of ICD-9

and ICD-1073. Cardiac specific re-admissions were ascertained using sets of the CIHI

most-responsible diagnosis codes (AMI: ICD-9 410, 412, 4141; ICD-10 I21, I22, and

I2382; angina: ICD-9 411, 413, 4140, and 4142-4149; ICD-10 I20, I241, I251, I252,

I253, and I258; CHF: ICD-9 428, 415, 4254, 4298; ICD-10 I50, I255, I420, I429) that a

previous study has shown have modest sensitivity, but high specificity73.

Statistical Analysis

47

The Mantel-Haenszel test for trend was used for categorical data and t-tests (or

nonparametric tests where relevant) were used for continuous data to detect unadjusted

differences in baseline characteristics. We estimated Poisson regression models for rates

of service utilization, as well as Cox proportional hazard models for mortality, recurrent

AMI, and time to first angina hospitalization, adjusting for age, sex, income, cardiac risk

factors, total medical comorbidities, prognostic index (GRACE score80

and DASI63),

drugs at discharge, and peri-infarction procedures using non-parsimonious modeling.

Formal diagnostic testing revealed no evidence of multi-collinearity in any of our

statistical models. We tested for violations of the proportionality assumption in all

proportional hazard model specifications. Finally, for the health service utilization

multivariate models, we used generalized estimating equations (GEE) to adjust for

hospital-level variations in patient care.

All analyses were performed using SAS statistical software, version 9.1 (SAS Institute,

Cary, NC).

Sensitivity analyses

We conducted a number of sensitivity and sub-group analyses. First, we

examined the relationship between the 3 replacement questions that were missing from

the BCDRS (Appendix Table 1) and health service utilization. Second, we explored

whether our results were consistent when using the GUSTO quality of life sub-study

depression measure, another psychometric instrument 72 and our modified 12-item

48

BCDRS scale 68. Finally, multiple imputation was used to impute depression measures

for the 888 missing depression values due to non-response57 to test whether our outcomes

were affected by systematic differences between survey responders and non-responders.

Variables used to model the missing data included demographic variables (age, sex,

SES), cardiac risk factors and illness severity (GRACE prognostic index and DASI),

revascularization procedures, drugs prescribed at discharge, and the 9-item depression

scale. There was no difference between aggregated results from 5 and 10 imputed data

sets; results from 5 imputed data sets are reported. The missing depression measures

were imputed using PROC MI and the results from the multiple, imputed data sets were

aggregated using PROC MIAnalyze in SAS version 9.1 (SAS Institute, Cary, NC).

Results

Baseline Characteristics

The 888 (31%) patients who did not respond to the one month follow-up survey were

similar to the 1941 subjects in terms of gender, cardiac risk factors, likelihood of

receiving revascularization procedures within 30 days of discharge, and the likelihood of

receiving ACE inhibitors, Beta blockers, and nitrates upon discharge. However, the

sample of 888 survey non-respondents were older (mean (SD) age of respondents 66.3

(13.7) years vs. 62.4 (12.8) years; P<0.001), more likely to have 3 or more non-cardiac

medical comorbidities (55.7% vs. 47.2%; P < 0.001), less likely to receive statins upon

discharge (48.7% vs. 54.9%; P = 0.002), and had a higher GRACE score (predictive of 6

month mortality)(4.1 vs. 3.2; P < 0.001) than the 1941 survey respondents who consented

and completed the 1 month post-myocardial infarction evaluation (see Table 2.1).

49

Among the 1941 subjects included in this study, the median age was 64 years (range, 26

to 96 years). 575 (29.6%) were women. Table 3.1 illustrates that the baseline

characteristics of patients reporting five or more depressive symptoms differed from

those reporting fewer depressive symptoms. Specifically, patients with depressive

symptoms were more likely to be female, were less affluent, and were more likely to

have diabetes, hypercholesterolemia, and non-cardiac comorbidities. There were no

significant differences in baseline cardiovascular prognosis (GRACE scores 109.5 vs.

113.0, p=0.22), but patients with depressive symptoms had poorer cardiovascular

functional status (DASI score 11.2 vs. 19.6, p<0.001; lower score indicates worse peak

oxygen uptake)(Table 3.1).

Health Service Consumption

Total number of hospitalizations, total number of hospitalization days, length of hospital

stays, and post-AMI ambulatory service use were greater for patients with depressive

symptoms (Table 3.2).

After adjustment for age, sex, income, risk factors, medical comorbidity, prognosis

(GRACE score), drugs at discharge, 30-day procedure use (percutaneous coronary

interventions and/or coronary artery bypass surgery), and symptom burden (DASI),

depressive symptoms remained an independent predictor of most service consumption

measures, with a 24% (Adjusted RR:1.24, 95% CI:1.19-1.30, P<0.001) increase in all

50

cause re-hospitalizations, a 9% (Adjusted RR:1.09; 95% CI:1.02-1.16, P=0.007) increase

in cardiac-related hospitalizations and a 43% (Adjusted RR:1.43; 95% CI:1.34-1.52,

P<0.001) increase in non-cardiac hospitalizations visits following AMI discharge (Figure

1). Depressive symptoms were associated with increases in cardiology, internal medicine

and family practice visits after adjustment for baseline variables (Figure 3.1).

Cardiac-specific re-admissions were comprised mostly of angina and CHF

hospitalizations. Depressed patients were substantially more likely to be admitted for

angina in the 18 months post-AMI than non-depressed patients (HR 1.75; 95% C.I. 1.44 –

2.14), even when the DASI is included in the model (HR 1.41; 95% C.I. 1.14-1.75).

However, after adjusting for emergency-room visits, depression was no longer a

significant predictor of hospital readmissions.

To evaluate whether depressive-associated increases in health service consumption were

consistent across cardiac illness risk severity levels, subgroup analyses were performed in

which the sample was stratified according to median GRACE prognostic index score and

median DASI scores. When stratifying patients around the median GRACE or DASI

scores, the increase in health service consumption associated with depressive symptom

burden was disproportionately higher among patients with lower GRACE scores (better

cardiac prognostic risk) and higher DASI scores (greater cardiac functional status) than

among their higher risk poorer functional capacity counterparts (Table 3.3). This effect

was also tested by including depression/illness severity interaction terms in the full

models. Depression/GRACE score interactions were statistically significant for the

51

following outcomes when death and recurrent AMI were included (General Practitioner

(GP) Visits (P<0.001) and when death and recurrent AMI were excluded: Total

hospitalization days (P<0.001); Total Cardiac-related hospitalization days (P<0.001);

Non-cardiac hospitalization days (P<0.001); Cardiologist Visits (P = 0.02); Internist

Visits (P = 0.05); GP visits (P<0.001); and Total ER Counts (P=0.002).

Depression/DASI interactions were significant for the following outcomes when death

and recurrent AMI were included: Total hospitalization days (P<0.001); Non-cardiac

hospitalization days (P<0.001); GP Visits (P<0.001); and Total Emergency Room (ER)

Visits (P<0.001). The depression/DASI interaction was also significant when death and

recurrent AMI were excluded for the following outcomes: Total hospitalization days

(P=0.02); Non-cardiac hospitalization days (P=0.004); GP Visits (P<0.001) and Total ER

Visits (P=0.03).

Mortality and Recurrent AMI

There was no significant difference in two-year mortality (HR = 0.92; 95% CI 0.59-

1.42), re-AMI (HR = 1.22; 95% CI 0.89-1.66), or in the composite risk of death or

recurrent AMI (Adjusted HR=1.12; 95% CI 0.80-1.57) between patients with and without

depressive symptoms.


First, the relationship between depression and health service consumption when using the

52

3 replacement items (Appendix Table 1) yielded similar outcomes as those generated

with the 9-item BCDRS (Table 3.4). Second variations in the cutoff scores used to

distinguish depressive from non-depressive patients did not alter our results. For

example, scores of 4 or 6 using the 9-item BCDRS generated similar results as a score of

5. Similarly, the use of the GUSTO quality of life sub-study depression measure scale

yielded similar results as did our modified, 12-item BCDRS. Finally, we used multiple

imputation to repeat the analyses to include 888 subjects with missing depression scores,

which did not substantively change the results from the complete data analyses (Figure

3.2).

Discussion

Depressive symptoms at one-month post-AMI were significantly associated with health

service consumption in the 18 months post-AMI. The relationship between health

service consumption and depressive symptoms persisted after adjusting for comorbidity

and cardiac illness severity. Indeed, the relationship between cardiac health service

consumption and depressive symptoms was even greater among those patients with lower

as compared with higher cardiac illness severity. In short, the increased likelihood of

health service consumption among AMI patients were over and above that expected

based on cardiac illness severity alone.

In our sample, depressive symptoms were not associated with increased mortality or

53

recurrent AMI after adjustment for potential confounding factors, which is contrary to

previous evidence 22 87. However, the association between depression and cardiovascular

outcomes remains controversial 26 28

and the results of this study are not sufficient to

address the controversial prognostic importance of depression definitively, especially

given inter-study differences in measurement and timing of depression. It is conceivable

that measuring depression at one-month post-AMI conferred a survival bias for which

multiple imputation methods were unable to fully address. A recent publication

measuring depression directly prior to one-month post-AMI found no association

between depression and mortality at either time point after adjustment for covariates 25.

Our study demonstrated a significant interaction between depression, prognostic severity

and health service consumption, such that higher health service use occurred among

patients of lower rather than higher cardiac severity - - a distinct contrast to those patients

without depressive symptoms whose health service experiences better mirrored their

cardiac illness severity levels. Available evidence has demonstrated a treatment-risk

paradox whereby service provision is paradoxically most intensive among patients with

lowest need 88,89. Our study suggests that the presence of depressive symptoms might

partially mediate the treatment-risk paradox.

Some have hypothesized that mental illness and/or psychological phenomena such as

somatization promote health seeking behaviors 48,90. Our results support such hypotheses.

Indeed, in our study, depression-associated increases in the risk of recurrent

hospitalization were no longer significant after adjusting for ER visits, suggesting that

54

depression-readmission rates may be attributable to patient health-seeking behaviors

more so than physicians’ propensity to admit for more discretionary indications91,92.

Our study has several important clinical and policy implications. Patients with depressive

symptoms post-AMI pose a serious challenge to physicians and health care systems. On

one hand, such patients are sicker and may be more likely to die following AMI,

underscoring the need for closer attentiveness and management21. On the other hand,

depressive patients consume significant amounts of cardiac health services and may do so

disproportionately in relation to objective measures of prognostic or symptomatic burden.

Integrated chronic cardiac care programs are well-established and effective interventions

for patients post-AMI93. Community-based depression interventions have also been

shown to be effective at treating depression and improving quality of life94. Given the

benefits of both chronic vascular disease management93, depression case-management94

and the high prevalence of depression in patients post-AMI95, our study suggests that

integrating depression screening and case-management into existing cardiac secondary

prevention programs may be effective in improving the quality of life of depressed post-

MI patients, and in reducing the apparent mismatch between need and service

consumption. The importance of systematic depression screening is further underscored

by the low detection rates of depression after AMI 96.

Our study has several noteworthy limitations. First, our depression measure was

designed to ascertain depressive symptoms rather than depressive disorders and was

missing 3 items from the original, validated depression rating scale. Our health service

55

consumption findings were similar whether the 9 items or 3 replacement items were used

to define depression. The comparability of results with these two depression measures

suggests that the two scales are measuring the same construct. Second, depressive

symptoms were measured cross-sectionally and at one-month post-AMI. As such, we

could not determine the persistence of depressive symptoms over the follow-up interval.

The delay in measuring depression post-AMI also imposes a survival bias on the sample.

However, survival bias is less of an issue with the main outcome of health service

consumption, since service consumption differences are only relevant in those patients

who survive long enough to use services. Finally, information regarding depression was

missing in just over 30% of the potentially eligible respondent population. However, our

sensitivity analyses utilizing model-based multiple imputation to include those patients

with missing depression values provide evidence that the missing values did not

materially affect our findings.

In conclusion, our study demonstrated the increased consumption of health services

among post-AMI patients with depressive symptoms is independent of comorbidity and

cardiac illness severity. Moreover, the increase in cardiovascular health service

consumption patterns among AMI patients with depressive symptoms was most

pronounced among those of lower rather than higher cardiac illness severity suggests that

cardiac health seeking behaviors among depressive patients may be mediated by

psychosocial factors in addition to objective measures of need. Future research must

evaluate whether systematic depression detection and integrated chronic disease

management systems improve health service efficiency and allow for better alignment

between illness severity and health service consumption among such patient populations.

56

This chapter has been published in BMC Health Services Research97.

57

Table 3.1: Baseline characteristics of depressed vs. non-depressed AMI patients.

Depressed* Non-Depressed P Value Characteristics (n = 494) (n = 1447) Age, y -- % 0.002

19-49 21 15 50-64 39 34 65-74 20 29 > 74 20 21

Male Sex -- % 63 73 <0.001 Income, Canadian $ -- % <0.001

Low (<$30,000) 33 24 Intermediate ($30,000 - $59,999) 34 35 High (>$59,999) 33 42

Coronary Risk Factors -- % Diabetes 29 22 0.001 Hypercholesterolemia 45 39 0.02 Hypertension 49 46 0.33 Smoking 43 39 0.08

Non-cardiac comorbidities -- % None 6 9 <0.001 1 16 23 2 27 23 3 or more 51 46

Prognostic Indicators -- mean (SD) GRACE 6-month prognostic index score 11.6 (30.3) 113.6 (28.5) 0.22 DASI Score 11.2 (8.4) 19.6 (11.8) <0.001

Processes of Care -- % Percutaneous transluminal coronary angiography 8 8 0.97 Coronary angioplasty bypass grafting 12 10 0.28 ACE Inhibitor 63 62 0.64

58

Beta Blocker 68 70 0.44 Statin 55 55 0.89 Nitrate 37 31 0.01

The depression measure is a depression scale containing 9 items from the Brief Carroll Depression Scale (BCDS)(cut-off score of 5). Abbreviations: GRACE - Global Registry of Acute Coronary Events scale; DASI - Duke Activity Status Index

59

Table 3.2: Health service consumption in depressed and non-depressed post-AMI patients.

Service Consumption Variable Depressed Non-depressed P ValueHospitalization -- mean (SD)

Total hospitalization days 8.5 (19.8) 5.5 (14.6) 0.002 Total cardiac hospitalization days 4.8 (10.2) 3.3 (7.6) 0.002 Total non-cardiac hospitalization days 3.7 (15.5) 2.3 (11.5) 0.06 Total number of hospitalizations 1.2 (1.7) 0.8 (1.2) <0.001 Total number of cardiac hospitalizations 0.7 (1.1) 0.5 (0.9) <0.001 Total number of non-cardiac hospitalizations 0.5 (1.1) 0.3 (0.7) 0.001

Ambulatory service consumption -- mean (SD) Cardiologist visits 12.5 (11.5) 9.8 (10.2) <0.001 General internist visits 14.1 (26.0) 10.8 (20.9) 0.01 Family doctor visits 36.5 (25.3) 31.6 (24.6) <0.001 Emergency department visits 1.7 (2.3) 1.3 (1.9) <0.001

Legend: The depression measure is a depression scale containing 9 items from the Brief Carroll Depression Scale (BCDS)(cut-off score of 5). Hospitalization days are a count of total days in hospital over the 18-month follow-up period and can accumulate from multiple hospitalizations.

60

Table 3.3: Multivariate service utilization rate for depression stratified across cardiac illness severity.

Complete Data Death and Recurrent AMI Excluded

Low risk High risk Low risk High risk Cardiac prognostic risk (GRACE Score) (n = 1079) (n = 862) (n = 995) (n = 709) Hospitalization

Total Hospitalization Days 1.80 (1.69-1.92) 1.32 (1.26-1.39) 1.49 (1.36-1.62) 1.15 (1.07-1.23) Cardiac-related Days 1.78 (1.65-1.92) 1.18 (1.10-1.26) 1.36 (1.22-1.51) 0.97 (0.87-1.08) Non-Cardiac Days 1.82 (1.62-2.04) 1.52 (1.42-1.63) 1.76 (1.51-2.05) 1.31 (1.19-1.43)

Ambulatory

Cardiologist Visits 1.29 (1.23-1.34) 1.24 (1.19-1.30) 1.17 (1.11-1.23) 1.11 (1.06 (1.18) Internist Visits 1.28 (1.22-1.34) 1.23 (1.18-1.28) 1.04 (0.99-1.09) 1.14 (1.09-1.20) GP Visits 1.23 (1.20-1.27) 1.05 (1.03-1.08) 1.17 (1.14-1.21) 0.92 (0.90-0.95) Total ER Visits 1.21 (1.07-1.37) 0.98 (0.86-1.11) 1.24 (1.08-1.43) 1.01 (0.87-1.17)

High capacity Low capacity High capacity Low capacity Functional capacity (DASI score) (n = 922) (n = 1019) (n = 848) (n = 856) Hospitalization

Total Hospitalization Days 1.96 (1.79-2.16) 1.09 (1.04-1.14) 1.53 (1.35-1.73) 1.24 (1.17-1.32) Cardiac-related Days 1.13 (1.07-1.20) 0.90 (0.76-1.05) 1.12 (0.94-1.35) 1.19 (1.09-1.30) Non-Cardiac Days 4.04 (3.53-4.56) 1.04 (0.97-1.12) 2.15 (1.80-2.56) 1.33 (1.22-1.45)

Ambulatory

Cardiologist Visits 1.23 (1.15-1.32) 1.14 (1.10-1.18) 1.26 (1.17-1.35) 1.16 (1.11-1.21) Internist Visits 1.33 (1.24-1.43) 1.02 (0.98-1.05) 1.37 (1.28-1.48) 1.09 (1.05-1.13) GP Visits 1.28 (1.23-1.33) 0.98 (0.96-1.00) 1.25 (1.21-1.30) 1.00 (0.97-1.02) Total ER Visits 1.49 (1.24-1.79) 1.00 (0.91-1.11) 1.66 (1.37-2.01) 1.02 (0.91-1.16)

61

Legend: All relative risks were adjusted for age, sex, income, pre-existing cardiovascular disease, pre-existing non-cardiovascular conditions, in-hospital processes of care, prognostic risk (GRACE and DASI). Abbreviation: AMI: Acute myocardial infarction. Interaction terms were included in models with the complete sample. Depression/GRACE score interactions were statistically significant for the following outcomes when death and recurrent AMI were included (General Practitioner (GP) Visits (P<0.001) and when death and recurrent AMI were excluded: Total hospitalization days (P<0.001); Total Cardiac-related hospitalization days (P<0.001); Non-cardiac hospitalization days (P<0.001); Cardiologist Visits (P = 0.02); Internist Visits (P = 0.05); GP visits (P<0.001); and Total ER Counts (P=0.002). Depression/DASI interactions were significant for the following outcomes when death and recurrent AMI were included: Total hospitalization days (P<0.001); Non-cardiac hospitalization days (P<0.001); GP Visits (P<0.001); and Total Emergency Room (ER) Visits (P<0.001). The depression/DASI interaction was also significant when death and recurrent AMI were excluded for the following outcomes: Total hospitalization days (P=0.02); Non-cardiac hospitalization days (P=0.004); GP Visits (P<0.001) and Total ER Visits (P=0.03). Depression status was based on the 9-item depression rating scale. Abbreviations: GRACE – Global Registry of Acute Coronary Events; DASI – Duke Activity Status Index. Hospitalization days are a count of total days in hospital over the 18 month follow-up period and can accumulate from multiple hospitalizations. Total and cardiac hospitalization results excluded recurrent AMI hospitalizations.

62

Table 3.4: Multivariate health service consumption rates for depression measure

using only 3 replacement items.

Total Hospitalization Days 1.25 (1.20-1.30) Cardiac Hospitalization Days 1.07 (1.01-1.13) Non-Cardiac Hospitalization days 1.54 (1.45-1.63) Total Number of Hospitalizations 1.10 (0.99-1.22) Total Number of Cardiac Hospitalizations 0.99 (0.89-1.13) Total Number of Non-Cardiac Hospitalizations 1.29 (1.09-1.53) Total ER count 1.02 (0.93-1.10) Cardiologist Visits 1.12 (1.09-1.16) Internist Visits 0.98 (0.95-1.01) GP visits 1.06 (1.04-1.08)

Legend: Results from multivariate Poisson regression models adjusted for age, sex, income, cardiac risk factors (diabetes, hypertension, hypercholesterolemia, smoking history), medical comorbidities, CABG, PTCA, drugs at discharge (ACE inhibitors, Beta blockers, statins, and nitrates), GRACE score and DASI score and are reported as point estimate with 95% confidence intervals. Hospitalization days are a count of total days in hospital over the 18 month follow-up period and can accumulate from multiple hospitalizations. Total and cardiac hospitalization results excluded recurrent AMI hospitalizations. The three depression measures are a depression scale containing 9 items from the Brief Carroll Depression Scale (BCDS)(cut-off score of 5), a 12-item depression scale with 3 items replacing the missing items from the BCDS(cut off score of 6), and the Duke Psychological Well-being scale (cut off of 25). Abbreviation: ER – Emergency Room

63

Figure 3.1: The adjusted relative rate of service consumption attributable to

depression.

64

Legend: All outcomes were adjusted for age, sex, income, cardiac risk factors, coronary artery bypass graft (CABG), percutaneous, transluminal coronary angiography (PTCA), drugs at discharge, GRACE prognostic index score, and DASI score. Hospitalization days are a count of total days in hospital over the 18-month follow-up period and can accumulate from multiple hospitalizations. The depression measure is a depression scale containing 9 items from the Brief Carroll Depression Scale (BCDS)(cut-off score of 5). Total and cardiac hospitalization results excluded recurrent AMI hospitalizations. Abbreviation: ER – Emergency Room.

65

Figure 3.2: The adjusted relative rate of service consumption attributable to

depression – multiple imputation results.

66

Legend: All outcomes were adjusted for age, sex, income, cardiac risk factors, coronary artery bypass graft (CABG), percutaneous, transluminal coronary angiography (PTCA), drugs at discharge, GRACE prognostic index score, and DASI score. Hospitalization days are a count of total days in hospital over the 18-month follow-up period and can accumulate from multiple hospitalizations. The depression measure is a depression scale containing 9 items from the Brief Carroll Depression Scale (BCDS)(cut-off score of 5). Total and cardiac hospitalization results excluded recurrent AMI hospitalizations. Abbreviation: ER – Emergency Room.

67

Chapter 4: Missing data: scope and impact on

outcomes.


1) Define missing data mechanisms and their potential impact on estimates

generated from statistical analyses.

2) Provide an overview of different ways to address missing data in statistical

analyses.

3) Outline the extent to which missing data are explained and addressed in the

current psychiatric literature.

4) Provide an example of the impact of different data analytic methods for handling

missing data on outcomes.

Abstract

Missing data occur frequently in clinical or population-based research where data

are acquired by survey or administered questionnaires. Typically, missing data are

ignored in statistically analyses, which can result in inaccurate variance estimation due to

reduced sample size or bias. Various methods have been proposed to address missing

data, and their utility depends on the missing data mechanism (reviewed in this chapter).

A review of the top three psychiatric journals in 2004 and 2005 revealed 222

articles that had either a cross-sectional or cohort design and relied on survey data. Of

those studies, 69.8% had more than 15% of the sample with missing data. Of those

studies, only 9 (5.8%) used specific analytic methods to estimate the impact of missing

data on outcomes, and only one study used model-based methods. Of the 222 studies,

68

only 68.9% of the studies provided description of missing data that was sufficient to

allow readers to determine whether the missing data are ignorable.

When SESAMI data were analyzed using complete case analysis (observations

with missing data are ignored), the adjusted mortality risk attributable to depression was

1.18 (95% CI 0.92-1.51). When the missing depression values were imputed using a

simple mean method, the mortality rate was reduced to 1.05 (95% CI 0.76-1.44). Using

multiple imputation, the mortality rate was similar to the simple mean imputation method

(HR = 1.02), but the 95% CI was substantially broader (0.64-1.62).

Missing data are frequently ignored in psychiatric literature, despite evidence

provided that estimates made by ignoring missing data have a high likelihood of being

biased. The paper concludes with recommendations for the routine description of

missing data in studies and, when the data are missing beyond a certain proportion (10%),

that model-based strategies be used to determine the impact of missing data on estimates

of outcomes.

Introduction

Clinical and population-based psychiatric mental health research studies involve large

samples and rely on diagnostic survey or rating instruments to generate diagnoses.

Whenever questionnaires are administered in a sample of subjects, missing data are

invariably generated. Researchers typically handle missing data by only analyzing the

data that are collected and complete for all subjects. Excluding missing data in this

manner leads to a loss of power at best and biased parameter estimates at worst

69

depending on the degree of similarity between missing and complete data for a given

variable.

The purpose of this study is to briefly review the problem of missing data and methods

that have been proposed to deal with them, and to ascertain how frequently those

methods are used in relevant studies published in high quality psychiatric journals. Once

such methods are reviewed, a review of articles from three high impact psychiatric

journals will determine how often such methods are being employed in studies with

substantial amounts of missing data. Finally, the implications of using three methods –

complete data analysis, simple mean imputation, and multiple imputation – will be

heuristically explored using real data.

Missing Data – The Problem

In studies where data are collected from a large number of subjects , missing data are the

norm rather than the exception. The goal of any analysis is to generate estimates that are

a valid reflection of the target population. Typically, statistical software packages

automatically remove subjects or observations with any missing data for variables

included in a multivariate analysis. Such an analysis is called a “complete case analysis”

57. With the goal of valid inferences for a target population in mind, complete case

analysis may be problematic for two reasons. First, the missing data may be missing in

ways that bias the results of complete case analyses, which offer no way to address such

biases. In other words, the complete case sample is often different from the target

70

population sample (complete plus missing data) in ways that are not completely random.

Second, even if the estimate generated from a complete case analysis is accurate, the

variance associated with such an estimate will be inaccurate due to the loss of

observations with missing data.

Modern statistical methods to manage missing data, which are based on an understanding

of the mechanism that generated the missing data, were developed in the 1970s98. A

detailed discussion of the mechanisms for missing data is beyond the scope of this paper

and has been reviewed elsewhere99; however a brief description is necessary in order to

understand the motivations for methods of handling missing data.

Three different mechanisms for missing data have been described: 1) Missing completely

at random (MCAR); 2) Missing at random (MAR); and 3) Not missing at random

(NMAR) 9957. These missingness mechanisms are defined by whether missingness is

dependent on the values of observed or missing data.

Missing Completely at Random (MCAR)

If the missing data depend neither on observed or missing values, then the missingness is

MCAR. The complete randomness of MCAR is analogous to the randomization

procedure of a randomized, controlled clinical trial, except in this case, observations

would be randomly assigned to be missing or recorded for a given variable. Given a

sufficiently large sample, there should be no bias in an estimate produced from a sample

71

with missing data following a MCAR pattern because there is no systematic difference,

known or unknown, between the missing and non-missing data. As long as the change in

the sample size is factored into the estimate of the variance associated with the estimate,

results from complete case analyses with missing data following a MCAR pattern will

produce unbiased estimates for the target population of interest.

Missing at Random (MAR)

Data that are MAR occur when the mechanism responsible for generating the

missingness is dependent on the data observed and collected, but not on missing data.

With a MAR missingness pattern, bias is likely to exist. However, since the missingness

is dependent on observed and collected data, these observed and non-missing data can be

used to address the missingness. The following has been provided as an example of

MAR100. Consider a survey that collected data on gender and income, and where females

are less likely to respond to questions about income. If the reduced likelihood for

females to respond to income questions is independent of the income (e.g. wealthy

women are no more or less likely to respond to income questions than less wealthy

women), then unbiased estimates for income can be made because the sex of every

subject is known and the income of some of the females is known. In other words, a

variable is MAR when subjects have missing values that are only randomly different

from other subjects, given the values of other available variables101.

Not Missing at Random (NMAR)

72

If the missingness is not missing at random (NMAR), then the missingness is dependent

on the missing observations. To continue with our analogy, a NMAR missingness

mechanism is like an observational study where the outcome variable is dependent on

unobserved variables. In such circumstances, risk adjustment cannot address potential

bias because the data required to risk adjust are not known. When data are NMAR,

unless the mechanism that generated the missing data is known the analyst is generally

restricted to conducting sensitivity analyses that assess the impact of alternate

assumptions about the missing data mechanism. For instance, assume that there is a

sample of subjects of whom 30% did not respond to a dichotomous (yes/no) variable. An

extreme sensitivity analysis would be to perform an analysis that assumes two scenarios:

1) that all 30% of the subjects would have responded “yes” to the missing variable; or

alternatively, 2) that all 30% of the subjects would have responded “no.” A sensitivity

analysis would involve creating two separate data sets, one with a “yes” value to the 30%

missing and one with a “no” value to the 30% missing and re-running multivariate

analyses on the two separate data sets to determine the impact of the two extreme

response types on the parameter estimates.

Ways to address missing data

There are four general approaches to analyses with missing data. They are complete case

analysis, analyses using weights, imputation, and model-based analyses. Complete case

analysis has already been discussed; it involves using only observations for which

completely recorded data are available102. The advantage of this approach is expedience

because the analysis can proceed with no need to address the missing data. The obvious

73

disadvantage, when missing data are not MCAR, is the distinct likelihood that the results

will be biased and the estimate of the variance will be imprecise.

A common example of the use of weighting to address non-response occurs in

population-based surveys. Typically, surveys are administered to a sub-sample of a

community. To get adequate representation, certain sub-samples of a population are

over-sampled. In order for surveys to produce accurate estimates, design weights – the

inverse of the probability for a particular observation to be selected – are used57. The use

of such weighting methods can be generalized to any type of data with missing values.

Using weighting assumes that the missing data follow a MCAR pattern. A MCAR

missingness pattern is likely quite uncommon and, therefore, is likely not a reasonable

assumption. However, if there is compelling evidence that such a mechanism is relevant,

then applying weights to analyses of samples with missing data can address the

underestimation of variance from a complete case analysis, but cannot address the

potential bias produced by missingness that is anything but MCAR. Another limitation is

the difficulty in computing valid standard errors for estimates using weighting103. More

recently, the use of inverse probability or nonresponse weighting has been proposed for

use with generalized estimating equations in situations where the missing data follow a

MAR pattern104.

Imputation is a technique whereby missing values are filled in with substitute values prior

to complete case analysis. Common ways to fill in the missing values are “hot deck”, in

which values from the complete data are used to substitute missing values105, and “mean

74

imputation” where the mean from a set of observed values are used to substitute for

missing values. Once a missing value is filled in with a substitute, it is treated as though

it was actually observed. Thus, the uncertainty associated with the missing value (e.g. its

value could fall within some range of values) is artificially reduced by the replaced value.

This uncertainty is further reduced with simple mean imputation, where a single value

(the observed mean for a given variable) is used to replace all missing values. Simple

imputation can address the bias of complete case analysis. Because simple imputation

under-estimates the uncertainty associated with missing values, the variance generated by

analyses from simple imputation will likely be artificially low.

Multiple imputation is a model-based imputation method and has the advantage of

addressing both the potentially biased estimates produced from data with missing values

and the sampling variability associated with missing data. Multiple imputation is a

procedure for replacing each missing value several times, thereby creating several

complete datasets containing the observed (non-missing) and filled-in data. By creating

several imputed data sets, each missing value is replaced by a set of plausible values

generated by the specified model. Each complete data set containing imputed values is

analyzed separately, but the results from these analyses are combined to produce one

overall estimate. In combining the results, there are two kinds of variance that are

accounted for: 1) the within-imputation variance (i.e. the variance created for each

imputed data set) and 2) the between-imputation variance (i.e. the variance arising from

the multiple, imputed data sets). By factoring in both types of variance, standard errors

associated with estimates generated from multiple imputation procedures more accurately

75

reflect the uncertainty associated with missing values that are, by definition, unknown.

The impact of both simple and multiple imputation on sample estimates and standard

errors will be demonstrated below.

Review of the Literature

The purposes of this section are as follows: 1) to determine the proportion of published

articles in psychiatric journals report missing data in such a way as to allow the reader to

ascertain whether missing data could be a source of bias; and 2) to determine the extent to

which missing data techniques are being employed.

The use of weighting of nonresponse in surveys has been in use for decades. The most

recent innovation, multiple imputation, was first described 30 years ago9899. Other

methods for handling missing data, such as hot deck imputation, have also been available

for more than a decade106. However, techniques to address missing data are likely to

appear in clinical research journals at a considerably later date. Indeed, statistical

software that incorporates robust procedures that allow users to easily perform

methodologies such as hot deck imputation or multiple imputation have only recently

been widely available. For example, hot deck imputation was available for Stata in 1999

107; SPSS introduced similar software in the same year 108. In 2002, multiple imputation

procedures were included in SAS version 9 109; for details see 110.

Assuming that a lag would have occurred between the release of software capable of

multiple imputation and the publication of research incorporating such methodologies,

76

psychiatric journals published in the years 2004 and 2005 were chosen to be included in

this review. To be included in this literature review, psychiatric journals with an impact

factor of 6 or higher were chosen. An impact factor of 6 is an arbitrary cut-off, but was

chosen to select for high quality clinical journals in which adoption of new

methodologies is most likely to occur. Three psychiatric journals had an impact factor of

6 or greater in those two years: the Archives of General Psychiatry, the American

Journal of Psychiatry, and Biological Psychiatry.

All original research articles published in these journals were assessed (excluding brief

reports because of brief methodology descriptions). Studies with data collection

involving administered or self-report surveys or rating scales occurring cross-sectionally

or longitudinally in clinical or population-based samples were included. Randomized-

controlled trials were excluded because there is an extensive literature on handling

missing data in this type of research design111. Where missing data were reported, I

ascertained whether one of three methods for handling missing data were executed: 1)

using weighting reflecting response probability; 2) simple imputation; or 3) multiple

imputation.

These three journals published 1236 articles in 2004 and 2005. Of those articles, 208

consisted of studies using either cross-sectional or longitudinal study designs. Of those

208, only 153 indicate whether the sample contained missing data and describe the

missing data sufficiently for a reader to ascertain whether excluding missing data from

the analysis is likely to affect the results. Of those 153 articles, 9 studies have made an

77

attempt to address the impact that missing data may have on parameter estimates (8 by

using weights and 1 by multiple imputation) (Fig. 4.1). As such, 5.9% of studies address

potential error introduced by missing data, and only one of the articles used a model-

based method.

A Demonstration Using Real Data


Data were obtained from the Socio-Economic and Acute Myocardial Infarction Study

(SESAMI) study, a prospective observational study of patients hospitalized for AMI

across 53 acute care institutions throughout Ontario, Canada between December 1, 1999

and February 28, 2003112. The study sample of AMI patients who had baseline data

linked to administrative data consisted of 2829 subjects. Of those 2829 patients with

successfully abstracted index AMI admissions, 888 subjects did not complete the one-

month follow-up survey either because of death prior to the survey (N=73) or a refusal to

participate (N=815). Since the depression measure was administered in the one-month

follow-up phone survey, 31.4% (N = 888) subjects had missing depression scores,

leaving a total of 1941 patients with complete data for all variables to be included in our

analysis. The only covariate with missing data in our examples is depression; all other

covariates were either abstracted from the patient’s chart or collected at baseline. The

SESAMI cohort has been described in detail elsewhere112.

78

For the purposes of this demonstration, the goal was to determine whether analysis with

and without the 888 missing subjects from the sample with successfully abstracted AMI

admissions were substantially different. The outcome used for this example was time-to-

death in the 18 months post-AMI. Cox proportional hazard survival analysis was used to

model the hazard ratio for depression independent of the other covariates included in the

model.

The first step is to compare the observations with missing depression values to

observations with a completed depression value. Subjects with missing depression scores

were slightly older and, in terms of cardiac risk factors and medical comorbidity, were

more like depressed subjects than non-depressed. Subjects with missing depression

scores also had a higher mean Grace prognostic index score, which is a measure of

cardiac severity (Table 4.1). The descriptive results from Table 4.1 reveal that subjects

with missing depression values are significantly different compared to those observations

with complete depression scores. The data do not appear to follow a MCAR pattern,

which means the estimates produced by ignoring the observations with missing data are

likely to be biased.

As indicated in the review above, if the data follow a MAR pattern for missingness, then

there are ways to model missingness. If the data follow a NMAR pattern, then less

rigorous methods to determine the impact of the missing data on outcomes are available.

To test whether the data follow a MAR pattern, the outcome was modelled twice using

the data with complete values for all variables. First, the only covariate to be included in

79

the model is the variable with the missing data – in our case, the depression variable.

Next, the analysis is repeated with all the other covariates included in the model.

Attenuation of the estimate with all the covariates included suggests that depression is

dependent on the covariates included in the model. In other words, an attenuation of the

association between depression and our outcome, mortality, suggests that the missingness

associated with depression follows a MAR pattern since the association between

depression and mortality is dependent on the other covariates. The hazard ratio for

depression in the model that includes only our depression variable is 1.27 (95% CI 0.86-

1.88). The hazard ratio for depression in the model that includes the remaining variables

is 1.18 (95% CI 0.92-1.51)(Figure 4.2).

Now that there is evidence that the data follow a MAR pattern, the impact of handling the

missing data in different ways can be assessed. The multivariate, proportional hazard

models were analyzed three different ways. First, complete case analysis was used, with

automatic removal of data from the 888 subjects for whom there was no depression rating

scale. Second, the missing depression score was replaced using simple mean imputation

– the mean depression score for the entire sample was assigned to those with missing

depression scores. Finally, multiple imputation was used to generate 5 complete data sets

with a model-based imputation. The missing depression scores were generated using a

logistic regression model within Proc MI (SAS, Cary, North Carolina). Proc MIAnalyze

(SAS, Cary, North Carolina) was used to aggregate the estimates from the 5 separate Cox

proportional hazard models using the multiply imputed data sets.

80

The complete case Cox proportional hazard analysis revealed a hazard ratio for

depression of 1.18 (95% CI 0.92-1.51). Simple imputation by imputing the mean

depression score (the mean depression score is simply the proportion of depression cases

within the complete sample, 0.25) resulted in a hazard ratio for depression of 1.05 (95%

CI 0.76-1.44) (Figure 4.3). For multiple imputation, the 5 imputed data sets were

analyzed separate to generate 5 Cox proportional hazard ratios. The 5 hazard ratios for

depression are then combined to account for both within and between data set variance.

The aggregate hazard ratio from the multiple imputation 1.02 (95% CI 0.64-1.62)(Figure

4.3).

Discussion

This literature review highlights a few issues regarding how missing data are reported

and handled in the three psychiatric journals reviewed. First, studies using research

designs likely to generate missing data (cross-sectional and cohort studies) frequently do

not report whether or not missing data have occurred. When missing data are neither

reported nor described, readers have an opportunity to speculate on the impact that

missing data may have had on the accuracy of the results. Second, when missing data are

reported, it is quite common to report descriptive similarity of observed value parameters

(i.e. means or proportions) amongst groups with missing and non-missing values. For

example, studies that reported the degree of similarity of variables collected at baseline of

a prospective cohort often concluded that the missing data likely are not a source of bias

in parameter estimates. These same studies typically proceed to conduct complete case

analyses where the observations containing missing values for any variable included in

81

the analyses are discarded. This reasoning is flawed for several reasons. Using complete

case analysis reduces the sample size and will affect the variance regardless of similarity

of missing data to non-missing data. Furthermore, bivariate analyses can misrepresent

relationships between independent and dependent variables in multivariate analyses.

Third, only rarely do studies systematically use techniques to test whether missing data

are a potential source of bias in their analyses. The most common way to handle missing

data was using weights. As reviewed above, using weights assumes that the missing data

follow a MCAR pattern; using weights simply adjusts the variance to reflect the entire

sample rather than the sample with complete data which may not be a realistic

assumption.

In 5 paperes the authors argued that using generalized estimating equation techniques was

a valid way to address missing data. Generalized estimating equation analyses do not

remove observations with missing data (unlike complete case analysis). They do not

estimate values for missing data. Generalized estimating equation methods allow for

inclusion of observations for which incomplete data are present, but do not provide a way

of estimating the impact of the missing data associated with these observations in the

absence of other techniques. None of these studies used non-response or inverse

proportional weights with generalized estimating equation analyses, which has been

shown to be a legitimate way to handle data with a MAR missingness pattern104. As

such, for the purposes of this literature review, they were not considered a systematic way

of handling missing data. Even if one accpepts generalized estimating equations a

legitimate way to address the problem of missing data, it does not change the conclusion

82

of our literature review that systematic handling of missing data does not occur with high

frequency (our proportion of articles that handle missing data would increase from 5.9%

to 9.2%) in most studies published in high-quality psychiatric journals.

In our example using real data, our three results demonstrate that the way missing data

are handled can result in both bias and underestimation of parameter variance. The hazard

ratio generated using complete case analysis (where any subjects with missing data were

discarded) reflects a stronger relationship between depression and mortality than the

hazard ratio generated using either of the two imputation techniques, suggesting that

discarding the subjects with missing data biased the hazard ratio attributable to

depression. The confidence intervals surrounding the hazard ratios generated from the

complete case analysis and the simple mean imputation analysis are narrower than the

confidence intervals for the hazard ratio generated from the multiple imputation

technique. For the complete case analysis, one would expect wider confidence intervals

given the reduced sample size. However, the narrow confidence intervals associated with

the simple mean imputation (which resulted in a larger sample size) are a reflection of the

reduced variance associated with this procedure. The simple imputation permitted

analysis using the entire data, but the variance is artificially narrow because the same

value (in this case, the mean) was assigned to all missing values. The wider confidence

intervals of the multiple imputation hazard ratio reflect the fact that multiple imputation

uses the entire data set and factors in the uncertainty related to the true value of the

missing depression values. While all three hazard ratio estimates were not statistically

significant in that the confidence interval for all three included unity, one can imagine a

83

scenario where the reduced variance and bias of a complete case analysis would lead to

false conclusions based on either biased estimates or artificially reduced variance. More

importantly, systematically addressing the impact of missing values adds robustness to

data analyses with missing data.

We have demonstrated that modern statistical methods for handling missing data occur

rarely in published studies in the psychiatric literature. We have also demonstrated that

ignoring missing data, as in complete case analysis, has the potential to introduce bias.

Given the frequent occurrence of missing data in the psychiatric literature, we advocate

adopting practices that would allow readers to determine the impact of missing data and

to analyze the data in a way that takes into account missingness. As such, all studies with

missing data should report whether or not missing data are present and provide

descriptive statistics to describe the pattern of missingness. Furthermore, efforts should

be made to test analytically whether missing data have an impact on estimates generated

from multivariate analyses. There are a number of techniques one can employ if faced

with missing data that is both MAR and is likely to introduce bias102. Multiple

imputation is a robust technique that has the potential to test whether missing data are a

source of bias, especially in the case of data from surveys. Now that multiple imputation

is included in standard software packages (SAS, Cary, North Carolina; SPSS,),

techniques for determining the impact of missing data should be a routine requirement in

studies where missing data are generated to improve the quality of scientific reporting.

Without this practise, readers of the psychiatric literature will not be able to determine the

presence of bias in clinical journals that publish papers with missing data.

84

Depressed

Non-Depressed Missing

P Value

N (%) 494

(17.4) 1447 (50.8) 907

(31.9) Past History of Depression (%) 7.3 4.9 7.6 0.4

Age (mean, SD) 61.4 (13.1) 63.5 (12.7) 66.3

(14.3) Male Sex (%) 62.8 73.1 67.0 0.5 Income, Can $ (%) Low (<30,000) 32.5 23.7 39.8 Intermediate (30,000-59,999) 34.4 34.5 31.3 High (>59,999) 33.1 60.0 24.6 Coronary Risk Factors (%) Diabetes 29.0 21.6 26.8 0.9 Hypercholesterolemia 44.7 38.9 39.8 0.2 Hypertension 49.0 46.4 49.5 0.6 Smoking History 42.9 38.5 41.4 0.9 Medical Comorbidities (%) 0 5.7 8.5 5.4 0.1 1 16.4 22.7 17.6 2 26.5 23.1 21.3 3 or greater 51.4 45.8 55.7

Processes of Care (%) Percutaneous transluinal coronary angiography 8.1 8.2 7.2 0.5

Table 4.1: Clinical and demographic characteristics amongst depressed, non-depressed and observations with

missing depression values.

85

Figure 4.1: Flow chart of literature review of psychiatric journals for missing data.

1236 Articles

9 articles address missing data

222 cross-sectional or cohort studies

1 article used multiple imputation

8 articles used weights

Less than 5% missing – 14.9%

5% to 15% missing – 15.3%

Greater than 15% missing – 69.8%

86

Figure 4.2: Test for Missing at Random (MAR).

Legend: This figure compares the hazard ratio (HR) (95% CI) attributable to depression as a single covariate and the HR for depression in a multivariate model adjusted for demographics (age, sex, income), cardiac risk factors, non-cardiac comorbidities, a prognostic severity index (Global Registry of Acute Coronary Events (GRACE) score), and processes of care (percutaneous, transluminal coronary angiography, coronary artery bypass graft, and drugs at discharge).

87

Figure 4.3: Mortality risk attributable to depression in three different analyses.

Legend: This figure compares the hazard ratio (HR) (95% CI) attributable to depression in three different models (complete case, simple mean imputation, and multiple imputation). The estimate for depression is adjusted for demographics (age, sex, income), cardiac risk factors, non-cardiac comorbidities, a prognostic severity index (Global Registry of Acute Coronary Events (GRACE) score), and processes of care (percutaneous, transluminal coronary angiography, coronary artery bypass graft, and drugs at discharge).

88

Chapter 5: Conclusions


1) Summarize the three studies comprising this thesis and argue that depression is

associated with both mortality and health service outcomes following AMI.

2) Discuss the limitations of the work.

3) Outline some suggestions for future research of depression in cardiovascular

illnesses.

4) Describe the implications of my thesis for clinicians, policy makers, and

researchers.

Summary of Research

The unifying postulate of this thesis is that depression is associated with outcomes

following AMI in specific ways that have implications for understanding comorbidity, for

clinicians, and health service planners. All three studies have contributed to developing

and demonstrating this concept.

First, I examined the relationship between depression and mortality following

AMI. In this study, I demonstrated that in comparison to patients with few depressive

symptoms, patients with a high number of depressive symptoms were at higher risk for

mortality in the two years following AMI. Decreased survival related to depression was

accounted for by self-reported cardiac functional status, as measured by the DASI. As

such, any increase in mortality risk following AMI attributable to depression appears to

be mediated by Self-reported cardiac functional status. These findings should provide

89

impetus for further research exploring the nature of the relationship between depression

and reduced cardiac functional status. Additionally, these findings provide a theoretical

explanation for a recent report indicating that depression interventions do not have any

impact on survival following AMI. The reason that depression interventions do not

improve survival may be due to the fact that depression treatment does not have any

direct impact on the physical health status of depressed patients such that improvements

in survival would be observed.

Second, the relationship between depression and health service consumption

following AMI was explored. In this study, health service consumption was increased

for patients with depression following AMI. Furthermore, the increase in health service

consumption appeared to be disproportionate to measures of cardiac illness severity.

This finding has implications for health service provision to patients with depression

following AMI. Specifically, depressed patients pose a challenge in that, as Chapter 2

revealed, they are at greater risk of dying following AMI due to reduced cardiac

functional status, but consume more health care disproportionate to cardiac illness

severity. Providing appropriate services to depressed patients while preventing excessive

use of services is likely a challenge for health care professional who provide care to

depressed patients and for a health care system in general.

Third, the data used for the first two studies included approximately 30% of

patients who did not respond to the one month follow-up phone survey to generate a

depression score. Many psychiatric studies rely on self-report or administered rating

scales for diagnoses. As such, missing data due to non-response is a recurrent problem.

A review of the literature revealed that a significant proportion of mental health studies

90

that have missing data do an inadequate job describing the data that are missing and using

methodological techniques to assess the impact of missing data on multivariate estimates.

This study advocates for more comprehensive and standardized methods for both

reporting and analyzing data that have a certain proportion of missingness.

These three studies demonstrate a better understanding of the relationship

between depression and outcomes following AMI. Two methodological advances

developed during the course of the thesis. First, competing risk survival analysis was

used. Competing risk survival analysis is a method used to estimate hazard ratios when

the normal assumptions of censoring are likely to be violated. When an outcome of

interest is something other than mortality, censoring a subject who has died prior to the

outcome of interest, mortality “competes” with the outcome of interest in that mortality is

not an independent occurrence. Competing risk survival analysis is an under-utilized

method one can use to test whether normal assumptions related to censoring are

applicable. Second, this study used multiple imputation to determine the impact of

missing depression values due to non-response. Using model-based methods to

determine the impact of missing data is rarely used, as shown by the literature review in

Chapter 4. Multiple imputation is a robust technique that could prove useful to

researchers who are confronted with missing data.

In order to place the findings from these studies into perspective, it would be

useful to return to the conceptual framework proposed in the Chapter 1 (Figure 1.1).

Unlike previous studies on the relationship between depression and mortality, this study

measured whether comorbid medical conditions and/or reduced cardiac functional status

mediated the relationship between depression and mortality following AMI. Including

91

medical comorbidity and reduced cardiac functional status as covariates in our study was

an important contribution because both medical comorbidity113 and reduced cardiac

functional status47 are more prevalent in depressed individuals and both predict mortality

following AMI45, 46, 114, 115.

Cardiac functional status gradients explained the relationship between depression

and mortality following AMI. Indeed, cardiac risk factors and medical comorbidities did

not predict mortality once cardiac functional status was included in the model. The

reduced cardiac functional status related to depression explains the increased mortality

previously observed in studies. Cardiac functional status reflects the capacity of the

cardiovascular system to maximally deliver oxygen to end-organ tissues. Depressed

individuals have predisposing risk factors for reduced cardiac functional status such as an

increased likelihood to smoke116, decreased likelihood to adhere to cardioprotective

lifestyles1, 42, 43, 117, and a higher load of cardiac risk factors118. Implications for this

finding will be discussed below.

The relationship between depression and health service utilization following AMI

is complicated. Unlike the relationship between depression and mortality, depression

was a significant predictor of health service consumption independent of demographics,

cardiac risk factors, socioeconomic status, cardiac illness severity and processes of care.

This finding is in keeping with previous studies measuring depression-related health

service consumption in primary care settings49, 50. However, the interaction between

depression and illness severity with respect to health service utilization is unique.

Specifically, depression is most prominently associated with increased health service

consumption when illness severity is lowest. There have been prior reports of treatment-

92

risk paradoxes, whereby the sickest patients do not receive the most intense services88, 89,

119. However, these studies measure the extent to which patients who should be receiving

health services are not. The study described in Chapter 4 revealed the extent to which

depression increases the likelihood of patients with objectively lesser need for services

are paradoxically more likely to receive them. What is driving this health-seeking

behaviour? Some speculate that amplification of somatic symptoms results in increased

health-seeking behaviour7, 11. Our study does not answer the causes of depression-related

health service consumption gradients. However, our survival analyses on time to first

angina hospitalization following AMI are of interest. Arriving at a diagnosis of angina

means determining the presence of cardiac-related symptoms without the presence of

objective evidence of myocardial infarction. As such, the diagnosis is based not on lab

values, but on the patient’s description of his/her cardiac symptoms and the physician’s

interpretation of these symptoms. In these analyses, the presence of depressive

symptoms strongly and independently predicted angina hospitalization. To test,

indirectly, whether the increase was due to patient-related vs. physician-related factors,

we measured the number of Emergency Department visits that occurred between the

index AMI and the first angina hospitalization. If the increase in angina hospitalizations

were unaffected by including number of Emergency Department visits between index

AMI and angina re-hospitalization, we would have concluded that the increased

likelihood for admission was related to physician factors. Including the number of ED

visits removed the association between depression and angina hospitalization, suggesting

that the increased likelihood is due to patient factors; specifically, depressed patients

frequent the Emergency Department with cardiovascular symptoms resulting in an

93

increased likelihood to be hospitalized. And since this relationship was independent of

two validated cardiac illness severity measures, the increased likelihood to be admitted

for angina was not explained by differences in cardiac illness severity.

One way to conceptualize the role of depression in health service utilization

following AMI is to see it as a modifier of clinical need. Andersen’s Socio-Behavioural

Model of health service utilization is useful for identifying predictors of use to assess

whether services are distributed based on need or other factors such as income and age

120. In particular, Andersen’s model is useful for exploring the interaction between

depression and other factors related to health service utilization. The model identifies

three sets of predictors: predisposing, enabling and need factors. Predisposing factors

include, but are not limited to, demographic variables like gender, marital status, race,

age and education. Andersen argues that factors such as attitudes and beliefs about

seeking care and social-psychological resources are also rightly placed in this category

although, typically, such variables are not included in analyses using this model 120. In

this proposal, predisposing factors will be included principally as confounding effects.

Enabling factors refer to both social and structural resources (e.g., social support from

friends and family and income) that facilitate or impede access to formal care. Like

predisposing factors, enabling factors are included as confounding effects. Need factors

are the symptoms or experience of illness (including severity) that prompt an individual

to seek help. The measurement of need can be based on evaluated (diagnostic) criteria or

perceived criteria.

The paradoxically high service utilization in depressed individuals with low

clinical severity suggests that depressive symptoms act as a need factor. Need is

94

consistently found to be the strongest predictor of service use in Andersen’s frame work

120. However, our findings suggest that depression modifies an individual’s perception of

need to such an extent that service utilization and objective measures of clinical need are

no longer correlated. This represents a unique extension of Andersen’s thesis, by

incorporating a co-morbid perspective into the framework. In other words, depression

interacts with the illness experience (such as pain symptoms) to modify an individual’s

perception of illness severity, prompting help-seeking. This type of interaction is in

keeping with previous suggestions that depression amplifies physical symptoms leading

to increased service utilization 121.

The obvious question raised by the depression-related increases in health service

consumption is what are the needs of a depressed patient following AMI, and to what

extent do the increased health service consumption meet those needs? The answer is

complicated, since depression is associated with having a greater number of cardiac risk

factors, reduced cardiac functional status, and more medical comorbidity, all of which

predict poorer outcomes. However, it is possible that the increased health service use

following AMI is a reflection of psychological distress rather than cardiac illness-related

clinical need. Previous studies looking at “high utilizers” in primary care settings, the

prevalence of depression and anxiety was very high in those amongst the top 10% of

health care utilizers12, and that the expression of medical symptoms without identified

pathology is very commonly associated with depression and anxiety122. At the same

time, depression is routinely undetected in cardiovascular health care settings96, 123,

providing little opportunity to address psychological distress. Preliminary evidence

indicates that systematic detection and treatment of depression may reduce health care

95

consumption78. Further study of interventions to detect and treat depression in this

population are promising and could be cost-effective.

Limitations

One limitation to all the studies is the use of depression measures that have not

been used in major studies assessing the association between depression and various

outcomes following AMI. Indeed, the 9-item depression measure include in the SESAMI

study includes an incomplete sample of the 12 items of the BCDS. The use of the three

depression measures limits the comparability of the findings from the studies presented in

this thesis to other studies using instruments such as the Beck Depression Inventory

(BDI) or the Centre for Epidemiologic Studies Depression Scale (CES-D). However, the

three studies had adequate agreement and high concordance for all of the outcomes in this

study. Similarly, the distribution of depressed and non-depressed across several

demographic and cardiac risk factors for the three scales is comparable to distributions

described in studies using more common depression measures30.

A second limitation was the use of a self-report measure of cardiac functional

status (the DASI 63) in studies outlined in Chapters 2 and 3. Any self-report measure of

cardiac functional status measures degree of effort required to perform tasks. As such,

the presence of depression could confound the presence of low cardiac functional status.

Having said that, the DASI is a valid measure of cardiac functional status in different

patient groups63, 64 Comparing the ability of the DASI to correlate with the gold standard

of exercise stress testing in both depressed and non-depressed subjects would be an

important future step to validate the findings outlined in Chapter 2.

96

All studies used administrative health data. Administrative health data are data

routinely collected when health services are provided. In Ontario, as with other Canadian

provinces, the major advantage of administrative health data is the universal health

insurance coverage. The universality of the health insurance coverage means that data

are routinely collected for the vast majority of Ontario residents. The use of

administrative health data has been criticized for issues of data quality 124. SESAMI uses

administrative health data for some of its measures. In all three chapters, outcomes such

as mortality and health service consumption were derived, at least in part, from

administrative health data. Important covariates such as medical comorbidity were also

derived from administrative health data. To mitigate against data quality criticisms 124,

administrative health data researchers have advocated for validation of administrative

health data 125. SESAMI data are derived from a number of different sources, including

chart abstraction, structured phone interview, and linked administrative health data.

Variables derived from administrative health data have been validated using these other

sources and have been shown to be accurate67.

Finally, while the outcomes were collected prospectively, the independent

variables used to risk adjust or explain the relationship between depression and the

various outcomes were largely collected cross-sectionally. Thus, while my theoretical

framework implies a temporal relationship between the various factors (Fig. 1.1), my

analyses can only demonstrate the degree to which these factors are correlated between

depression and the outcomes of mortality and health service consumption.

97

Directions for Future Research

One issue raised by this thesis is the prognostic impact of cardiac functional status

on the relationship between depression and mortality following AMI. The DASI, a self-

report measure of cardiac functional status, has been validated in a general cardiac illness

sample63. The DASI has also been validated in various patient populations59-65. It has

not been specifically validated in depressed patients with cardiovascular illnesses.

Validating the DASI would require comparing the DASI in a sample of depressed and

non-depressed patients following AMI in whom an objective measure of cardiac

functional status (e.g. VO2peak measured by exercise testing) had been obtained. This

validation exercise would be an important addition to the literature because depressed

individuals may report poor cardiac performance because of the lack of motivation

inherent to depression rather than a true reflection of reduced cardiac functional status.

Additionally, comparing the prognostic significance of the DASI with the objective

measure of VO2peak would be important. It would also be valuable to re-assess the

paradoxical relationship between health service consumption and cardiac illness severity

using an objective measure of cardiac functional status.

Assuming that cardiac functional status is the causal link between depression and

increased mortality following AMI, it is not obvious that depression treatment will

improve survival following AMI (this will be explored in more detail in Implications).

Two depression intervention trials involving AMI patients have not shown improved

survival amongst patients receiving depression treatment15, 76. Preliminary evidence

suggests that depressed patients may benefit from participation in a cardiac rehabilitation

program both in terms of increased survival and improvement in depression symptoms79.

98

In the study by Milani et al. 79, patients were not randomly assigned to participate in

cardiac rehabilitation, so selection bias could affect the interpretation of results.

Given the demonstrated benefits of cardiac rehabilitation and the questionable

benefit of depression-specific treatments on survival following AMI, it would be

unethical to randomize depressed patients to either depression treatment or cardiac

rehabilitation. A more reasonable approach would be to randomize depressed patients

participating in a cardiac rehabilitation program to either depression treatment or to

motivational strategies aimed at improving adherence with a cardiac rehabilitation

program and medications. This type of intervention study would help determine whether

treatment of depression per se improves survival compared with efforts to target the

aspects of depression such as low motivation that detract from the potential benefit

gained from participating in cardiac rehabilitation following AMI. There is evidence to

suggest that depressed patients are not as likely as non-depressed patients to engage in

processes of care that could potentially improve survival following AMI, including

participation in cardiac rehabilitation 42, 77 and adherence to medications117, 126.

Another outcome of interest in this study would be impact of interventions

(cardiac rehabilitation, depression treatment/cognitive therapy, or both) on health service

consumption following AMI. Systematic assessment and treatment of depression in

patients with diabetes has been shown to be both effective in improving depression

outcomes and cost-effective (particularly when depression-free days are given a value in

the analyses) 78. In Chapter 3, among cardiovascular hospitalizations, depression

increased the likelihood of hospitalizations for angina, but not for recurrent AMI. One

reason for this may be that the diagnosis of angina is discretionary and is made after a

99

diagnosis of AMI is ruled out. Depressed patients are more likely to report physical

symptoms and to have greater severity of symptoms compared to non-depressed patients

122. Depressed patients may both be more likely to seek help and less likely to be

reassured when investigations rule out objective signs of clinical illness. All these

factors: physical symptom amplification, increased help-seeking behaviour, and distress

may result in increased likelihood to be hospitalized for increasingly discretionary

indications. Addressing distress and providing coping strategies in a cardiac

rehabilitation setting might be an effective way to reduce discretionary and costly health

service consumption associated with depression following AMI.

Finally, AMI is an acute event that can dramatically change an individual’s life.

Other medical conditions, such as diabetes mellitus or osteoarthritis, are more chronic in

nature. It would be of interest to determine the relationship between depression and

service consumption is a more chronic illness like osteoarthritis to determine whether

increased health service consumption in a distress response to an acute event or whether

it is also seen during the course of a chronic illness that a patient struggles with over the

course of years. Similarly, retrospective assessment of health service consumption pre-

dating AMI using SESAMI data could determine whether the increased health service

consumption is in response to an acute event or more of a longstanding pattern of

behaviour in patients prone to depression.

Implications

Cardiac functional status appears to underlie the relationship between depression

and mortality. In order for treatment of depression to improve survival following AMI,

100

the following are required: 1) factors related to reduced cardiac functional status should

be attributable to depression; 2) the depression-specific factors that explain a reduced

cardiac functional status need to be expressed simultaneously with the expression of

depression symptoms; and 3) The factors are reversed or substantially reduced by

treatment of depression. There is evidence that SSRIs reverse pro-thrombotic platelet

activity associated with depression and the possibility that this activity translates into

improved survival 36. Two randomized controlled trials have not found any benefit of

depression treatment on survival following AMI 76.

Alternatively, there is no biological association between depression and reduced

cardiac functional status beyond that explained by traditional lifestyle and cardiac risk

factors. If this is the case, then improving mortality in depressed patients following AMI

has less to do with identifying factors specific to depression and more to do with ensuring

meaningful participation in interventions that are known to be effective in patients

following AMI, such as cardiac rehabilitation 127, as has been preliminarily demonstrated

79.

If the findings of Chapters 2 and 3 are borne out, namely that reduced cardiac

functional status explains the relationship between depression and decreased survival

following AMI and that depression-related health service consumption following AMI is

disproportionate to measures of cardiac illness severity, the implications for both

physicians and health care policy makers are complex. For physicians, the findings

suggest that proven secondary prevention strategies are likely to benefit patients with

depression following AMI as long as depressed patients are able to participate in such

prevention strategies. Furthermore, relying on objective measures of cardiac illness such

101

as exercise testing will provide a validated means to allocate resources to patients in need

regardless of depression status. Addressing both the reduced cardiac functional status

and increased service consumption using objective measures of clinical need is likely an

overly-simplistic strategy, as discrepancies between clinical need and the likelihood to

receive treatments exist across a number of patient variables in addition to depression 67

81 128 despite the existence of good, objective measures of clinical need. The robust

presence of a treatment-risk paradox indicates that clinical need is not a reliable predictor

of service use 88, 89, 119.

Depression likely plays a role in decreasing survival following AMI through

increased cardiac risk factors and decreased adherence to interventions such as lifestyle

modification, exercise and cardiac rehabilitation. Depression might be viewed as a

cardiac risk factor to the extent that depressed patients are more likely to have traditional

cardiac risk factors and less likely to modify behaviour to reduce the impact of such

traditional cardiac risk factors.

The increased health service consumption related to depression is in keeping with

many other studies suggesting that depression is a costly illness. The health care cost

increases attributable to depression are not limited to mental health service consumption.

Rather, increased costs have been shown to be independent of severity of medical illness

and to be due to increased general medical health care utilization 49. As with the

relationship between depression and mortality following AMI, the mechanisms by which

depression increases service consumption following AMI are not self-evident. A recent

study showed that somatization, the expression of psychological distress as bodily

symptoms, is independently associated with increased health care utilization and costs 90.

102

Somatization is a common feature of depression 48, particularly in depressed patients with

comorbid chronic medical conditions. The increased likelihood for depressed patients to

be hospitalized for angina following AMI could be an example of somatization increasing

health care consumption. If somatization is the driving force behind depression-related

health care consumption, then, as with depression interventions with mortality as an

outcome, depression treatment may not result in substantial reductions in health care

consumption, because it is unclear whether somatization responds to depression treatment

in the same way that depressed mood, sleep disturbance, decreased energy, and other

symptoms of depression respond to treatment.

The findings from these studies are relevant contributions to our understanding of

the relationship between depression and cardiovascular illness and chronic medical

illnesses in general. They have confirmed the cost of depression and suggested an

explanation for the decreased survival following AMI attributable to depression. It

remains to be seen whether interventions aimed at addressing reduced cardiac functional

status and depression-related patient characteristics can both increase survival and better

titrate health care use with need for services.

103

Appendix 1: Overview of SESAMI Depression Measures

Three depression measures were included in the SESAMI follow-up phone

survey. The first two measures were a 9-item and modified 12-item questionnaire both

adapted from the Brief Carroll Depression Rating Scale (BCDRS). The BCDRS is a

depression rating scale that has been validated in the hospitalized medically ill and has a

sensitivity of 92% and a specificity of 89% with a cut-off score of 668. The third was a

psychological well-being scale used in the Global Utilization of Streptokinase and Tissue

Plasminogen Activator for Occluded Coronary Arteries (GUSTO) quality of life sub-

study72 - - a study which served as the template for the SESAMI study design. All

surveys were telephone administered by standardized trained health care personnel

(nurses).

The original validation of the BCDRS was based on a 12-item questionnaire68.

Our primary depression instrument included only nine of the original twelve BCDRS

items. However, we also constructed a modified 12-item BCDRS questionnaire by

supplementing our nine-item questionnaire with three additional questions obtained from

other survey instruments to ensure that all psychosocial domains evaluated in the original

12-item BCDRS items also comprised our modified 12-item questionnaire. This type of

imputation of missing items has been shown to be feasible and valid in other settings129.

The three questions comprising the 12-item BCDRS not included in the SESAMI survey,

related to sleep disturbance, concentration, and self-esteem, and were as follows:: “my

sleep is restless and disturbed”; “I can concentrate easily when reading the paper”; and “I

feel worthless and ashamed about myself”. However, these missing questions were

replaced with the following: “During the past week, how much time have you had trouble

104

sleeping for example, having trouble falling asleep or waking up too early and unable to

get back to sleep”; and “you are a burden on others” (Appendix Table 1). These three

items were abstracted from two data sources: The first two replacement items originated

from a depression measure incorporated within the Global Utilization of Streptokinase

and t-PA for Occluded Coronary Arteries (GUSTO) trial quality of life substudy 72,

whose design served as the foundation for the SESAMI130. The remaining replacement

item originated from the SF-12. For scoring purposes, responses related to the first 2

replacement questions were dichotomized around the median (3+ days per week vs. < 2

days per week). The last replacement question elicited a binary response (Appendix

Table 1).The correlation between our 9-item BCDRS and the modified 12-item BCDRS

questionnaire was high (spearman correlation of 0.95, P<0.0001) (Appendix Table 1).

For the purposes of the study outlined in Chapter 2, depressive symptoms were

stratified into minimal, moderate, and severe depressive symptoms as follows: 1) a score

of 3 or less (minimal), 4 to 6 (moderate), and 7 or greater (severe) on the 9-item scale

with items from the BCDRS; 2) a score of 0 to 4 (minimal), 5 to 8 (moderate), and 9 or

greater (severe) on the modified 12-item BCDRS; and 3) a score between 31 and 40

(minimal), 21 to 30 (moderate) and 20 or less (severe) on the GUSTO psychological

well-being scale. In Chapter 3, the depressive symptoms were dichotomized in the

following way: 1) a score of 5 or greater on the 9-item scale with items from the BCDRS;

2) a score of 6 or greater on the modified 12-item BCDRS; 3) a score of 25 or less on the

GUSTO psychological well-being scale (minimum score of 10, maximum score of 40);

The kappa agreement between the three scales using the dichotomous cut-offs above are

as follows: 9-item Depression scale vs. BCDRS – kappa (95% CI) = 0.81 (0.78-0.84); 9-

105

item Depression Scale vs. GUSTO psychological well-being – kappa (95% CI) = 0.37

(0.32-0.41); and BCDRS vs. GUSTO psychological well-being – kappa (95% CI) = 0.44

(0.39-0.48)(Appendix Table 2). Finally, Appendix Table 3 shows the degree to which

the baseline characteristics are similarly distributed according to the three different

depression measures, further demonstrating the comparability in domains captured by

each of the three depressive instruments. The prevalence of post-AMI depression using

each of these rating scales is similar to those reported using other depression rating

scales16, 22.

106

Appendix Table 1: BCDRS* items (included and missing) and replacement items from SESAMI survey.

Depression Questions Item Questionnaire Depression Construct

I am losing weight. BCDRS Loss of appetite/weight loss

I have dropped many of my interests and activities.

BCDRS Loss of interest

It must be obvious that I am disturbed and agitated.

BCDRS Psychomotor agitation.

I am miserable or often feel like crying.

BCDRS Low mood

I often wish I were dead. BCDRS Suicidal thoughts.

I feel in good spirits. BCDRS Low mood.

I still enjoy my meals. BCDRS Loss of appetite.

I get hardly anything done lately. BCDRS Low motivation.

I am exhausted much of the time. BCDRS Low energy.

My sleep is restless and disturbed. BCDRS (missing) Disturbed sleep

I can concentrate easily when reading the papers.

BCDRS (missing) Poor concentration

I feel worthless and ashamed about myself.

BCDRS (missing) Low self-esteem

Replacement Items

107

During the past week how much of the time have you had trouble sleeping for example, having trouble falling asleep or waking up too early and unable to get back to sleep?*

GUSTO Trial Psychological Well-being Scale[14]

Disturbed sleep

During the past week, how much of the time have you had trouble concentrating or keeping your mind on what you’re doing?*

GUSTO Trial Psychological Well-being Scale[14]

Poor concentration You are a burden on others SF-12

Low self-esteem *The BCDRS and SF-12 items were dichotomous (yes/no). The GUSTO trial items had 4 options from None of the time, some of the time (1-2 days/wk), some of the time (3-4 days/wk), and All/most of the time (5 or more days/wk). The GUSTO items were dichotomized around the median of 3 or more days/wk.

108

Appendix Table 2: Measures of agreement for SESAMI depression measures.

Agreement measures for depression measures

Agreement Measure kappa (95% CI) 9-Item Depression Inventory 12-Item BCDRS12-Item Brief Carroll Depression Inventory 0.81 (0.78-0.84) Duke Psychological Well-being Scale 0.37 (0.32-0.41) 0.44 (0.39-0.48) Correlation of Raw Scores 12-Item Brief Carroll Depression Inventory 0.95 (P<0.001) Duke Psychological Well-being Scale 0.56 (P<0.001) 0.66 (P<0.001)

Legend: The three depression measures are a depression scale containing 9 items from the Brief Carroll Depression Scale (BCDS)(cut-off score of 5), a 12-item depression scale with 3 items replacing the missing items from the BCDS(cut off score of 6), and the Duke Psychological Well-being scale (cut off of 25).

109

Appendix Table 3: Baseline characteristics for three SESAMI depression measures.

Characteristics Depressed Non-Depressed P Value9-Item Depression Scale (n = 494) (n = 1447) Age, mean (SD) 61.4 (13.1) 63.5 (12.7) 0.003 Male Sex -- % 63 73 <0.001Income, Canadian $ -- % <0.001



Non-cardiac comorbidities -- % None 6 9 <0.0011 16 23 2 27 23 3 or more 51 46


Processes of Care -- % Percutaneous transluminal coronary angiography 27 28 0.71 Coronary angioplasty bypass grafting 12 10 0.28 ACE Inhibitor 63 62 0.64 Beta Blocker 68 70 0.44 Statin 55 55 0.89 Nitrate 37 31 0.01

12-Item Depression Scale (n=407) (n=1534) Age, mean (SD) 61.2 (13.2) 63.4 (12.6) 0.21 Male Sex -- % 61 73 <0.001Income, Canadian $ -- % <0.001


Coronary Risk Factors -- % Diabetes 31 21 <0.001Hypercholesterolemia 45 39 0.05 Hypertension 47 47 0.97 Smoking 44 38 0.04

110




Duke Psychological Well-Being Scale (n=553) (n=1388) Age, mean (SD) 62.8 (13.6) 63.0 (12.5) 0.72 Male Sex -- % 61 74 <0.001Income, Canadian $ -- % <0.001






111

Legend: The three depression measures are a depression scale containing 9 items from the Brief Carroll Depression Scale (BCDS)(cut-off score of 5), a 12-item depression scale with 3 items replacing the missing items from the BCDS (cut off score of 6), and the Duke Psychological Well-being scale (cut off of 25). Abbreviation: GRACE: Global Registry of Acute Coronary Events scale. Abbreviation: DASI: Duke Activity Status Index

112

References

1. Evans DL, Charney DS, Lewis L, et al. Mood disorders in the medically ill:

scientific review and recommendations. Biol Psychiatry 2005; 58, 175-189.

2. Wells KB, Golding JM, Burnam MA. Psychiatric disorder in a sample of the

general population with and without chronic medical conditions. Am J Psychiatry

1988; 145, 976-981.

3. Murray CJ, Lopez AD. Global mortality, disability, and the contribution of risk

factors: Global Burden of Disease Study. Lancet 1997; 349, 1436-1442.

4. Wells KB, Stewart A, Hays RD, et al. The functioning and well-being of depressed

patients. Results from the Medical Outcomes Study. JAMA 1989; 262, 914-919.

5. Unutzer J, Patrick DL, Diehr P, Simon G, Grembowski D, Katon W. Quality

adjusted life years in older adults with depressive symptoms and chronic medical

disorders. Int Psychogeriatr 2000; 12, 15-33.

6. Diagnostic and statistical manual of mental disorders : DSM-IV-TR. Washington,

DC: American Psychiatric Association, 2000.

7. Katon W, Lin EH, Kroenke K. The association of depression and anxiety with

medical symptom burden in patients with chronic medical illness. Gen Hosp

Psychiatry 2007; 29, 147-155.

8. Callahan CM, Hui SL, Nienaber NA, Musick BS, Tierney WM. Longitudinal study

of depression and health services use among elderly primary care patients. J Am

Geriatr Soc 1994; 42, 833-838.

9. Simon GE, VonKorff M, Barlow W. Health care costs of primary care patients with

recognized depression. Arch Gen Psychiatry 1995; 52, 850-856.

113

10. Unutzer J, Katon W, Callahan CM, et al. Collaborative care management of late-life

depression in the primary care setting: a randomized controlled trial. JAMA 2002;

288, 2836-2845.

11. Katon W, Ciechanowski P. Impact of major depression on chronic medical illness. J

Psychosom Res 2002; 53, 859-863.

12. Katon W, Von Korff M, Lin E, et al. Distressed high utilizers of medical care.

DSM-III-R diagnoses and treatment needs. Gen Hosp Psychiatry 1990; 12, 355-

362.

13. Levenson JL, Hamer RM, Rossiter LF. Relation of psychopathology in general

medical inpatients to use and cost of services. Am J Psychiatry 1990; 147, 1498-

1503.

14. Saravay SM, Pollack S, Steinberg MD, Weinschel B, Habert M. Four-year follow-

up of the influence of psychological comorbidity on medical rehospitalization. Am

J Psychiatry 1996; 153, 397-403.

15. Berkman LF, Blumenthal J, Burg M, et al. Effects of treating depression and low

perceived social support on clinical events after myocardial infarction: the

Enhancing Recovery in Coronary Heart Disease Patients (ENRICHD) Randomized

Trial. JAMA 2003; 289, 3106-3116.

16. Schleifer SJ, Macari-Hinson MM, Coyle DA, et al. The nature and course of

depression following myocardial infarction. Arch Intern Med 1989; 149, 1785-

1789.

114

17. Kessler RC, McGonagle KA, Zhao S, et al. Lifetime and 12-month prevalence of

DSM-III-R psychiatric disorders in the United States. Results from the National

Comorbidity Survey. Arch Gen Psychiatry 1994; 51, 8-19.

18. Lauzon C, Beck CA, Huynh T, et al. Depression and prognosis following hospital

admission because of acute myocardial infarction. CMAJ 2003; 168, 547-552.

19. Frasure-Smith N, Lesperance F. Depression and anxiety as predictors of 2-year

cardiac events in patients with stable coronary artery disease. Arch Gen Psychiatry

2008; 65, 62-71.

20. Parashar S, Rumsfeld JS, Spertus JA, et al. Time course of depression and outcome

of myocardial infarction. Arch Intern Med 2006; 166, 2035-2043.

21. Lett HS, Blumenthal JA, Babyak MA, et al. Depression as a risk factor for coronary

artery disease: evidence, mechanisms, and treatment. Psychosom Med 2004; 66,

305-315.

22. Frasure-Smith N, Lesperance F, Talajic M. Depression and 18-month prognosis

after myocardial infarction. Circulation 1995; 91, 999-1005.

23. Irvine J, Basinski A, Baker B, et al. Depression and risk of sudden cardiac death

after acute myocardial infarction: testing for the confounding effects of fatigue.

Psychosom Med 1999; 61, 729-737.

24. Ladwig KH, Kieser M, Konig J, Breithardt G, Borggrefe M. Affective disorders and

survival after acute myocardial infarction. Results from the post-infarction late

potential study. Eur Heart J 1991; 12, 959-964.

115

25. Dickens C, McGowan L, Percival C, et al. Depression is a risk factor for mortality

after myocardial infarction: fact or artifact? J Am Coll Cardiol 2007; 49, 1834-

1840.

26. Lane D, Carroll D, Ring C, Beevers DG, Lip GY. Mortality and quality of life 12

months after myocardial infarction: effects of depression and anxiety. Psychosom

Med 2001; 63, 221-230.

27. Lane D, Carroll D, Ring C, Beevers DG, Lip GY. In-hospital symptoms of

depression do not predict mortality 3 years after myocardial infarction. Int J

Epidemiol 2002; 31, 1179-1182.

28. Sorensenf C, Friis-Hasche E, Haghfelt T, Bech P. Postmyocardial infarction

mortality in relation to depression: a systematic critical review. Psychother

Psychosom 2005; 74, 69-80.

29. Parakh K, Thombs BD, Fauerbach JA, Bush DE, Ziegelstein RC. Effect of

depression on late (8 years) mortality after myocardial infarction. Am J Cardiol

2008; 101, 602-606.

30. Frasure-Smith N, Lesperance F, Talajic M. Depression following myocardial

infarction. Impact on 6-month survival. JAMA 1993; 270, 1819-1825.

31. Barefoot JC, Schroll M. Symptoms of depression, acute myocardial infarction, and

total mortality in a community sample. Circulation 1996; 93, 1976-1980.

32. Frasure-Smith N, Lesperance F, Juneau M, Talajic M, Bourassa MG. Gender,

depression, and one-year prognosis after myocardial infarction. Psychosom Med

1999; 61, 26-37.

116

33. Welin C, Lappas G, Wilhelmsen L. Independent importance of psychosocial factors

for prognosis after myocardial infarction. J Intern Med 2000; 247, 629-639.

34. Lesperance F, Frasure-Smith N, Talajic M, Bourassa MG. Five-year risk of cardiac

mortality in relation to initial severity and one-year changes in depression

symptoms after myocardial infarction. Circulation 2002; 105, 1049-1053.

35. Nemeroff CB, Musselman DL. Are platelets the link between depression and

ischemic heart disease? Am Heart J 2000; 140, 57-62.

36. Sauer WH, Berlin JA, Kimmel SE. Effect of antidepressants and their relative

affinity for the serotonin transporter on the risk of myocardial infarction.

Circulation 2003; 108, 32-36.

37. Shimbo D, Child J, Davidson K, et al. Exaggerated serotonin-mediated platelet

reactivity as a possible link in depression and acute coronary syndromes. Am J

Cardiol 2002; 89, 331-333.

38. Carney RM, Howells WB, Blumenthal JA, et al. Heart rate turbulence, depression,

and survival after acute myocardial infarction. Psychosom Med 2007; 69, 4-9.

39. Carney RM, Saunders RD, Freedland KE, Stein P, Rich MW, Jaffe AS. Association

of depression with reduced heart rate variability in coronary artery disease. Am J

Cardiol 1995; 76, 562-564.

40. Stein PK, Carney RM, Freedland KE, et al. Severe depression is associated with

markedly reduced heart rate variability in patients with stable coronary heart

disease. J Psychosom Res 2000; 48, 493-500.

41. Marrugat J, Sala J, Masia R, et al. Mortality differences between men and women

following first myocardial infarction. RESCATE Investigators. Recursos

117

Empleados en el Sindrome Coronario Agudo y Tiempo de Espera. JAMA 1998;

280, 1405-1409.

42. Blumenthal JA, Williams RS, Wallace AG, Williams RBJ, Needles TL.

Physiological and psychological variables predict compliance to prescribed exercise

therapy in patients recovering from myocardial infarction. Psychosom Med 1982;

44, 519-527.

43. DiMatteo MR, Lepper HS, Croghan TW. Depression is a risk factor for

noncompliance with medical treatment: meta-analysis of the effects of anxiety and

depression on patient adherence. Arch Intern Med 2000; 160, 2101-2107.

44. Rieckmann N, Kronish IM, Haas D, et al. Persistent depressive symptoms lower

aspirin adherence after acute coronary syndromes. Am Heart J 2006; 152, 922-927.

45. Kavanagh T, Mertens DJ, Hamm LF, et al. Prediction of long-term prognosis in 12

169 men referred for cardiac rehabilitation. Circulation 2002; 106, 666-671.

46. Laukkanen JA, Lakka TA, Rauramaa R, et al. Cardiovascular fitness as a predictor

of mortality in men. Arch Intern Med 2001; 161, 825-831.

47. Spertus JA, McDonell M, Woodman CL, Fihn SD. Association between depression

and worse disease-specific functional status in outpatients with coronary artery

disease. Am Heart J 2000; 140, 105-110.

48. Katon W. The impact of major depression on chronic medical illness. Gen Hosp

Psychiatry 1996; 18, 215-219.

49. Simon G, Ormel J, VonKorff M, Barlow W. Health care costs associated with

depressive and anxiety disorders in primary care. Am J Psychiatry 1995; 152, 352-

357.

118

50. Unutzer J, Patrick DL, Simon G, et al. Depressive symptoms and the cost of health

services in HMO patients aged 65 years and older. A 4-year prospective study.

JAMA 1997; 277, 1618-1623.

51. Katzelnick DJ, Simon GE, Pearson SD, et al. Randomized trial of a depression

management program in high utilizers of medical care. Arch Fam Med 2000; 9,

345-351.

52. Jiang W, Alexander J, Christopher E, et al. Relationship of depression to increased

risk of mortality and rehospitalization in patients with congestive heart failure. Arch

Intern Med 2001; 161, 1849-1856.

53. Sullivan M, Simon G, Spertus J, Russo J. Depression-related costs in heart failure

care. Arch Intern Med 2002; 162, 1860-1866.

54. Frasure-Smith N, Lesperance F, Gravel G, et al. Depression and health-care costs

during the first year following myocardial infarction. J Psychosom Res 2000; 48,

471-478.

55. Bush DE, Ziegelstein RC, Patel UV, et al. Post-myocardial infarction depression.

Rockville, MD: Agency for Healthcare Research and Quality, 2005.

56. Alter DA, Chong A, Austin PC, et al. Socioeconomic status and mortality after

acute myocardial infarction. Ann Intern Med 2006; 144, 82-93.

57. Little RJA, Rubin D. Statistical Analysis with Missing Data, second edition. New

York: Wiley, 2002.

58. Thombs BD, de Jonge P, Coyne JC, et al. Depression screening and patient

outcomes in cardiovascular care: a systematic review. JAMA 2008; 300, 2161-

2171.

119

59. Carter R, Holiday DB, Grothues C, Nwasuruba C, Stocks J, Tiep B. Criterion

validity of the Duke Activity Status Index for assessing functional capacity in

patients with chronic obstructive pulmonary disease. J Cardiopulm Rehabil 2002;

22, 298-308.

60. Eisenberg MJ, Teng FF, Chaudhry MR, et al. Impact of invasive management

versus noninvasive management on functional status and quality of life following

non-Q-wave myocardial infarction: a randomized clinical trial. Am Heart J 2005;

149, 813-819.

61. Hamilton DM, Haennel RG. Validity and reliability of the 6-minute walk test in a

cardiac rehabilitation population. J Cardiopulm Rehabil 2000; 20, 156-164.

62. Hedeshian MH, Namour N, Dziadik E, Stewart RD, Campos CT. Does increasing

age have a negative impact on six-month functional outcome after coronary artery

bypass? Surgery 2002; 132, 239-244.

63. Hlatky MA, Boineau RE, Higginbotham MB, et al. A brief self-administered

questionnaire to determine functional capacity (the Duke Activity Status Index).

Am J Cardiol 1989; 64, 651-654.

64. Shaw LJ, Olson MB, Kip K, et al. The value of estimated functional capacity in

estimating outcome: results from the NHBLI-Sponsored Women's Ischemia

Syndrome Evaluation (WISE) Study. J Am Coll Cardiol 2006; 47, S36-43.

65. Wessel TR, Arant CB, Olson MB, et al. Relationship of physical fitness vs body

mass index with coronary artery disease and cardiovascular events in women.

JAMA 2004; 292, 1179-1187.

120

66. Koch CG, Li L, Lauer M, Sabik J, Starr NJ, Blackstone EH. Effect of functional

health-related quality of life on long-term survival after cardiac surgery. Circulation

2007; 115, 692-699.

67. Alter DA, Iron K, Austin PC, Naylor CD. Socioeconomic status, service patterns,

and perceptions of care among survivors of acute myocardial infarction in Canada.

JAMA 2004; 291, 1100-1107.

68. Koenig HG, Meador KG, Cohen HJ, Blazer DG. Self-rated depression scales and

screening for major depression in the older hospitalized patient with medical illness.

J Am Geriatr Soc 1988; 36, 699-706.

69. Fox KA, Dabbous OH, Goldberg RJ, et al. Prediction of risk of death and

myocardial infarction in the six months after presentation with acute coronary

syndrome: prospective multinational observational study (GRACE). BMJ 2006;

333, 1091.

70. Alter DA, Venkatesh V, Chong A. Evaluating the performance of the Global

Registry of Acute Coronary Events risk-adjustment index across socioeconomic

strata among patients discharged from the hospital after acute myocardial infarction.

Am Heart J 2006; 151, 323-331.

71. Sytkowski PA, Kannel WB, D'Agostino RB. Changes in risk factors and the decline

in mortality from cardiovascular disease. The Framingham Heart Study. N Engl J

Med 1990; 322, 1635-1641.

72. Mark DB, Naylor CD, Hlatky MA, et al. Use of medical resources and quality of

life after acute myocardial infarction in Canada and the United States. N Engl J

Med 1994; 331, 1130-1135.

121

73. Austin PC, Daly PA, Tu JV. A multicenter study of the coding accuracy of hospital

discharge administrative data for patients admitted to cardiac care units in Ontario.

Am Heart J 2002; 144, 290-296.

74. Myers J, Prakash M, Froelicher V, Do D, Partington S, Atwood JE. Exercise

capacity and mortality among men referred for exercise testing. N Engl J Med

2002; 346, 793-801.

75. Gulati M, Pandey DK, Arnsdorf MF, et al. Exercise capacity and the risk of death in

women: the St James Women Take Heart Project. Circulation 2003; 108, 1554-

1559.

76. Glassman AH, O'Connor CM, Califf RM, et al. Sertraline treatment of major

depression in patients with acute MI or unstable angina. JAMA 2002; 288, 701-709.

77. Glazer KM, Emery CF, Frid DJ, Banyasz RE. Psychological predictors of

adherence and outcomes among patients in cardiac rehabilitation. J Cardiopulm

Rehabil 2002; 22, 40-46.

78. Simon GE, Katon WJ, Lin EH, et al. Cost-effectiveness of systematic depression

treatment among people with diabetes mellitus. Arch Gen Psychiatry 2007; 64, 65-

72.

79. Milani RV, Lavie CJ. Impact of cardiac rehabilitation on depression and its

associated mortality. Am J Med 2007; 120, 799-806.

80. Granger CB, Goldberg RJ, Dabbous O, et al. Predictors of hospital mortality in the

global registry of acute coronary events. Arch Intern Med 2003; 163, 2345-2353.

122

81. Chen J, Rathore SS, Radford MJ, Wang Y, Krumholz HM. Racial differences in the

use of cardiac catheterization after acute myocardial infarction. N Engl J Med 2001;

344, 1443-1449.

82. Act CH. Chapter C-6.

83. Vanhees L, Fagard R, Thijs L, Staessen J, Amery A. Prognostic significance of

peak exercise capacity in patients with coronary artery disease. J Am Coll Cardiol

1994; 23, 358-363.

84. Gravelly-Witte S, Stewart DE, Suskin N, Higginson L, Alter DA, Grace SL.

Cardiologists' charting varied by risk factor, and was often discordant with patient

report. J Clin Epidemiol 2008; In press,

85. Zhang JX, Iwashyna TJ, Christakis NA. The performance of different lookback

periods and sources of information for Charlson comorbidity adjustment in

Medicare claims. Med Care 1999; 37, 1128-1139.

86. Kerr EA, Heisler M, Krein SL, et al. Beyond comorbidity counts: how do

comorbidity type and severity influence diabetes patients' treatment priorities and

self-management? J Gen Intern Med 2007; 22, 1635-1640.

87. van Melle JP, de Jonge P, Spijkerman TA, et al. Prognostic association of

depression following myocardial infarction with mortality and cardiovascular

events: a meta-analysis. Psychosom Med 2004; 66, 814-822.

88. Ko DT, Mamdani M, Alter DA. Lipid-lowering therapy with statins in high-risk

elderly patients: the treatment-risk paradox. JAMA 2004; 291, 1864-1870.

89. Lee DS, Tu JV, Juurlink DN, et al. Risk-treatment mismatch in the

pharmacotherapy of heart failure. JAMA 2005; 294, 1240-1247.

123

90. Barsky AJ, Orav EJ, Bates DW. Somatization increases medical utilization and

costs independent of psychiatric and medical comorbidity. Arch Gen Psychiatry

2005; 62, 903-910.

91. Guadagnoli E, Hauptman PJ, Ayanian JZ, Pashos CL, McNeil BJ, Cleary PD.

Variation in the use of cardiac procedures after acute myocardial infarction. N Engl

J Med 1995; 333, 573-578.

92. Muller DW, Topol EJ, Califf RM, et al. Relationship between antecedent angina

pectoris and short-term prognosis after thrombolytic therapy for acute myocardial

infarction. Thrombolysis and Angioplasty in Myocardial Infarction (TAMI) Study

Group. Am Heart J 1990; 119, 224-231.

93. Clark AM, Hartling L, Vandermeer B, McAlister FA. Meta-analysis: secondary

prevention programs for patients with coronary artery disease. Ann Intern Med

2005; 143, 659-672.

94. Ciechanowski P, Wagner E, Schmaling K, et al. Community-integrated home-based

depression treatment in older adults: a randomized controlled trial. JAMA 2004;

291, 1569-1577.

95. Thombs BD, Bass EB, Ford DE, et al. Prevalence of depression in survivors of

acute myocardial infarction. J Gen Intern Med 2006; 21, 30-38.

96. Amin AA, Jones AM, Nugent K, Rumsfeld JS, Spertus JA. The prevalence of

unrecognized depression in patients with acute coronary syndrome. Am Heart J

2006; 152, 928-934.

97. Kurdyak PA, Gnam WH, Goering P, Chong A, Alter DA. The relationship between

depressive symptoms, health service consumption, and prognosis after acute

124

myocardial infarction: a prospective cohort study. BMC Health Serv Res 2008; 8,

200.

98. Rubin DB. Formalizing subjective notions about the effect of nonrespondents in

sample surveys. Journal of the American Statistical Association 1977; 72, 538-543.

99. Rubin DB. Multiple imputation after 18+ years. Journal of the American Statistical

Association 1996; 91, 473-488.

100. Harrell FE. Missing Data. ed. Regression Modeling Strategies. New York: Springer,

2001:41-52.

101. Rubin DB, Schenker N. Multiple imputation in health-care databases: an overview

and some applications. Stat Med 1991; 10, 585-598.

102. Little RJA. Regression with missing X's: a review. Journal of the American

Statistical Association 1992; 87, 1227-1237.

103. Robins JM, Rotnitzky A, Zhao LP. Analysis of semiparametric regression models

for repeated outcomes in the presence of missing data. Journal of the American

Statistical Association 1995; 90, 106-121.

104. Preisser JS, Lohman KK, Rathouz PJ. Performance of weighted estimating

equations for longitudinal binary data with drop-outs missing at random. Stat Med

2002; 21, 3035-3054.

105. Rao JNK, Shao, J. Jackknife variance esimation with survey data under hot deck

imputation. Biometrika 1992; 79, 811-822.

106. Reilly M, Pepe M. The relationship between hot-deck multiple imputation and

weighted likelihood. Stat Med 1997; 16, 5-19.

125

107. Mander A, Clayton D. HOTDECK: Stata module to impute missing values using

the hotdeck method. 1999;

108. SPSS. SPSS Press Releases. 1999;

109. SAS. SAS System, Version 9: Installation and Administration Information. 2007;

110. Yuan YC. Multiple imputation for missing data: concepts and new development.

111. Gueorguieva R, Krystal JH. Move over ANOVA: progress in analyzing repeated-

measures data and its reflection in papers published in the Archives of General

Psychiatry. Arch Gen Psychiatry 2004; 61, 310-317.

112. Alter DA, Naylor CD, Austin P, Tu JV. Effects of socioeconomic status on access

to invasive cardiac procedures and on mortality after acute myocardial infarction. N

Engl J Med 1999; 341, 1359-1367.

113. Wells KB, Golding JM, Burnam MA. Chronic medical conditions in a sample of

the general population with anxiety, affective, and substance use disorders. Am J

Psychiatry 1989; 146, 1440-1446.

114. Chen J, Radford MJ, Wang Y, Krumholz HM. Care and outcomes of elderly

patients with acute myocardial infarction by physician specialty: the effects of

comorbidity and functional limitations. Am J Med 2000; 108, 460-469.

115. Krumholz HM, Chen J, Chen YT, Wang Y, Radford MJ. Predicting one-year

mortality among elderly survivors of hospitalization for an acute myocardial

infarction: results from the Cooperative Cardiovascular Project. J Am Coll Cardiol

2001; 38, 453-459.

126

116. Wilhelm K, Wedgwood L, Niven H, Kay-Lambkin F. Smoking cessation and

depression: current knowledge and future directions. Drug Alcohol Rev 2006; 25,

97-107.

117. Carney RM, Freedland KE, Eisen SA, Rich MW, Jaffe AS. Major depression and

medication adherence in elderly patients with coronary artery disease. Health

Psychol 1995; 14, 88-90.

118. Katon WJ, Lin EH, Russo J, et al. Cardiac risk factors in patients with diabetes

mellitus and major depression. J Gen Intern Med 2004; 19, 1192-1199.

119. McAlister FA, Oreopoulos A, Norris CM, et al. Exploring the treatment-risk

paradox in coronary disease. Arch Intern Med 2007; 167, 1019-1025.

120. Andersen RM. Revisiting the behavioral model and access to medical care: does it

matter? J Health Soc Behav 1995; 36, 1-10.

121. Katon WJ. Clinical and health services relationships between major depression,

depressive symptoms, and general medical illness. Biol Psychiatry 2003; 54, 216-

226.

122. Katon W, Sullivan M, Walker E. Medical symptoms without identified pathology:

relationship to psychiatric disorders, childhood and adult trauma, and personality

traits. Ann Intern Med 2001; 134, 917-925.

123. Ziegelstein RC, Kim SY, Kao D, et al. Can doctors and nurses recognize depression

in patients hospitalized with an acute myocardial infarction in the absence of formal

screening? Psychosom Med 2005; 67, 393-397.

124. Iezzoni LI. Assessing quality using administrative data. Ann Intern Med 1997; 127,

666-674.

127

125. Williams J, Young W. A summary of studies on the quality of health care

administrative databases in Canada. In: Goel V, Williams J, Anderson G,

Blackstien-Hirsch P, Fooks C, Naylor D, eds. Patterns of health care in Ontario. 2nd

ed. Ottawa: Canadian Medical Association, 1996

126. Wang PS, Bohn RL, Knight E, Glynn RJ, Mogun H, Avorn J. Noncompliance with

antihypertensive medications: the impact of depressive symptoms and psychosocial

factors. J Gen Intern Med 2002; 17, 504-511.

127. Ades PA. Cardiac rehabilitation and secondary prevention of coronary heart

disease. N Engl J Med 2001; 345, 892-902.

128. Filate WA, Johansen HL, Kennedy CC, Tu JV. Regional variations in

cardiovascular mortality in Canada. Can J Cardiol 2003; 19, 1241-1248.

129. Eng K JA, Feeny D. Missing data and methods for imputing missing depression

scores: experience in a study of hop fracture patients. 2006;

130. Alter DA, Brandes S, Irvine J, Iron K. Impact of socioeconomic status on

cardiovascular outcomes in Canada. Expert Review of Pharmacoeconomics

Outcomes Research 2003; 3, 691-702.

the impact of depression on outcomes ......iv acknowledgements when embarking on a phd dissertation,...

Documents