modeling sources of random and systematic error
TRANSCRIPT
Modeling Sources of Random and Systematic Error
Hans Baumgartner Penn State University
Sources of random and systematic error
Overview when a researcher conducts a survey, the
expectation is that the information collected will yield a faithful representation of reality;
unfortunately, this is often not the case, and survey researchers have identified many different sources of error in surveys;
these errors may contaminate the research results and limit the managerial usefulness of the findings;
Sources of random and systematic error
Sources of error in survey research coverage (frame) error: the sampling frame used
does not match the target population; sampling error: the survey is based on a sample of
respondents from the entire population; nonresponse error: only a subset of all respondents
who are contacted actually responds to the survey; measurement error: the obtained response does not
fully reflect the “true” response; □ components due to the respondent, the method, and
the context; □ random or systematic;
Sources of random and systematic error
The survey measurement process
Construct
Measure
Response
Sources of random and systematic error
Constructs, measures and responses Construct: the phenomenon of interest or focal concept
that the survey researcher wishes to measure; □ can be observable or unobservable; □ requires the specification of the conceptual meaning of the
construct; □ all facets of the target construct have to be represented and
overlap with other constructs has to be avoided;
Measure: observed indicators of the construct (e.g., a question in a survey designed to tap the target construct); □ generation of items that represent the construct □ reflective vs. formative indicators
Response: the answer to a survey question provided by the respondent;
Sources of random and systematic error
focal construct
Examples of different reflective or formative measurement models
focal construct
focal construct
focal construct
Sources of random and systematic error
Criteria for distinguishing between reflective and formative indicator models
Are the indicators manifestations of the underlying construct or defining characteristics of it?
Are the indicators conceptually interchangeable? Are the indicators expected to covary? Are all of the indicators expected to have the
same antecedents and/or consequences?
Based on MacKenzie, Podsakoff, and Jarvis (2005)
Sources of random and systematic error
The relationship between observed measurements and constructs of interest:
Single-item measures The observed single-item LS
score is a perfect measure of “true” LS. All of the variability in observed
scores is trait (substantive) variance.
Life satisfaction
Measure of Life satisfaction
(e.g., I am satisfied with my life.) T1 T2
Sources of random and systematic error
When single-item measures might be sufficient (Rossiter 2002; Bergkvist and Rossiter 2007)
measures of marketing constructs often involve two things: □ the object of the construct (e.g., ad, brand, company) □ an attribute of the construct (e.g., attitude, quality, liking)
if the object is “concrete singular” (i.e., easily and uniformly imagined) and the attribute is also “concrete” (i.e., easily and uniformly imagined) – in which case the construct is “doubly concrete” – single-item measures are sufficient;
Thinking about the ad for /BRAND/, which of the following
statements best describes your feelings about the ad? ______ ______ ______ ______ ______ I disliked it I disliked it. I neither liked it I liked it. I liked it very much. nor disliked it. very much.
Sources of random and systematic error
When single-item measures are insufficient a construct is “abstract” if
□ the object of the construct consists of more than one dimension (e.g., materialism) or several sub-objects (e.g., elements of job satisfaction) [e.g., How materialistic are you? How satisfied are you with your job?], and/or
□ the attribute of the object consists of more than one dimension (e.g., service quality) [e.g., How good is the company’s service?]
for abstract constructs, multiple items are needed to capture the multiple objects and/or multiple attributes;
Sources of random and systematic error
The relationship between observed measurements and constructs of interest:
Random measurement error The observed life satisfaction
score is contaminated by random measurement error.
If only a single measure is available, random measurement error cannot be taken into account.
Life satisfaction
Measure of Life satisfaction
ε
Sources of random and systematic error
T
E
The relationship between observed measurements and constructs of interest (cont’d)
The total variability of observed scores consists of both trait (substantive) variance and random error variance.
This results in unreliability of measurement and the attenuation of observed correlations.
T1 T2 E2 E1
Sources of random and systematic error
The relationship between observed measurements and constructs of interest (cont’d)
Life satisfaction
Life satisfaction measure 1
ε1
Life satisfaction measure 2
ε2
Life satisfaction measure 3
ε3
λ1 λ3 λ2
Solution: Use multiple indicators to measure the focal construct, in which case we can assess reliability and correct for attenuation.
Sources of random and systematic error
ls-pa ls-na pa-na
Correlations between Individual measures of ls, pa, and na (ls3, pa2, na2)
Correlations between averages of five ls, pa, and na measures
Correlations between averages of ls, pa, and na (corrected for attenuation with α)
Disattenuated correlations based on CFA
Correlations between ls, pa and na
Sources of random and systematic error
Reliability and average variance extracted of LS, PA and NA measures
LS: o Coefficient alpha: o Average variance extracted:
PA: o Coefficient alpha: o Average variance extracted:
NA: o Coefficient alpha: o Average variance extracted:
Sources of random and systematic error
T
E
M
T1 T2
E2 E1
M1 M2
The relationship between observed measurements and constructs of interest:
Systematic measurement error The total variability of observed
scores consists of trait (substantive), random error, and systematic error (method) variance.
This is likely to confound the assessment of reliability and relationships with other constructs.
It also complicates the comparison of means.
Sources of random and systematic error
Sources of systematic error in surveys Respondent factors:
response styles (acquiescence and disacquiescence, extreme responding, midpoint responding), social desirability, consistency bias, implicit theories, leniency bias, positive or negative affectivity, etc.
Item characteristics: item reversal and negation, common scale formats and anchors, item demand, etc.
Item context effects: item arrangement, scale length, etc.
General context effects: time pressure, mood effects, etc.
Sources of random and systematic error
Procedural remedies for common method bias
Use different sources to measure different constructs
Separate the measurement of different constructs (time, position, cover story)
Eliminate common scale properties Improve item wording (item ambiguity, item social
desirability, balancing scales)
Sources of random and systematic error
Statistical control of systematic error explicit vs. implicit control of systematic error
(depending on whether or not the source of the bias can be identified and measured);
correction at the scale level or individual item level specification of a single source of systematic error or
multiple sources (e.g., one or more method factors) for implicit control and correction at the individual
item level, specification of a method factor or correlated errors
for explicit control, measurement error in the method factor is or is not taken into account
Sources of random and systematic error
explicit vs. implicit control of ARS (at the individual item level):
Statistical control of systematic error in the case of ARS
p1 p2 n1 n2 p3 p4 n3 n4
A B
“ARS”
+ + - + + - - -
p1 p2 n1 n2 p3 p4 n3 n4
A B
ARS
+ + - + + - - -
Based on a direct measure of ARS
Based on an inferred common ARS factor
Sources of random and systematic error
correction at the scale level vs. individual item level (using a direct ARS measure):
Statistical control of systematic error in the case of ARS
p1 p2 n1 n2 p3 p4 n3 n4
A B
ARS
+ + - + + - - -
Item level Scale level
ARS
A B
Sources of random and systematic error
specification of multiple sources of inconsistency bias (at the individual item level):
Statistical control of systematic error in the case of ARS
p1 p2 n1 n2 p3 p4 n3 n4
A B
INCON
+ + - + + - - -
ARS IMC
Sources of random and systematic error
method factors vs. correlated errors:
Statistical control of systematic error in the case of ARS
p1 p2 n1 n2 p3 p4 n3 n4
A B
“ARS”
+ + - + + - - -
Correlated errors Method factor
p1 p2 n1 n2 p3 p4 n3 n4
A B + + - + + - - -
Sources of random and systematic error
taking into account measurement error in a directly measured method (ARS) factor:
Statistical control of systematic error in the case of ARS
p1 p2 n1 n2 p3 p4 n3 n4
A B
ARS
+ + - + + - - -
ars1 ars2 ars3
Sources of random and systematic error
Statistical control of systematic error in the case of ARS
modeling more complicated effects of systematic error:
ARS
A B
ζγγγγ ++++= )()(3210 ARSAARSAB
Sources of random and systematic error
Exercise: Statistical remedies for common method bias
Harman’s single-factor test Partial correlation procedures (with control at scale level)
□ Implicit control (e.g., general method factor, marker variable) or explicit control (e.g., SD)
Single-method scale score approaches (with control at item level) □ Implicit control or explicit control (no correction for
measurement error in control variable) □ Implicit control or explicit control (with correction for
measurement error in control variable) Multi-method factor approaches (with control at item level)
□ Implicit control or explicit control (with or without control for measurement error)
Sources of random and systematic error
Recommendations for the statistical control of systematic error
if a survey is known or expected to be susceptible to specific biases, try to measure the source of the bias directly (e.g., social desirability);
correct for systematic error at the individual item level, if possible;
consider multiple potential sources of systematic error; use method factors rather than correlated errors; measurement error in the method factor can be
ignored if reliability is adequate; controlling for the linear effects of systematic error is
not always sufficient;