Download - Nonindependence David A. Kenny
![Page 1: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/1.jpg)
NonindependenceDavid A. Kenny
October 1, 2012
![Page 2: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/2.jpg)
Types of Linkages– Temporal– Spatial– Network– Group or Dyad
![Page 3: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/3.jpg)
Temporal Linkage
![Page 4: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/4.jpg)
Spatial Linkage
![Page 5: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/5.jpg)
Network Linkage
![Page 6: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/6.jpg)
6
![Page 7: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/7.jpg)
7
![Page 8: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/8.jpg)
Defining Nonindependence• In general –When two linked scores are more similar to (or
different from) two unlinked scores, the data are nonindependent.– The strength of the links is the measure of
nonindependence.• In the Standard Dyadic Design– The two scores on the same variable from the
two members of the dyad are correlated.
8
![Page 9: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/9.jpg)
Negative Nonindependence• Nonindependence is often
defined as the proportion of variance explained by the dyad.• BUT, nonindependence can be
negative• Proportions of variance cannot.
9
![Page 10: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/10.jpg)
I. How Might Negative Correlations Arise?
• Compensation: If one person has a large score, the other person lowers his or her score. For example, if one person acts very friendly, the partner may distance him or herself.
• Social comparison: The members of the dyad use the relative difference on some measure to determine some other variable. For instance, satisfaction after a tennis match is determined by the score of that match.
10
![Page 11: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/11.jpg)
• Zero-sum: The sum of two scores is the same for each dyad. For instance, the two members divide a reward that is the same for all dyads.
• Division of labor: Dyad members assign one member to do one task and the other member to do another. For instance, the amount of housework done in the household may be negatively correlated.
II. How Might Negative Correlations Arise?
11
![Page 12: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/12.jpg)
Sources of Nonindependence
• Compositional Effects–The dyad members are similar
even before they were paired.• Partner Effects:–A characteristic or behavior of
one person affects his or her partner’s outcomes. 12
![Page 13: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/13.jpg)
Sources of Nonindependence (cont.)
• Mutual Influence–One person’s outcomes directly affect the
other person’s outcomes and vice versa (reciprocity or feedback).
• Common fate–Both members are exposed to the same
causal factors.
13
![Page 14: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/14.jpg)
Measures of Nonindependence
• Distinguishable Dyads –Pearson r
• Indistinguishable Dyads –Intraclass r•Pairwise method•ANOVA method
14
(See davidakenny.net/webinar/nonindependence/com_icc.pdf for details.)
![Page 15: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/15.jpg)
Measures of Nonindependence: Distinguishable Dyads
• If one person’s score is X and other person’s score is X’– Compute Pearson correlation using a dyad dataset– Does not assume equal means or variances for X or X’– Effect sizes • .5 large • .3 moderate • .1 small
15
![Page 16: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/16.jpg)
Measures of Nonindependence: Indistinguishable Dyads
• Pairwise Method: Intraclass r–Compute a Pearson correlation with a
pairwise data set (i.e., the “doubled” data)• ANOVA: Intraclass r–Dyad as the “independent variable” •number of levels of dyad variable: n• Each dyad has two “participants.”•Pairwise or individual dataset
16
![Page 17: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/17.jpg)
Example: Marital Satisfaction
Distinguishable Dyad Members: r = .624Distinguishable Dyad Members: r = .624Indistinguishable Dyad Members (Intraclass r: ICC)Indistinguishable Dyad Members (Intraclass r: ICC)
Pairwise Method (r = .616)Pairwise Method (r = .616)ANOVA Method ANOVA Method ICC = (MSB – MSW)/ (MSB + MSW)ICC = (MSB – MSW)/ (MSB + MSW) .618 = [.39968 - .09422]/ [.39968 .618 = [.39968 - .09422]/ [.39968
+ .09422]+ .09422]
17
![Page 18: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/18.jpg)
Effect of Nonindependence
Effect Estimates UnbiasedStandard Errors Biased–Sometimes too large–Sometimes too small–Sometimes hardly biased
Loss of Degrees of Freedom18
![Page 19: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/19.jpg)
Direction of Bias Depends onDirection of Nonindependence
Positive: linked scores more similarNegative: linked scores more dissimilar
Between and Within DyadsBetween Dyads: linked scores in the same
conditionWithin Dyads: linked scores in different
conditions
19
![Page 20: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/20.jpg)
20
![Page 21: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/21.jpg)
21
![Page 22: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/22.jpg)
Effect of Ignoring Nonindependence
on Significance TestsPositive Negative
Between Too liberalToo
conservative
Within Tooconservative
Too liberal
22
![Page 23: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/23.jpg)
ICC of Between and Within-Dyads Variables
Between-dyads: +1Within-dyads: -1
23
![Page 24: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/24.jpg)
Mixed VariablesMeasure the intraclass correlation.If positive, the effect on significance
testing is like a between-dyads independent variable.
If negative, the effect on significance testing is like a within-dyads independent variable.
If near zero, relatively little bias in the standard errors.
24
![Page 25: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/25.jpg)
![Page 26: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/26.jpg)
Consequences of Ignoring Nonindependence
• As seen, there is bias in significance testing
• More importantly: Analyses that focus on the individual as the unit of analysis lose much of what is important in the relationship (e.g., reciprocity or mutuality).
26
![Page 27: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/27.jpg)
What Not To Do!– Ignore it and treat individual as unit • significance tests are biased
– Discard the data from one dyad member and analyze only one members’ data • loss of precision • might have different results if different dyad members were
dropped.– Collect data from only one dyad member to avoid the problem • (same problems as above)
– Treat the data as if they were from two samples (e.g., doing an analysis for husbands and a separate one for wives)• May be no distinguishing variable to split the sample• Presumes differences between genders (or whatever the
distinguishing variable is)• Loss of power 27
![Page 28: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/28.jpg)
What To Do
• The presence of nonindependence invalidates the use of “person as unit.”
• The analysis must include dyad in the analysis.–Make dyad the unit (not always possible)–Consider both individual and dyad in one
analysis•Multilevel Modeling• Structural Equation Modeling
28
![Page 29: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/29.jpg)
Traditional Model: Random Intercepts
• Model for Persons 1 and 2 in Dyad j– Y1j = b0j + aX1j + bZj + e1j
– Y2j = b0j + aX2j + bZj + e2j
–Note the common intercept which captures the nonindependence.–Works well with positive nonindependence,
but not negative.
29
![Page 30: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/30.jpg)
Alternative Model: Correlated Errors
• Model for Persons 1 and 2 in dyad j– Y1j = b0 + aX1j + bZj + e1j
– Y2j = b0 + aX2j + bZj + e2j
–where e1j and e2j are correlated; if no predictors, then that correlation is an intraclass correlation
• Estimation–Maximum Likelihood (ML)–Restricted Maximum Likelihood (REML)
30
![Page 31: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/31.jpg)
Estimating Nonindependence in Dyadic Data Analysis
Estimate a Estimate a modelmodel in which each score has multiple in which each score has multiple predictors.predictors.
Nonindependence: Correlations of errorsNonindependence: Correlations of errorsModel with no predictors is called the Empty Model with no predictors is called the Empty
ModelModelMultilevel: r = .618, REML; r = .616, MLMultilevel: r = .618, REML; r = .616, ML
SEM: r = .616, MLSEM: r = .616, ML
31
![Page 32: Nonindependence David A. Kenny](https://reader035.vdocuments.us/reader035/viewer/2022062217/56813ba8550346895da4d8d6/html5/thumbnails/32.jpg)
Additional Readings
Kenny, D. A., Kashy, D. A., & Cook, W. L. Dyadic data analysis. New York: Guilford Press, Chapter 2.
See also papers listed at davidakenny.net/webinar/nonindependence/nibib.pdf
32