Transcript Slide 1

The Importance of Accounting For Covariates and Prior
Measurements in Longitudinal Gerontology Studies
Jason T. Newsom
Institute on Aging, Portland State University
Portland, OR
Karen S. Rook
University of California, Irvine
Kathleen J. Bonn
Institute on Aging, Portland State University
Portland, OR
This project is funded by AG14130 from the National Institute on Aging. Email:
[email protected]
Cross-lagged Panel Models
 Interest and availability of longitudinal designs in gerontology is increasing
 Cross-lagged panel models are useful for questions about causal
directionality and change over a discrete interval with passive observational
designs
 Cross-lagged panel models examine the predictive association between two
variables over time, each controlling for effects at earlier time points
Xt0
Yt0
Xt1
Yt1
Causality
 Kenny (1979) summarized three criteria for concluding causality:
• Correlation
• Time precedence
• Nonspuriousness
 Cross-lagged panel models address correlation and time precedence criteria
 These models, however, rarely address nonspuriousness by controlling for
third variables which may covary with either variable
 Omission of covariates can lead to biases in the stability of a variable over
time or the longitudinal association of two variables over time
 Time invariant or time-specific covariates can be included
Background
 Negative social exchanges involve criticisms, intrusiveness, demands, or
rejection from family or friends
 Large body of work suggests a strong relationship between negative social
exchanges and psychological distress among older adults (Rook, 1992)
 Most research has been cross-sectional, but longitudinal studies support
causal hypothesis that negative exchanges affect distress (e.g., Finch & Zautra,
1992; Newsom, Nishishiba, Morgan & Rook, 2003)
 Question often raised about whether psychological distress or mood may
affect reports of negative exchanges
Study Design and Methods
 Participants are from the Later Life Study of Social Exchanges (LLSSE), a
national sample of adults 65-90 years old.
 3 in-person interviews at 12-mo intervals (T1, T3, T5)
 N =916 completed interviews at Time 1, N = 667 completed interviews at
Time 5
 At baseline, average age = 74.16; 62% female; 63% HS degree or less
education; 83% Caucasian, 11% African American, 5% Latino, 1% other
groups.
Negative Social Exchanges Measure
 12-item measure, 4 domains, 3 items each.
 Frequency of occurrence in the past month
 Subscale scores for 4 factors were indicators of negative social
exchanges latent variable
• Unsympathetic/insensitive behavior (e.g., “… act unsympathetic or
critical about your personal concerns?”)
• Failure to provide needed aid (e.g., “…let you down when you
needed help?”)
• Unwanted advice (e.g., “… give you unwanted advice?”)
• Neglect/rejection (e.g., “…forget or ignore you?”)
Depression Measure
 Center for Epidemiologic Studies Depression scale (CES-D; Radloff,
1977)
 Brief 9-item version developed by Santor and Coyne (1997)
 Subscale scores for 3 factors (following McCallum, MacKinnon, Simons, &
Simons, 1995) were indicators of depression latent variable:
• Positive affect (e.g., “You were happy”)
• Negative Affect (e.g., “You felt sad”)
• Somatic Symptoms (e.g., “Your sleep was restless”)
Health
Health was measured by three indicators at Time 1:

Self-rated health: How would you describe your health at the
present time? Would you say it is …”(0 = poor, 4=excellent)

The number of chronic conditions out of 12 (e.g., heart disease,
cancer, chronic lung disease)

15 Activities of daily living (e.g., climb stairs, use the telephone,
bathe or dress)
Analyses
 Structural equation models using Mplus, version 3.11 (Muthen & Muthen,
2004).
 Sample size based on Time 5 complete interviews (N=667)
 FIML missing data estimation used for missing responses not due to attrition
 Because of concerns about multivariate non-normality, Yuan-Bentler (2000)
estimation for non-normal missing data were used.
Model Specification Precautions

Using latent variables can reduce bias in autoregressive paths and cross-lag
paths

Correlated measurement errors over time reduce bias in autoregressive
paths (Finkel, 1995)

At least partial longitudinal measurement invariance (loadings) should be
established through chi-square difference tests

Equality constraints in autoregressive paths, cross-lagged paths, or
synchronous correlations
Model Specification Precautions

Using latent variables can reduce bias in autoregressive paths and cross-lag
paths

Correlated measurement errors over time reduce bias in autoregressive
paths (Finkel, 1995)

At least partial longitudinal measurement invariance (loadings) should be
established through chi-square difference tests

Equality constraints in autoregressive paths, cross-lagged paths, or
synchronous correlations
Correlated Measurement Errors
Depression
T1
PAT1
NAT1
Depression
T2
SomT1
PAT2
NAT2
SomT2
Model Specification Precautions

Using latent variables can reduce bias in autoregressive paths and cross-lag
paths

Correlated measurement errors over time reduce bias in autoregressive
paths (Finkel, 1995)

At least partial longitudinal measurement invariance (loadings) should be
established through chi-square difference tests

Equality constraints in autoregressive paths, cross-lagged paths, or
synchronous correlations
Measurement Invariance Equality Constraints
Depression
T1
PAT1
NAT1
Depression
T2
SomT1
PAT2
NAT2
SomT2
Model Specification Precautions

Using latent variables can reduce bias in autoregressive paths and cross-lag
paths

Correlated measurement errors over time reduce bias in autoregressive
paths (Finkel, 1995)

At least partial longitudinal measurement invariance (loadings) should be
established through chi-square difference tests

Equality constraints in autoregressive paths, cross-lagged paths, or
synchronous correlations
Autoregressive Equality Constraints
Depression
T1
Depression
T3
Depression
T5
NSE
T1
NSE
T3
NSE
T5
Cross-lag Equality Constraints
Depression
T1
Depression
T3
Depression
T5
NSE
T1
NSE
T3
NSE
T5
Synchronous Correlation Equality Constraints
Depression
T1
Depression
T3
Depression
T5
NSE
T1
NSE
T3
NSE
T5
.418***
Basic Cross-lagged Panel Model Between
Depression and Negative Social Exchanges (NSE’s)
Depression
T1
NSE
T1
.626***
.652***
Depression
T3
NSE
T3
.680***
Depression
T5
.669***
c2 (172, N= 667) = 262.673, IFI = .976 , SRMR = .044
NSE
T5
.434***
Cross-lagged Panel Model Between Depression and
Negative Social Exchanges (NSE’s) with Covariate
Depression
T1
NSE
T1
.566***
.659***
Depression
T3
NSE
T3
.610***
.657***
Depression
T5
NSE
T5
Health
T1
c2 (229, N= 667) = 386.526, p < .001 IFI = .964, SRMR = .043
Summary and Conclusions

Results suggested that negative social exchanges did not significantly
predict depression when health was omitted from the model

When health was included, findings also support an effect of negative
social exchanges on depression

Conclusions about the causal direction of the relationship between two
variables can differ if the correct covariates are not included in the model

Results may be revealing about the role of health. One reason for the
correlation between NSEs and depression may be that NSEs affect
depression among those with worse health by contributing to health-related
conflicts and disappointments (Margolin & McIntyre-Kingsolver, 1988
Newsom, 1999;; Skelton & Dominian, 1973)
General Comments

Effects of covariates can be complex, because covariate is controlled in
autoregressive paths as well as cross-lagged paths

These analyses only included a measure of health, but other important
covariates likely

Stable health was assumed, but health or other variables can be included as
time-specific covariates

Including time-specific covariates requires additional cross-lagged paths
and autoregressive paths, adding substantially to the complexity of the
model
Recommended Readings

Ferrer, E., & McArdle, J.J. (2003). Alternative structural models for multivariate
longitudinal data analysis. Structural Equation Modeling, 10, 493-524.

Finkel, S.E. (1995). Causal analysis with panel data. Thousand Oaks: Sage.
QASS #105

Joreskog, K.G. (1979). Statistical estimation of structural models in longitudinaldevelopmental investigations. In J.R. Nesselroade & P.B. Baltes (Eds.),
Longitudinal research in the study of behavior and development (pp. 303-352).
New York: Academic.

Kessler, R.C., & Greenberg, D.F. (1981). Linear panel analysis: Models of
quantitative change. New York: Academic.

McArdle, J.J., & Hamagami F. (2001). Latent difference score structural models
for linear dynamic analyses with incomplete longitudional data. In L. Collins & A.
Sayer (Eds.), New methods for the analysis of change (pp. 139-175). Washington,
D.C.: American Psychological Association.

Wheaton, B., Muthen, B., Alwin, D.F., & Summers, G.F. (1977). Assessing
reliability and stability in panel models. In D. R. Heise (Ed.), Sociological
Methodology 1977 (pp. 84-136). San Francisco: Jossey-Bass.
A copy of this talk will be available at the GSA Measurement,
Statistics, and Research Design Interest group website:
http://www.ioa.pdx.edu/newsom/gsaquant