PS Methods with SEM Flashcards
Why is it important to consider latent covariates in propensity score analysis?
Propensity score analyses may use latent variables (i.e. constructs) rather than observed variables as covariates in the propensity score model in the design stage and/or as the outcome in the analysis stage.
How can confirmatory factor analysis be used in the design stage of propensity score analysis?
Confirmatory factor analysis (CFA), which is a structural equation model to measure latent variables, can be used to control for measurement error in latent confounding variables.
What is the consequence of ignoring measurement error of latent covariates?
Propensity scores will not balance the latent confounding variables across treated and control groups if they are measured with substantial error.
What is the role of structural equation modeling in the analysis stage of a propensity score analysis?
Second, in the analysis stage, SEM can adjust for unreliability of measurement of latent outcomes.
Why is measurement invariance important in propensity score analysis with latent covariates and latent outcomes?
Third, SEM can provide validity evidence for the measurement of the latent outcome. In particular, SEM can be used to examine whether the measurement of the construct of interest is equivalent across treated and untreated groups, which is known in the SEM literature as measurement invariance testing.
What level of measurement invariance is required for the use of latent covariates and latent outcomes in propensity score analysis?
? SEM can be used to examine whether the measurement of the construct of interest is equivalent across treated and untreated groups.
What is the consequence of measurement error in the estimate of a treatment effect on a latent outcome?
This is important because ignoring the measurement error in latent outcomes results in biased standard errors.
What is the consequence of lack of measurement invariance in the estimate of a treatment effect on a latent outcome?
Estimated treatment effects on latent variables measured with error will be attenuated.
How can multiple-group structural equation models be used to estimate treatment effects with propensity score matching?
Multiple-group Analysis: estimate parameters of the two groups simultaneously with and without constraints for parameter equality across groups. Perform chi-square difference tests between constrained and unconstrained models.
(Hoshimo, Kurata and Shigemasu (2006) proposed a propensity score weighted estimator for multiple-group structural equation modeling to estimate the effects of single, multiple or continuous treatments, and show through a simulation study that their proposed method is more robust than maximum likelihood estimation when the relationships between covariates and the outcome is misspecified.)
How can the multiple-indicators multiple-causes model be used to estimate treatment effects with propensity score methods?
The multiple indicator and multiple causes model (MIMIC) is a single-group model where the treatment indicator is added as a covariate.
(The use of SEM with propensity score analysis was pioneered by Kaplan (1999), who demonstrated combining propensity score stratification with the multiple indicator multiple cause (MIMIC) model, which is a SEM where a dummy-treatment indicator predicts a latent variable.)
