|
Abstract:
|
In propensity score analysis, latent/mismeasured covariates lead to bias in treatment effect estimates. The strategy of using factor scores (FS, estimated from multiple error-prone measurements) to represent the covariate has been used, but with limited success. We compare the FS based on a model with the measurements only (the simple FS) to two other types of FSs generated from structural equation models (or substitute factor models) that also include the treatment variable (treatment-inclusive FSs) and the other covariates from the propensity score model (treatment-&-covariates-inclusive FSs) -- in the context of propensity score weighting analysis, with continuous/ordinal measurements with classical errors. Logit, probit and identity links are used for the treatment variable and ordinal measurements. The simple FS method is as biased as using the measurements directly. The treatment-inclusive FS method is less biased but still biased. The treatment-&-covariates-inclusive FS method effectively removes bias, and is thus the recommended method. We discuss situations where this correction is most useful and present options for obtaining confidence intervals for the treatment effect.
|
