Abstract
Causal inference techniques, such as inverse propensity weighting (IPW), are increasingly used in medical, social, and behavioral research. When data is collected with an observational study design rather than in a randomized controlled trial, treatment effects estimates will be confounded. However, causal inference provides a toolbox for accounting for these confounding effects and to estimate average treatment effects (ATE) based on observational data. IPW can be easily combined with standard statistical models such as generalized linear models or survival analysis.
However, sometimes the outcome of interest is not directly observable, and a measurement model is needed, e.g., when analyzing patient-reported outcome measures data. Latent class analysis (LCA) and its extensions are particularly suited for analyzing such data as they explicitly model the multidimensionality of these constructs. Recently, a one-step approach (Lanza, Coffman, & Xu; 2013) and a three-step approach (Clouth, Pauws, Mols, & Vermunt; 2021) have been proposed to incorporate IPW in LCA. While these approaches work well when the latent class model is correctly specified, differential item functioning (DIF) often prohibits estimating the ATE correctly. DIF occurs when treatment or confounding variables have a direct effect on some of the indicator variables, which violates the assumption that indicator variables and auxiliary variables are independent conditional on class membership. This can lead to biased estimates of the ATE or even to the detection of spurious classes.
In this talk, I will present an analysis strategy that allows for the correct estimation of the ATE in the presence of DIF
However, sometimes the outcome of interest is not directly observable, and a measurement model is needed, e.g., when analyzing patient-reported outcome measures data. Latent class analysis (LCA) and its extensions are particularly suited for analyzing such data as they explicitly model the multidimensionality of these constructs. Recently, a one-step approach (Lanza, Coffman, & Xu; 2013) and a three-step approach (Clouth, Pauws, Mols, & Vermunt; 2021) have been proposed to incorporate IPW in LCA. While these approaches work well when the latent class model is correctly specified, differential item functioning (DIF) often prohibits estimating the ATE correctly. DIF occurs when treatment or confounding variables have a direct effect on some of the indicator variables, which violates the assumption that indicator variables and auxiliary variables are independent conditional on class membership. This can lead to biased estimates of the ATE or even to the detection of spurious classes.
In this talk, I will present an analysis strategy that allows for the correct estimation of the ATE in the presence of DIF
Original language | English |
---|---|
Title of host publication | IMPS 2022 International Meeting of the Psychometric Society |
Number of pages | 1 |
Publication status | Published - 2022 |
Event | IMPS 2022 International Meeting of the Psychometric Society - University of Bologna, Bologna, Italy Duration: 11 Jul 2022 → 15 Jul 2022 |
Conference
Conference | IMPS 2022 International Meeting of the Psychometric Society |
---|---|
Country/Territory | Italy |
City | Bologna |
Period | 11/07/22 → 15/07/22 |