Instrumental variables and inverse probability weighting for causal inference from longitudinal observational studies

Inferring causal effects from longitudinal repeated measures data has high relevance to a number of areas of research, including economics, social sciences and epidemiology. In observational studies in particular, the treatment receipt mechanism is typically not under the control of the investigator; it can depend on various factors, including the outcome of interest. This results in differential selection into treatment levels, and can lead to selection bias when standard routines such as least squares regression are used to estimate causal effects. Interestingly, both the characterization of and methodology for handling selection bias can differ substantially by disciplinary tradition. In social sciences and economics, instrumental variables (IV) is the standard method for estimating linear and nonlinear models in which the error term may be correlated with an observed covariate. When such correlation is not ruled out, the covariate is called endogenous and least squares estimates of the covariate effect are typically biased. The availability of an instrumental variable can be used to reduce or eliminate the bias. In public health and clinical medicine (e.g., epidemiology and biostatistics), selection bias is typically viewed in terms of confounders, and the prevailing methods are geared toward making proper adjustments via explicit use of observed confounders (e.g., stratification, standardization). A class of methods known as inverse probability weighting (IPW) estimators, which relies on modeling selection in terms of confounders, is gaining in popularity for making such adjustments. Our objective is to review and compare IPW and IV for estimating causal treatment effects from longitudinal data, where the treatment may vary with time. We accomplish this by defining the causal estimands in terms of a linear stochastic model of potential outcomes (counterfactuals). Our comparison includes a review of terminology typically used in discussions of causal inference (e.g., confounding, endogeneity); a review of assumptions required to identify causal effects and their implications for estimation and interpretation; description of estimation via inverse weighting and instrumental variables; and a comparative analysis of data from a longitudinal cohort study of HIV-infected women. In our discussion of assumptions and estimation routines, we try to emphasize sufficient conditions needed to implement relatively standard analyses that can essentially be formulated as regression models. In that sense this review is geared toward the quantitative practitioner. The objective of the data analysis is to estimate the causal (therapeutic) effect of receiving combination antiviral therapy on longitudinal CD4 cell counts, where receipt of therapy varies with time and depends on CD4 count and other covariates. Assumptions are reviewed in context, and resulting inferences are compared. The analysis illustrates the importance of considering the existence of unmeasured confounding and of checking for ‘weak instruments.’ It also suggests that IV methodology may have a role in longitudinal cohort studies where potential instrumental variables are available.