Panel Regression Models for Causal Analysis in Structural Equation Modeling: Recent Developments and Applications

Establishing causal relationships is arguably the most important task of the social sciences. While the relationship between the social sciences and the concept of causality has been rocky, the randomized experiment gives us a concrete definition of a causal effect as the difference in outcomes due to the researcher's intervention. However, many interesting questions cannot be easily examined using experiments. Feasibility and ethics limit the use of randomized experiments in some situations and retrospective questions, i.e., working from the observed outcome to uncover the cause, require a different logic. Observational studies in which we observe pairs of variables without any intervention lend themselves to such situations but come with many difficulties. That is, it is not immediately clear whether an observed relationship between two variables is due to a true causal effect, or whether the relationship is due to other common causes. Panel data describe repeated observations of the same units over time. They offer a powerful framework for approaching causal questions with observational data. Panel analysis allows us to essentially use each unit as their own control. In an experiment, random assignment to either treatment and control group makes both groups equal on all characteristics. Similarly, if we compare the same individual pre- and post-treatment, then the two are equal at least on the things that do not change over time, such as sex, date of birth, nationality, etc. Structural equation modeling (SEM) is a group of statistical methods for assessing relationships between variables, often at the latent (unobserved) variable level. The use of SEM for panel analysis allows for a great deal of flexibility. Latent variables can be incorporated to account for measurement error and rule out alternative models. This dissertation focuses on the use of panel data in SEM for causal analysis. It comprises an introduction, four main chapters and a conclusion. After a short introduction (Chapter 1) outlining the goals and scope of the dissertation, Chapter 2 provides an overview of the topic of causality in the social sciences. Since the randomized experiment is often not feasible in social research, special emphasis has been placed on non-experimental, i.e., observational data. The chapter outlines some competing views on causality with non-experimental data, then discusses the two currently dominant frameworks for causal analysis, potential outcomes and directed graphs. It goes on to outline empirical methods and notes their compatibility with SEM. Chapter 3 discusses how panel data can be used to deal with unobserved time-invariant heterogeneity, i.e., stable characteristics that might normally confound analyses. It attempts to show in detail how basic panel regression in SEM works. It also discusses some issues that are not normally addressed outside of SEM, e.g., measurement error in observed variables, effects that change over time, model comparisons, etc. This discussion of the more basic panel regression setup provides a sort of basis for the more complex discussion in the following chapters. Chapter 4 compares and contrasts several ways to model dynamic processes, where the outcome at a particular point in time may affect future outcomes or even the presumed cause later on. It shows that popular recently proposed modeling techniques have much do to with their older counterparts. In fact, the newer modeling techniques do not seem to offer benefit with regards to estimating the causal effects of interest. The chapter focuses on arguably common situations in which the newer techniques may have serious drawbacks. Chapter 5 provides an applied example. It looks to better assess the causal effect of environmental attitudes on environmental behaviour (mobility, consumption, willingness to sacrifice). It touches on many of the aspects from the previous chapters, including the use of latent variables for constructs that are not directly observable, unobserved time-invariant confounders, state dependence (feedback from outcome to outcome), and reverse causality (feedback from outcome to cause). It shows that failure to account for time-invariant confounders leads to biased estimates of the effect of attitudes on behaviour. After controlling for these factors, the effects disappear in terms of mobility and consumption behaviour: when a person's attitudes become more positive, their behaviour does not become more environmentally-friendly. There is, however, a fairly robust effect of attitudes on willingness to sacrifice, even after controlling for unobserved time-invariant confounders, state dependence and reverse causality. This suggests changing attitudes do affect willingness to make sacrifices, holding potential time-invariant confounders, outcome to outcome feedback (essentially habits), as well as some time-varying confounders constant. Finally, Chapter 6 summarizes the previous chapters and provides an outlook for future work.:1. Introduction 2. Causal Inference in the Social Sciences 3. A Closer Look at Random and Fixed Effects Panel Regression in Structural Equation Modeling Using lavaan 4. Equivalent Approaches to Dealing with Unobserved Heterogeneity in Cross-Lagged Panel Models? 5. Re-Examining the Effect of Environmental Attitudes on Behaviour in a Panel Setting 6. Conclusio