of associations between health lifestyle and health status due to diagnosed conditions that affected lifestyles in middle adulthood.
Answering Descriptive Research Questions
Many questions of interest to medical sociologists are descriptive in nature. Their answers provide descriptions of circumstances without necessarily providing a sense of how they might be altered. For instance, we might be interested in knowing the extent to which immigrants exhibit better health than non-immigrants, trends in life expectancy among members of different communities over time, or differences in health care expenditures across countries. Answers to such questions not only provide valuable information for medical sociologists but are also critical for policymakers and other constituencies that rely on data to inform decision-making.
In some cases descriptive research questions can be addressed in a straightforward manner through calculating simple statistics such as means and proportions. In other cases, finding answers to descriptive research questions requires the use of a statistical model that takes into account multiple factors. For instance, in estimating trends in death rates due to cancer for different communities over time we might want to adjust for different age structures across the communities. A statistical model provides the means to make such an adjustment.
Various forms of regression models are the primary statistical models used in quantitative medical sociology research (Gelman and Hill 2007; Kalbfleisch and Prentice 2011; Long 1997). Regression models take the form of regressing an outcome or dependent variable on a set of predictors or independent variables, one of which is often considered the focal independent variable. The estimates from fitting a regression model provide a means for assessing the relationship between a focal independent variable and an outcome (e.g. years of schooling and self-rated health) while adjusting or controlling for one or more other independent variables. The outcome variable may be continuous (e.g. health care expenditures in constant dollars), categorical (e.g. an indicator for smoking in the past month), or even unobserved (e.g. the risk of dying in a given year). The independent variables may take any level of measurement.
Answering Causal Research Questions
Despite the value of answers to descriptive research questions, we often want to talk about the causes of the statistical associations we observe. In the physical sciences, causal research questions are addressed via experiments. As noted above, experiments are also used in medical sociology research, but more frequently we rely on observational data. Because observational data lack the statistical properties of experimental data, conclusions regarding causal processes are often tentative. However, over the past 20 to 30 years, methodologists have developed a framework for causal analysis with observational data referred to as the counterfactual framework (Morgan and Winship 2015; Pearl 2009).
The counterfactual framework has two components, the potential outcomes model and causal graphs. The potential outcomes model provides a precise statement of what we mean when we say that a focal independent variable has a causal effect on an outcome (Rubin 1974). In particular, we mean the difference in the outcome that would be observed if a given case experienced an alternative exposure of a focal independent variable than the exposure that was observed. Let us consider smoking during pregnancy as our focal independent variable and the child’s birthweight as our outcome. Then the causal effect of smoking would be defined as the difference in a child’s birthweight for a mother who smoked during pregnancy had she not smoked. It’s impossible, however, to observe both of these states simultaneously (i.e. the birthweight of a child from a mother who smoked, and the birthweight of the same child had the same mother not smoked). Instead, we do the best we can to construct a comparison group to estimate what birthweight would have been observed had the mother not smoked.
Figure 3.1 Causal graph of the relationships between adult children‘s education(ACE), a vector of mediators (M), and mortality(MOR). X represents a vector of pretreatment confounders (e.g., respondent education), and U1 and U2 represent potential unobserved pretreatment and posttreatment confounders, respectively.
The second component, causal graphs, provide a systematic approach to determining what strategies for causal analysis are available with a given set of data and which variables need to be included in the analysis (Pearl 2009). Causal graphs depict relevant variables for an analysis as nodes and causal relationships between the variables using directed edges from the predictor to the outcome. In addition, bidirected edges can be used to indicate that two variables share a common cause. The relationships indicated in a causal graph are non-parametric and include all possible interactions. Figure 3.1 illustrates an example of a causal graph of mortality. The graph indicates key confounders of the effect of education on mortality (X and U1) as well as confounders of the effect of mediators on mortality (U2). Confounders are variables that affect both the focal independent variable and the outcome. If confounders are not addressed, estimates of the causal effect of the focal independent variable on an outcome will be biased. In contrast, mediators are variables that are thought to transmit the effect of a focal independent variable to an outcome. A mediation analysis seeks to identify the mechanisms that underlie a causal process. In this causal graph, we see that some confounders are observed (e.g. education) while others are unobserved (e.g. non-cognitive resources). The relationships depicted in this graph and the presence of observed and unobserved confounders have implications for strategies for estimating causal effects.
Strategies for Estimating Causal Effects
In broad terms, there are three strategies for estimating causal effects that involve (1) conditioning on confounders, (2) identifying instrumental variables, or (3) specifying mechanisms. The first, and by far the most commonly used, involves conditioning on all confounders in a statistical model. For instance, it is likely that mother’s education predicts both whether a mother smokes and the birthweight of her child. In this case, if we do not adjust for mother’s education in our analysis of smoking and birthweight, then we would likely overestimate the causal effect of smoking on birthweight, as part of the effect is likely due to mother’s education as a common cause of both. As noted above, a failure to incorporate all confounders in an analysis leads to biased estimates of causal effects. Well-articulated causal graphs allow researchers to identify the confounders that need to be included in an analysis. In practice, we never have measures of all possible confounders available, and therefore we need to rely on sensitivity analyses and/or more modest statements about our findings with respect to estimating causal effects.
The second strategy involves identifying an exogenous source of variation in the focal independent variable and often using statistical models based on instrumental variables. An exogenous source of variation refers to some feature of the world that induces a change in the focal independent variable for at least some cases but is otherwise unrelated to the outcome. Quasi-experiments fit into this category. Natural disasters can certainly cause changes in people’s lives that may be otherwise unrelated to an outcome or policy-oriented changes can fulfill the same role. To give an example, some studies in medical sociology have leveraged changes in compulsory schooling laws to try to estimate the effect of education on health (Courtin et al. 2019). A source of exogenous variation can be treated as an instrumental variable, i.e. a variable that has an effect on the focal independent variable but not the outcome.
The third strategy, and least commonly used, involves (1) specifying all of the mechanisms that link a focal independent variable to an outcome, (2) estimating the effects of each mechanism, and then (3) adding up the effects to get a total causal effect. A mechanism refers to the process through which a variable has an effect on an outcome. If we return to the education and health example, there are a number of mechanisms thought to account for the positive relationship. To name a few, higher levels of education lead to better and higher paying jobs,
