Earlier in this chapter, we suggested that it is common to focus on associations in quantitative research, and that a statistical effect of one variable on another variable is not the same as a causal effect. Yet, we use concepts throughout the guide that sort of imply causality, such as “exposure”, “outcome”, “mediator”, and “pathway”. In this section, we will discuss the issue of causality – or causal inference – in a bit greater depth. Of course, there is not enough space here to address the full complexity of the issue.  

Data analysis is often concerned with causal questions. For example, can a given intervention program improve program participant’s outcomes? Can a given sickness be prevented? Why do girls typically outperform boys in the educational system? In contrast to statistical inference, in which information obtained from various forms of random sample of observations are used to draw conclusions about the value of some parameter (e.g. a mean or a regression coefficient) in the population from which the sample was drawn (see Chapter 3), causal inference typically refers to the process where multiple sources of information are used to draw reasonable inference about cause and effect.

Causal inference taps into important discussions related to ontology and epistemology, which will not be addressed here. For the purposes of this guide, three broadly defined (and partly overlapping) perspectives on causal inference will be outlined (based on Goldthorpe’s “Causation, statistics, and sociology” from 2001), all of which to some extent may relate to the empirical methodologies detailed above: causation as robust association, causation as manipulation, and causation as generative mechanisms.  

While it remains true that the widely recognized statement that association (i.e. correlation) does not imply causation, causation must in some way imply association. Causation as robust association comes in many versions but a common denominator is that it emphasizes efforts to ensure that estimated associations are not spurious, i.e. the association cannot be eliminated through one or more other variables being introduced in the analysis. In practice, this approach typically proposes a set of criteria such as temporality and predictive power to assess causal connections between variables.

Causation as manipulation appears to some extent to have emerged in reaction to that of causation as robust association. Here, attention centres on establishing causation through experimental methods. In short, the key idea is that causes can only be those factors that, at least theoretically, can serve as treatments (or more generally exposures) in experiments. This means that causes must in some sense be manipulable, and that causation is determined by comparing what would happened to an observational unit in regard to an outcome if this unit have been vs. not have been exposed to the addressed factor. Since it is not possible for the same unit to be both exposed and not exposed, the solution for estimating a causal effect is to compare the average response for those units exposed to the average response for those units that were not exposed. For this solution to be viable, however, a number of conditions and assumptions need to be met. These conditions are ideally those of randomised controlled trials, but substantial efforts have been made to develop statistical analyses that to a large extent mimic the conditions of such trials (e.g. propensity score matching, endogenous treatment effect regression). Causation as manipulation also comes in different versions and the main difference lies in to what extent there is an emphasis on designing a study or on analysis of already collected data (cf. the potential outcome framework). While the former focusses on removing threats to internal validity by using appropriate experimental designs, the latter focusses on strategies for estimating causal effects using observational data (often in a longitudinal design).

In contrast to the above, causation as generative mechanisms does not focus on relationships between variables but rather on what needs to be added to any criteria before a reasonable argument for causation can be made, namely the agentic capabilities of observational units (typically individuals). Here, actors, their relationships, and the (un)intended outcomes of their actions are emphasized. The properties of actors and their environments can be measured and thereby represented by variables, but the causality does not operate at the variable level. According to this perspective, the actors are the agents of change and the causal process should therefore be specified at the actor level which means that in order to move from association to causation it is not sufficient to just establishing that a given factor precedes the outcome (rather than the other way around), it is also necessary to specify the mechanisms that explain why actors do what they do and how these actions translate into outcomes. In order to do so, proponents of this approach typically suggest that various theories of rational action can be utilized. 

Research questions often relates to issues of causality and the perspectives outlined above all have their pros and cons. In any event, estimates of associations alone cannot be used for causal inference. Various sources of information, which includes a theoretical framework of causality combined with the best available research design and data for the research question at hand, are imperative in the process of evaluating whether our estimates may allow for reasonable causal interpretations.