4.2. Causal inference

When you use terms like ‘causes’ and ‘consequences’, ‘impact’, ‘effects’, or even when you say that something ‘increases’, ‘amplifies’, ‘enhances’, ‘reduces’ or ‘hinders’ something else, you make statements about causality. The question you should ask yourself when you use such terms is: does the research design actually allow for causal inference? Some types of research designs are better suited to answer causal questions. Let’s discuss the possibilities for causal inference in the four designs, starting from right to left.

Experiments. Strictly speaking, causal inference is only possible by using experiments in which participants are randomly assigned to treatment and control conditions. Think of an experiment as a randomized control trial (RCT) in which drugs are tested. If you have the possibility to design an experiment to answer your research question, seize it. Make sure you use a manipulation that is effectively altering the cause you are after – and only this cause, not others. You do not want to conclude that something in your experiment worked, but you do not know what. So use a clean manipulation of only one cause.

Also you want to do better than conclude that your manipulation worked, but you do not know why. Think about the variables mediating the effects of your manipulation, and measure these. When to measure the mediating variables is a bit tricky though. If you measure something, you easily make participants in your experiment aware of it, and you want to avoid this awareness to affect the behavior of participants. You would rather have your participants remain blind to your hypotheses. This is also the reason why the manipulation check – which tells you whether the participants actually swallowed your manipulation – is generally conducted after all other relevant measures in the experiment have been taken.

For the majority of explanatory research questions you cannot randomize participants in different conditions. For instance, if you have a question like ‘What is the influence of divorce on happiness?’ you will not be able to make some people experience the treatment (divorce) and withhold the treatment from others (remaining married or unmarried) and then observe the consequence (happiness). In these cases, you will need to make some strong assumptions to use non-experimental research designs to answer your question. The assumptions are for instance that assignment to treatment is not correlated with the outcome variable. In the example of divorce and happiness you need to assume that people who will go through a divorce at some later point in time were equally happy before the divorce as people who will remain married. If less happy people are more likely to divorce, it could be that the association between divorce and happiness reflects an influence of happiness on divorce, and not the reverse.

Other designs. If you cannot conduct or find an experiment, a longitudinal design or panel study is better than a cross-sectional design. Case studies are notoriously difficult for causal inference, especially when they have been selected on the dependent variable (i.e., a collection of ‘best practices’). Remember that correlation is not causation. There is much more to say about the potential for causal inference in various research designs and methods of analysis. These issues go beyond the scope of this text. One particularly helpful and insightful introductory book discussing rules for research is Glenn Firebaugh’s (2008) book Seven Rules for Social Research. More advanced discussion is provided by Shadish, Cook & Campbell (2002).