9.6 Criticisms of Mediation

If we think of causality, we usually think of a process in which one thing leads to another thing, which leads to something else, and so on. This is apparent if we want to explain why we think that one phenomenon causes another (see Section 9.1.2). Mediation, however, is difficult to establish with regression analysis and, as some argue, perhaps impossible to establish.

9.6.1 Causal order assumed

It is paramount to note that the regression approach to mediation and path models does not tell us anything about the causal order of the variables. The causal order is purely an assumption that we make. The plausibility of the assumptions depends on how well we can justify the time order of the variables and the absence of common causes for cause-consequence pairs (see 9.1.1).

9.6.2 Time order

To establish the time order of variables, we must think about the time at which the behaviours or opinions that we measure took place. This is what matters, not the time at which we measure them. We can collect information on behaviour a long time after the fact, for example by asking respondents when they started using news sites or checking internet use logs.

If cause and consequence appear very closely in time, it may be difficult to argue that one variable precedes the other. This may also apply to the time at which measurement takes place. If we measure cause and consequence nearly at the same time, it can be difficult to establish the time order of the two.

9.6.3 Causality or underlying construct?

For causes and consequences that appear nearly simultaneously, we should take into account that the two variables may measure the same underlying construct. Think of the way we construct a scale from items: We assume that the items measure the same underlying attitude, for instance, political cynicism.

The indicators of a scale are correlated because they have a common cause, namely, the underlying attitude. But it does not make sense to interpret the correlation as a sign of mediation. One item does not trigger another item, and so on. A mediator must be theoretically and conceptually different from both the predictor and outcome. We have to provide arguments that they are really different.

9.6.4 Every effect in a path model can be confounded

In Chapter 8, you learned that the estimated regression coefficients can be too small, too large, or have the wrong sign (direction) if there are confounders: variables not included in the regression model that are correlated with the predictor and outcome variable. If we analyze a path model with a series of regression models, there can be confounders for each regression model. Every estimated direct effect can be wrong. As a consequence, every indirect effect, which is the product of direct effects, can be wrong.

The surest way to get rid of a confounder is adding it to the regression model. In a path model, we can add a variable that we expect to be a confounder as a covariate (Section 9.3) or as an additional mediator. If a confounder comes after the outcome variable in the causal order of the path model, it cannot be a common cause to both a predictor and the outcome variable. In this situation, the confounder can be ignored. This underlines the importance of choosing a correct causal order when we construct a path model. Unfortunately, we can never be sure about this.

In practice, we do not know all confounders and we cannot include all of them in our regression models. We can minimize the risk of having confounders if we use randomization in an experiment. Section 8.2.3 explained how randomization of the experimental treatment variable helps to eliminate confounders for the effect of the experimental treatment (predictor variable) on the dependent variable. We expect that this effect is not confounded.

In a path model, a mediator also serves as a predictor, so we also have to randomize the mediator variable to get correct estimates for the causal effect of the mediator on the outcome variable. With randomized predictor and mediator variables, the direct effects are probably not confounded, so the indirect effects calculated from the direct effects are also unlikely to be confounded.

It is difficult to manipulate a mediator in an experiment (Bullock & Ha, 2011). If we hypothesize, for example, that political interest mediates the effect of age on newspaper reading time, how can we assign a random level of political interest to a participant in an experiment? By the way, it will also be impossible to randomize participant age in this example.

9.6.5 Recommendations

All in all, mediation is an intuitively simple and appealing concept. Unfortunately, it is very difficult to substantiate the claim that indirect effects in path models represent mediation. Mediation assumes causal effects and causality is difficult to establish.

If you plan to investigate mediation:

  1. Justify that the mediator is theoretically and conceptually different from the predictor and outcome.

  2. Motivate the time order of variables in the model.

  3. Include variables that are likely to confound the effects of the predictor or mediator(s) in your research project and in the regression models that you are going to estimate.