| Time | Chapter |
|---|---|
| 9:10 | WelcomeToday is all about an analysis called mediation analysis, and we're also going to be discovering a new JASP module, the Process module. A mediator comes in between two entities, so we see whether the association between two variables is mediated by a third variable. |
| 13:21 | Multiple regression as a path model (applet)We can visualize our regression model with a path model, with boxes for the predictors and a box for the outcome, connected with a regular line. Multiple regression is just based on correlations, and a causal interpretation needs to stem from your methodology, where you experimentally manipulate your predictor variable. |
| 17:33 | MediationYou start with an association between two variables, and then you ask to what extent this relationship is mediated by a third variable. The predictor can directly influence the outcome (the direct relationship), or influence the mediator, which then influences the outcome (the indirect relationship). |
| 20:58 | Shared varianceAs in ANCOVA and partial correlations, this is about the overlap between what your predictor variables explain. The overlap between all three variables (box C) is the indirect effect, and mediation separates it from the direct contribution of the predictor (box A). |
| 25:10 | Example (presumed media influence)The example is about how likely people are to buy sugar after reading an article about an incoming sugar crisis, placed either on the front page or somewhere in the middle. We ask whether the association between condition and intention is mediated by perceived media influence, how much people think other people will buy sugar after that article. |
| 27:47 | Total effectThe total effect is the simple relationship between our one predictor, condition, and our one outcome variable, intention. We are not yet dissecting it into indirect and direct effect, and we could test it with a linear regression (or an ANOVA). |
| 29:32 | Mediation pathsTo get an idea of the indirect effect, we also look at the regression of our mediator, media influence, on our predictor, condition. The combined effect models the outcome from the predictor and the mediator at the same time, so that is a multiple regression. |
| 31:14 | The data and the three regression modelsWe run a series of regressions that reflect the extent to which the predictor is associated with the outcome, but also with the mediator. If your predictor and your mediator are completely uncorrelated, then you will not have any indirect effect. |
| 38:12 | Path plot and the indirect effectBased on those regression coefficients, we can fill in our path diagram of mediation. The indirect effect is the product of the a and b pathways, and if we add the indirect effect to the direct effect, we get back the total effect. |
| 44:59 | Effect sizes of mediationThe most important effect size is just the indirect effect, reported with a p-value, or a confidence interval if you are more interested in estimating it. To test for mediation, we test the indirect effect a × b itself. |
| 46:43 | Checking the assumptionsThe one thing to take away in terms of assumptions for mediation, but also moderation, is the bootstrap. It is a robust tool to get confidence intervals that do not rely on the normality and homoscedasticity assumptions. |
| 48:27 | Is the indirect effect significant?The p-value of the indirect effect answers how likely it is to observe this estimate, or more extreme, if on the population level there is no mediation effect. Now that we have multiple p-values in our output, always be specific about which p-value you are referring to. |
| 53:20 | Break |
| 1:04:41 | Caution with Process modelsProcess models answer causal questions, but they are very, very often misused (Julia Rohrer's 2022 paper outlines the pitfalls). If you simply throw observational data into a process model you get all sorts of estimates, but numbers alone don't give you enough ground to make causal claims. |
| 1:07:14 | Equivalent modelsWith observational data on three variables, we cannot distinguish between these four models, with arrows pointing in different directions. If there is no temporal precedence, no manipulation, don't use these models, just use regression. |
| 1:13:11 | Mediation in JASPOn the exam you have a fresh JASP, so click the blue plus and select Process to add it to the top bar. In the Models tab you pick a Hayes model number (mediation is number 4, moderation number 1) and tell JASP which variable is the predictor and which is the mediator. |
| 1:20:01 | Path coefficients and the effects tablesThe path coefficients table lists the a, b and c paths, the pairwise associations between your variables. The direct and indirect effects table splits the total effect, and here the indirect effect is not significant, so we fail to reject the null hypothesis that there is no indirect effect. |
| 1:29:32 | Building the path model yourselfThere's also the way described in the book, where you build your path model from the bottom up by creating the different pathways yourself. It is a more flexible approach that can better match your theoretical question, so choose your favorite. |
| 1:31:13 | Interpreting the resultsA significant association between mediator and outcome, with no significant association between predictor and outcome, leaves not much to mediate. Always do hypothesis testing with your p-value, because confidence intervals are for estimation. |
| 1:38:00 | Associated predictors and multicollinearity (applet)In the applet you can induce a correlation between the two predictor variables, and that starts to mess up your multiple regression results. A very common pattern is a significant effect in the simple model that disappears the moment you add a second, correlated predictor, because the standard error increases. |
| 1:43:17 | Associated predictors versus interactionAn association between predictors is about the value of one variable affecting the value of another variable. An interaction effect is about one predictor influencing the effect of another predictor, and you can have either one without the other. |
| 1:47:27 | RecapIn mediation analysis we model dependence in a causal way, so be sure that your methodology, your experimental manipulation, allows that. We test mediation by looking at the p-value for the indirect effect, a × b. |
22. Mediation
Summary and chapters of the recording
28 October 2026 · Recording · Slides · Handout
This lecture is about mediation, where the association between a predictor and an outcome runs (partly) through a third variable, the mediator. We start from multiple regression drawn as a path model, split the shared variance into a direct and an indirect part, and work through the presumed media influence example (does a front-page article change people’s intention to buy sugar, through how much media influence they perceive?). We fit the three regressions ourselves, combine the a and b paths into the indirect effect (a × b), and test it. After the break we discuss why a mediation model needs an experimental design before it supports causal claims (equivalent models), run the analysis in JASP’s Process module, and return to the applet to separate associated predictors from an interaction.
Open the recording and skip to the chapter you need. The chapters come from the lecture transcript, so one may start a few seconds early or late.