Can the calculators on this page be used to analyze mediation analysis results from published research?

In principle the calculators can be used for this purpose. One motivation behind implementing these particular methods is that only summary information from an analysis is required in order to compute inference and intervals for the indirect effect. The raw data is not required as it is for resampling methods such as bootstrapping. If there is missing data or non-normality, it is desirable that the original analyst has chosen an appropriate modeling technique to account for this. In addition, one is not tied to a particular proprietary statistical package in order to use these calculators; you may use whatever statistical program you choose to estimate the appropriate regression models, structural equation models, or hierarchical linear models.

Can the calculators on this page be used if there is missing data in my dataset?

Yes, provided that the missingness is appropriately handled through a technique such as multiple imputation, or so-called full information maximum likelihood (using the term in the structural equation modeling literature). If this is done, then you can just use the point estimates and standard errors from such an analysis and use directly in the above calculators. Biesanz et al (2010) conducted simulations using the partial posterior and Hierarchical Bayesian approaches with similar techniques, and R code is also provided below. If you have a substantial amount of missing data, be wary of approaches that use so-called listwise deletion as this is not an ideal approach. For instance, Andrew Hayes’ PROCESS macro for SPSS and SAS can apparently only handle complete data at this time and will do listwise deletion by default, e.g., see the FAQ for PROCESS.

Can the calculators on this page be used with nonlinear models (e.g., dichotomous mediator and outcome)?

No. They were not designed for such use. The R package mediation may also be of help, as is Mplus code on Andrew Hayes’ website in this FAQ section.

Can the calculators on this page do sensitivity analysis?

Can I use these calculators with multilevel models?

It depends a little on the design and at what level are the measured predictor, mediator, and outcome. If mediation is all at level 1, then usually I would start with reading this paper: Bauer, Preacher, Gil (2006), and look here if your slopes are all random: Kris Preacher’s website. If you need to calculate the asymptotic covariance matrix of random effects and you’re doing this with the lme4 package in R, it may be possible to obtain this with the bootmlm package (vcov_vc function). Otherwise vcov should be able to obtain this for the fixed effects.