Statistical methods
Need help?
- Do not hesitate to ask Jeanette Mumford for help with any statistical analyses.
Multiple testing
- Methods for multiple testing correction should be chosen prior to data analysis and reported in the preregistration document.
- The nature of grouping for multiple testing corrections (e.g., parameter within single models, sets of models, etc) should be pre-specified.
- For fMRI data, if using cluster-based thresholding the cluster forming threshold must be Z ≥ 3.1 if using parametric thresholding. There are no restrictions on threshold when using nonparametric thresholding, but the threshold should be pre-specified.
Cross-validation
- Proper cross-validation practices should be considered. Refer to https://pubmed.ncbi.nlm.nih.gov/31774490/ for more detail.
- In-sample model fit indices should not be reported as evidence for predictive accuracy
- The cross-validation procedure should encompass all operations applied to the data
- Prediction analyses should not be performed with samples smaller than several hundred observations
- Multiple measures of prediction accuracy should be examined and reported
- The coefficient of determination should be computed using the sums of squares formulation and not the squared correlation coefficient
- Leave-one-out cross-validation should be avoided in favor of shuffle-split or K-fold cross-validation.
Mixed-effects models
- Mixed effects models should include all possible random effects, subject to the constraint of adequate model convergence.
- Convergence should be checked for all models, and models should be checked for highly influential observations.
- Simplification or random effects structure can be done if the p-value > 0.2 (https://www.google.com/url?q=https://www.sciencedirect.com/science/article/pii/S0749596X17300013&sa=D&source=editors&ust=1623857125982000&usg=AOvVaw2E_twuab7e56Nn_i1yTdmn) assuming convergence.
Model suitability and diagnostics
- Does the model properly test the hypothesis of interest?
- Do the data conform to the assumptions of the model?
- All model fits should be criticized to identify violation of any assumptions or presence of outliers.
Reporting results
- Effect size, parameter estimate, standard error, p-value and confidence interval should all be reported.
- In cases where null effects are reported, some measure of evidence for the null hypothesis (e.g. Bayes factors, equivalence tests) should be reported.
- Results for all analyses that were preregistered should be reported.