Dr. Guy Prochilo 🏳️‍🌈

Many power analyses I read in published manuscripts are not reported with sufficient transparency to understand what on earth they are doing! What is the effect estimate? Where did it come from? For what test does it apply? Etc. Here is a reporting guide I give students #phdchat

On the lookout for academics who want to pick up some psychology marking work at ISN Psychology—research proposals, theses, and reports—at $72/hour. Must hold a PhD in Psychology and be based in Australia. DM for details. #PsychologyJobs #PhDJobs #PhDChat #RemoteWork

Almost ready to submit my paper on a randomized trial! Every number, figure, table, and even the CONSORT diagram will be fully reproducible and can be generated on the fly. Can't wait to share it! #phdchat #rstats

note: all these images are just random screenshots of papers where the methods section makes claims about Little's test that it doesn't actually provide. Some even claim it provides support for MAR...

To sum it up: Missing data is complicated. Little’s test probably isn’t doing what you think it is, and MCAR is often unlikely. Don’t rely on Little's test as an excuse to avoid properly handling your missing data.

🧵 Little's MCAR test is often used to check if missing data is "Missing Completely at Random" (MCAR). However, it might not actually do what many researchers think it does. Let's look at the details! #phdchat #statstwitter

Enders, C. K. (2022). Applied Missing Data Analysis, Second Edition (2nd ed.). The Guilford Press.

Missing at Random (MAR) means that the probability of data being missing is purely random after conditioning on observed data (e.g., in multiple imputation). That is to say, it depends only on the observed data, not the missing values themselves. #phdchat #statstwitter

@AJThurston time to throw out all this fancy dataviz and return to the comforting embrace of old trusty barplot✨we were naïve to abandon rectangle and stick

Do you think raincloud plots are too much detail for lay audiences? When would they be most appropriate?

Just wrapped up today's lecture on missing data! Covered item- and construct-level missingness, detection, mechanisms, analysis (using Welch's t-tests with missingness indicators), solutions, and a theory intro to multiple imputation (since #jamovi lacks modules). #phdchat

Borenstein, M. (2019). Common Mistakes in Meta-Analysis and How to Avoid Them. Biostat Inc.

Don't use a significance test of heterogeneity (eg Cochran's Q) to choose between fixed or random effects meta-analysis. If studies estimate a single common effect size, use fixed effects. If studies differ in populations, measures, or protocols, opt for random effects. #phdchat

Borenstein, M., Hedges, L. V., Higgins, J. P. T., & Rothstein, H. R. (2009). Introduction to Meta-Analysis (1st ed.). Wiley.

A fixed effects meta-analysis assumes a single common effect size across all studies, which is often unrealistic given differences in interventions, populations, and measurements. In psychology especially, a random effects meta-analysis is usually the more plausible #phdchat

Legha, A., Riley, R. D., Ensor, J., Snell, K. I. E., Morris, T. P., & Burke, D. L. (2018). Individual participant data meta‐analysis of continuous outcomes: A comparison of approaches for specifying and estimating one‐stage models. Statistics in Medicine, 37(29), 4404–4420.

A one-stage IPD meta-analysis aggregates data across all studies and can be analyzed using a linear mixed model with the study ID as a random effect. This aligns with a random effects meta-analysis where we assume the true effect size differs across studies #phdchat #statstwitter