EDF9850_Projects_Review
Watch on YouTube →
Overview
Lesa Hoffman reviews students’ multilevel-model project plans, explains how to diagnose R warnings such as singular fits and convergence problems, and clarifies expectations for final presentations. She emphasizes a 12 to 17 minute presentation focused on data, model-building decisions, results, and lessons learned, then distinguishes longitudinal residual-correlation structures from lagged predictors using an eight-week stress and psoriasis study.
Key takeaways
- A near-zero random-slope variance can make a multilevel model difficult to fit; inspect the variance and random-effect correlations, and consider removing the slope or rescaling its predictor.
- Project presentations should spend more time on sampling, centering, model-building decisions, and analysis roadblocks than on a conventional conference-style literature introduction.
- A 12 to 17 minute presentation should use readable visuals rather than dense tables or AIC, BIC, and likelihood-ratio statistics.
- In longitudinal models, residual correlations describe dependence in the outcome over time, whereas lagged predictors test how earlier measurements relate to a later outcome; the two structures do not need matching lags.
- The psoriasis example required residual correlations through three weekly lags, but the tested stress predictors showed a small prior-week effect and a separate association between average stress and average severity.
Chapters
- Lesa Hoffman encourages students to keep exchanging informal project updates, including recorded walkthroughs when work schedules prevent meetings.
- In R, singularity generally signals redundant random effects; convergence or tolerance warnings can occur when random-slope variances approach zero or random-effect correlations approach -1 or 1.
- Inspect model output alongside likelihood-ratio tests, and consider removing near-zero random slopes or rescaling predictors to improve numerical precision.
- Cross-level interactions can explain away a small random-slope variance and trigger convergence warnings, a process students can describe in their presentations.
- The project deliverable is a presentation, not a paper; students may revise a presentation and submit a recording for regrading if a live presentation goes poorly.
- Plan for 12 to 17 minutes, ideally about 15, with roughly two introductory slides and enough methods detail to explain the sampling design and variables.
- Describe centering choices, model-building steps, roadblocks, and what was learned rather than presenting only final results.
- Avoid AIC, BIC, likelihood-ratio statistics, and dense number tables; use clear diagrams and pictures to explain models and answer research questions.
- Every member of a presentation pair must speak, with work divided reasonably evenly and not split so one person handles only the introduction and discussion.
- For Zoom presentations, pairs should coordinate who shares and advances slides; students should use a good microphone and appear on camera when possible.
- Presentations are intended to be recorded in Zoom and shared on Canvas, not YouTube, with privacy and FERPA concerns in mind.
- Each student completes an independent peer review, sends it to the assigned presenter, and CCs Hoffman; reviews are due the Wednesday after Thanksgiving.
- Lecture 6, an eight-slide lesson with a 25-minute recording, is available in Canvas.
- The upcoming class will discuss formative assessment 6 and example 6A on cross-classified random effects, while example 6B on repeated measures will be provided as a separate recording.
- Students may email Hoffman about whether they prefer asynchronous examples or more class time for project feedback and partner work.
- Project check-in revisions should be emailed to Hoffman; Thanksgiving week has no class, followed by one final meeting after break.
- Hoffman says students do not need to master the full matrix mathematics yet; the main goal is to recognize additional correlation structures in longitudinal data.
- Random-effect variances and covariances capture some within-person relationships, while level-one residuals may require their own correlation structure.
- Separating these components gives longitudinal models more flexibility than placing all covariance patterns into one matrix.
- This decomposition connects multilevel longitudinal models to traditional MANOVA and repeated-measures ANOVA.
- A study measured weekly stress and psoriasis severity over eight weeks; its spaghetti plot showed limited within-person change and fluctuation.
- The outcome model used a random intercept, heterogeneous weekly variances, and residual correlations at lags 1, 2, and 3 to represent dependence across weeks.
- Modeling the outcome’s residual covariance is separate from testing whether current or prior-week stress predicts psoriasis severity.
- Prior-week stress had a small association with this week’s psoriasis severity, while people with higher average stress also tended to have higher average disease severity.
Transcript excerpt
0:00 I'm officially starting recording since we're doing that's here. So, projects. All of you should have gotten feedback from me via email or questions from me about what you're doing via email, one of the two. And I would say I'm sorry but not sorry for sending them back already. I know how that feels when you work on something and it's a big thing and you finally get it off your plate so you can tackle all the other things that you didn't do while you were doing that only to have it come right back. Um I will give you a brief story of my own about that. When I was in graduate school, I wrote a grant application for my comprehensive exams and I was allowed to get feedback on preliminary versions from my committee and it was a very long undertaking and I was so happy to finally get a draft together. I sent it
0:49 to someone whose opinion I valued. I was like, "Oh, thank goodness, it's over." Eight hours later, she sends me an email. "I've read it. Come to my office tomorrow and let's chat." I was like, "Okay, well, I had one day of feeling relief but then the next day I popped into her office and I'm like, 'Hi, you wanted to see me?'" and she has it printed out and she goes, "Here's what's wrong with this." Thunk. Which was maybe not the reaction that I was hoping for. But as she was talking, I was like, "She's right. She's right. She's right. She's right." Like and so I appreciated having more time to fix the things that she had identified as issues as opposed to her dragging her feet and not get not getting me feedback. And that grant was funded, by the way. I
This is the start of the transcript. Save this video free for the whole transcript, searchable and time-coded.
Related notes
Summary, takeaways, and chapters were generated by AI from the video's transcript and may contain errors. The video belongs to its creator, Lesa Hoffman.