EDF850_Example5b_Part2
Overview
Lesa Hoffman uses a fictional longitudinal study of children’s depression to show how multilevel models can separate starting age, change over time, and gaming behavior. Centering base age at 10 reveals that older children start with lower depression but increase faster, while separating gaming into person averages and occasion-specific deviations shows that gaming more than usual is associated with higher depression at that occasion, not that children who game more on average have worse outcomes.
Key takeaways
- In longitudinal data, starting age can represent cohort differences as well as age-related development, so Hoffman models base age separately from time rather than treating age as the whole explanation.
- With time centered at the first occasion and base age centered at 10, the reference child’s predicted depression is 8 and the expected increase is 1 unit per year.
- Older children in the example start with lower depression but have a more positive increase over time; base age explains 11% of intercept variance, while its interaction with time explains 57% of slope variance.
- Gaming’s roughly 47% between-child variance means a single occasion-level score risks conflating children who usually game more with occasions when a child games more than usual.
- In Hoffman’s simulated example, average gaming is not significantly associated with baseline depression or change, but an occasion-specific increase above a child’s usual gaming level is associated with higher depression at that occasion; this is an association, not evidence of causation.
- For plots of multilevel interactions, using color and line style for distinct predictors makes effects easier to interpret, and the plotted directions and parallelism should be checked against the model.
Chapters
- Age alone can smoosh developmental change together with cohort effects: being a seventh grader after COVID may differ from being one before COVID.
- The example codes children’s starting age, from 8 to 12, as a level-two, time-invariant predictor centered at 10.
- The reference child is age 10 at the first occasion, with predicted depression of 8 and an increase of 1 unit per year.
- Older starting age predicts lower depression at baseline but a more positive rate of change; the model explains 11% of intercept variance and 57% of slope variance.
- Gaming has an intraclass correlation of about 46% to 47%, so a single gaming score would conflate differences between children with changes within a child.
- Hoffman treats gaming as an outcome and checks categorical occasion means, variances, and covariances before choosing a centering strategy.
- Average gaming is about 3 hours per day, with no clear omnibus time effect and relatively stable variance and correlations.
- A random linear change model shows a near-zero slope variance and no significant random slope, supporting person-mean centering for this example.
- The level-one predictor is current gaming minus the child’s person mean, capturing whether gaming at an occasion exceeds that child’s usual level.
- Person-mean gaming is centered at 3 hours per day and included as a level-two predictor of baseline depression and possible change over time.
- The reference growth curve is for a 10-year-old who games 3 hours per day on average: predicted depression is 8 at baseline, increasing by 1 unit per year.
- In the simulated example, average gaming does not significantly predict baseline depression or the rate of increase, while gaming more than usual is associated with significantly higher depression at that occasion.
- The within-person gaming deviation explains residual variance because it is a level-one predictor without a random slope.
- A level-one predictor with a random slope can explain that slope’s variance; a cross-level interaction targets the random slope when the level-one effect varies across people.
- Adding a nonsignificant person-mean predictor can increase estimated intercept variance as an artifact, so Hoffman cautions against treating that increase as substantive.
- Potential next questions include whether the within-person gaming effect differs by child, starting age, or average gaming level.
- The age plot uses blue for younger starters and red for older starters, showing older children’s lower baseline scores and steeper depression trajectories.
- Dashed versus solid lines distinguish person-mean gaming levels; the small separation reflects the minimal average-gaming effect.
- A separate plot uses deviation from usual gaming on the x-axis, not time, to show the positive association with depression and the lack of a clear difference by average gaming.
- Graph code is in the supplementary materials; Hoffman says Claude helped refine plotting code and recommends checking that plotted slopes, parallelism, colors, and line styles match the model.
Transcript excerpt
0:00 Since we're here, here's my plan, folks. We've got a few more models to do, not a whole lot, and then we will end as a group after that. If folks want to hang out and chat with me or chat with each other about your projects, I'm happy to do that. Otherwise, if you uh have decided you're done for the day, then that's fine, too. Fair. Fair. All right. So, where we've been is focusing just on time and the modeling of time in this fictitious study of changes in depression as children age. Now the question is what happens to age? So it may have struck you as a little bit weird that up to this point we've ignored how actually old they are. It's a little weird but the reason for that is because if you just put in age and you track depress depression as a
0:48 function of age that is a smooshing problem because it may be the case that it matters when you were 10. Not just that you were 10. So when you have these overlapping cohorts of people and they start at different ages, you can have what are known as cohort effects where it's different to be at that age or in that grade at a different time. So if you think about like COVID is the biggest example of this, right? It's very different to be a seventh grader after COVID than before COVID in terms of what you're expected to do with computing and whatnot. Um my son doesn't have lockers anymore. I don't understand why that's the case, but that went away after COVID and I'm really confused as to why they expect 12-y olds to keep all their stuff with them all day long. It seems impossible
This is the start of the transcript. Save this video free for the whole transcript, searchable and time-coded.
In this course
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.