EDF9850_Unit4_Review_Part2

Overview

Lesa Hoffman explains how centering choices in multilevel models determine whether predictor slopes represent within-cluster, between-cluster, or contextual effects, and how those interpretations change when predictors and interactions enter the same model. She also gives practical guidance on cross-level moderation, random-slope model building, multilevel data structure, and ways to learn the material through real data, simulation, and teaching.

Key takeaways

A cluster-mean-centered level-one predictor estimates a within-cluster slope; if its cluster mean is omitted, the model assumes the between-cluster slope is zero.
Entering an uncentered or constant-centered level-one predictor by itself combines within and between information and assumes the two slopes are equal.
A contextual effect measures the added association of a cluster's average predictor value after controlling for the individual's own value; it equals the between effect minus the within effect.
For level-one predictors with cross-level interactions, a lone interaction can combine within- and between-slope moderation; including matching predictor and cluster-mean terms separates those effects.
A sound model-building order is fixed slope, random slope, then cross-level interaction, because the interaction explains variation in the random slope.
Working with meaningful data, simulating multilevel data, and teaching the model to someone else are practical ways to deepen understanding of its assumptions and components.

Chapters

0:00 Cluster-Mean and Constant Centering of Level-One Predictors

Cluster-mean centering subtracts each cluster's mean from a level-one predictor, isolating within-cluster variation.
Including only the centered level-one predictor estimates a within effect while assuming the between effect is zero; omitting it while including the cluster mean assumes the within effect is zero.
Constant centering subtracts a chosen value; grand-mean centering is the special case that uses the sample mean. For binary predictors, retaining the original 0/1 coding is often more interpretable than cluster-mean centering.

6:30 How Predictor Combinations Define Within, Between, and Contextual Effects

A level-one predictor entered alone without separating its cluster mean has a smooshed slope: the model assumes within- and between-cluster effects are equal.
With constant centering, entering both the level-one predictor and its cluster mean separates the level-one within effect from the level-two contextual effect.
A between effect equals the within effect plus the contextual effect; equivalently, contextual equals between minus within, which can be tested using linear combinations of fixed slopes.
A contextual effect asks whether the group's average predictor value matters after controlling for the individual's predictor value—for example, whether a student's school-average SCES adds an effect beyond the student's own SCES.

13:00 Cross-Level Interactions and Moderation of Within and Between Slopes

A level-two moderator Z can interact with the cluster-mean-centered level-one predictor to let the within slope change across clusters.
Interacting Z with the cluster mean lets the between slope vary with Z; both interactions can be included when both forms of moderation matter.
With constant-centered level-one predictors, an interaction entered alone combines the within and between moderation effects; adding the corresponding cluster-mean interaction separates them into within and contextual components.
When using constant centering, Hoffman advises applying the same transformation to a level-one predictor and its level-two mean to avoid unintentionally combining effects.

17:30 Why a Cluster-Mean Moderator Creates Quadratic Terms

If the moderator is the cluster mean of the predictor itself, interactions between the individual-level predictor and that mean produce quadratic-style terms.
These terms allow the within effect or between effect to depend on the cluster's average predictor level.
The cluster-mean interaction is needed to separate moderation under the constant-centering parameterization, but not when the model has already separated within and between components.

20:00 Choosing a Model and Ordering Random Slopes and Interactions

Choose the parameterization based on the target: cluster-mean centering plus cluster means gives between effects directly, while constant centering can yield contextual effects.
For categorical predictors, constant centering is often easier to interpret; continuous predictors can also be cluster-mean centered to estimate within and between effects.
Add the fixed slope before its random slope, then add cross-level interactions, because the interaction is intended to explain variation in the random slope.
To assess variance explained, build models from a baseline and add random components before predictors that explain them; when effect sizes are not the goal, a fuller model may be fitted at once if its random structure is appropriate.

26:00 Studying Multilevel Models, Higher Levels, and Next Steps

Hoffman recommends applying models to data that matter to the researcher, simulating data to understand how level-one and level-two variables and random effects fit together, and explaining the model to someone else.
For a possible six-level model, assess variance at each level with empty models; higher levels may contribute less, and a level with only about six groups may be better represented with fixed categorical effects.
Cross-level interactions can use binary or continuous level-two moderators; they answer why a slope varies across clusters rather than being required automatically.
Hoffman closes by encouraging students to persist with the difficult material, noting applications in workshops, research collaborations, and authorship, and reminds them that Homework 2 is due the following week.

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.