Mixed-Effects Models

You use Mixed-Effects Models whenever your data comes in natural groups—students in classrooms, patients in hospitals, repeated measurements from the same person, or sales across different stores. After this lesson, you will be able to recognise these grouped structures in your work and understand why standard analysis tools lie to you when they ignore them. This skill lets you correctly estimate how a treatment works across different contexts, such as whether a new teaching method helps all students equally or only those in specific classes, without being misled by the hidden connections between people in the same group.

Consider a dataset from a school where you measure test scores for many students across several classrooms. Some of the variation in scores comes from each student's individual ability, but some comes from the classroom environment: a great teacher lifts everyone’s score, while a noisy room drags everyone down. If you ignore the classroom structure and simply average scores as if every student were independent, you are pretending that students in the same room do not influence or resemble each other more than strangers do. That is false. Mixed-effects models admit there are two kinds of influences: those that apply to everyone equally (fixed effects), such as 'each extra hour of sleep adds 2 points to a score', and those that vary by group (random effects), such as 'Classroom 3 gets a +5 point bump on average, while Classroom 7 gets a -3 point penalty'. The model estimates both the overall rule for everyone and the specific tweaks for each group. These group-specific tweaks are not just noise to be removed; they tell you which groups are truly different and by how much. Crucially, this approach allows you to make reasonable predictions for new groups that have little or no data by borrowing strength from all the other groups.

Imagine you are a cricket coach assessing the batting ability of players from different clubs. Each player has a baseline talent level—a fixed effect—but each club also provides different conditions: a specific pitch, a particular coach, and a unique practice routine—these are the random effects for that club. You take a few measurements of each player's score. Some players have played very few matches, so their raw average is highly uncertain. A player who averaged 10 in two matches might actually be talented but unlucky; a player who averaged 90 might be mediocre but lucky. A smart coach does not trust those raw averages blindly. Instead, she adjusts the extreme averages toward the overall average of all players in the league. This adjustment is called 'shrinkage' or 'partial pooling'. The coach uses information from the entire league to correct the noisy estimates for individual players. In mixed-effects models, the group-specific effects do exactly this: they pull small-group estimates towards the global mean (the average of all groups), pulling more strongly when a group has very little data. This analogy holds until you ask what defines a 'group'. In real data, groups can be repeated measurements from the same person, patients in a hospital, or lanes on a highway. The amount of shrinkage is not arbitrary; it is determined by the data itself, specifically by how much variation exists between groups compared to within groups. If there is no difference between groups, the adjustment disappears; if the differences are huge, the model treats each group as entirely separate.

A mixed-effects model is a statistical tool that combines general rules that apply to all observations with specific adjustments that vary by group, describing how an outcome depends on input variables while acknowledging that data points within the same group are related rather than independent.

You have notes on ANOVA (Analysis of Variance), which splits total variation into differences between groups and differences within groups; a mixed-effects model does this but also quantifies the uncertainty of those group differences correctly. It extends your work with Hypothesis Testing, because standard tests assume data points are independent—a condition violated by clustered data—whereas mixed-effects models correct for this by explicitly modelling the correlation within groups.