Understanding levels of variation and mixed models

Data that has some kind of hierarchical structure to it is very common in many fields, but is rarely discussed in introductory statistics courses. Terms used to describe this kind of data include hierarchical data, multi-level data, longitudinal data, split-plot designs or repeated measures designs. Statistical models used for these types of data include mixed-effects models (often abbreviated to just mixed models), repeated measures ANOVA and generalised estimating equations (GEEs).

All of these terms and models arose from different contexts, but they share a common feature. Observations are not independent, as many classical statistical methods assume, and there is structure to the dependence which can be incorporated into a statistical model.

Using a statistical model which doesn’t account for the hierarchical nature of this data will give incorrect results.

Common examples of hierarchical data

Data with two levels of variation often arise when multiple measurements are made on the same units of observation. In the case of designed experiments, treatments may also be assigned to different levels of the hierarchy. Factors are commonly described as within-subject (varying at the lowest level) or between-subject (varying at a higher level). Some examples:

  • Measurements made on the same people at several points in time. This is often called longitudinal data. In this example, time would be a within-subject factor and most other variables of interest—e.g., treatment, age or gender—would be between-subject.
  • Measurements made at different depths of a number of rock core samples. In this example, the depths would be a within-subject factor and the location where the sample was obtained would be a between-subject factor.
  • Assigning different treatments to different legs, arms or eyes of a number of people. For example, one eye may be given a new drug and the other eye a placebo. In this example, treatment is a within-subject factor.
  • A split-plot experiment in agriculture: splitting plots of land into sections, planting different crops in each section, and using different irrigation methods on different plots. In this example, variety is a within-plot factor and irrigation is a between-plot factor.
  • A split-mouth design in dentistry: assigning different treatments to different parts of participants’ mouths.

It is possible to have more than two levels of variation. Some examples from different fields of research:

  • Students within classrooms within schools.
  • Repeated surveys administered to individuals within organisations.
  • Glands within lesions within patients.
  • Blocks of land divided into plots divided into subplots.

In situations with multiple levels, it is common to describe variables based on the level at which they are measured or assigned, e.g. student-level, classroom-level or school-level.

Random effects

Multi-level data is commonly modelled using mixed-effects models, which get their name because they have both fixed effects and random effects. Fixed effects are the kind of explanatory variables you may be used to in ANOVA or linear regression: you would like to directly estimate the effect of these variables on your outcome. For example: treatment (drug or placebo) and time; crop variety and irrigation; depth and location of rock core samples. In these examples, the random effects would be the variables which group together correlated observations: participants in a trial; plots of land; rock core samples.

Here are two different ways to think about random effects:

  1. Random effects are factors where the individual levels are random samples from a larger population, or can be thought of in this way.
  2. Random effects are factors where you don’t care about the actual effect they have on your outcome, just their ability to account for correlation between observations.

For a random effect, instead of estimating the effect of each specific level of the factor (e.g. each individual in a study), the model estimates the variance explained by that factor. This is sometimes reported as the proportion of variation explained at each level, e.g. 63% of variance was at the individual level.

Random effects can be “nested” inside other random effects if there are more than two levels of variation. For example, “classroom within school” could be specified as a random effect.

Simpler analysis options

It is sometimes possible to simplify an analysis if there are no variables which distinguish individuals at a particular level. For example, consider an experiment in which treatments were randomly assigned to litters of pigs but measurements were made on individual pigs, with no pig-level variables (e.g. sex) of interest in the analysis. In this situation, analysing litter averages would be a simpler analysis providing the same results as the mixed model.

Another common situation is when there is a single within-subjects factor with two levels, for example before and after measurements. This kind of design can be analysed with a paired t-test or Wilcoxon signed rank test.

As a practical consideration, random effects work best when they have a reasonable number of different levels. It may be better to treat a factor which is conceptually random as a fixed effect instead in some cases; e.g. if you are studying students from two or three schools, school should probably be a fixed effect rather than a random effect. There are also situations where this approach is not appropriate.

If in doubt about how to analyse your multi-level data, consult a statistician.

Study design considerations

Effects at the lowest level of the hierarchy (e.g. within-subject) are usually estimated more precisely than effects at higher levels (e.g. between-subject). Or equivalently, tests of within-subject effects tend to be more powerful than tests of between-subject effects. One intuition about this is that there are more observations at the lowest level (e.g. number of subplots) than there are at higher levels (e.g. number of plots). Another way to look at this is that for within-subject factors, each individual unit of observation is effectively their own control.