3 Multiple random effects
Previously we used (1|group)
to fit our random effect. Whatever is on the right side of the |
operator is a factor and referred to as a “grouping factor” for the term.
3.0.1 Crossed or Nested
When we have more than one possible random effect these can be crossed or nested - it depends on the relationship between the variables. Let’s have a look.
A common issue that causes confusion is this issue of specifying random effects as either ‘crossed’ or ‘nested’. In reality, the way you specify your random effects will be determined by your experimental or sampling design. A simple example can illustrate the difference.
3.0.1.1 1. Crossed Random Effects:
Crossed random effects occur when the levels of two or more grouping variables are crossed or independent of each other. In this case, the grouping variables are unrelated, and each combination of levels is represented in the data.
Example 1: Let's consider a study examining the academic performance of students from different schools and different cities. The grouping variables are "School" and "City". Each school can be located in multiple cities, and each city can have multiple schools. The random effects of "School" and "City" are crossed since the levels of these variables are independent of each other.
lmer(y ~ x + (1 | School) + (1 | City), data = dataset)
Example 2: Imagine there is a long-term study on breeding success in passerine birds across multiple woodlands, and the researcher returns every year for five years to carry out measurements. Here "Year" is a crossed random effect with "Woodland" as each Woodland can appear in multiple years of study. Imagine a researcher was interested in understanding the factors affecting the clutch mass of a passerine bird. They have a study population spread across five separate woodlands, each containing 30 nest boxes. Every week during breeding they measure the foraging rate of females at feeders, and measure their subsequent clutch mass. Some females have multiple clutches in a season and contribute multiple data points.
lmer(y ~ x + (1 | Year) + (1 | Woodland), data = dataset)
3.0.1.2 2. Nested Random Effects:
Nested random effects occur when the levels of one grouping variable are completely nested within the levels of another grouping variable. In other words, the levels of one variable are uniquely and exclusively associated with specific levels of another variable.
Example 1. Consider a study on the job performance of employees within different departments of an organization. The grouping variables are "Employee" and "Department". Each employee belongs to one specific department, and no employee can be part of multiple departments. The random effects of "Employee" are nested within the random effects of "Department" since each employee is uniquely associated with a specific department.
lmer(y ~ x + (1 | Department/Employee), data = dataset)
Example 2: In the same bird study female ID is said to be nested within woodland : each woodland contains multiple females unique to that woodland (that never move among woodlands). The nested random effect controls for the fact that (i) clutches from the same female are not independent, and (ii) females from the same woodland may have clutch masses more similar to one another than to females from other woodlands
lmer(y ~ x + (1 | Woodland/Female ID), data = dataset)
or if we remember year we have a model with both crossed and nested random effects
lmer(y ~ x + (1 | Woodland/Female ID) + (1|Year), data = dataset)
For more on designing models around crossed and nested designs check out this amazing nature paper Krzywinski et al. (2014)