Skip to contents

Splits the variance of each trait among levels, such as regions, sites, genera, species and populations, with a linear mixed model of random intercepts fitted by restricted maximum likelihood (Messier et al. 2010). The levels can form one nested hierarchy (factors) or two hierarchies crossed with each other (factors and crossed), such as a spatial one (region > site) and a taxonomic one (genus > species).

Usage

nested_variance(data, traits, factors, crossed = NULL, interaction = TRUE)

Arguments

data

Data frame with the traits and the factors.

traits

Names of the columns with trait values.

factors

Names of the columns with the nested factors, from the broadest level to the finest.

crossed

Names of the columns of a second hierarchy of nested factors, from the broadest level to the finest, crossed with factors (for instance c("genus", "species") when factors are spatial), or NULL (the default) for a single nested hierarchy.

interaction

With crossed, whether to add a level for the combination of the finest level of factors and the finest level of crossed (for instance the population, species by site). Without it, differences among populations of a species go to within.

Value

A data frame with one row per level and one for the variance within the finest level (within), and one column per trait with the share of variance of each level. The attribute variance holds the variances, singular whether each fit was singular and engine the package used. It has class nested_variance, with a plot() method.

Details

Nested or crossed. In a nested design each level of a factor lies within a single level of the factor before it, as sites within regions or individuals' populations within sites. Species are rarely nested within sites: the same species occurs at several sites. Put after site in factors, species then stands for each species at each site, that is, the population, and its share mixes two things: how much species differ from one another and how much populations of a species differ among sites. When species occur at several sites, give the spatial levels in factors and the taxonomic ones in crossed. The model then has a term for each level of both hierarchies and, by default, one for their interaction (the finest spatial level by the finest taxonomic level, for instance site:species, the population), which separates differences among species from differences among populations of the same species. The variance left (within) is among individuals of a population. A rule of thumb: if a column of factors has labels that recur under several levels of the factors before it and these labels are the same units (the same species), the design is crossed; if they are only reused names (site 1, 2, ... in each region), it is nested.

Checks. Within each hierarchy, a factor that does not split any group of the factors before it (for instance species after populations, since each population holds a single species) adds no level of variation, and the function stops, because the order of the factors is then most likely reversed. A factor of factors whose labels recur under several groups of the factors before it gets a warning that points to crossed.

Fitting. Nested models are fitted with nlme::lme(). Crossed models, and nested models for which nlme::lme() fails to converge, are fitted with lme4::lmer(), which needs the package lme4. When a level cannot be told apart from the others (for instance genera that hold a single species at most sites), lme4::lmer() sets its variance to zero (a singular fit) instead of failing; the function then warns, and the attribute singular says which traits were affected. A zero share in a singular fit means that the data do not separate that level, not that it does not vary. With few levels of a factor (two regions, say), the variance of that factor is estimated with very few degrees of freedom and is very uncertain.

References

Messier, J., McGill, B. J., & Lechowicz, M. J. (2010). How do traits vary across ecological scales? A case for trait-based ecology. Ecology Letters, 13, 838–848. doi:10.1111/j.1461-0248.2010.01476.x

Examples

set.seed(1)
d <- data.frame(region = rep(c("north", "south"), each = 60),
                site = rep(paste0("s", 1:6), each = 20),
                species = sample(c("a", "b", "c"), 120, replace = TRUE))
d$population <- paste(d$site, d$species)
d$trait <- c(a = 10, b = 20, c = 30)[d$species] + rnorm(120)
res <- nested_variance(d, "trait", c("region", "site", "population"))
res
#>        level        trait
#> 1     region 1.741575e-08
#> 2       site 1.125857e-17
#> 3 population 9.862820e-01
#> 4     within 1.371800e-02
if (requireNamespace("ggplot2", quietly = TRUE)) plot(res)

# the same species at every site: species crossed with sites, and populations differ
d <- expand.grid(individual = 1:5, species = paste0("sp", 1:8), site = paste0("s", 1:12))
d$region <- ifelse(d$site %in% paste0("s", 1:6), "north", "south")
species_effect <- rnorm(8, 0, 3)
population_effect <- rnorm(8 * 12)
d$trait <- species_effect[d$species] +
  population_effect[interaction(d$species, d$site)] + rnorm(nrow(d))
nested_variance(d, "trait", c("region", "site"), crossed = "species")
#> Warning: singular fit for `trait`: the variance of at least one level is estimated at zero because the data do not separate it from the others
#>          level     trait
#> 1       region 0.0000000
#> 2         site 0.0000000
#> 3      species 0.6959101
#> 4 site:species 0.1795086
#> 5       within 0.1245813