1
Supplementary Materials for
The comparative strength of biodiversity trends in marine and terrestrial
5
assemblages
Shane A. Blowes and Sarah R. Supp, Laura H. Antão, Amanda Bates, Helge
Bruelheide, Jonathan M. Chase, Faye Moyes, Anne Magurran, Brian McGill,
Isla Myers-Smith, Marten Winter, Anne D. Bjorkman, Diana Bowler, Jarrett
10
E.K. Byrnes, Andrew Gonzalez, Jes Hines, Forest Isbell, Holly Jones, Laetitia
M. Navarro, Patrick Thompson, Mark Vellend, Conor Waldock, Maria Dornelas
Correspondence to: sablowes@gmail.com, supps@dension.edu, maadd@st-andrews.ac.uk
15
This PDF file includes:
Materials and Methods
Figs. S1 to S18
Table S1
20
Materials and Methods
25
Data description and pre-processing
The BioTIME database represents the largest global effort mobilizing assemblage time series.
It includes 386 studies, and currently holds over 12 million records of abundance for over 45
thousand species across plants, invertebrates, fish, birds and mammals (37). Analyses
30
presented in this study used only time series of abundance data (i.e., studies that recorded
counts of the number of individuals for each species in an assemblage).
2
As we were interested in quantifying biodiversity change at the local scale, studies with
multiple sampling locations and extents greater than 71.7 km
2
(n = 126) were partitioned into
96 km
2
equal area icosahedron hexagonal grid cells (39). This threshold was determined using
the spatial extents of studies that reported sampling in only a single location, and was
calculated as the mean plus one standard deviation of the extent in these studies. Studies with
5
a single location, and those with extents < 71.7 km
2
were assigned to the grid cell in which
their centre latitude and longitude were located. The sample locations from all other studies
were assigned to cells based on the latitude and longitude of individual samples, gridding the
initial sampling extent of a study into multiple different cells. Each cell-level time series was
given a unique identifier that was the concatenation of the study ID and the cell reference
10
number, allowing the integrity of each study within each grid cell to be retained for analyses.
We then collated species within each unique study-cell combination for each year, resulting in
new assemblage time series within grid cells. Most grid cells contained only a single time
series (n = 32,878). For those that have more than one time series (n = 6,487), our
concatenation of study ID and the cell reference number means that each cell-level time series
15
in our analysis was comprised of samples from only one study, and that important study-level
considerations (e.g., sampling method) were consistent within each time series. In total, this
process yielded 51,932 time series distributed among 39,365 cells.
To minimize the effect of unobserved species on our estimates of biodiversity change, we
20
calculated the abundance-based coverage (60) of each (annual) sample (mean = 0.95, sd =
0.11) within each cell-level time series, and removed all samples with coverage less 0.85.
This means that for the remaining time series included in our analysis, there was a <15%
chance that another sample in any given year of one more individual would represent a new
species.
25
3
Finally, and before calculating our measures of biodiversity, we used sample-based
rarefaction to standardize the number of samples per year within each cell-level time series.
This procedure prevents temporal variation in sampling effort from affecting diversity
estimates (61). Within each time series, we counted the number of samples taken in each year,
and identified the minimum. This minimum was then used to randomly resample each year
5
down to that number of samples, after which dissimilarity metrics of community composition
and species richness were calculated. We repeated this process 199 times for each time series,
recorded the values and took the median for dissimilarity and species richness in each year.
Using the median instead of the mean reduced the effect of any outlier samples on our
estimate, and meant we did not need to round to get an integer value for species richness. We
10
calculated community dissimilarity using pairwise Jaccard dissimilarity measured between
the first year and each subsequent year in the time series; and, species richness for each year
of the time series. To assess if changes in community composition were driven by species
replacement or changes in species richness, we partitioned total Jaccard dissimilarity in the
additive components of turnover and nestedness (45, 62). Additionally, to examine whether
15
any trends in compositional change were sensitive to our choice of comparisons with the
initial assemblage, we also calculated pairwise turnover and nestedness components of
dissimilarity between consecutive assemblages (i.e., t
1
compared with t
0
, t
2
compared with t
1
,
t
3
compared with t
2
, etc.).
20
The times series resulting from the gridding, filtering and standardization processes had a
mean duration of 5.5 years. To maximize the spatial and temporal coverage of the data, we
chose to include time series with only two samples, and the minimum duration (i.e., the time
elapsed between the samples) for these two-point time series was three years, but many have
longer durations (median = 6 years, maximum = 55 years). The time series span from the late
25
1800s to the present, with most data collected in the past 40 years (Fig. S2). The data set
4
includes 9013 time series with duration spanning more than two decades and 27,619 spanning
more than one. Temporal extent and start date varied substantially in the data, and because
time series have such varied start dates, the data includes more than 1500 time series at any
one point since the mid 1960s (Fig. S2).
5
Models of biodiversity change
We examined geographic patterns of biodiversity change using three complementary
hierarchical linear models. All models nested the cell-level time series into the original
studies from which they originated at the lowest levels, but differed in how these studies were
grouped geographically. Grouping the cell-level time series into studies accounts for the non-
10
independence of time series from within studies, and the higher-level groupings (described
below) allowed us to characterize biodiversity change for different levels of the data. For
concision, in the main text we focus on the biome-taxon model that has the richest detail and
the realm-latitude-taxon that is the simplest. However, we used all three models to examine
the robustness of our results to how the spatial groupings were defined.
15
Our first and most detailed model nested cells and studies first into taxonomic-habitat groups,
and secondly into ecological biomes; we refer to this model as the biome-taxon (BT) model.
The taxonomic-habitat groupings reflect the BioTIME metadata for each study and were
amphibians, benthos, birds, fish, invertebrates, mammals, marine plants/invertebrates, plants,
20
and multiple taxa (indicating studies that measured more than one taxonomic group). Cell-
level time series were assigned ecological biomes using the geographic center of the samples
within each cell. Specifically, where samples within a cell came from a single location, that
coordinate was used to assign the biome; for cells with samples from multiple locations,
biome was assigned based on the center of a convex hull drawn around all the samples within
25
the cell. We used biomes from the published Ecoregions of the World (EOW) datasets
5
available from The Nature Conservancy website (http://maps.tnc.org/gis_data.html, (40-43).
Specifically, for the terrestrial realm, we used the EOW biome classification that is based
largely on vegetation types (40). For freshwater regions, we used the Major Habitat Types
(MHTs) that are considered to be roughly equivalent to the EOW biomes for terrestrial
systems (42). Our marine biomes used the province-level from the Marine Ecoregions of the
5
World; these are coastal and shelf areas expected to be of relatively distinct biota (41). Where
cell locations fell outside these EOW classifications, biomes were assigned using the nearest
appropriate (terrestrial biome, freshwater MHT, marine province) neighboring EOW group,
specifically: terrestrial: n
study
= 11, n
time series
= 27; freshwater: n
study
= 3, n
time series
= 52; marine: n
study
=
12, n
time series
= 4526). Our data span 10 of the 14 terrestrial EOW biomes defined globally, 5 of
10
the 12 freshwater MHTs, and 33 of the 62 marine provinces. This model allowed us to
characterize variation at the level of biomes, taxon groups within biomes, and among studies
of the same taxon group within biomes.
Our model of intermediate complexity again first nested cells into studies. However, as
15
consistent biome classifications across realms were not available (e.g., both terrestrial and
freshwater biomes incorporate more detail of specific habitats compared to the marine biomes
that are based more strongly on geography), we replaced the terrestrial and freshwater biomes
with broader spatial groupings. Specifically, studies in terrestrial and freshwater systems were
grouped into continents, and we retained the marine biomes for studies in the marine realm.
20
Cells and studies were nested into a new grouping covariate that was the concatenation of
realm, region and taxon, and we refer to the model as the realm-region-taxon (RRT) model.
Realm was one of marine, terrestrial or freshwater; region was a continent (i.e., Africa, Asia,
Australia, Europe, North America, and South America) for studies in the terrestrial and
freshwater realms, and biome (defined above) for marine studies; taxonomic-habitat
25
groupings were the same as the BT model described above. This model allows to us to