Summary of activities -- Workshop:
"Habitat and Climate Inference from the Structure of Mammal Communities"
August 21-27, 1998
John Damuth (Convenor)
Jan van Dam
Blaire Van Valkenburgh
Paleomammalogists have explored in recent years a
variety of techniques for inferring past climate and vegetation from structural
characteristics of fossil mammal communities and faunas. These techniques
hold enormous promise for expanding temporal and geographic ranges over
which we can make accurate, objective reconstructions of terrestrial habitat
characteristics (chiefly, vegetation) and local or regional climate. However,
we badly need a comparison, evaluation, and synthesis of the various methods.
This working group developed a global dataset of mammal faunas, by pooling
and standardizing the data participants contributed, as well as by compiling
additional information at the workshop. This dataset will allow us to evaluate
different variables and methods by checking them against the same global
dataset, in order to generate the most effective means of using mammalian
community data to infer climate and habitat. In future work we will apply
the techniques to Quaternary and Tertiary data, both to test out the techniques
in practice and to shed light on selected paleoecological questions.
Initial Presentations and Definition of the Problems
The group spent the first day and a half presenting
brief descriptions of previous and ongoing research, describing their datasets,
and discussing general topics and approaches. The wide range of subjects
considered defies easy summary, but several themes and issues emerged that
were discussed at greater length.
The Nature and Adequacy of the Variables ---
Our basic aim is to be able to use metrics that are
likely to be available from fossil mammal local faunas or individual localities
to reconstruct past vegetation and/or climate. Traditionally, this has
been approached by assigning fossil mammals to a number of functional categories,
such as diet classes, locomotor classes, and so forth, on the basis of
morphometric measurements (ecomorphology). Likewise, vegetation is usually
expressed in terms of some kind of classification scheme. Clearly, any
statistical results will be affected by the particular classifications
used. An appealing alternative is to use directly morphometric measures
that are considered to represent important ecological attributes such as
diet or locomotion, and to use combinations of continuous climatic variables
rather than classes. Both class-variables and continuous variables have
theoretical strengths and weaknesses. Because the group contains a number
of researchers at the forefront of ecomorphological research with fossil
mammals, we are able to compare an extensive array of such continuous morphometric
and categorical data in our combined dataset.
The Adequacy of Coverage of Extant Ecosystems --
We need to establish that we have covered as large a
range of extant climatic conditions as possible, with data of similar quality,
if we want our results to apply as broadly as possible. Also, we need to
know that our vegetational classifications are related coherently to climatic
variables if we want to use similar methods to predict each.
What we know about the Relationship between Climate and Mammals --
Andrews and Badgley both presented their unpublished
research in progress on the relationship between mammal species diversity
and climate. Working independently, they have converged on very similar
methods for treating southern Africa and North America, respectively, and
find significant relationships between mammal diversity and temperature
and moisture variables. Madden also finds similar relationships between
species richness and rainfall in South America.
The participants then broke down into three task groups for most of the
Global Dataset Preparation -- Integration of the various datasets,
to prepare a single, global dataset in "flat" file form for analysis. This
required standardization of species nomenclature, addition of much new
information from literature sources, and the addition of new variables
(particularly for diet and vegetation classes).
African Data Integration & Analysis -- The available morphometric
data are most extensive and complete for African species, which also represent
a wide range of habitat types. This group wanted to experiment with the
effectiveness of various morphometric measures in representing climate
or vegetation, using some for the first time, and compare the results with
those using categorical approaches.
Visualization of Locality & Climate Data -- Simply to have
baseline information about the nature of our climatic and vegetational
coverage, this group wanted to map and produce summary statistics on these
attributes of the localities in the global dataset.
Global Dataset -- The compilation and integration of the
global dataset turned out to be the most time-consuming task, as the participants
had brought a very large number of data. Using Badgley's global dataset
as the basis, we ended up with a dataset with 141 variables: Locality geographic,
vegetational and climatic variables (Badgley, the working group, and Madden
[not yet fully integrated]); Species diet and behavioral variables (Andrews,
Badgley, Janis); Postcranial morphometric variables (Hixson); "Mesowear"
dental variables (Fortelius & N. Solunias); Craniodental morphometric
variables (Reed, Janis).
There are a total of 265 localities worldwide. Madden's
data including 80 localities and 556 species are not yet fully integrated
into the global dataset. Not counting Madden's data, there are 185 localities
and 2016 species, with 8954 distinct species occurrences. This means that
the flat-file representation of these data is an Excel spreadsheet with
1.2 million cells! This spreadsheet had repeatedly to be transformed from
one file format to another, modified, and proofread. The dataset still
needs additional entries, particularly an expansion of the morphometric
data, with, ideally, additional information for small mammals.
Figure 1. The 256 localities, by vegetation type (variable VEGNO).
African data integration -- This group's work was also
integrated into the global dataset. In preliminary multivariate analyses
it was discovered that continuous locomotor and dental characters separate
species extremely well according to their habitat preferences -- and represent
open-closed and wet-dry dichotomies on separate axes. This suggests that
such variables will have considerable predictive power in the global dataset
once the data become available for a wider range of species.
Locality data visualization -- Since this activity was
dependent upon the global dataset, time constraints were severe. With Frank
Davis' help, the group was able to plot our localities on a global map
of vegetation zones -- a good check on both our coverage and vegetation
assignments. Figure 2 shows mean annual rainfall plotted against mean annual
temperature for dataset localities identified to vegetation type. It can
be seen that for the majority of our vegetation types there is a relatively
restricted range of climatic variables within which they occur. For grasslands
and deserts, there is an upper limit of moisture (green and red lines),
and bushlands and woodlands overlap considerably (polygons). It is evident
that our major vegetation categories are in fact representing climatic
signals as well.
Figure 2. Climatic distribution of locality types.
Multivariate Analysis of the Global Dataset -- Example: Forests vs non-forests
worldwide, predicted from faunal structure
As an example of the results anticipated from the global
dataset once it is complete and in a form that allows more effective data
manipulation, I present a logistic regression* analysis, predicting vegetation
type from six faunal variables (such as percent hypsodont species). These
six variables jointly properly classify 73% of the localities overall (Table
1). We expect ultimately to do better than this, but the details here are
instructive. In 90% or more of the cases we can distinguish forests from
all other vegetation types, and temperate forests are never confused with
tropical forests. The next step, then, is to discover -- among the large
number of variables yetr to be evaluated -- ways to distinguish among grassland,
bushland and woodlant vegetation types. This result is consistent with
what previous results from Africa suggested, but this is the first time
it has been statistically shown in a global context. Also, note
that this is consistent with an interpretation of Figure 2, where bushlands
and woodlands overlap considerably on major climatic variables and clearly
will be the most difficult to tell apart.
Table 1. Logistic regression analysis. 97 Localities (119
Desert, 5 Grassland, 12 Bushland, 17 Woodland, 14 Temperate Forest, 30
||Predicted Temperate Forest
||Predicted Tropical Forest
*Logistic regression is a relatively new technique that
has not been used in this field before. It is similar to canonical discriminant
analysis, in that its dependent variable is a categorical one with two
or more classes (in this case, vegetation types). As a likelihood-based
technique, though, it is much more flexible in the kinds of independent
variables it allows, and its assumptions about the data are much less restrictive
It is anticipated that we will produce a large joint paper, thoroughly
exploring and documenting the potential for reconstructing past climates
and habitats using fossil mammal faunas. This may require a monographic
treatment, or may be published as a series of papers in a special issue
of a journal. Plans were also made for production of a "News & Views"
type of piece for Nature or Science.
The group needs to continue to fill in the holes in the dataset, and to
refine the ecomorphic and vegetation categories. In addition, the group
recognized the difficulty of working with the dataset in flat-file format.
Instead, we propose to store the data in a proper relational database,
from which it will much easier to generate subset data files for analysis.
An outline of the schema for such a database is depicted in Figure 3.