Sampling and Experimental Design in Community Ecology


To consult the statistician after an experiment is finished is often merely to ask him to conduct a post mortem examination. He can perhaps say what the experiment died of.  –Ronald Fisher

Designing an experiment properly will not only help you in analyzing data – it may determine whether you can analyze data at all! 

The topic of experimental design is difficult to address – since community ecologists study such a wide range of systems and with unique qualities.  Thus, sampling is largely idiosyncratic with respect to the community and organisms being studied.  In addition, the optimal study design will largely depend on:

·         The question(s) being addressed.

·         Whether it is desirable to compare results to that of previous research

·         Availability of suitable habitat or study sites

·         Availability of suitable resources (time, money, equipment, helpers).

The first point above is the most important component of experimental design.  However, it is also one of the most difficult ones for which to provide general advice – since novel questions will typically require novel sampling designs.

Nevertheless, there are some general principles I will quickly describe below.

Sampling in the field

While much of ecology has moved towards manipulative experiments, community ecology still depends upon observational experiments (note that some scientists incorrectly use the word ‘experiment’ to refer only to manipulative experiments.  In reality, an experiment is a series of observations intentionally planned to address a scientific question).

One common criticism of observational experiments is that ‘correlation does not imply causation’.  I disagree with this.  Correlation implies causation.  The problem is that we cannot infer the direction of causation without some scientific sleuthing.  A correlation between soil organic matter and tree cover could be due to trees preferring organic soils, or leaf litter contributing to organic matter in the soil, or another variable simultaneously affecting tree distributions and organic matter. 

It should be noted that causality is improperly inferred from many manipulative studies.  A classic example (I hope it is fictional) is the scientist who trained frogs to jump on voice command, and incorrectly concluded that frogs are unable to hear when you remove their legs.

A key advantage of observational experiments is that they address how nature actually is, not what factors can be important. 

In community ecology, the discipline of gradient analysis specifically addresses how species respond to spatial variation in the natural environment.  Because of this, we need to pay special attention to how to place the samples in a spatial context.

With some exceptions below I will call an individual sample a ‘plot’, but note that many other types of samples are possible.  I will also assume we are sampling on a 2-dimensional landscape, but note some exceptions, each with their own set of sampling issues:

·         For some aquatic or soil organisms, we might more likely be sampling in 3 dimensions.

·         For organisms associated with streams, coasts, or other more or less linear features, we are sampling along 1 dimension.

·         Organisms dependent on other organisms (e.g. mosses on trees, phyllosphere bacteria) will require some form of point sampling

Despite all of these unique considerations, most of the principles below are still relevant.

Location of Plots


Subjective locations

If we are interested in observing species response to gradients, we might find it useful to insert our own natural history knowledge or intuition into plot locations.  For example, if we wish to understand how species respond to gradients, it would be valuable to have the gradient extremes (e.g. very wet and very dry, or very steep and very level, or very acidic to very basic) if we wish to fully describe the responses of individual species.  We may also want to have good representation of the intermediates.    Having good representation of the gradients is not guaranteed with objective sampling.  Subjective locations are also far easier to implement than objectively-located samples.

Typically plots are located to ensure within-plot homogeneity.  This presumes the investigator’s intuition appropriately assesses homogeneity.  There is a risk of ‘reification’ – that is, believing a community type is a real entity, sampling in such a way to maximize the distinctness and homogeneity of that community type, and then analyzing the data to highlight and demonstrate the existence of that community type.  This is one of the main drawbacks of phytosociology.

Nevertheless, some kinds of heterogeneity can make inferred gradients messy and effectively decrease measured beta diversity (see Palmer, M. W. and P. M. Dixon. 1990. Small scale environmental variability and the analysis of species distributions along gradients. Journal of Vegetation Science 1:57-65.)  An extreme case would be if a plot straddled the border between an oldfield and an old-growth forest, the ability to recover a successional pattern would be diminished.


Another drawback to subjective locations is that inferential statistics are invalid (though they tend nevertheless to be used in the literature).  Also, the infusion of the investigator’s intuition into locating the plots makes the research non-replicable.



Often (but by no means always) we are interested in species responses to an environmental gradient that varies directionally in space.  For example, we might be interested in bird species distributions as a function of elevation on a mountain, invertebrate species as a function of distance from a roadside, or mammal communities as a function of latitude.  In such studies, we subjectively choose the gradient we are interested in a priori.  It is often most productive then to locate our samples in a transect, systematically in a direction of maximum change in the gradient.  Such transects are known as gradsects.

Below, I illustrate two kinds of gradsects along an elevational gradient.  The blue squares represent contiguous quadrats in what we call a belt transect.  They would perhaps be suitable for plant communities.   The red dots represent point samples such as pitfall traps, and optimally would be placed at regular elevational intervals.