The idea of a specific course on “wildlife data analysis” implies that the ordinary methods of analysis taught in a basic statistics course and available in standard statistical packages (eg. SPSS) do not suffice for wildlife data.

What is special about the data we typically gather on wildlife which sets them apart from most scientific data?

1. The variable of interest (the response variable) is usually not a continuous variable, and sampling errors are not normally distributed. It is usually:

• a probability based on binary events (detected/not detected, survived/died, marked/unmarked, …) binomial distribution, often over-dispersed, or
• a positive integer based on counts (number of individual organisms, occupied sites, species present, …) (over-dispersed) Poisson distribution.

While it is true that with large numbers both the binomial and Poisson distributions approximate the normal distribution, small numbers typify wildlife datasets.

2. The observation process invariably affects the data gathered. In particular, detection is hardly ever perfect, so that some animals or species or occupied sites are not recorded. Assuming perfect detection leads to erroneous and often quite misleading inferences

So we need to jointly estimate probability of detection and the parameter(s) of interest, and studies must be designed to make this possible. Joint likelihoods need to be explicitly calculated and maximised, as simple ‘plug-in’ estimators are not available; fortunately, specialist software packages have been developed for most such situations.

3. Ecological systems are complex, with uncertainty about the mechanisms operating. Limiting inferences to “accept” or “reject” for individual effects ignores this complexity and uncertainty.

An approach which considers multiple models and the relative support for each provided by the data is preferable for wildlife science. Information Theory, based on Kullback-Leibler divergence, provides a framework for comparing models and is implemented in several specialist software packages (eg. PRESENCE, DISTANCE, MARK).

4. The “p-values” generated by the usual null hypothesis significance testing methods are of little use (or are actually misleading) for management decision-making. Their use reflects proper scientific caution, but this runs counter to the precautionary principle enjoined on environmental managers.

In contrast, Bayesian posterior probabilities are easily incorporated into formal decision-making processes. In the past, Bayesian methods could only be used for very simple problems, but with the advent of powerful personal computers and ingenious algorithms they are now routinely used in wildlife data analysis.