Statistics for Wildlifers: How much and what kind?
What should a wildlifer know about statistics?
There are some basic things that come from statistical training. We illustrate some of these, not to suggest that they form an exhaustive list, but simply to exemplify key points. Again, these are based on our own experiences and represent common sources of confusion among biologists. Some technical facts that should be known are:
- What a confidence interval is, and is not
- The difference between an estimator and an estimate
- What a least-squares estimator is. What a maximum-likelihood estimator is. And that sometimes they are the same, but not always
- The meaning of bias, precision, and accuracy
- That some biased estimators actually may be better than unbiased estimators.
- That most data are not distributed normally
- That, nonetheless, most means are distributed nearly normally, even with modest sample sizes
- That parametric tests do not always require that data be normal
- In most situations, estimation is more useful and appropriate than hypothesis testing
- Determining what are valid sample units is sometimes challenging. What units are really independent? Distinguishing true replication from pseudoreplication (Hurlbert 1984)
- Sophisticated methods are not always better than simpler ones. Being complicated may confuse more than clarify
- A random sample may not be representative of the population from which the sample was drawn
- Results from unfamiliar statistical packages or statistical procedures should be viewed with a certain amount of skepticism. This is true for familiar packages, actually, but is especially important for unfamiliar ones. When using new software or methods, it is often worthwhile to make up some data with known properties, or find trusted data with known properties, and use them to test the new tools.
- Data dredging can be dangerous (Anderson et al. 2001). Avoid "beating the data till they confess." Think of the questions you want to ask before looking at the data. If you find some new and unexpected patterns in the data, that is great, but use that occasion to develop a question to ask of a fresh set of data rather than testing to see if the pattern is "significant" with the data already in hand.
- Focus on analyzing the problem, not the data.
- There likely is no single right analysis, or a single right model to use (Burnham and Anderson 1998). Different analyses, or different models, may appropriately be used in any situation. If different analyses give substantially the same results, one has greater confidence in those results. This is true especially if the analyses are based on different sets of assumptions.
Previous Section -- The U-shaped function
Return to Contents
Next Section -- A Final Thought