Background/Question/Methods
Ecological research often begins with a pilot study. Pilot studies reduce design-based uncertainties before investments are made in full-scale research. If the pilot study's results indicate a need for major changes to experimental design, then pooling pilot and non-pilot data can be infeasible. But ignoring pilot study data after a more comprehensive study has been completed is suboptimal. Either statistical power is forgone by not including the pilot data, or an extra cost is incurred of sampling additional data equivalent to the pilot study's sample size. Bayesian data analysis can be used to incorporate information from multiple sources, such as pilot and non-pilot study data. A Bayesian model generates posterior estimates of parameters which are weighted averages of the information from the input dataset and the prior probabilities of the parameters. Prior probability distributions for model parameters can be posterior estimates of models built from a different and otherwise incompatible data type to the data for the model they inform. With a Bayesian approach, pilot study data can be used as an informative prior for a model built from the main study dataset. We demonstrate Bayesian methods' utility for recovering information from otherwise unusable pilot study data with a case study on eucalypt seedling mortality. A pilot study of eucalypt seedling mortality was carried in south-eastern Australia in 2005.
Results/Conclusions
The pilot study results indicated that the experimental design required significant changes. A larger scale study with a modified design was carried out the following year. The two datasets were of a significantly different form that they could no easily be incorporated into a single dataset. Using Bayesian data analysis we show how the posterior estimates of pilot dataset model parameters are used to inform a model for second larger dataset. Using numerical and graphical posterior predictive checks we show that a model with pilot study data included as an informative prior has greater predictive power and more precise parameter estimates.