STATISTICAL CONSIDERATIONS IN THE DESIGN AND ANALYSIS OF PAIRED-CHOICE ASSAYS

In paired-choice assays, two treatments are presented simultaneously to each subject. Data from such arrays should not be considered to be independent, and correct statistical-analysis must account for the correlation. A statistical test that often is appropriate for these assays is the paired-sample t-test. I present curves showing the extent to which statistical power of this test is affected by sample size, effect size (i.e., magnitude of treatment differences), and correlation. For a given effect size and replication, positive correlation between paired observations substantially improves power of the test, whereas negative correlation reduces power. I conducted a literature sun ev of paired-choice assays to determine whether there are patterns in effect sizes and correlation that might assist in designing studies or in predicting minimum sample sizes necessary to achieve reasonable statistical power; experiments were categorized according to whether they were feeding or oviposition assays. The review indicated that correlation was highly variable and ranged between strongly negative and strongly positive values. Oviposition assays showed larger positive correlations than did feeding assays, resulting in larger effect sizes (adjusted for correlation); however, feeding assays tended to use larger sample sizes than oviposition assays, hence estimated statistical power was similar between the two types of assays. Oviposition assays often used multiple insects per arena, apparently sacrificing replication, whereas feeding assays tended to use a single insect per arena. Approximately 45% of experiments failed to detect significant treatment effects. The majority of nonsignificant assays had too few replications to detect even a large effect size with a reasonable statistical power. Literature examples are presented to show that assay methodology (specifically number of insects per arena, distance between paired choices, and assay duration) can affect correlation, effect size, and statistical power. Finally, scatter plots of Che data, although rarely presented, are shown to provide insight into methodological, statistical, and biological aspects of paired-choice assays.