Analytical Strategies for the Marble Burying Test: Avoiding Impossible Predictions and Invalid P-values

Background: The marble burying test is used to measure repetitive and anxiety-related behaviour in rodents. The number of marbles that animals bury are count data (non-negative integers), which are bounded below by zero and above by the number of marbles present. Count data are often analysed using normal linear models, which include the t-test and analysis of variance (ANOVA) as special cases. Linear models assume that the data are unbounded and that the variance is constant across groups. These requirements are rarely met with count data, leading to 95% confidence intervals that include impossible values (less than zero or greater than the number of marbles present), misleading p-values, and impossible predictions. Transforming the data or using nonparametric methods are common alternatives but transformations do not perform well when many zero values are present and nonparametric methods have several drawbacks.

Findings: The problems with using normal linear models to analyse marble burying data are demonstrated and generalised linear models (GLMs) are introduced as more appropriate alternatives.

Conclusions: GLMs have been specifically developed to deal with count and other types of non-Gaussian data, are straightforward to use and interpret, and will lead to more sensible inferences.