Count data in biology-Data transformation or model reformation?

Area of Science:

• Ecology
• Evolutionary Biology
• Genetics
• Biostatistics

Background:

• Traditional statistical methods in biology often require data transformation to meet assumptions of F and t tests.
• Generalized linear models (GLM) offer an alternative approach, allowing for non-normal error structures and parameter estimation on the original data scale.
• While data transformation can control Type I error, model reformation provides more consistent coefficient estimates.

Purpose of the Study:

• To investigate temporal trends in statistical analysis methods (data transformation vs. model reformation) in biological research over 35 years.
• To compare the practical outcomes of data transformation and GLM approaches using published datasets.

Main Methods:

• Literature review of statistical recommendations in biology textbooks and primary research articles from the past 35 years.
• Comparative analysis of 12 published datasets using square root transformation, log transformation, and GLM.
• Evaluation of Type I error rates, residual plot diagnostics, and coefficient estimates from each analytical approach.

Main Results:

• A significant increase in the use of GLM (model reformation) in primary biological literature since 1996, contrasting with static textbook recommendations.
• All tested methods produced acceptable diagnostic plots and similar p-values, but coefficient estimates varied substantially between transformation and reformation.
• Lack of a standardized method for back-transforming coefficients from linear models applied to transformed data was identified.

Conclusions:

• Model reformation using GLM is increasingly adopted in biological research for its ability to provide consistent parameter estimates on the original scale.
• The inconsistency in back-transformed coefficients from data transformation methods presents a significant drawback compared to model reformation.
• Biologists should consider adopting GLM for statistical analyses to improve the reliability and interpretability of their findings.