Count data in biology—Data transformation or model reformation?

Abstract

Statistical analyses are an integral component of scientific research, and for decades, biologists have applied transformations to data to meet the normal error assumptions for F and t tests. Over the years, there has been a movement from data transformation toward model reformation—the use of non‐normal error structures within the framework of the generalized linear model (GLM). The principal advantage of model reformation is that parameters are estimated on the original, rather than the transformed scale. However, data transformation has been shown to give better control over type I error, for simulated data with known error structures. We conducted a literature review of statistical textbooks directed toward biologists and of journal articles published in the primary literature to determine temporal trends in both the text recommendations and the practice in the refereed literature over the past 35 years. In this review, a trend of increasing use of reformation in the primary literature was evident, moving from no use of reformation before 1996 to >50% of the articles reviewed applying GLM after 2006. However, no such trend was observed in the recommendations in statistical textbooks. We then undertook 12 analyses based on published datasets in which we compared the type I error estimates, residual plot diagnostics, and coefficients yielded by analyses using square root transformations, log transformations, and the GLM. All analyses yielded acceptable residual versus fit plots and had similar p‐values within each analysis, but as expected, the coefficient estimates differed substantially. Furthermore, no consensus could be found in the literature regarding a procedure to back‐transform the coefficient estimates obtained from linear models performed on transformed datasets. This lack of consistency among coefficient estimates constitutes a major argument for model reformation over data transformation in biology.

Exploring the Patterns of Job Satisfaction for Individuals Aged 50 and over from Three Historical Regions of Romania. An Inductive Approach with Respect to Triangulation, Cross-Validation and Support for Replication of Results

“…We did not perform yet the skewness tests of variables in the case of these regional datasets. Still, asymmetry was not a matter of concern for the corresponding models, since both the binary logistic regression [109] and the Poisson one with robust standard errors as a generalized linear model form of regression analysis [110] are proven to be robust when using non-normal data.…”

A generalized Bayesian nonlinear mixed‐effects regression model for zero‐inflated longitudinal count data in tuberculosis trials

“…These models can be extended to relatively complex situations, including regression models, mixed-effects models, and models that accommodate overdispersion, underdispersion, and excess zeros. [1][2][3] However, in particular when very large counts are to be modeled, such as counts of viral or bacterial load in medical applications, [4][5][6][7] or large counts in biology and ecology, 8,9 researchers often log-transform the data. After doing so, the data might be expected to be normally distributed, at least approximately, and standard statistical procedures, such as normal linear models, including mixed-effects linear models, are applied.…”

Eucalyptus cloeziana seed count data: a comparative analysis of statistical models

“…ANOVA requires that the following assumptions be met: homogeneity of variances, normally distributed residuals and independent distributions. However, the data produced by experiments involving count data in the biological sciences often do not meet such assumptions (St-Pierre;Shikon;Schneider, 2018;Kosma et al, 2019), as seen in studies of seeds of genetically impoverished species, which present a high variability rate, such as forest seeds Araújo, 2018). Sileshi (2012) showed that in 429 studies with seed viability data published from 2002 to 2012, 70% of researchers used ANOVA to analyse the data, and among those who did, only 20% performed tests to verify the homogeneity of variances and normality of residuals.…”

“…However, a particular transformation is not always available to satisfy all the assumptions of the analysis. In addition, data transformation creates difficulty in interpreting results due to a change in scale (St-Pierre;Shikon;Schneider, 2018). Thus, the reformulation of the statistical model at the expense of data transformation is an advantageous solution.…”

“…Although the theory of generalized linear models was first presented in 1970, by 2012, only 13% of germination studies used these models. In contrast, the use was approximately 52% for non-normal count data in other areas of biological science (Sileshi, 2012;St-Pierre;Shikon;Schneider, 2018).…”