Generalized additive models to analyze nonlinear trends in biomedical longitudinal data using R: Beyond repeated measures ANOVA and linear mixed models

Abstract

In biomedical research, the outcome of longitudinal studies has been traditionally analyzed using the repeated measures analysis of variance (rm‐ANOVA) or more recently, linear mixed models (LMEMs). Although LMEMs are less restrictive than rm‐ANOVA as they can work with unbalanced data and non‐constant correlation between observations, both methodologies assume a linear trend in the measured response. It is common in biomedical research that the true trend response is nonlinear and in these cases the linearity assumption of rm‐ANOVA and LMEMs can lead to biased estimates and unreliable inference. In contrast, GAMs relax the linearity assumption of rm‐ANOVA and LMEMs and allow the data to determine the fit of the model while also permitting incomplete observations and different correlation structures. Therefore, GAMs present an excellent choice to analyze longitudinal data with non‐linear trends in the context of biomedical research. This paper summarizes the limitations of rm‐ANOVA and LMEMs and uses simulated data to visually show how both methods produce biased estimates when used on data with non‐linear trends. We present the basic theory of GAMs and using reported trends of oxygen saturation in tumors, we simulate example longitudinal data (2 treatment groups, 10 subjects per group, 5 repeated measures for each group) to demonstrate their implementation in R. We also show that GAMs are able to produce estimates with non‐linear trends even when incomplete observations exist (with 40% of the simulated observations missing). To make this work reproducible, the code and data used in this paper are available at: https://github.com/aimundo/GAMs-biomedical-research.

Change in absolute neutrophil count after COVID-19 infection in patients using clozapine versus other antipsychotics

“…The restricted maximum likelihood (REML) scores for each model were compared using the chi-square test. Second, the difference between the two fitted smooth curves at each time point was computed and the corresponding empirical Bayesian confidence interval (EBCI) around this difference was calculated according to the method of Mundo et al (2022). The EBCI provides a range of values within which the true difference is believed to lie with a certain degree of confidence (95%).…”

Antibacterial plant combinations prevent postweaning diarrhea in organically raised piglets challenged with enterotoxigenic Escherichia coli F18

“…The models included treatment as parametric term, independent smooths over time for each treatment, interaction smooths over time for each pig, and random effects of pen and sow. Pairwise differences between treatment groups against the PC group were assessed via 95% simultaneous empirical Bayesian confidence intervals (sEBCI) as described by Mundo et al ( 36 ).…”

Influence of Biological Sex and Fitness on Core Temperature Change and Sweating in Children Exercising in Warm Conditions

“…A Bayesian hierarchical generalized additive model (HGAM) (33) was fitted with T gi as the response variable and time, sex (male/female), and VO 2peak (mL•kg −1 •min −1 ) as the predictor variables. The multilevel model analysis method was chosen as the T gi -time relationship was nonlinear and varied between participants (34,35). An HGAM implements penalized smoothing splines that allow the form of the relationship to be determined from the data (36).…”

“…An HGAM implements penalized smoothing splines that allow the form of the relationship to be determined from the data (36). Penalization of the splines reduces the effect of overfitting to the data (34). A thin-plate regression spline allowed for two continuous predictor variables (time and V̇O 2peak ) to be analyzed.…”