Evaluating statistical models for overdispersed multiomics data: a multiplex immunofluorescence case study

Multiomic data analysis poses statistical challenges. We evaluated statistical models for our multiplex immunofluorescence study of T cell subset densities in colorectal cancer. Using 1235 cases, we compared 7 models-ordinal logistic regression, Poisson, quasi-Poisson, quadratic negative binomial (NB), linear NB, zero-inflated NB, and hurdle NB models-assessing associations with a strong (microsatellite instability, MSI) and a weak (calcium intake) exposure. Simulation studies assessed type I error and power. Effect estimates were generally consistent for the strong exposure (MSI) but varied for the weaker exposure (calcium). Simulations revealed inflated false-positive rates for the Poisson and NB-based models, including quadratic NB, zero-inflated, and hurdle, but not for ordinal logistic regression or the linear NB model. The quasi-Poisson model showed modest inflation of low P-values, but the overall P-value distribution remained approximately uniform under the null. Ordinal logistic, linear NB, and quasi-Poisson models achieved the best or near-best power across a range of zero proportions, dispersion levels, and distributions. The ordinal logistic, linear NB, and quasi-Poisson models are useful and robust options for epidemiologic analyses of overdispersed, right-skewed multiomic data with a nontrivial proportion of zero counts.