Signal detection remains central to pharmacovigilance. The dominant technique for spontaneous reporting systems (SRS) is disproportionality analysis (DA), but SRS data are often sparse and contain many zeros. Excess zeros reflect both the rarity of adverse drug reactions (ADRs) and substantial under-reporting, which can violate standard count-model assumptions and distort inference.
This study evaluated whether explicitly modelling excess zeros with zero-inflated count models could improve fit and signal detection for SRS-derived counts.
The analysis used publicly available Medicines and Healthcare products Regulatory Agency (MHRA) interactive Drug Analysis Profiles (iDAPs) for selective serotonin reuptake inhibitors (SSRIs). From these iDAPs the authors derived adverse event counts aggregated at the system organ class (SOC) level. The analysis therefore focused on SOC-level event counts as the unit for model fitting and signal exploration.
Initial goodness-of-fit testing assessed whether SOC counts conformed to common count distributions. The SOC counts were variably consistent with either a Poisson or a negative binomial distribution across SOCs. Because neither standard Poisson nor standard negative binomial models address excess zeros explicitly, the authors considered zero-inflated extensions as plausible alternatives when excess zeros were present.
Two classes of zero-inflated models were evaluated: zero-inflated Poisson (ZIP) and zero-inflated negative binomial (ZINB). These models combine a point mass at zero (an explicit zero-generating process) with a conventional count distribution (Poisson or negative binomial) for the non-structural counts. The ZINB additionally accommodates overdispersion beyond that explained by zero inflation.
Model adequacy was assessed using two complementary approaches. Absolute fit was evaluated via Pearson-residual dispersion, which examines residual variation relative to model expectations. Relative fit across competing models used Akaike’s Information Criterion (AIC). These metrics together informed whether zero-inflated formulations offered improved representation of the observed SOC counts compared with standard Poisson or negative binomial models.
Zero-inflated models provided an improved fit for 10 out of 27 SOCs. The SOCs for which zero-inflated models performed better were:
For these SOCs, accounting explicitly for excess zeros reduced Pearson-residual dispersion and improved AIC relative to non-zero-inflated counterparts, indicating better absolute and relative fit in the contexts reported.
Where zero-inflated models identified departures from expected counts, the resulting signals were reported to be broadly concordant with findings from traditional disproportionality analysis. The authors specifically note concordance for metabolism-related events among the signals examined. The study therefore positions zero-inflated modelling as a complementary approach that can reproduce many DA signals while providing an explicit modelling framework for excess zeros.
The findings support the use of zero-inflated count models as a complementary signal detection method when working with iDAPs-derived count data, especially in settings where excess zeros are present and an internal, model-based validation is desirable. Zero-inflated models may be particularly useful for SOCs or event groupings prone to sparse reporting, allowing analysts to separate structural zeros from sampling zeros and to address overdispersion where necessary.
The authors do not claim replacement of DA but rather an additional methodological option that can yield concordant signals and improved model diagnostics for selected SOCs.
The preprint reports that data used for the analysis are publicly available from the MHRA iDAPs and in a University of Birmingham research data repository. The project was approved in line with University of Birmingham research ethics processes (approval record cited as ERN4560Jun2025). The authors declare one competing interest: Prof. Anthony Cox holds an honorary Pharmacovigilance Pharmacist position at the Yellow Card Centre West Midlands. The preprint was posted September 22, 2026, and supporting information and code are referenced in the source record.