Ordinal regression increases statistical power to predict epilepsy surgical outcomes

• Authors: Adam Dickey, Emory University; Robert Krafty, Emory University; Nigel Pedersen, Emory University
• Journal Title: EPILEPSIA OPEN
• Volume: Volume 7, Number 2
• Publisher: WILEY | 2022-02-23, Pages 344-349
• Type of Work: Article | Final Publisher PDF
• Publisher DOI: http://doi.org/10.1002/epi4.12585
• Subjects: Engineering, Biomedical; Biology, Biostatistics
• Keywords: Science & Technology; Life Sciences & Biomedicine; Clinical Neurology; Neurosciences; Neurosciences & Neurology; Engel outcome; epilepsy surgery; logistic regression; seizure freedom; statistical power; TEMPORAL-LOBE EPILEPSY; LASER-ABLATION

Abstract

Studies of epilepsy surgery outcomes are often small and thus underpowered to reach statistically valid conclusions. We hypothesized that ordinal logistic regression would have greater statistical power than binary logistic regression when analyzing epilepsy surgery outcomes. We reviewed 10 manuscripts included in a recent meta-analysis which found that mesial temporal sclerosis (MTS) predicted better surgical outcomes after a stereotactic laser amygdalohippocampectomy (SLAH). We extracted data from 239 patients from eight studies that reported four discrete Engel surgical outcomes after SLAH, stratified by the presence or absence of MTS. The rate of freedom from disabling seizures (Engel I) was 64.3% (110/171) for patients with MTS compared to 44.1% (30/68) without MTS. The statistical power to detect MTS as a predictor for better surgical outcome after a SLAH was 29% using ordinal regression, which was significantly more than the 13% power using binary logistic regression (paired t-test, P <.001). Only 120 patients are needed for this example to achieve 80% power to detect MTS as a predictor using ordinal regression, compared to 210 patients that are needed to achieve 80% power using binary logistic regression. Ordinal regression should be considered when analyzing ordinal outcomes (such as Engel surgical outcomes), especially for datasets with small sample sizes.