A novel weighting method to remove bias from within-subject exposure dependency in case-crossover studies

• Main Authors: Kubota, Kiyoshi, Kelly, Thu-Lan, Sato, Tsugumichi, Pratt, Nicole, Roughead, Elizabeth, Yamaguchi, Takuhiro
• Format: Article in Journal/Newspaper
• Language: unknown
• Published: figshare 2021
• Subjects: Medicine; Genetics; FOS Biological sciences; Biotechnology; 39999 Chemical Sciences not elsewhere classified; FOS Chemical sciences; Immunology; FOS Clinical medicine; 19999 Mathematical Sciences not elsewhere classified; FOS Mathematics; Greenland
• Online Access: https://dx.doi.org/10.6084/m9.figshare.c.5666215.v1; https://springernature.figshare.com/collections/A_novel_weighting_method_to_remove_bias_from_within-subject_exposure_dependency_in_case-crossover_studies/5666215/1

Abstract Background Case-crossover studies have been widely used in various fields including pharmacoepidemiology. Vines and Farrington indicated in 2001 that when within-subject exposure dependency exists, conditional logistic regression can be biased. However, this bias has not been well studied. Methods We have extended findings by Vines and Farrington to develop a weighting method for the case-crossover study which removes bias from within-subject exposure dependency. Our method calculates the exposure probability at the case period in the case-crossover study which is used to weight the likelihood formulae presented by Greenland in 1999. We simulated data for the population with a disease where most patients receive a cyclic treatment pattern with within-subject exposure dependency but no time trends while some patients stop and start treatment. Finally, the method was applied to real-world data from Japan to study the association between celecoxib and peripheral edema and to study the association between selective serotonin reuptake inhibitor (SSRI) and hip fracture in Australia. Results When the simulated rate ratio of the outcome was 4.0 in a case-crossover study with no time-varying confounder, the proposed weighting method and the Mantel-Haenszel odds ratio reproduced the true rate ratio. When a time-varying confounder existed, the Mantel-Haenszel method was biased but the weighting method was not. When more than one control period was used, standard conditional logistic regression was biased either with or without time-varying confounding and the bias increased (up to 8.7) when the study period was extended. In real-world analysis with a binary exposure variable in Japan and Australia, the point estimate of the odds ratio (around 2.5 for the association between celecoxib and peripheral edema and around 1.6 between SSRI and hip fracture) by our weighting method was equal to the Mantel-Haenszel odds ratio and stable compared with standard conditional logistic regression. Conclusion Case-crossover studies may be biased from within-subject exposure dependency, even without exposure time trends. This bias can be identified by comparing the odds ratio by the Mantel-Haenszel method and that by standard conditional logistic regression. We recommend using our proposed method which removes bias from within-subject exposure dependency and can account for time-varying confounders.