Burstyn Igor, Miller David
Department of Environmental and Occupational Health, Dornsife School of Public Health, Drexel University, Philadelphia, PA 19104, USA.
US Environmental Protection Agency (Retired), USA.
Glob Epidemiol. 2024 Jul 8;8:100154. doi: 10.1016/j.gloepi.2024.100154. eCollection 2024 Dec.
Bias away from the null in odds ratios (OR), aggravated by low power, is a well-known phenomenon in statistics (sparse data bias). Such bias increases in presence of selection of "significant" results on the basis of null hypothesis testing (effect size magnification, ESM).
We seek to illustrate these issues and adjust for suspected sparse data bias in the context of a reported more than doubling of the odds of new onset type 2 diabetes in presence of occupational trichlorfon insecticide exposure reported in the Agricultural Health Study.
We performed ESM analysis on the crude ORs extracted from the contingency table in the published report, which is done by simulating selected OR given a posited true OR. Next, we applied easily accessible methods that adjust for sparse data bias to the extracted contingency tables, including data augmentation, bootstrap, Firth's regression, and Bayesian methods with weakly informative priors.
During the ESM analysis, we observed that there was a reasonable chance that a "statistically significant" OR of around 2.5-2.6 would be observed for true OR of 1.2. Adjustment for sparse data bias revealed that Bayesian methods outperformed alternative approaches in terms of yielding more precise inference, while not making unjustified distributional assumptions about estimates of OR. The OR in the original paper of about 2.5-2.6 was reduced on average to OR of 1.9 to 2.2, with 95% (Bayesian) credible intervals that included the null.
It is reasonable to adjust ORs for sparse data bias when the reported association has societal importance, because policy must be informed by the least biased estimates of the effect. We think that such adjustment would lead to a more appropriate evaluation of the extent of evidence on the contribution of occupational exposure to trichlorfon pesticide to risk of new onset diabetes.
在统计学中，比值比（OR）远离无效值的偏差，在低效能的情况下会加剧，这是一种众所周知的现象（稀疏数据偏差）。在基于无效假设检验选择“显著”结果时（效应大小放大，ESM），这种偏差会增加。
在农业健康研究报告中提到职业性接触敌百虫杀虫剂会使新发2型糖尿病的几率增加一倍多的背景下，我们试图阐述这些问题并对疑似稀疏数据偏差进行调整。
我们对已发表报告中列联表提取的粗OR进行了ESM分析，即通过给定假定的真实OR来模拟选定的OR。接下来，我们将易于获取的针对稀疏数据偏差进行调整的方法应用于提取的列联表，包括数据增强、自助法、Firth回归以及使用弱信息先验的贝叶斯方法。
在ESM分析过程中，我们观察到对于真实OR为1.2的情况，有合理的可能性会观察到“统计学显著”的OR约为2.5 - 2.6。对稀疏数据偏差的调整表明，贝叶斯方法在产生更精确的推断方面优于其他方法，同时不会对OR估计做出不合理的分布假设。原论文中约为2.5 - 2.6的OR平均降至1.9至2.2，其95%（贝叶斯）可信区间包含无效值。
当报告的关联具有社会重要性时，对OR进行稀疏数据偏差调整是合理的，因为政策必须基于对效应的偏差最小的估计。我们认为这种调整将导致对职业接触敌百虫农药对新发糖尿病风险贡献的证据程度进行更恰当的评估。
