Department of Biostatistics, Harvard T.H. Chan School of Public Health, Boston, Massachusetts, USA.
These authors contributed equally as co-first authors.
Ear Hear. 2024;45(5):1165-1172. doi: 10.1097/AUD.0000000000001503. Epub 2024 Mar 28.
Hypothesis tests for hearing threshold data may be challenging due to the special structure of the response variable, which consists of the measurements from the participant's two ears at multiple frequencies. The commonly-used methods may have inflated type I error rates for the global test that examines whether exposure-hearing threshold associations exist in at least one of the frequencies. We propose using both-ear methods, including all frequencies in the same model for hypothesis testing.
We compared the both-ear method to commonly used single-ear methods, such as the worse-ear, better-ear, left/right-ear, average-ear methods, and both-ear methods that evaluate individual audiometric frequencies in separate models, through both theoretical consideration and a simulation study. Differences between the methods were illustrated using hypothesis tests for the associations between the Dietary Approaches to Stop Hypertension adherence score and 3-year change in hearing thresholds among participants in the Conservation of Hearing Study.
We found that (1) in the absence of ear-level confounders, the better-ear, worse-ear and left/right-ear methods have less power for frequency-specific tests and for the global test; (2) in the presence of ear-level confounders, the better-ear and worse-ear methods are invalid, and the left/right-ear and average-ear methods have less power, with the power loss in the left/right-ear much greater than the average-ear method, for frequency-specific tests and for the global test; and (3) the both-ear method with separate analyses for individual frequencies is invalid for the global test.
For hypothesis testing to evaluate whether there are significant associations between an exposure of interest and audiometric hearing threshold measurements, the both-ear method that includes all frequencies in the same model is the recommended analytic approach.
由于反应变量的特殊结构，即由参与者在多个频率下双耳的测量值组成，因此听力阈值数据的假设检验可能具有挑战性。常用的方法可能会使检验暴露与听力阈值之间关联是否至少存在于一个频率中的全局检验的Ⅰ型错误率膨胀。我们建议使用双耳方法，即将所有频率纳入同一模型进行假设检验。
我们通过理论考虑和模拟研究，将双耳方法与常用的单耳方法（如最差耳、最佳耳、左/右耳、平均耳方法）以及分别在单独模型中评估个体听力频率的双耳方法进行了比较。使用高血压饮食法（DASH）依从评分与听力研究中参与者的 3 年听力阈值变化之间的关联的假设检验，说明了这些方法之间的差异。
我们发现，（1）在不存在耳间混杂因素的情况下，最佳耳、最差耳和左/右耳方法对于特定频率的检验和全局检验的效力较低；（2）在存在耳间混杂因素的情况下，最佳耳和最差耳方法无效，左/右耳和平均耳方法的效力较低，且左/右耳方法的效力损失大于平均耳方法，对于特定频率的检验和全局检验；（3）对于个体频率的单独分析的双耳方法对于全局检验无效。
对于假设检验，以评估感兴趣的暴露与听力阈值测量值之间是否存在显著关联，建议采用包括所有频率的同一模型的双耳方法。
