Bayesian correction for measurement error following group-based exposure assessment in a case-referent study
本研究将贝叶斯分析应用于职业噪声暴露与缺血性心脏病死亡的病例对照数据，对比了是否校正分组噪声暴露估计中测量误差的分析结果，发现测量误差未导致未校正结果出现偏倚。
<h3>Objectives</h3> We applied Bayesian analysis to case-referent data on occupational noise exposure and death from ischaemic heart disease (IHD) and contrast analyses with and without correction for measurement error in group-level noise exposure estimates. <h3>Methods</h3> A 1:1 matched case-referent study nested in an industrial cohort in England resulted in 117 matched sets; 7225 area noise measurements in dBA from 215 buildings were the basis of modeling building-specific average exposures during the decade of in service IHD death. An additive quasi-Berkson error model was assumed. Bayesian analysis was conducted under varying assumptions about magnitude of error (with SD of error (SDe) up to 10 dBA) and a prior strength of hypothesis (flat vs informative -- 98% range (1.00,1.02) -- prior on OR). All analyses ignored matching and were conducted without adjustment for confounders to estimate log(OR)/dBA in a logistic disease model. <h3>Results</h3> Analysis not corrected for measurement error with flat prior yielded OR 0.99 (95% CrI 0.96–1.02). With flat prior on OR with measurement error correction, OR had 95% CrI 0.97–1.02; with informative prior on the association, the corresponding OR is 1.01, 95% CrI 1.00–1.02, same as prior. The posterior distribution of SDe had median 1.6 (95% CrI 1.2–2.0) dBA. <h3>Conclusions</h3> Measurement error did not bias the uncorrected results. Conditional logistic regression with adjustment for confounders is congruent with Bayesian analysis (115 pairs, OR 0.98, 95% CI 0.94 to 1.02). Analysis provided insights into plausible magnitudes of measurement error in the study. Extension of this methodology to consider matching, confounders and retrospective nature of data is required.