Impact of measurement error on estimating risk factors for diseases: An example from parasitic infections in the Philippines

Introduction: Failure to account for measurement error can produce significantly biased estimates of associations between risk factors and diseases. As proof of principal, we show the impact of adjustments for measurement error on odds ratio (OR) estimates for the association between hookworm and Schistosoma japonicum infections.

Methods: Longitudinal data from 50 villages in Samar province, Philippines in 2004-2005 with 1-3 stool samples collected at baseline and follow-up were used to detect infections with hookworm and S. japonicum. 2276 participants with ≥1 stool sample and who were treated during mass treatment activities were included. Four Bayesian logit models were used to investigate the impact of measurement error on the association between these two infections. Informative priors for Sensitivity and Specificity of the diagnostic test were used. All results were adjusted for age, sex, and village's irrigation status.

Results: Absent correction for measurement error, the OR between hookworm and S. japonicum infections was 1.34 (95% Bayesian confidence intervals (BCI): 0.98, 1.84). The ORs were 1.54 (95% BCI: 1.01, 2.35) and 1.60 (95% BCI: 1.02, 2.53) when measurement error was corrected for hookworm infection alone and S. japonicum infection alone, respectively, and the OR was 2.18 (95% BCI: 1.19, 4.14) when both hookworm and S. japonicum infection status were corrected for measurement error.

Discussion: In this case, the estimated strength of association between hookworm and S. japonicum nearly doubled when measurement error for both infections was taken into account. The impact of measurement error in epidemiological studies should be routinely estimated and corrected.