EXAMPLE 1

• Impact of random measurement error in an exposure variable: the assessment of bacterial vaginosis (BV) in a hypothetical case-control study of maternal bacterial vaginosis and low birthweight deliveries.

BV based on Amsel (clinical) criteria is measured imperfectly (with 70% sensitivity and 95% specificity) with respect to Nugent (microbiological) criteria. The result is an attenuation of the true association (odds ratio in table 1) towards a smaller association observed in the measured study data (table 2).

To complete the contingency table with measurement error (table 2), 30% (100%−70%) of the individuals with BV in the “true” table were reclassified as not having BV, while 5% (100%−95%) of the individuals without BV were reclassified as having BV. Cell values have been rounded as necessary. Note that this example presumes that low birthweight status is measured perfectly, and as a result, case-control classification does not change between the “true” and measured tables.

EXAMPLE 2

• Impact of random measurement error in a confounding variable: the assessment of multiple sexual partners as a confounder in a hypothetical cohort study of hormonal contraceptive use and risk of HIV infection.

High risk sexual behaviour is measured as a dichotomous variable with the measured values having 75% sensitivity and 80% specificity compared to true values. In the hypothetical “truth” (scenario 1), true high risk sexual behaviour confounds completely the HIV hormonal contraception association (as demonstrated by the comparison of the unadjusted and adjusted odds ratios resulting from table 3 versus tables 4 and 5). However, non-differential errors in the measurement of sexual behaviours (giving rise to the observed data in scenario 2) result in an inability to adjust completely for the confounding effect. The result is the appearance of an appreciable “adjusted” association in the observed data (tables 7 and 8) when none is truly present.

Scenario 1: True data

Mantel-Haenszel adjusted relative risk for tables 4 and 5 = 1.0 (95% CI: 0.71 to 1.40). The Mantel-Haenszel relative risk, based on a weighted average of the stratum specific relative risks, is used to present a single measure of association that is “adjusted” for the measured confounding variable. In scenario 1, this adjusted measure shows that the unadjusted measure is truly confounded by high risk sexual behaviours; in scenario 2, it shows that statistical adjustment does not remove the entire confounding effect when the confounding variable is measured with error.

Scenario 2: Observed data, with measurement error (sexual behaviour measure is 75% sensitive and 80% specific)

Mantel-Haenszel adjusted relative risk for tables 7 and 8 = 1.82 (95% CI: 1.34 to 2.48). To complete tables 7 and 8, 25% ( = 100%−75%) of individuals in each exposure/disease category in table 7 (high risk sexual behaviours) were reclassified as having low risk behaviours transferred to the corresponding exposure/disease category in table 8. At the same time, 20% ( = 100%−80%) of those in each exposure/disease category in table 8 (low risk sexual behaviours) were reclassified as having high risk behaviours, and transferred to the corresponding exposure/disease category in table 7.