Skip to main content

Table 1 The proportion of 10,000 simulated adjusted analyses where a hypothesized null exposure-outcome association (RRED|W) is indicated, after adjustment for a confounder W that is measured with different degrees of measurement error (Simulated change-in-estimate cutoff for Type I error <0.05 = 0.06 %)

From: Identification of confounder in epidemiologic data contaminated by measurement error in covariates

Noise-to-signal ratioa

Proportion of results where RRED|W p >0.05 (%)

RR ED | W (adjusted for confounder W)

Average

95 % Confidence Interval

0.10

100

0.99

0.97–1.01

0.25

100

0.98

0.96–1.00

0.50

91

0.96

0.94–0.98

0.55b

78

0.95

0.94–0.97

0.60

56

0.95

0.93–0.97

1.0

0

0.91

0.90–0.92

10c

0

0.83

0.83–0.83

  1. aratio of variance of measurement error relative to variance of confounder
  2. bin practice, it is not possible to have such precise knowledge about the extent of measurement error, so any calculation of this sort is necessarily approximate and is meant as a guideline for selection of suitable method to measure a confounder, but we can say that error variance should be closer to 0.5 than to 0.6
  3. cempirically determined to correspond to error in confounder hypothesized to exist in the data (F obs ) and resulting in failure to control cofounding effect of RRED|Fobs = 0.83 for exposure to mercury