A multivariate method for measurement error correction using pairs of concentration biomarkers

Measurement error is a pervasive problem in behavioral epidemiology, and available methods of correction all have generally untenable assumptions. We propose a multivariate method with more realistic assumptions.

The method uses two concentration biomarkers for each nutritional variable of interest and structural equation modeling. This produces corrected estimates of the effects on an outcome variable of changing the true exposure variables by one standard deviation, a standardized regression calibration. However, hypothesis testing in original units is preserved. The main assumptions are that certain error correlations between dietary estimates and biomarkers or between biomarkers be close to zero.

Two illustrative models used simulated data with the covariance structure of a real data set. The corrections produced often were very substantial. A sensitivity analysis allowed error correlations to depart from zero over a modest range. Root mean square biases show the advantage of the corrected approach. Relatively large calibration studies are needed for adequate precision.

As long as concentration biomarkers are selected carefully, error-corrected multivariate hypothesis testing and standardized effect estimation is possible. With the deviations from assumptions that were tested, the corrected method usually produces much less biased results than an uncorrected analysis.