Background/objectives: To compare linear regression coefficients adjusted for random errors with true coefficients.
Subjects/methods: Three hundred and two individuals from the city of Rio de Janeiro, Brazil answered 20 non-consecutive 24-hr. Means of 20 24-hr were used as an approximation of the usual dietary intakes. It was simulated outcomes with pre-defined linear regression coefficient (β=1.0, referred as 'true coefficient') for usual coffee and soft-drink intakes as explanatory variables controlled for sex and age. Regression calibration was applied in each 1000 random combinations of j days of intake (j=2, 4 and 6), and adjusted coefficients were compared with true one.
Results: Mean-adjusted coefficients were 1.06 to 1.03 (coffee) and 1.17 to 1.11 (soft drink). The association was not detected (95% CI included zero) in 33 to 23% (coffee) and 37 to 23% (soft drink) when using two and six collection days, respectively, compared with 20% when using observed usual intake. Frequency of consumption as covariate in the regression calibration model increased the precision of the adjusted coefficients.
Conclusions: Adjustment for random errors de-attenuates the association but its precision depends mainly on the number of collection days and sample size.