Precision of Usual Food Intake Estimates According to the Percentage of Individuals with a Second Dietary Measurement

BackgroundStatistical methods for estimating usual intake require at least two short-term dietary measurements in a subsample of the target population. However, the percentage of individuals with a second dietary measurement (replication rate) may influence the precision of estimates, such as percentiles and proportions of individuals below cut-offs of intake.ObjectiveTo investigate the precision of the usual food intake estimates using different replication rates and different sample sizes.Participants/settingAdolescents participating in the continuous National Health and Nutrition Examination Survey 2007-2008 (n=1,304) who completed two 24-hour recalls.Statistical analyses performedThe National Cancer Institute method was used to estimate the usual intake of dark green vegetables in the original sample comprising 1,304 adolescents with a replication rate of 100%. A bootstrap with 100 replications was performed to estimate CIs for percentiles and proportions of individuals below cut-offs of intake. Using the same bootstrap replications, four sets of data sets were sampled with different replication rates (80%, 60%, 40%, and 20%). For each data set created, the National Cancer Institute method was performed and percentiles, CI, and proportions of individuals below cut-offs were calculated. Precision estimates were checked by comparing each CI obtained from data sets with different replication rates with the CI obtained from original data set. Further, we sampled 1,000, 750, 500, and 250 individuals from the original data set, and performed the same analytical procedures.ResultsPercentiles of intake and percentage of individuals below the cut-off points were similar throughout the replication rates and sample sizes, but the CI increased as the replication rate decreased. Wider CIs were observed at 40% and 20% of replication rate.ConclusionsThe precision of the usual intake estimates decreased when low replication rates were used. However, even with different sample sizes, replication rates >40% may not lead to an important loss of precision.