Eduardo Villamor1, Ronald J Bosch2. 1. a Department of Epidemiology , University of Michigan School of Public Health , Ann Arbor , MI , USA and. 2. b Center for Biostatistics in AIDS Research, Harvard School of Public Health , Boston , MA , USA.
Abstract
BACKGROUND: Anthropometric studies often include replicates of each measurement to decrease error. The optimal method to combine these measurements is uncertain. AIM: To identify the optimal method to combine replicate measures for analysis. METHODS: The authors carried out 10 000 Monte Carlo simulations to explore the effect of six approaches to combine replicate measurements in a hypothetical two-group intervention study (n = 100 per arm) in which the outcome, infant length at age 1 year, was measured two or three times. One group had a true value with a normal distribution N (mean = 76, SD = 2.4 cm). Statistical power was estimated to detect a 1 cm difference between the groups, based on a t-test. RESULTS: Under a realistic scenario with a measurement error distribution N (0, 0.8), highest power was reached by use of the mean and the median of pairwise averages. However, when a portion of the data (≥2%) were contaminated by greater error (e.g. due to data entry), the median of three measurements outperformed all other methods while the mean had the lowest performance. CONCLUSION: Obtaining three rather than two measures and using the median of the three replicates is a safe and robust approach to combine participants' raw data values for use in subsequent analyses.
BACKGROUND: Anthropometric studies often include replicates of each measurement to decrease error. The optimal method to combine these measurements is uncertain. AIM: To identify the optimal method to combine replicate measures for analysis. METHODS: The authors carried out 10 000 Monte Carlo simulations to explore the effect of six approaches to combine replicate measurements in a hypothetical two-group intervention study (n = 100 per arm) in which the outcome, infant length at age 1 year, was measured two or three times. One group had a true value with a normal distribution N (mean = 76, SD = 2.4 cm). Statistical power was estimated to detect a 1 cm difference between the groups, based on a t-test. RESULTS: Under a realistic scenario with a measurement error distribution N (0, 0.8), highest power was reached by use of the mean and the median of pairwise averages. However, when a portion of the data (≥2%) were contaminated by greater error (e.g. due to data entry), the median of three measurements outperformed all other methods while the mean had the lowest performance. CONCLUSION: Obtaining three rather than two measures and using the median of the three replicates is a safe and robust approach to combine participants' raw data values for use in subsequent analyses.
Authors: Izzuddin M Aris; Jonathan Y Bernard; Ling-Wei Chen; Mya Thway Tint; Wei Wei Pang; Wai Yee Lim; Shu E Soh; Seang-Mei Saw; Keith M Godfrey; Peter D Gluckman; Yap-Seng Chong; Fabian Yap; Michael S Kramer; Yung Seng Lee Journal: Int J Epidemiol Date: 2017-04-01 Impact factor: 7.196
Authors: Martina Persson; Sven Cnattingius; Eduardo Villamor; Jonas Söderling; Björn Pasternak; Olof Stephansson; Martin Neovius Journal: BMJ Date: 2017-06-14