Repository hosted by TU Delft Library

Home · Contact · About · Disclaimer ·

Improved accuracy when screening for human growth disorders by likelihood ratios

Publication files not online:

Author: Buuren, S. van
Institution: TNO Kwaliteit van Leven
Source:Statistics in Medicine, 9, 27, 1527-1538
Identifier: 240765
doi: doi:10.1002/sim.3031
Keywords: Health · Classification · Height · Screening · Turner's syndrome · article · body height · calculation · child · controlled study · diagnostic accuracy · female · growth disorder · human · infant · intermethod comparison · major clinical study · normal distribution · prevalence · probability · roc curve · screening test · sensitivity and specificity · standard · statistical model · Turner syndrome · Biometry · Child · Child, Preschool · Female · Growth Disorders · Humans · Infant · Infant, Newborn · Likelihood Functions · Male · Mass Screening · Netherlands · Probability · Turner Syndrome · Healthy for Life · Healthy Living


The standard deviation score (SDS) is a powerful tool for screening for growth-related problems. However, referral rules of the type 'if SDS(Y)<d, then refer' (for some constant d) are not optimal for answering the question: 'Does this child with measurement Y belong to the reference or to the diseased population?'. If the growth standard for the diseased population is known, then the likelihood ratio (LR) and the log-likelihood ratio (LLR) can be calculated for individual measurements. Rules of the type 'if LLR(Y)<e, then refer' are uniformly the most powerful test for any constant e, implying that their receiver operating characteristic curves are above those for all other possible tests based on Y. As an empirical demonstration, both types of rules are applied to longitudinal growth data comparing a group with diagnosed Turner syndrome and a reference group from birth to 10 years of age. Conforming with theory, the LR rules were found to be superior to the SDS rules in terms of sensitivity and specificity. We conclude that the LR is the natural measure for two-group studies that can be easily calculated for individual measurements. The LR is firmly rooted within both statistical and decision theory and can be used to estimate the absolute probability of disease. Copyright © 2007 John Wiley & Sons, Ltd.