Home > Research > Publications & Outputs > Influence of Missing Observations on Estimated ...
View graph of relations

Influence of Missing Observations on Estimated Rates of Decline in Health Status

Research output: Contribution to conference - Without ISBN/ISSN Posterpeer-review

Published
Close
Publication date2003
Number of pages0
<mark>Original language</mark>English
EventAmerican Thoracic Society - Seattle, United States
Duration: 16/05/200321/05/2003

Conference

ConferenceAmerican Thoracic Society
CountryUnited States
CitySeattle
Period16/05/0321/05/03

Abstract

Hierarchical linear model (HLM) estimates of rates of change are used increasingly in COPD. We tested the effect of differential dropout between treatment arms on HLM and simple linear regression (REG) estimates of decline in health status measured using SGRQ data obtained from 149 patients 6-monthly for 3 years (7 observations/patient). 50% of patients were randomly allocated as 'dropouts', from whose data the last 1 or last 4 observations were removed to model different times in the 'study'. All data were retained in the remaining patients 'completers'. There were no differences in HLM or REG decline rates from complete data sets of the two groups (P>0.05). Mean decline rates are tabulated by group and number of missing observations (high score=faster decline).
[Table]
In 'dropouts' decline rates increased as missing data increased (i.e. period in "study" decreased), but the distribution of HLM decline rates (the SD) decreased while SD of REG estimates increased. In 'completers', the REG decline rates were stable, but the HLM decline rates increased as missing data increased in the 'dropouts'. These two methods of estimating decline rates respond differently to differential study dropout with important implications for longitudinal studies in COPD.
This abstract is funded by: GlaxoSmithKline

Sunday, May 18, 2003 8:15 AM