THE CAMBRIDGE STATISTICS DISCUSSION GROUP
Tuesday 3rd February 2004 7:15 for 7:45
Department of Applied Mathematics and Theoretical Physics,
Centre for Mathematical Sciences,
Wilberforce Road,
Cambridge
Dealing with Missing Outcome Data
Angela Wood,
M.R.C. Biostatistics Unit
Abstract:
Randomised controlled trials almost always have some individuals with
missing outcomes. Inadequate handling of these missing data in the
analysis can cause substantial bias in the effect estimates. In the last
few years the statistical literature has been loaded with suggestions to
explore the nature of missing data and methods to minimise potential bias,
however, results from a recent survey of the medical literature show these
are rarely put into practice. The results from the survey will be
presented, focusing on statistical procedures for handling missing data
and sensitivity analyses.
We will then review various methods to handle missing outcomes in single
and repeated measures data, covering a wide spectrum of missing data
assumptions. Using data from a randomized controlled trial to compare two
interventions for increasing physical activity, we compare the following
methods: complete-case analysis; ad-hoc imputation techniques such as last
observation carried forward and worst-case; model-based imputations;
random effects longitudinal models; and recently proposed joint models for
longitudinal data and non-ignorable dropout. Results from ad hoc
imputation methods vary widely. Standard multiple imputation and repeated
measures modelling agree closely. Modifying the modelling method to allow
for informative dropout did not affect our conclusions, but imputing using
a common model in both arms gave more conservative results. We conclude
results from ad hoc imputation methods should be interpreted with caution.
Standard multiple imputation methods and repeated measures modelling
methods are equivalent. However, the imputation and modelling families
suggest different ways of relaxing the missing-at-random assumption, and
the choice between them depends on contextual knowledge.
Speaker:
Angela has been working as a research statistician for 2 years, funded by the
Medical Research Council, based at the Biostatistics Unit in Cambridge.
Prior to this she completed her doctorate in the Mathematics and Statistics
Department at Lancaster University. Current research interests include
methods for dealing with missing data and joint modelling of longitudinal and
time-to-event data.
Directions: The main entrance is reached from Clarkson Road by going along the
footpath to the right of the Newton Institute, and turning left through the
gatehouse towards the main building (Pavilion A), which has a glass front and
a curved grassed roof. The main entrance is in the middle of the glass front.
Free Parking is available after 5pm on Clarkson and Wilberforce Roads and by
entering the site off Wilberforce Road. Admittance may be difficult after 7:45.
Next Meetings:
3rd March - Haidee Philpott & John Law (NIAB) on 'A case study in Winter
Wheat Fungicide Trials in the UK 1991-2001'.
5th April - Anthony Edwards (Gonville and Caius).
13th May - Melanie Cooray (Statwood Partnership).
Supper: Some members eat regularly in the University Centre before
each talk meeting in the downstairs bar at 6pm. Feel free to join them.
Subscriptions: of 4 pounds per member are now due for 2003-2004 session.
Secretary: Peter Watson, MRC Cognition and Brain Sciences Unit,
15 Chaucer Road, Cambridge CB2 2EF;telephone 01223 355294 Extension 801;
E-mail [log in to unmask]
|