FREE UK WORKSHOP Missing data in health research 5th December 2001 organised by the Department of Psychiatry, University of Cambridge with support from the Stanley Foundation Target audience includes generalists in health research, not just statisticians. Open to all. Free of charge to those employed by Universities or NHS. Pre-registration by email is required (to [log in to unmask]) since the venue restricts attendance to 100 persons. Time and Place: 10.15am-4.15pm on Weds 5th December 2001 Alice Fisher Lecture Theatre, Addenbrooke's Hospital site, Cambridge http://www.addenbrookes.org.uk/directcontact/maps/locationmap1.html Detailed joining instructions will be sent in response to email registration. Any further enquiries please contact: Tim Croudace 01223 336599 [log in to unmask]) Please forward this message to interested parties thank you Tim Croudace ************************************** Programme 10.15 Coffee on arrival 10.30 Welcome Prof Peter Jones, Head of Dept of Psychiatry University of Cambridge 10.35 Introduction Simon Thompson Director MRC Biostatistics Unit, Cambridge 10.45 Health research warning: ignoring missing data can seriously bias your estimates Margaret Ely, Anglia Polytechnic University, Cambridge 11.45 Missing information in practice: a model based approach applied to asthma trials James Carpenter, Medical Statistics Unit London School of Hygiene and Tropical Medicine ********************* 12.45-2.00 Buffet LUNCH ********************** Chair: Prof Glenys Parry, University of Sheffield 2.00 The modelling approach to missing data: a re-analysis of the UK700 trial Ian White MRC Biostatistics Unit, Cambridge _____________________________________________________________ 3.00 KEYNOTE SPEAKER: Mike Kenward, SB Professor of Biostatistics, London School of Hygiene and Tropical Medicine ***** Perspectives on missing data: an overview *** ____________________________________________________________ 4.00 Summing up / Close Tim Croudace, Dept of Psychiatry, Univ of Cambridge 4.15 TEA and departure ______________________________________________________________ Abstracts of talks *Missing Values: an Introduction and Overview * MIKE KENWARD SB Professor of Biostatistics, London School of Hygiene and Tropical Medicine The problem of missing values will be considered from a statistical perspective. Some standard definitions and concepts in the topic will be introduced and related to practical problems. The relationships between aims and methods of analysis will be stressed, and contrasts drawn between simplicity of analysis and simplicity of assumptions. The ideas behind some some common methods such as simple imputation and multiple imputation and model based analyses will be will be sketched. ______________________________________________________________ *Health Research Warning: ignoring missing data can seriously bias your estimates* MARGARET ELY Centre for Research in Health and Social Care, Anglia Polytechnic University, Cambridge Objectives -To demonstrate the serious bias resulting from ignoring missing data values or using naïve imputation methods, and the advantages of using multiple imputation (MI) in an epidemiological study of alcohol consumption. Design - birth cohort study Subjects - Men and women in the MRC National Survey of Health and Development, a national cohort study of 5362 births in 1946, of whom 3262 were interviewed in 1989 at the age of 43. Outcome measures - Alcohol consumption is derived from a seven day diet diary. Measures of excessive consumption are drinking in excess of 3U(f) or 4U(m), and of double this recommended limit, 6U(f) or 8U(m), on any day of the week. Methods - Only 2002 (61%) of the 3262 study members interviewed in 1989completed the diet diary. Using this complete data MCAR, MAR and MNAR mechanisms of missingness were simulated and listwise deletion (LD), group mean imputation (GM) and MI were applied. Multiple imputed data sets were generated using SOLAS software, based on regression models using covariates gender, reported weekly consumption, smoking, CAGE score, systolic blood pressure, day of the week and consumption on recorded days. The model was applied to the 3262 study members. Results - LD and GM produced progressively greater biased estimates with departure from MCAR, whereas those using MI were unbiased even when the missingness of the data was related to the amount people drank (MNAR). Average estimates of the proportion of men drinking >8U, known to be 38.0%, was 27.7% for LD, 38.8% for MI. Further, MI is more efficient than LD (RMSE 1.2, 10.4). Applying MI to the full data gives 42.5% (95% CI 39.4 to 45.6). Conclusion - Missing data poses a problem for epidemiological studies in which the reason for missing data is not known and in which it is unlikely to be missing at random. Multiple Imputation provides insurance against bias in even when the data is not missing at random. ______________________________________________________________ *Missing information in practice: a model based approach applied to asthma trials* JAMES CARPENTER Medical Statistics Unit, London School of Hygiene and Tropical Medicine In most clinical trials, some patients do not complete their intended follow-up according to protocol, for a variety of reasons, and are often described as having 'dropped out' before the conclusion of the trial. Their subsequent measurements are missing, and this makes the analysis of the trial's repeated measures data more difficult. We briefly review the reasons for patient dropout, and their implications for some commonly used methods of analysis. We then propose a selection models for modelling both the response to treatment and the dropout process. Such models are readily fitted in a Bayesian framework using non-informative priors with the software BUGS. The results from such models are then compared with the results of standard methods for dealing with missing data in clinical trials, such as last observation carried forward. We further propose the use of a time transformation to linearise an asymptotic pattern of repeated measures over time and therefore simplify the modelling. All these ideas are illustrated using data from a 5-arm asthma clinical trial. ______________________________________________________________ *The modelling approach to missing data: a re-analysis of the UK700 trial* IAN WHITE MRC Biostatistics Unit, Cambridge The UK700 trial was a large randomised controlled trial in community psychiatry.The proportion of patients who were interviewed at follow-up differed between the two arms, and there was therefore concern that any analysis which excluded those not followed up would be biased. I will describe a sequence of analyses which included successively more information in the model: first baseline values of the outcome variables, then values of the outcome variables measured mid-way through follow-up, and finally other outcome variables which had less missing data. This sequence of analyses aims (1) to reduce bias by making the underlying assumptions about the missing data more plausible, and (2) to increase precision by including more information. The simpler models were fitted in standard software but the more complex models were fitted in MLWin software.In the UK700 trial the estimated treatment effect differed little between analyses and we conclude that bias due to missing data is probably small. ______________________________________________________________ Sponsor: This workshop is sponsored by the Stanley Foundation. http://www.stanleyresearch.org/about/ The Stanley Foundation is a nonprofit organization that supports research on the causes and treatment of schizophrenia and bipolar disorder (manic-depressive illness). The Dept of Psychiatry, University of Cambridge is a Stanley Foundation European Research Centre.This funding supports population-based epidemiological studies of schizophrenia and bipolar disorder, including MRI and neuropsychological components, and collaborative work on treatment trials.