Background: Missing data are an unavoidable feature in longitudinal studies and inad- equate handling can result in bias. The South London Stroke Register (SLSR) follows up participants at three months and annually after stroke. The majority of data collected are categorical in nature. Typically a third of survivors miss each follow-up and the impact of, and `best' methods for, dealing with these missing data are not clear. The aim of the thesis is to compare and determine the most appropriate methods for handling non-continuous missing data in the SLSR. Methods: Exploratory analyses identi ed predictors of incomplete follow-up and in- formed a simulation study in which the biases associated with prevalence rates of four indicators of poor outcome were estimated and analysis methods compared across four scenarios. Models making di ering assumptions about the missing data assessed the im- pact of missing data on associations between baseline characteristics and outcomes. Results: Missing data were strongly associated with disability and activity level after stroke and likely missing not at random (MNAR). Estimates of prevalence of poor out- comes from available case analyses were relatively unbiased apart from when a strong MNAR assumption was made and outcomes were strongly associated with dropout, with prevalence underestimated by up to 7% points. Bias was reduced after using multiple imputation (MI) with maximum bias of 5% points. There was no evidence that missing data in uenced associations between baseline characteristics and outcome. Conclusions: Some subgroups of the SLSR are at greater risk of non-participation than others but the resulting bias is likely to be minimal. When summarising population out- comes using rates MI should be used in addition to available case analysis. Future work will seek to further quantify potential biases using routinely collected data from GPs to compare responders and non-responders.
Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:724276 |
Date | January 2016 |
Creators | Crichton, Siobhan Laura |
Contributors | Peacock, Janet Lesley ; Wolfe, Charles David Alexander ; Toschke, Andre Michael |
Publisher | King's College London (University of London) |
Source Sets | Ethos UK |
Detected Language | English |
Type | Electronic Thesis or Dissertation |
Source | https://kclpure.kcl.ac.uk/portal/en/theses/methods-for-handling-missing-data-in-a-population-based-cohort-study(78bf4d3c-43e5-4a2d-ab08-a6ce453f5cc1).html |
Page generated in 0.0014 seconds