Home Page >
Research areas >
Methodology >
Applied stat methodology >
Avoiding bias in the analysis of longitudinal data
Avoiding bias in the analysis of longitudinal
data
Data from repeated measurements or events
are ubiquitous in trials and observational studies.
In many studies, data are collected on
biomarkers, such as CD4 count in HIV infection, CA125 in ovarian
cancer and PSA in prostate cancer. Biomarkers are often used to
define different types of outcome measure, for example, the marker
value or its change from baseline at fixed times after
randomisation.
A single model for all the measurements
gives a more complete picture of the effect of treatment over
follow up, may reduce bias associated with missing data and may
improve efficiency. However, we need to adapt or develop
methodologies in this area as current methods are susceptible to
bias. This is because they make strong and frequently unrealistic
assumptions, and the bias may be severe in some circumstances.
Another methodological gap is the analysis
of repeated events, such as adverse clinical events, in trials and
observational studies. We typically wish to examine predictors of
more serious events. The number of adverse events experienced by
the patient varies and is frequently related to their severity
(i.e. informative cluster size).
Standard methods such as random effects
models or generalised estimating equations (GEE) can be appreciably
biased and the methodology developed by others for informative
cluster size data has not been explored thoroughly. The CTU
collaborates with the MRC Biostatistics Unit in this area.
Key
projects
- Better analysis of repeated
measurements data: better understanding of disease
evolution and treatment comparisons in unit studies, investigation
of bias in standard methods, development and promotion of methods
that avoid bias and answer scientific questions of interest
- Methods to analyse repeated events
data where the number of events experienced by patients
varies: better understanding of adverse events in studies,
development and promotion of unbiased methods for analysis of these
key outcome measures
Selected
publications
- Copas AJ, Seaman SR Bias from the use of generalised estimating
equations to analyse incomplete longitudinal binary data J App
Stats 2010 37: 911–922
- Seaman S, Copas A. Doubly robust generalized estimating
equations for longitudinal data. Statistics in Medicine 2009;
28:937-955
- Harrison L, Dunn D, Green H, Copas AJ The Analysis of Change
Over Time Accounting for Baseline Differences Stats in Med 2009 28:
3260-3275
