ArticlesSubgroup analysis and other (mis)uses of baseline data in clinical trials
Introduction
For most randomised clinical trials, substantial baseline data are collected on each patient at randomisation. These data relate to demographics, medical history, current signs and symptoms, and quantitative disease measures (including some measured again later in the study as outcomes). Gathering of such baseline data seems to have four main aims. First, baseline data are used to characterise the patients included in the trial, and to show that the treatment groups are well balanced. Second, randomisation may include some means of balancing or stratification on a few key factors. Third, for analysis of outcome by treatment group, covariate adjustment may be used to take account of certain baseline factors. Fourth, subgroup analyses may be carried out to assess whether treatment differences in outcome (or lack thereof) depend on certain characteristics of patients.
Statistical reporting of clinical trials has improved, many journals have statistical refereeing, and clearer guidelines to authors1 may further improve reporting quality. However, insufficient attention is paid to the quality and extent of reporting on these uses of baseline data.
We aimed to describe and critically evaluate current practice on the use of baseline data in clinical-trial reports in major medical journals, and to make recommendations to enhance the quality of future reporting, especially on the dangers of overemphasising subgroup analyses.
Section snippets
Methods
We handsearched all reports of clinical trials with individual randomisation of patients during July to September, 1997, in BMJ, JAMA, The Lancet, and New England Journal of Medicine. Crossover trials and cluster-randomised trials were excluded, as were small trials with less than 50 patients per group. Any trials with non-random allocation would also have been excluded, but none were identified. We decided that a sample size of 50 trials was large enough to provide representative and reliable
Results
The 50 randomised trials surveyed had the following characteristics: 39 trials had two randomised treatments, five had three treatments, and six (two with a factorial design) had four treatments. The number of patients ranged from 100 to 8803 with a median 494, and 40 trials had multiple centres. Follow-up of patients ranged from 11 days to 15 years (median 1 year). 36 trials had a predefined primary outcome, and 34 trials claimed an overall treatment difference with p<0·05, 13 of which had
Baseline comparability
Only four trials lacked a table of baseline characteristics by treatment allocation, two of which were previously published. The number of baseline features varied widely with a median 14 features and a maximum 41 features (table 1). The largest table of baseline data occupied nearly a whole journal column.2
Half the trials assessed imbalances between treatment groups by significance tests. The investigators of 17 trials reported baseline imbalances. Such declarations of imbalance were based on
Randomisation methods
The statistical method for randomisation was not mentioned in 27 reports (table 2). Most of the rest used randomised permuted blocks within strata.3 Most multicentre trials did balance randomisation by centre, usually with separate randomised blocks for each centre. Other baseline features were balanced for in many trials, usually for just one or two factors by use of random permuted blocks within strata. The few trials balancing for more factors used minimisation or a similar dynamic method.4,
Covariate adjustment
Most trial reports (38 of 50) emphasised simple outcome comparisons between treatments, unadjusted for baseline covariates (table 3). 14 such trials gave only unadjusted results. The other 24 gave covariate-adjusted results as a back-up to the unadjusted analyses. The remaining 12 reports gave covariate-adjusted analyses primary (or equal) emphasis; six of these 12 gave no unadjusted results. The number of covariates adjusted for varied substantially, with a median three and a maximum 14
Subgroup analyses
Most trial reports did include subgroup analyses, that is treatment outcome comparisons for patients subdivided by baseline characteristics (table 4). Many trials confined subgroup attention to just one baseline factor, but five trials examined more than six factors. The number of outcomes subjected to subgroup analysis also varied substantially: many reports studied one outcome for subgroup differences, but six reports explored six or more outcomes.
The total of subgroup analyses is the product
Discussion
We have identified some key shortcomings and controversies in the uses of baseline data in clinical trials in current reporting practice.
The CONSORT statement,1 which is used by many journals for the reporting of controlled trials, does recommend documentation of randomisation methods, but such information is still lacking in many trial reports. Inadequate reporting of randomisation was identified a few years ago6, 7 and as yet there seems little improvement. The achievement of allocation
Randomisation methods
Randomisation procedures, both the practical means of treatment allocation and the statistical methods used, need clearer explanation. In particular, reports should state which baseline factors the randomisation balanced for, and by what method. In design, balancing should be confined to centre and factors known to be strong predictors of outcome.
Baseline comparisons
Although reports should show in appropriate detail the types of patient included, the baseline comparisons across treatments need not be so extensive.
References (30)
- et al.
Randomised study of risk of fetal loss related to early amniocentesis versus chorionic villus sampling
Lancet
(1997) - et al.
Randomisation and baseline comparisons in clinical trials
Lancet
(1990) - et al.
Beneficial effects of azathioprine and prediction of prognosis in primary biliary cirrhosis: final results of an international trial
Gastroenterology
(1985) - et al.
Randomised trial of home-based psychosocial nursing intervention for patients recovering from myocardial infarction
Lancet
(1997) - et al.
Improving the quality of reporting of randomised controlled trials: the CONSORT statement
JAMA
(1996) Adaptive allocation in randomised controlled trials
Control Clin Trials
(1985)- et al.
On the use of Pocock and Simon's method for balancing treatment numbers over prognostic factors in the controlled clinical trial
Biometrics
(1976) - et al.
Assessing the quality of randomisation from reports of controlled trials published in obstetrics and gynecology journals
JAMA
(1994) - et al.
Practical experience of randomisation in cancer trials: an international survey
Br J Cancer
(1982)