Background: An intention-to-treat analysis of all randomised patients, is the only unbiased way to quantify the results of randomised controlled trials (RCTs). However, after randomisation, trial investigators frequently exclude patients from subsequent analyses of trial for reasons which could be related to treatment. These exclusions may bias trial estimates of treatment effect in unpredictable ways and combining these estimates in a meta-analyes could aggregate this bias. Indeed, there is evidence to suggest that exclusion of patients contributes to differences between the results of individual patient data (IPD) and published data meta-analyses. In IPD meta-analyses, we seek information on all randomised patients, even those excluded from the investigator's own analysis. We have examined the impact of trial investigator patient exclusions on IPD meta-analysis results, while factors, such as the type of analysis and length of follow-up were held constant.

Methods: 7 IPD meta-analysis were used, and because some studies analysed more than one outcome, a total of 13 comparisons are available. For each, we carried out an intention-to-treat analysis of all randomised patients and an analysis based on just those patients included in the investigators' own analyses and compared the results.

Results: The meta-analyses vary by cancer site (ovary, lung, and soft tissue), number of trials (9-19), number of patients (1569-3146), outcomes studied (recurrence, survival) and estimated effect of treatment (HR=0.68-1.21). Patient exclusions varied considerably by meta-analysis (3-22%) and between trials. Also, in 2 meta-analyses, exclusions were markedly unbalanced across treatment arms; in 1 case the proportion of exclusions on treatment was double that on control. 12 of the 13 analyses of 'included' patients had HRs and associated p-values more favourable toward the experimental treatment than the analyses of all patients, irrespective of the direction of the effect. In 1 case the results changed from non-significance to borderline significance by the exclusion of patients. Furthermore, statistical heterogeneity was consistently increased when patients were excluded from the analyses and in 4 cases, suggested that of pooling trial results was inappropriate, whereas when all patients were included, this was not so. These completed and further ongoing comparisons will be presented.

Conclusions: Exclusion of patients from published analyses of trials is widespread even although it is well known that it can potentially introduce bias to the analysis and conclusions. In these results, which are based on a large number of cancer trials, exclusion of patients makes the treatment appear more beneficial than when all randomised patients are analysed. This is even although we have used IPD, thereby already correcting for other biases associated with extracting data from published reports (e.g. publication bias, reporting bias, follow up bias). Reinstating excluded patients in the analyses also reduced heterogeneity between trials making pooled analyses justified. Where substantial numbers of patients are excluded from published reports of trials, systematic reviews using data from these publications could be biased. Reviews should therefore contact authors to obtain information on excluded patients, so that they can carry out an appropriate intention to treat analysis of all randomised patients.