arino et al.' have performed a meta-analysis using data extracted M from published randomized trials comparing radiotherapy (RT) to RT plus chemotherapy (CT) in Stage IIIA and IIIB unresectable nonsmall cell lung cancer (NSCLC). They identified 14 trials published between 1980 and 1994. The odds ratio (OR) summarizing the results was computed for 1887 patients. The NSCLC Collaborative Group (NSCLCCG) has recently published a meta-analysis using updated data from individual patients on the same topic,' including 22 trials which accrued patients between 1965 and 1991. The analysis was performed on 3033 patients (2814 deaths). These two approaches differ not only in the trials and patients included and the method of analysis used, but also in the results that they produced.
The two studies had nine trials in common. Five trials included in the Marino et al. meta-analysis were not eligible for inclusion in the NSCLCCG meta-analysis. Two were considered confounded, one used different RT doses per arm (Wils et al., trial 2), and in the other, concomitant cisplatin was given in both arms (Wolf et al., trial 2). A further three trials gave cisplatin only during RT. The NSCLCCG meta-analysis excluded all trials in which CT was given only during RT, to avoid mixingup trials studying the radiosensitizing effect of CT with those studying the cytotoxic effect. The trials included in the NSCLCCG meta-analysis, but not in the one by Marino et al. included 9 published between 1985 and 1991 (5 full papers, 4 abstracts), 1 trial published before this date, and 3 unpublished trials. It is not clear why the 9 trials (1106 patients) published between 1985 and 1991 were not included in the Marino et al. study. In the NSCLCCG meta-analysis, the proportion of Stage 111 patients was 83%.
For the 9 trials common to both studies, the number of patients included in the analysis by Marino et al. was 1332, whereas 1439 patients were included in the NSCLCCG meta-analysis. In the latter, the analysis was performed on an intent-to-treat basis using data from all randomized patients, whereas Marino et al. were restricted to the material available in the publication. Owing to the limited amount of information in the publications, a point estimate of survival on all eligible trials was only possible at 2 years in the Marino et al. meta-analysis. The NSCLCCG meta-analysis used individual patients' survival data in a time-to-event (log rank) test stratified by trial. This approach provides summary statistics relating to the whole survival experience rather than analyses based at selected fixed points in time.
In the Marino et al. meta-analysis, the pooled OR for death at 2 years was 0.70 (95% confidence interval (CI): 0.5-0.9) for the 11 cisplatin-based trials and 0.82 (0.5-1.3) for the 4 noncisplatin trials. No conclusion could be drawn from the 3-and 5-year results. In the NSCLCCG meta-analysis,