A comparison of checked versus unchecked individual patient data (IPD)

Article type
Burdett S, Stewart L
Introduction/Objectives: IPD meta-analyses (MAs) of randomised controlled trials (RCTs) usually involve detailed data checking. This can be done at different levels, from simply finding and querying missing or inconsistent data, to detailed investigation of the integrity of the randomisation and follow-up procedures. Detailed data-checking is resource-intensive and time consuming. Our aim was to ascertain whether the lengthy checking procedures performed on incoming IPD, before inclusion in our meta-analyses, are worthwhile.

Methods: Using data from our meta-analysis of post-operative radiotherapy in non-small cell lung cancer (PORT), checked and unchecked data were compared. Unchecked, non-updated IPD-A (10 RCTs, 2302 patients) were compared with checked, non-updated IPD-B (9 RCTs, 2128 patients). IPD-A was as received, except for minor recoding, required to make the individual trial datasets comparable. One trial was excluded from the PORT meta-analysis as irregularities were found when detailed checks were carried out on the data. As the trial was excluded only at this point, it was included in IPD-A. These results were compared with updated, thoroughly checked IPD-C which were used in the final analysis of PORT.

Results: IPD-C showed a significant detriment of PORT (HR = 1.21, p = 0.001) whereas IPD-A showed less evidence of a detriment (HR = 1.09, p = 0.097). IPD-B, despite not having been updated, gave the same result as IPD-C

Discussion: The difference in results between the checked and unchecked IPD is largely due to the exclusion of one RCT, the results of which are out of line with the other RCTs. In a meta-analysis with a less robust result, using only unchecked data such as this, could potentially give misleading results. In this case, checking data made little difference to the results of the meta-analysis of the 9 included RCTs. However, it provided clear reasons for excluding the 10th RCT, prior to analysis, thus improving the integrity of the database. Although time consuming (increasingly less so as automated procedures are used), checking IPD remains an important part of IPD MAs.