Although it is desirable to include trial results from intention to treat analyses, this is not always possible given the data provided in published reports. Reports commonly omit participants who do not comply, receive the wrong treatment, or who drop out of the study. All of these individuals can easily be included in intention to treat analyses if follow-up data are available, and it is most important that they are included if the reasons for exclusion relate to the treatment that they received (such as drop-outs due to side-effects and poor tolerability of treatment). Occasionally full details of the outcomes of those excluded during the trial may be mentioned in the text of the report, but in many situations assumptions must be made regarding their fate. By inventive use of sensitivity analysis (using worst case, best case and most likely case scenarios for every trial) it is possible to assess the influence of these excluded cases on the final results. The issue is more problematic for continuous outcomes, where there is a continuum of possible scenarios for every excluded participant.
Other problems can occur when trials have no events in one or both arms. In these situations inverse variance, Mantel-Haenszel and DerSimonian and Laird methods require the addition of a small quantity (usually 0.5) to the cell counts to avoid division by zero errors. (Many software implementations of these methods automatically add this correction to all cell counts regardless of whether it is strictly needed.) When both groups have event rates of zero (there being no events in either arm) odds ratios and relative risks are undefined, and such trials must be excluded from the analysis. The risk difference in such situations is zero, so the trials will still contribute to the analysis. However, both inverse variance and Mantel-Haenszel methods perform poorly when event rates are very low, underestimating both treatment effects and statistical significance.18 Peto's odds ratio method gives more accurate estimates of the treatment effects and their confidence intervals providing the sample sizes of the arms in the trials are not severely unbalanced.
Was this article helpful?