Low-dose aspirin has been shown to be effective in preventing atherothrombotic vascular events (nonfatal myocardial infarction, nonfatal stroke, or vascular death) in a meta-analysis including 9 primary prevention trials (50,868 subjects were treated with aspirin and 49,170 received placebo or control). Because meta-analyses aim to identify the effect of an intervention as early as possible, they are usually updated when new trials are published, so that repeated analyses are made on accumulating data. These multiple analyses induce repeated significance testing (usually with p value criterion, α = 5% for a 2-sided test), which leads to an increased risk of random error and subsequent misleading evidence in results. To distinguish real effects from random errors, it is necessary that the required number of participants in a meta-analysis (i.e., information size) should be at least as large as an adequately powered single trial. Trial sequential analysis (TSA) is a new statistical technique that aims to improve the interpretation of meta-analyses, reducing the risk of random errors due to repetitive testing by adopting trial sequential monitoring boundaries to evaluate the accumulating data obtained from updates in meta-analyses.
When applied even to apparently conclusive conventional meta-analyses, TSA may reveal that a relevant number of them (≤49%) do not reach the required information size and do not cross the monitoring boundaries, yielding inconclusive results because of random error and repetitive testing.
We selected the apparently conclusive meta-analysis previously mentioned to evaluate whether its conclusive results remained conclusive when accounting for risk of random error. Our analysis considered the end point of occurrence of serious vascular events (myocardial infarction, stroke, or vascular death). Our main assumptions included 2-sided testing, risk of type 1 error = 5%, and power = 80%. For stroke, the intervention effect was set at an event frequency of 4.2% for the control group (equal to the event frequency of the control group of the meta-analysis) and a relative risk reduction of 10%. The main result of TSA was expressed presenting the graph of cumulative z curve, in which the boundaries for concluding superiority or inferiority or futility were calculated according to the O’Brien-Fleming alpha spending function.
Our analysis was undertaken using a specific statistical software (User Manual for TSA, Copenhagen Trial Unit 2011, Copenhagen, Denmark, software downloadable at http://www.ctu.dk/tsa ).
Figure 1 shows our results. Applying TSA to the placebo-controlled randomized trials, the final part of the z curve was far beyond both the boundary of superiority and the line of “optimal” information size (92,379 patients).
