Woodhouse, Lisa Jane
(2022)
Statistical reanalysis of vascular event outcomes in primary and secondary vascular prevention trials.
PhD thesis, University of Nottingham.
Abstract
Randomised controlled trials in vascular prevention typically use binary events as the primary outcome. However, this practice is statistically inefficient and gives no indication of the severity of the event. Hypothetically, outcomes could be polychotomised, with ordering determined by severity.
The primary aim of this study was to identify the most efficient type of outcome and analysis technique using individual patient data from randomised controlled trials. Another aim was then to review the use of these approaches on other vascular event outcomes; done using data from the TARDIS trial, the first trial to use an ordinal event measure as the primary outcome. An additional consideration was that some trials use a composite outcome as the primary measure. Therefore, a third aim was to determine if severity data can be incorporated here as well and whether to analyse these novel outcomes using standard methods or global analysis techniques, which can analyse multiple outcome measures at once to determine an overall effect.
Chief investigators of vascular prevention trials were asked to share individual participant data. Binary and ordinal extensions of vascular outcomes were then developed and analysed using methods including binary logistic regression, bootstrapping, Chi-Square, Cochran-Armitage trend test, Cox proportional hazards, Mann-Whitney U, median test, multiple linear regression, ordinal logistic regression, t-test and the Win Ratio test. Regressions were performed adjusted and unadjusted. Global analysis techniques included the Wei-Lachin test (for ordinal/continuous outcomes), the Wald test (for binary outcomes) and the Hotelling’s T2 test (for continuous outcomes). Once the appropriate tests has been performed the resulting p-values were ordered within each trial set and given a rank, with the smallest p-value given the top rank. For the primary and global analyses the Freidman and Duncan’s multiple range tests were then performed to identify significant differences between statistical methods. In the case of the TARDIS trial analysis, the ranks of the p-values were compared using heat maps.
Data from 35 trials (257,749 patients) were shared, thirteen trials had multiple comparator arms, resulting in 59 datasets. The most commonly shared outcomes were stroke, myocardial infarction (MI) and bleeding. When compared against analysis methods performed on their binary counterparts (event/no-event), the analyses performed on the extended outcomes tended to be more efficient; e.g. Mann-Whitney U for stroke and MI outcomes (both p<0.0001) and multiple regression for bleeding event outcomes (P<0.0001).
As for the TARDIS analysis, ordinal versions of vascular event outcomes were created in 3096 participants for stroke, MI, major cardiac events, bleeding events, serious adverse events and venous thromboembolism (VTE), with 32 outcomes being created overall (29 in the subgroup population due to the absence of VTE events). Overall, tests run on the ordinal outcome appeared more efficient for 20/32 (63%) of the analyses in comparison with binary tests. 764 (24.7%) participants were recruited within 24 hours of a mild stroke/TIA; again, tests run on ordinal outcome were superior in 27/29 (93%) of the comparisons.
For the analysis including global analysis techniques, of the 35 trials shared, 24 (158,901 patients) had outcome data for both stroke and MI. From these 24 trials it was possible to develop 35 comparator datasets, as 5 of the trials had multiple intervention arms. With the introduction of further levels based on severity (that is, 4 or 5-level stroke with 3-level MI), the Wei-Lachin test had the smallest average rank and therefore appeared to be more efficient statistically than those that analysed the composite outcomes (p<0.0001).
This study found that methods that are used when analysing polychotomised outcomes appear to be more efficient than those used for binary outcomes. This was shown for both the primary analysis and the secondary TARDIS analysis. Based on this, there could be a potential role for using ordinal outcomes in vascular prevention trials. Furthermore, the global analysis suggested that, where there are multiple outcomes of interest, there could also be benefits to basing the primary measure on global techniques, as they appeared to be more efficient statistically.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
![[img]](/style/images/fileicons/application_pdf.png)