A reanalysis of PLATO

PLATO

PLATO NEJM 2009 was a multinational RCT of antiplatelets, ticagrelor vs. clopidogrel published in 2009(Wallentin et al. 2009) and has had > 5,000 citations. The primary end point was a composite of death from vascular causes, myocardial infarction, or stroke — occurred in 9.8% of ticagrelor subjects vs. 11.7% of clopidogrel subjects (HR, 0.84; 95% CI 0.77 to 0.92; P<0.001). 

There have been several publications questioning the integrity of the data and in December the BMJ published an excellent summary of these issues here.  

I think there are some pertinent statistical issues which deserve further comment.  

The FDA initially refused approval despite PLATO’s HR, 0.84; 95% CI 0.77 to 0.92; P<0.001 with concerns were centered on data integrity and the discordance between the overall and US results, \(HR_{US}\) 1.27, 95%CI 0.92 - 1.75. One year later with no change in the dataset, the FDA approved the drug. Apparently corporate reassurance prevailed and instead of looking at subgroups, the FDA accepted the pooled results presented in the NEJM publication1. Pooling the data implies that all ancillary post MI care is identical whether the patient was randomized in Bulgaria, South Africa or any of the other 41 participating countries. This is a strong assumption. Ignoring clustering within specific geographic countries or regions will underestimation the sampling errors resulting in falsely precise confidence intervals. On the other hand, examining only the US subgroup is wasteful as it excludes about 90% of the randomized patient data.

Hierarchical modeling, a compromise between and the no and complete pooling, with their optimal statistical properties(Gelman 2006) offers a sortie from this quandary. Individual country / region effects are shrunken towards the global mean and their variance is improved by this borrowing of information from other regions. For example, this model gives a \(HR_{US}\) 1.05 95% CIs 0.79-1.35. As predicted, the regional US estimate is shrunk towards the global mean with improved precision due to this borrowing of information. However, the overall summary mean will have wider confidence intervals (\(HR_{mean}\) 0.89 95% CIs 0.72-1.10) as this now includes both within and between region sampling variations. The next, as yet unperformed, study(Riley, Higgins, and Deeks, n.d.) is predicted to fall in the 95% CI 0.46 – 1.37. (see figure)

In other words, even accepting the PLATO data as is, simply by choosing a more realistic statistical model with a better estimate of the underlying the uncertainty we see that conventional statistical significance for ticagrelor is not met. Ignoring this uncertainty leads to publications and guidelines proclaiming ticagrelor superiority, when a more reasonable interpretation is that while there appears to be a positive ticagrelor signal this needs confirmation in further studies, especially for US patients. Interestingly, this second US study was never done. Moreover, later randomized trials have not been kind to ticagrelor. An Asian population who for genetic reasons were believed to have inferior clopidogrel activation showed not the expected improved but worse outcomes with ticagrelor (HR, 1.47; 95% CI: 0.88-2.44)(Goto et al., n.d.). Another RCT of elderly Dutch patients also has showed worse outcome with ticagrelor compared clopidogrel (absolute risk difference -4%, 95% CI -10.0 to 1.4; p=0.03 for non-inferiority)(Gimbel et al., n.d.).

Thus, even ignoring the questionable data integrity, ticagrelor superiority in PLATO is seen to be attributed to the choice of statistical model, specifically one that ignores a large portion of the study variability. While one may choose to reject the proposed hierarchical model as a better alternative, there should be a consensus that when conclusions differ drastically according to the choice of statistical model, the data can hardly be considered robust.

How then did ticagrelor attain its overwhelming supremacy? As I was once offered $750 by the ticagrelor sponsor to attend a dinner extolling PLATO’s benefits by a leading writer of the Canadian guidelines (I declined), one must wonder about the role of marketing in overcoming any scientific shortcomings.

These observations have been published as a rapid response letter in the BMJ

References

Gelman, & Hill, A. 2006. Analytical Methods for Social Research: Data Analysis Using Regression and Multilevel/Hierarchical Models. Book. Cambridge University Press.

Gimbel, M., K. Qaderdan, L. Willemsen, R. Hermanides, T. Bergmeijer, E. de Vrey, T. Heestermans, et al. n.d. “Clopidogrel Versus Ticagrelor or Prasugrel in Patients Aged 70 Years or Older with Non-ST-Elevation Acute Coronary Syndrome (POPular AGE): The Randomised, Open-Label, Non-Inferiority Trial.” Journal Article. Lancet 395 (10233): 1374–81.

Goto, S., C. H. Huang, S. J. Park, H. Emanuelsson, and T. Kimura. n.d. “Ticagrelor Vs. Clopidogrel in Japanese, Korean and Taiwanese Patients with Acute Coronary Syndrome – Randomized, Double-Blind, Phase III PHILO Study.” Journal Article. Circ J 79 (11): 2452–60.

Riley, R. D., J. P. Higgins, and J. J. Deeks. n.d. “Interpretation of Random Effects Meta-Analyses.” Journal Article. BMJ 342: d549.

Wallentin, L., R. C. Becker, A. Budaj, C. P. Cannon, H. Emanuelsson, C. Held, J. Horrow, et al. 2009. “Ticagrelor Versus Clopidogrel in Patients with Acute Coronary Syndromes.” Journal Article. N Engl J Med 361 (11): 1045–57. https://doi.org/10.1056/NEJMoa0904327.