This blog site has been archived

The Dangers of Selective Analysis: Has stroke treatment been misguided for a decade?


Alteplase is widely recommended for treatment of stroke occurring within 3-4.5 hours. Brian Alper discusses their recent publication that reanalysed the Third European Cooperative Acute Stroke Study (ECASS III) trial through adjustment for baseline imbalances.

Blog entry written on: Thrombolysis with alteplase 3–4.5 hours after acute ischaemic stroke: trial reanalysis adjusted for baseline imbalances (bmjebm-2020-111386)

Author: Brian S. Alper

In 2008, ECASS III was the first and only randomized trial to report significant clinical benefits for treatment with alteplase 3-4.5 hours after onset of acute ischemic stroke. Critical appraisal for point-of-care application, as done for DynaMed, found it insufficient to recommend such use due to risk of bias (including baseline imbalances favoring the alteplase group) and inconsistent results in other randomized trials.

In 2009, the American Heart Association/American Stroke Association (AHA/ASA) made strong recommendations to extend alteplase use to 4.5 hours after stroke onset and critical commentary did not persuade the AHA to reconsider their recommendations.

By 2014 multiple countries (including the UK and Australia) extended marketing authorization for alteplase to 4.5 hours and 12 of 13 clinical practice guidelines recommended alteplase at 3-4.5 hours after stroke onset. Systematic reviews and meta-analyses reporting comprehensive views of the evidence suggested alteplase at 3-4.5 hours was more beneficial than harmful. We considered these findings improbable, published our interpretations as “Thrombolysis in acute ischaemic stroke: time for a rethink?” and suggested the key to resolving the uncertainty was publishing the underlying data and reanalyzing the data transparently.

In 2015 we asked the ECASS III investigator if we could collaborate on a multivariable reanalysis (adjusting for baseline imbalances) or could access the original trial data for such a reanalysis. The ECASS III leadership stated their reported adjusted analysis was sufficient, but the data were available from the sponsor. The process took some years to obtain the data from the study sponsor, analyze it, share findings with the study sponsor to clarify understandings in the data, and ultimately report the reanalysis (bmjebm-2020-111386.R1 “Thrombolysis with Alteplase 3 to 4.5 Hours after Acute Ischemic Stroke: Trial Reanalysis Adjusted for Baseline Imbalances”).

Although the primary results followed an explicit clinical trial protocol and statistical analysis plan, the adjusted analysis reported with the primary results (and providing corroborating supportive evidence) did not. We were able to reproduce the authors’ reported adjusted analysis, but it took a combination of three distinct deviations from what would have been an optimal protocol if a protocol for adjusted analysis had existed. If each of these deviations were considered to have a 30% chance of occurring when a group approaches statistical analysis, then the combination of all three deviations would have a 2.7% chance of occurring. The selection of the approach to adjusted analysis does not appear likely to be due to chance. This “selective analysis” approach has had a substantial influence on supporting the reliability of the ECASS III trial findings since 2008.

In evidence-based medicine we are highly familiar with the concepts of publication bias (and thus have methods developed to report the likelihood of it) and selective outcomes reporting bias (and include it in some risk of bias assessments), but we rarely consider the “selective analysis reporting bias” that can occur in the selection of methods for statistical analysis. Many otherwise critical thinkers who evaluate multiple domains in assessing the validity of research findings may be prone to accept the statistical methods without critical appraisal.
If this example of selective analysis reporting bias leads to re-assessment of the underlying evidence base and in turn changes in recommendations for healthcare decision making, then it will have taken more than a decade to overcome the influence of this bias. If we have not been looking for this bias across common approaches to evidence-based medicine where else may it be influencing our thinking?

Author: Brian S. Alper

Brian Alper
Author Brian Alper

Short bio: Brian is the Founder of DynaMed and the Chief Medical Knowledge Officer of EBSCO Information Services.

COI: Dr. Alper has no financial interests in alteplase or any therapeutic or diagnostic intervention for stroke care. Dr. Alper is employed in an information services company providing clinical decision support, patient decision support, and evidence-based medicine products. Dr. Alper has previously published on subjects related to the ECASS III trial results.

Twitter: @BrianAlperMD


The views and opinions expressed on this site are solely those of the original authors. They do not necessarily represent the views of the BMJ and should not be used to replace medical advice. All information on this blog is for general information, is not peer-reviewed, requires checking with original sources and should not be used to make any decisions about healthcare. No responsibility for its accuracy and correctness is assumed by us, and we disclaim all liability and responsibility arising from any reliance placed on such commentary or content by any user or visitor to the Website, or by anyone who may be informed of any of its content. Any reliance you place on the material posted on this site is therefore strictly at your own risk.

(Visited 1,748 times, 1 visits today)