A Major Win for Open Science
In five short chapters, the Study of Week shows you a shining example of how open science could improve knowledge. It's a beautiful story of humility and generosity.
In August, Sensible Medicine described a unique study from the group of Brian Nosek at the University of Virginia. They showed that choices made in how data was analyzed could greatly affect the results.
It was a provocative study because Nosek’s team had brought together 29 teams of professional statisticians to analyze one data set to answer one question. The teams chose 29 different ways to analyze the data. This led to two-thirds of the groups finding a statistically significant result and one-third of the group finding no significant difference.
What struck me as a reader of medical studies is that no study uses 29 different analytic methods. They use one.
So, now, when reading a study with a result that is pretty close, I think: what if they had analyzed the data in different ways? Would the result hold up?
Nosek’s study never really got much traction. Perhaps because it was answering a social science question using soccer statistics not a hard science or medical question.
This week, I will show you a shining example of how open science and different analytic methods clarified a potentially major finding in cardiac surgery.
Here is the back story. There is a debate in cardiac surgery as to what are the best conduits to use for the bypass. All surgeons agree that the left internal mammary artery (now called left internal thoracic artery or LITA) is the absolute best conduit.
But nearly all bypass surgeries require more than one conduit. Surgeons can use a radial artery (from the arm), saphenous vein (from the leg) or the right internal thoracic artery (RITA) for the additional conduits.
It turns out that there have been many studies comparing these conduits. This summer, a group led by surgeon Mario Gaudino, from Weill Cornell Medicine in New York, did what’s called a meta-analysis, wherein they combined these studies hoping to learn which conduit was best.
The Gaudino study wasn’t a typical meta-analysis. That’s because the included trials had many different combinations. So, what they did was to take the 10k plus patients in the trials and use a statistical method called propensity matching to make triplet groups.
They found similar patients in each of three groups—RA, SVG and RITA. Their method for finding these matches—with propensity matching—was complicated. I don’t pretend to understand it. But it turns out that it may have been quite important.
Their major findings were shown in this figure. The group that used the radial artery had a nearly 40% lower death rate than the other groups.
Wow. That was a major finding.
Before this study, surgeons had a huge debate about which conduit was best. Previous studies really did not show a clear winner.
Now they had this amazingly strong finding, published in a big journal, from a respected group of authors.
Gaudino and his co-authors wrote candidly about the limitations of their study, mainly, that it was not a randomized comparison, but still, a 40% lower death rate could not be ignored.
This part of the story begins with appropriate questions regarding the massive effect size. It just didn’t seem plausible.
Here comes the beautiful part.
The editors asked Prof Gaudino if he would share the dataset and code.
He said yes.
That generosity allowed another statistics group, led by Prof Nick Freemantle, from the University College London, to re-analyze the same dataset.
The first thing Freemantle’s team did was to analyze the data in the same way that Gaudino’s group did. Here they found that the results were identical. This exercise showed there were no errors in the analysis.
But now, as it was in the original Nosek study of 29 different teams, Freemantle’s team made different analytic choices from Gaudino’s team. They felt that there were more traditional ways to do propensity matching. Again, I can’t say which is better.
Freemantle’s analytic method yielded no significant differences in outcomes with the different conduits. No differences. None.
Again, it’s the same dataset. Just different choices in analyzing it.
Now you may be wondering…I don’t know a damn thing about statistics or bypass surgery, so who do we believe?
I will get to that in the final chapter.
The lack of difference in outcomes better fits the prevailing thinking amongst surgeons. In fact, another academic heart surgeon, David Taggert, along with colleagues, published an editorial on the original Gaudino paper and argued that a 40% reduction from a choice of conduit was implausibly large.
Taggert did not mention Freemantle’s group re-analysis. Perhaps he wrote it before Freemantle’s team had finished.
Taggert’s team had a super-important way to explain the implausibly large mortality reduction. It’s something all readers should plug into their memory banks.
He noticed that the reduction in major cardiac adverse outcomes (we abbreviate MACE) was less than the overall mortality reduction. Think about that. If a cardiac surgical technique lowers death rate, it must do so by reducing cardiac outcomes.
In any study, when the reduction in overall mortality is greater than the reduction in the specific outcomes of the treatment, this suggests bias—specifically, healthier patients got one type of intervention.
This is one of the year’s most important stories.
Had Prof Gaudino not shared his data set, and had it not been re-analyzed with different choices, surgeons may have moved toward more radial artery usage.
Based on the re-analysis, and the Taggert-led editorial, there looks to be great uncertainty regarding the choice of conduit.
So the answer to this important question should be obvious:
You do a randomized controlled trial wherein patients who require bypass are randomized to three groups—one with each conduit.
Randomization, not surgeons, choose the conduit. This balances the known (and unknown) patient characteristics, and in some years, there will be answer.
Imagine a scientific world where authors of major medical studies were as generous as Professor Gaudino and his team.
Many, if not most, of the practice-changing studies in medicine are never re-analyzed. Or, if they are, it’s very late, after therapeutic fashions have been established.
I hope I live long enough to see a new approach to judging medical evidence. One in which any important findings are verified independently and transparently.