Stick with randomization
Some have suggested that loosening the standards for evidence generation in medicine to include observational data, big data, artificial intelligence, and alternative trial strategies, such as Mendelian randomization and causal inference of nonrandomized data, might help drive new treatments to the clinic faster. To this, Dr. Nissen and Dr. Cannon offer an emphatic no.
“The idea that you can use big data or any kind of nonrandomized data to replace randomized control trials is a bad idea, and the reason is that nonrandomized data is often bad data,” Dr. Nissen said in an interview.
“I can’t count how many bad studies we’ve seen that were enormous in size, and where they tried to control the variables to balance it out, and they still get the wrong answer,” he added. “The bottom line is that observational data has failed us over and over again.”
Not to say that observational studies have no value, it’s just not for determining which treatments are most efficacious or safe, said Dr. Cannon. “If you want to identify markers of disease or risk factors, you can use observational data like data collected from wearables and screen for patients who, say, might be at high risk of dying of COVID-19. Or even more directly, you can use a heart rate and temperature monitor to identify people who are about to test positive for COVID-19.
“But the findings of observational analyses, no matter how much you try to control for confounding, are only ever going to be hypothesis generating. They can’t be used to say this biomarker causes death from COVID or this blood thinner is better than that blood thinner.”
Concurring with this, the ESC, AHA, ACC, and WHF statement authors acknowledged the value of nonrandomized evidence in today’s big data, electronic world, but advocated for the “appropriate use of routine EHRs (i.e. ‘real-world’ data) within randomized trials, recognizing the huge potential of centrally or regionally held electronic health data for trial recruitment and follow-up, as well as to highlight the severe limitations of using observational analyses when the purpose is to draw causal inference about the risks and benefits of an intervention.”