by Carl V Phillips
A few days ago, I thought I would write about one particular bit of ANTZ junk epidemiology as an example of how easy it is for dishonest authors of epidemiology (some of my fellow epidemiologists resent them being called epidemiologists — a fair point) to cheat to get what results they want. But the more I thought about it, and the more news that came in this week about an error in an influential economics paper, the more I realized I needed to go broader and start earlier than that. So…
Once upon a time, there was a child taking a science class in school. He thought he was learning science, but really he was forever stunting his intellectual growth and ability to really understand science. Fortunately, he was an industrious and hard-working kid, so he still made it through medical school, but he still never learned how to understand how science works, nor even realize that he did not understand. The end. (Or maybe he could not handle the math in the mandatory pre-med science classes, which are generally no better than high school classes, and so went into health journalism instead, or maybe he always wanted to be a sociologist — doesn’t really matter.)
I am not talking about some backwater school teaching the Jewish creation myth as if it were science. That meme is comparatively easy to rid oneself of. I am talking about typical classes that supposedly teach someone science, but really teach only the results of scientific research (e.g., the causes of weather patterns; the laws of planetary motion; the way a seahorse reproduces) and analysis and do random bits of unstructured exploring of the physical world (e.g., cutting up dead critters). When they do an “experiment” it is usually just a demonstration (e.g., making something that goes boom or changes colors in a beaker). When they really do an experiment, it is so set-piece or targeted that it is easy to know exactly what is right (e.g., identifying a reagent using a set list of tests; measuring a pendulum’s motion to show it follows the Newtonian laws that it is supposed to follow).
The result of this mal-education is that most people do not realize how tricky, uncertain, and ugly actual scientific discovery and analysis is. Indeed, they are taught exactly the opposite. They do not realize that science is, well, as much art as science. Instead, the teaching focuses on theories and trivia that are established beyond doubt, and so the dangerous subliminal message is that when you read a scientific result, it is Right. If a historical example that was Wrong is ever mentioned (alchemy and leech therapy seems to be favorites), there is almost no insight offered into why it was convincing to many at the time, and its wrongness is just offered as an unintentionally ironic bit of support for the claim that we now get everything Right.
Even once you rise above those problems a bit, there is the further problem that physics is taught as if it were the representative science, when in reality it is quite different from most other sciences. The small number of new discoveries in physics are carefully poured over by lots of people, are based on repeated observations to the extent possible, and are parsed against a lot of theory. Even for the physics that requires bleeding edge technology, there is still a lot of replication, and it has long been thus (e.g., the Michelson-Morley experiment and its then-surprising result was aggressively refined and replicated). Perhaps this is because physicists are just better scientists, but the field also has the advantage that the number of topics to explore is relatively limited, and so replication and thus embarrassment for claiming something that is never replicated are inevitable.
On the other hand, consider one classic bit of physics and imagine how it looked at the time and not in hindsight. When Galileo peered through his telescope and identified moons orbiting Jupiter, parsing that with Copernicus’s theories, there were only a few telescopes powerful enough to see the moons and other people who looked through those instruments could not make out what Galileo claimed to see (almost, though not quite, everyone — there was actually a contemporary independent discovery). His contribution to science relied on his scrupulousness and honesty because the claim was, for a while, entirely dependent on his work. Had he claimed to see angels carrying signs that said “terra est centrum” (or whatever the right grammar is for “Earth is the center of the universe”), that might have remained the “evidence based” reality for a while longer.
So, to circle back to epidemiology (the quantitative analysis of health outcomes and exposures in people, both observational and experimental): Doing epidemiology is trivial, thanks to plug-and-play software, but doing it right is extremely difficult. A three-year-old can draw a picture and a first semester student can crank out some epidemiology statistics. This characteristic is not what sets it apart from other sciences, however; a grade school student can do the physics experiment with the pendulum. The contrast appears because the drawing sees no gallery other than the family refrigerator and the pendulum data is graded and discarded, but the equivalent work in epidemiology becomes a “peer-reviewed publication”.
We rely on researchers to be expert in their science as a start, and then to be scrupulous and careful in their research, and to recognize uncertainty and to examine it to the extent possible and be epistemically modest because it exists. But in epidemiology, almost none of what is published meets those standards. But like cheap wine put in a fancy bottle, the average consumer of the science has no idea how much like the child’s drawing it really is.
But it is even worse than that. Because the standards of the field are so weak, it is trivially easy to not just do bad science, but to do intentionally misleading “science”. Epidemiology is made bad enough by the deluge of publications by inadequately competent authors, but it is made far worse by its hijacking by “public health” (the political movement, not to be confused with real public health scientists, including epidemiologists, who are at least trying to do good research). There are oh so many ways to get the result you want by doing bad epidemiology. Doing intentionally-misleading epidemiology is far easier than doing good epidemiology.
Sometimes it is incredibly obvious what someone has done, as with this example where “researchers” with obvious biases have picked which periods of data to consider to intentionally gin up the incredible claim that banning smoking in pubs results in measurable reductions in heart attacks. There has been an enormous amount written about this junk science because the manipulations are so obvious — I just linked to the most recent such analysis. But sometimes the manipulations are subtle enough that only a few experts notice them, as is the case with the series of publications from the Karolinska Institute in Sweden that claimed to find an association of snus use and various diseases. In that case, they clearly fished for statistical models that would show the strongest association (and thus were biased upward from the true value), something that is almost impossible to spot (a bit more about this here and here). This model fishing characterizes a large portion of epidemiology — even the bits that are not intentionally dishonest like these examples, but merely result from researchers not having enough expertise to even understand that this is bad science.
But, in all of these cases, the cheap wine is packaged up in the nice bottle, and so health reporters, policy makers, medics, and most everyone up just blindly lap it up.
So is this much different from, say, economics, the science that is most similar to epidemiology? Or from toxicology, another “public health”-influenced science that some consider more sciencey because it takes place in a laboratory? The answers, I would argue, are yes (economics shares many of the same problems, but has better mechanisms for fixing them) and no (toxicology is just as manipulated and misleading as epidemiology in practice). But those points will have to wait for another Sunday.