7 ms·
It's worth distinguishing between different types of bad science. 1. The first and most egregious type is outright fraud. This is when you intentionally manipu
by majos 7y ago
It's worth distinguishing between different types of bad science.
1. The first and most egregious type is outright fraud. This is when you intentionally manipulate or fake data. Everyone agrees this is bad, and honest actors are enough to prevent it. In some cases, other honest actors are sufficient to determine if the claims are fishy.
2. The second more subtle type is not paying attention to adaptivity. For example, maybe an investigator wants to look at the data before coming up with a hypothesis to test. This is dangerous because the investigator is already overfitting, so any p-values the investigator computes afterward do not mean what they're supposed to mean. This is less egregious because it's easy to do this just by not being careful or not knowing your statistics very well. A scientist can be honest, but imperfect, and do this. It's also not easy to sniff this out as a reviewer -- the scientist might just omit all the stuff that didn't work. But there appears to be growing awareness of this kind of problem.
3. The third, and hardest to solve, problem is not factoring in the whole population of experiments. This is where 100 labs independently try an idea and one of them gets a genuine (from their limited view) result with a genuine p-value. It's novel and that lab has (in the limited view) been careful about adaptivity and keeping their hypotheses carefully generated. Maybe they've even used carefully generated noise to ensure their conclusions generalize [1] (which would definitely cut down on this problem). So it's pretty much impossible for a reviewer to tell there's a problem, because they don't know about the 100 other people that tried this and failed because the randomness didn't go their way. Short of a public experiment registry, this one is hard to fix, especially because it may be that nobody's being malicious or ignorant.
[1] https://arxiv.org/abs/1411.2664 https://arxiv.org/abs/1411.2664
- marcus_holmes 7y ago>> Short of a public experiment registry, this one is hard to fix, especially because it may be that nobody's being malicious or ignorant. That seems like a really good idea, that would solve a lot of problems. Can we do that, please?
- ars 7y agoThere is: https://clinicaltrials.gov/ https://clinicaltrials.gov/ I have a vague memory or reading somewhere that if you don't register your desired success condition in the trial ahead of time, even if you find something it's not considered a real result.
- Fomite 7y agoFor some studies this works really well - clinical trials, for example. But a lot of studies don't fit into a clean "Single experiment to measure a single outcome" paradigm, and pre-registration becomes somewhere between difficult and impossible.
- kradroy 7y agoMy advisor was guilty of pushing me to do #2. "Your original hypothesis was disconfirmed? Find one that the data does not disconfirm and that'll lead to a follow-up paper!" I dropped out of grad school because I assumed this was the norm.
- Aperocky 7y agoThat may not be a wrong assumption.
- unlinked_dll 7y agoI feel like evidence dis-confirming a hypothesis is as valuable as confirming it. How many other grad students are going to attempt the same experiment/concept if null results are never published? I wish there was more praise for negative results in publication, because whether confirmed or not, the knowledge has value.
- bonoboTP 7y agoI think it's partially embarrassing when that happens. Hindsight is 20/20, and people will think it was obvious that it wouldn't be the way you expected. It's hard to argue that I had justification to think that novel intervention X would have an effect. It turns out it doesn't. Science is often very specialized, there's little chance others would have the very same idea. If it works out, the argumentation would have to be reversed: my idea was very novel and non-obvious but as I show it actually works, which no one would have guessed. The negative result story only works if the research community would have very strongly expected to see the effect, almost reversing the role of the null and the alternative.
- thfuran 7y ago>The negative result story only works if the research community would have very strongly expected to see the effect, almost reversing the role of the null and the alternative Depends what you mean by "works". If you mean "is reasonably publishable in the current academic climate, then I agree. If you mean "has value", then I disagree.
- bhk 7y agoI would think that #3 is actually easy to solve by increasing the statistical standards. #1 and #2 are difficult to solve, given the near-exponential growth of funding for academic research. https://www.researchgate.net/figure/Evolution-of-the-number-of-papers-published-per-year-from-1872-to-31-st-of-May-2014-The_fig2_263515300 https://www.researchgate.net/figure/Evolution-of-the-number-...
- tikej 7y ago> #1 and #2 are difficult to solve, given the near-exponential growth of funding for academic research. Could you show any data that supports the claim about nearly exponential growth of funding for academic research? AFAIK the funding, especially considering growing number of people in the field, gets rather worse than better.
- bhk 7y agoThe astounding growth in the rate of publications (the chart I linked to) could not have arisen without a tremendous growth in the number of people, and in the funding. I don't think growth can be achieved without a decrease in overall quality of contributors / contributions, and indeed I wouldn't be surprised to see lower funding per author.
- Gatsky 7y agoPoint 2 is fine if you perform orthogonal validation. In biology, it is not possible to formulate hypotheses prior to gathering data in all cases. What was the hypothesis for the human genome project? What hypothesis could you conceive before ‘looking’ at the genome?
- syrrim 7y agoI don't think people understand how bad the third one is. If 100 labs try an idea and only one publishes, then clearly we can get a spurious p-value of .01. But what if 10 labs try it, and 7 publish their non-significant results. Then, a meta-analysis takes the results of those 7 studies and combines them together. Here the meta-analysis can get a spurious p-value of .01 with an order of magnitude less original studies and with most of them publishing their results. The culture of selective publishing and of parading any significant result no matter how small the effect size is incompatible with correct science.
- FabHK 7y agoHow would a meta-analysis wrongly arrive at a spurious p=0.01 when 7 of 10 labs publish their (non-significant) results?
- aaavl2821 7y agoA good example of no. 2 happened last week. Biogen, a large biotech company, announced it was resurrecting a previously failed drug for Alzheimers based on a post hoc analysis. The drug has potential to be the biggest drug of all time, so the stakes are quite high for the company. But the data they have to support their product is quite inconclusive. I wrote an analysis of their data here: https://www.baybridgebio.com/blog/aducanumab-analysis.html https://www.baybridgebio.com/blog/aducanumab-analysis.html
- neumann 7y ago3. That's why you should not rely on p-values!