Wednesday, October 31, 2018

Should we believe the p-value of a study result more when the finding fits with (vs. challenges) our expectations?

I keep seeing the same debate about preregistration. (There are several, but one in particular seems to repeat over and over again, at conferences and online, between preregistration advocates and skeptics.) It goes something like this:

Advocate: Preregistration is really important for science because preregistering a study makes the findings more trustworthy.

Skeptic: This is ridiculous! A finding is not more likely to be true just because you happened to correctly predict it ahead of time!

Advocate: Nobody is saying that.

Skeptic: Let’s say you and I both run the exact same study, but we make opposing predictions: I predict A, you predict B. My study shows A, as I expected. Your study shows A, contrary to your expectation. My study’s finding isn’t somehow truer than yours just because I happened to call it correctly ahead of time!


I thought I saw, in this repeating debate, a simple but crucial miscommunication: Preregistration advocates were using the term “preregistration” to refer to pre-analysis plans, which constrain researcher degrees of freedom and can help ensure that p-values are interpretable as diagnostic about the likelihood of an outcome. But advocates would also sometimes talk about preregistrations as involving prediction, even though making a directional prediction isn’t necessary for constraining researcher degrees of freedom (the clearest illustration of this confusion is probably this Data Colada blog post in which the researchers who started AsPredicted point out that no prediction is necessary for preregistration, and that in retrospect they probably should have called their website AsPlanned).* And so skeptics would hear the term “preregistration” and think that it meant prediction, even though it often meant pre-analysis plan.

So I wrote this short piece to make what I thought was a minor but important point about clarifying our terminology. I framed it as a reply to one particular article that uses the term “prediction” a lot while advocating for pre-analysis plans, but I tried to emphasize that I was making a broader point about the language many of us use when promoting preregistration.

But I was in for a surprise: It turned out that what I thought was a simple miscommunication was in fact a deeper disagreement. It turns out that yes, some people are indeed saying that a finding is more likely to be true when you correctly predict it ahead of time.

I find this position unnerving because it’s hard for me to see where the line is between it and, say, a person who thinks vaccines cause autism deciding that they don’t believe the scientific evidence to the contrary because it challenges their personal beliefs or expectations. (Presumably, there is a line, but I have yet to see it clearly articulated.)

I had an opportunity to discuss this difference of opinion with two of the authors on the original PNAS paper linked above, Brian Nosek and Charlie Ebersole. The full email discussion is here. I’ll pull out some highlights:

Alison: …I think we might disagree here—or at least, I think we need to distinguish between beliefs or confidence in a THEORY versus beliefs or confidence in a study RESULT. I agree that we should update our confidence in a THEORY based on what it was able to predict ahead of time. I disagree that we should base our confidence in a RESULT on whether it was predicted ahead of time.

I think that in your example [here], the three researchers with different predictions should believe the study result equally based on the statistical evidence.…The strength of the EVIDENCE doesn't change depending on what was predicted ahead of time, but our beliefs about the THEORIES that gave rise to the predictions can and should. 

What do you think?

Brian: …If by RESULT you mean knowing something about what the finding is, then there is definitely disagreement.  If RESULT A is sequential priming is stronger when the prime precedes compared to follows the targets and RESULT B is sequential priming is stronger when the prime follows compared to precedes the targets, a p-value showing RESULT B is less believable than a p-value showing RESULT A.  I don't need any information about the theories that anticipate A versus B to have priors about the results themselves and therefore update those priors based on the statistical evidence. …. priors can be very relevant with no theory.  A mouse can have strong priors that pushing a lever produces a pellet and when the light produces the pellet instead, the mouse will not update the priors as much as one that did not have the prior experience.  We don't need to assert that the mouse has any theory at all--priors can be based entirely on contingency expectations without any model or explanation for how those contingencies emerged.

Alison: …"How confident am I in my theory or [prior] belief?"…is separate from the question "How confident am I in this study result?" Because the amount/strength/quality of the evidence provided by a single study does not depend on the researcher's prediction. A study is not stronger if a researcher guesses or predicts the result ahead of time (which is what I think you imply when you equate "prediction science" with pre-analysis plans in your original PNAS paper). The quality of the evidence depends on things like whether there was a pre-analysis plan and whether construct validity and internal validity and external validity were high. If all those things are in place, the study might provide very strong evidence. Whether that strong evidence is sufficient to change a researcher's mind about a theory [or prior belief/expectation] may depend on the researcher's degree of confidence in the theory before the study was run. If they are very confident in the theory, then even strong evidence may not be enough to change their mind. But they can't call it weak evidence or a poorly conducted study just because the results turned out to be different from what they expected.

Charlie: …I agree with your concerns about the threats to falsifiability that come from differential interpretations of studies. In my earlier email, I was mainly trying to simulate the reaction that someone might have to learning the results a preregistered study that goes way against their priors/theories/expectations....Someone who believes in ESP and I might be able to agree on how p-values work (although ESP does have some really interesting implications for the ability to construct data-independent analysis plans) but we are likely to not agree on how time works (where our theories disagree). Even if I agree that the p-value from their study is diagnostic and that their study is high quality (high validity and all that), I may still think it's more likely than not that their results represents a true false positive, and not reality, because it's so against my theory and prior beliefs. Again, I'm not saying that I'm being rational or fair in this situation, but it does represent the gap between believing in statistical results, judging the implications of a given result, and then revising theories/beliefs.

Alison: I agree that it's tempting (and human nature) to think: This new information contradicts my existing attitudes and beliefs (including my favorite theoretical predictions), and so I don't think it's as good quality as I would if the same kind of information supported my existing attitudes and beliefs. In fact, I have run studies on exactly this kind of motivated reasoning (e.g., participants believe a scientific study is better in quality when its conclusions support vs. contradict something they want to believe). But this kind of irrational reasoning is NOT good scientific reasoning, and it's arguably a big part of what landed us in our current mess. Scientists need to seek consensus about what constitutes good quality evidence independently from whether that evidence happens to support or oppose their preferred conclusions. Indeed, this is one of the major benefits of reviewed preregistrations (a.k.a. registered reports) like those at Cortex and CRSP and JESP and other journals: Reviewers evaluate the soundness of the study's methods before the results are known.

In the ESP example, I think what you need to say is: "It would take a LOT of high quality evidence to change my belief that people are unable to predict the future," rather than "evidence is high quality to the extent that it confirms my belief about people's ability to predict the future." 

Charlie: …I agree that that's not good scientific reasoning. I wasn't trying to display good scientific reasoning, merely just trying to highlight a situation where someone could have different views on 1) the diagnosticity of the statistical evidence (namely the p-value), 2) their interpretation of a study, and 3) their resulting shifts in beliefs/theories.

I do agree with your line:  "It would take a LOT of high quality evidence to change my belief that people are unable to predict the future," rather than "evidence is high quality to the extent that it confirms my belief about people's ability to predict the future." Based on my priors, a true false positive (they will happen from time to time, even with preregistration) may seem more likely in a single instance (judging a single study) and thus be my interpretation of the study ("it wasn't a bad study, I think the results were a fluke"). Multiple observations of the effect would then make them being false positives less likely and would force me to confront my beliefs/theories….

[We ended our discussion of this point soon after, so that we could move on to debate our second point of disagreement, which I will turn to in the next post.]


*Unless you’re planning one of a small handful of statistical tests in a NHST framework that do care about the direction of your prediction, like a one-tailed t-test. And of course, Bayesian statistics provide a formal way of integrating a researcher’s prediction (or lack of prediction), their confidence in that prediction/prior belief, and their commitment to update their beliefs based on the strength of evidence observed a given study. But we’re not talking about one-tailed tests or Bayesian statistics here.


Response from Charlie:

First, I want to thank Alison for engaging with us on this topic and for sharing a draft of this blog post before posting it. We had a very interesting conversation about these and other issues, and I’d encourage interested readers to read through the whole exchange. I’d also like to state upfront that I’m speaking just for myself here, as I was in our conversation. I don’t claim to know what Brian thinks (finishing my dissertation would be much easier if I did).

The main point Alison raises is that some folks (much to her surprise) think that findings are more likely to be true if they are predicted ahead of time. I’ll admit that I had some meta-surprise in response to this, as I have apparently missed the hypothetical argument that starts this blog post (I guess I’ve spent my time shouting about other aspects of preregistration [1]). However, that might be more a reflection of different people using terms to mean different things. I tried to explain my reasoning for this in the following paragraph, taken from our exchange:

“Jumping to your last email, Alison, I think I disagree with the statement that "A study is not stronger if a researcher guesses or predicts the result ahead of time" because pre-analysis plans, at least to me, specify a prediction and do provide stronger evidence when provided ahead of time. The variables in our models represent the conceptual relationships we are interested (in) and we can have more confidence in the inferences we make from those models if we've specified them data-independent. This feels to me a little bit like our bank shot metaphor in our response (referring to this: The bank shot is the model that the shooter plans to use to make the basket. The inference that I'm trying to draw from this scenario is whether or not I think the shooter is good at basketball (or at least shooting in basketball?). Whether or not they call bank, I can certainly agree that they made that particular shot. If that's the only shot I care about, I don't much care whether they called it or not (no need to use inferences if you've sampled the entire population of interest). But if I care about drawing further conclusions about the shooter's ability, I will have greater confidence in them if I knew their planned model ahead of time. In that sense, it's better evidence because I've got a more accurate representation of what happened (or a more accurate representation of the relation between prediction and result)”

When we use inferential statistics, we’re trying to infer something about the broader world, broader populations, or future events from what we observed in the data. If a researcher surveyed the political views of 100 undergraduates and only wanted to draw conclusions about those 100 students at that one time, there’d be no need to calculate p-values – they’ve sampled their entire population of interest. However, that’s not the kind of question we typically ask in research (and certainly not how we write our discussion sections). P-values give us a way of thinking about the likelihood of a result given a particular null hypothesis and are a tool we use to judge the likelihood of a finding. P-values also lose their diagnosticity if they come from data-dependent analyses, which isn’t a worry if you’ve called your shot (or model) ahead of time.

So yes, I think a finding is more likely to be true if it is predicted ahead of time. Our predictions are manifested in our statistical models and the results from those models inform our confidence in a finding. As long as we’re using p-values as a way to calibrate that confidence, it’s important to know if we’ve called our shots or not.

[1] Such as “should there be a hyphen in preregistration?” Answer: No. 

1 comment:

  1. Some confusion may stem from phrases like "believe a result" or "finding is true". What does it mean to believe a finding vs. believe a conclusion that is made contingent on that finding? I can believe a result while also believing it to be a type I error.