We were asked by a reviewer to provide p-values as to better understand the model estimates in our bayesian multilevel model. The model is a typical model of multiple observations per participant in an experiment. We estimated the model with Stan, so we can easily compute additional posterior statistics. Currently, we are reporting (visually and in tables) the mean estimate and the 0.025 and 0.975 quantiles.

My response so far would include:

  1. P-values are inconsistent with bayesian models, i.e. $P(X|\theta) \neq P(\theta|X).$
  2. Based on the posterior, we can calculate the probability of parameters being larger (smaller) than 0. This looks a bit like a traditional p-value.

My question is whether this is a response that can satisfy a reviewer or will it only cause more confusion?

Update 10-Oct: We rewrote the paper with the advice in the answer in mind. The paper is accepted so I will reiterate my earlier comment that this was really helpful advice!

  • 2
    $\begingroup$ Maybe the reviewer is asking for the usual frequentist p-values, independently of your Bayesian model ? $\endgroup$ Apr 28 '15 at 9:37
  • 1
    $\begingroup$ That is so obvious I had not even thought about it. The model has some complications (weakly informative priors, some missing values on the explanatory variables) which makes it fairly difficult to run a frequentist version but I will think about reporting a stripped down model with the frequentist p-values. $\endgroup$
    – stijn
    Apr 28 '15 at 9:50

First, a quick clarification: Although the likelihood is indeed not the posterior, p-values are not so much inconsistent with Bayesian inference as usually just a different thing, for all the reasons that confidence intervals may or may not line up with credible intervals. (Although not necessarily an entirely different thing, as shown by posterior predictive checking, which really does involve p-values.)

However I'm guessing that this level of sophistication is not what the reviewer has in mind. I'd guess they just 'know' that statistical models are meant to have p-values, so they've asked for them. So the question remains: how to respond?

When 'reviewer wants an X' I have found it useful to ask myself two related questions:

  1. Motivation: What do they want X to do for them?

  2. Rational reconstruction: What would be most similar-sounding sensible thing they could have asked for instead of X if they wanted to do that?

Then give them that instead.

The advantage of an ignorant reviewer (who may nevertheless be smart and right about the paper) is that they seldom have a clear idea of what they mean when they ask for X. This means that if you reconstruct them asking a better question, they'll be content to see you answer it instead.

In your case, it's quite possible that the reviewer wants a parallel frequentist analysis, though I doubt it. What I think you want to work with is the reviewer's hint that they want p-values to 'to better understand the model'. Your job, I think, is to parse this in a way that makes the reviewer sound wise. Presumably there were a few following sentences noting what was unclear from the paper. Perhaps there were some effects of interest to the reviewer that could not be reconstructed from your parameter marginals, or some quantities that would illuminate what the model would say about cases of interest to them, or a lack single number summaries...

If you can identify these concerns then you can wrap up your response in the following forms (original request in square brackets):

"the reviewer [demands a p-value for an interaction term] was concerned that it was unclear from our presentation how A varied with B, so in Figure 2 we show..." or "the reviewer wondered [whether we could reject the hypothesis that the effect of A is zero] about the direction of the effect of A. Table 3 shows that this model gives a 99% probability that this is negative" or "the reviewer wonders [whether our model is significantly better fitting than a model containing only A] how our model compared to one containing only A. We address this question by comparing it to ... by using DIC / computing a Bayes Factor / showing our inferences about A are robust to the inclusion of B" etc.

In each case there is a close translation of the original request and an answer.

Caveats: this strategy seems to work best when the reviewer is a subject matter expert with a relatively weak understanding of statistics. It does not work with the self-identified statistically sophisticated reviewer who actually does want an X because they like Xs or read about them somewhere recently. I have no suggestions for the latter.

Finally, I would strongly recommend not saying anything even faintly religious about Bayes being a different paradigm and the reviewers questions making no sense within it. Even if this is true, it makes everybody grumpy for no real gain.

  • 1
    $\begingroup$ It will probably be tempting to note that p-values are a little bit difficult to define in frequentist multilevel models. Don't do that either ;-) $\endgroup$ Apr 28 '15 at 14:38
  • 4
    $\begingroup$ If you really give the reviewer $P$-values, it will help her understand the problem less well. People claim to understand $P$-values when the only reason they accept their use is because they do not understand them. $\endgroup$ Apr 28 '15 at 15:56
  • $\begingroup$ So we're on the same page then. Or this is a response to @stéphane-laurent $\endgroup$ Apr 28 '15 at 16:00
  • $\begingroup$ Yes sorry I didn't make that clear $\endgroup$ Apr 28 '15 at 18:43
  • $\begingroup$ @conjugateprior Thanks! Huge help. I already realised that mentioning the p-values in multilevel models problem would be obnoxious. But my point 1 could equally come across "faintly religious" as you say. $\endgroup$
    – stijn
    Apr 29 '15 at 0:36

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.