Lately, I’ve been trying to learn more about open science and how it relates to research I’ve done, research I’d like to do, and how it relates to sociolinguistics in general. One topic that comes up regularly when talking about open science is pre-registration. For those who aren’t familiar with this process, pre-registration refers to publishing a detailed, time-stamped description of your research methods and analyses on some repository before ever actually looking at your data. Doing so increases transparency for the research and helps the researcher avoid P-hacking, aka data fishing1. There are apparently some arguments against pre-registering research, but I’ve yet to see any that don’t mischaracterize what pre-registration actually is, so it seems like a no brainer to do it.

But in looking into the actual mechanics behind producing a pre-registration, I ended up watching the following webinar from the Center for Open Science (COS) about using their Open Science Framework (OSF) to publish pre-registrations, which included this curious description of how to interpret P-values in different kinds of research2:

Basically, the claim is that pre-registration makes it clear which analyses are confirmatory3 and which are exploratory, which is great, but the other part of the claim is that P-values are uninterpretable in exploratory research. In other words, any P-values that are generated through analyses that weren’t pre-registered, i.e. through data fishing, are meaningless.

I can understand why this point is made, but I think it’s a bad point. Pre-registration does seem to create another level in the hierarchy of types of research — i.e. exploratory (observational, not pre-registered) > confirmatory (observational, pre-registered) > causal (experimental) — but I see no reason why P-values are uninterpretable at the exploratory level. It would seem that P-values are perfectly valid at all levels, and all that changes is how they should be interpreted, not whether they can be interpreted at all. To me, in experimental research, a P-value helps one argue for a causal relationship, whereas in confirmatory observational studies, a P-value helps one argue that some relationship exists, though not necessarily a causal one, and in exploratory observational research, a P-value simply suggests that there might be a relationship and so that potential relationship should be explored further in future research.

In the case of my thesis, I did employ P-values via Fisher’s exact test of independence, but I didn’t pre-register my analyses. That’s not to say that all my analyses were exploratory, just that I have no proof that I wasn’t data fishing. Indeed, I included variables that didn’t make any sense to include at all4, but still somehow turned out to be statistically significant, such as whether there was a relationship between the person who coded each token of my linguistic variable, (lol), and how that variable was realized. The webinar initially made me panic a bit, asking myself if it was irresponsible to have included P-values in my analyses, but after further reflection, I think it was completely justified. Most of my analyses were confirmatory anyway, even though I don’t have proof of that, and those that were arguably exploratory were still more useful to report with P-values as long as an explanation for how to interpret those P-values was also included, which is perhaps the one place where I could’ve done better.

Ultimately, while I can understand why there’s so much focus on data fishing as a negative thing, I think it’s important to not overshoot the mark. P-values can certainly be misused, but that misuse seems to come down to not providing enough information to allow the reader to properly interpret them, not to whether they were included when they shouldn’t have been.

1. I prefer the term data fishing, which can be more easily taken in both a negative and a positive way, whereas P-hacking sounds like it’s always negative to me. The Wikipedia article on data fishing gives a pretty clear explanation of what it is, for those who are unaware.
2. The webinar is really good, actually. I would suggest that anyone who’s new to open science watch the whole thing.
3. In this case, the speaker seems to be using the term “confirmatory research” as something different from “causal research”, otherwise their description doesn’t make any sense.
4. In fact, my thesis advisor didn’t see the point in me including these variables at all.