Friday, July 17, 2015

Hyman’s Maxim: The most important principle in observational sciences?

I grew up in a house of mathematicians. Among other things, this means that throughout my childhood, I heard a lot of jokes about physicists (the mathematicians’ equivalent of blondes jokes). As a child, I used to find those jokes mildly amusing. When I started learning about research methods in psychology, I started finding them really funny, but sad at the same time – after all, if a hard science like physics is a laughing stock because of their dodgy scientific practices, what does this mean for a soft science like psychology?
Here’s an example of a physicist joke:

A physicist is doing a talk at a conference. He is holding up a graph, and is explaining what the data on it means. After half an hour, a graduate student timidly raises her hand and politely notes that the graph is held upside down. The physicist stops, looks at the graph, turns it the right way up, and says: “Why, you’re right! Well, in this case the data is even easier to explain!”

The moral of the story is that, for a reasonably intelligent and creative person, it is almost always possible to come up with a plausible-sounding explanation for any set of results. This, of course, is already well-known: for this reason, the scientific method entails explicitly stating a hypothesis before the data is collected and analysed. However, in psychological research this principle is not straight forward, for a simple reason: it is very rare for the data to behave in the way that was anticipated.
Like probably everyone else in the field, I learnt this the hard way during my PhD. I conducted a study to look an effect in three conditions (let’s call the size of the effect in the three conditions A, B, and C, respectively). Two theories (let’s call them X and Y) made opposing predictions:
If X is true, A < B < C.
If Y is true, A > B > C.
The result? B = 0 < A = C.
I think this scenario is familiar to anyone who has ever done an experiment in the so-called soft sciences, and it’s a PhD student’s worst nightmare. What does one do with a set of results like these? One of my advisors said I can do one of the following: (1) Figure out why we got this unexpected set of results, (2) write up a paper with our initial predictions in the introduction, our results, and conclude that ‘more research is needed’ to understand this unexpected set of results, or (3) forget about the whole thing.
In retrospect, I should have done (3), but due to my stubbornness I went for (1) instead. I had numerous meetings with my advisors to discuss how any theory could account for the obtained results – but in this case, we could not even come up with a reasonable-sounding explanation. Then I decided to collect some more data. I conducted four more studies with larger samples, and eventually performed a meta-analysis of all the data on this effect that I could get my hands on (which, aside from the data that I had collected, was not much). Thus having maximised the power to obtain a potentially true effect, I found that A = B = 0, and C is only slightly larger than zero. On the bright side, this finally allowed me to conclude that the data is more compatible with Theory X than Theory Y, but at this stage I had wasted hours of my advisors’ time and most of my PhD trying to understand the results of the first experiment, which were basically just random noise.
This is where Hyman’s Maxim comes in. I came across it in this blogpost by chance, after I had already submitted my PhD thesis. The maxim says: “Do not try to explain something until you are sure there is something to be explained.” Ray Hyman started off as a magician, but later became a skeptic and a psychologist. Aside from the blogpost, I have not found any publications on Hyman’s Maxim, but in my opinion, this is the most important principle in psychological science, and possibly any science that involves drawing inferences from data. As a scientist’s main job is to obtain data that can support or refute theories, it is easy to get carried away with drawing the link between data and theory, and to forget how important it is to ensure that the data actually tells you what you think it tells you. In psychology, with generally small effects and noisy data, the non-zero probability that a statistically significant effect reflects random noise is often forgotten. Consequently, any statistically significant result is in the danger of being interpreted as ‘meaningful’: if the a priori theory did not predict it, we must be missing something, there must be some explanation, or moderating factor, which should explain this unexpected result.

In conclusion, unless we have ensured that an unexpected result is replicable, drawing inferences from a single study with a statistically significant result that was not predicted a priori is a lot like telling someone’s future from the stars or their tea leaves. In fact, if the null hypothesis happens to be true, it is literally like telling someone’s future from the stars or their tea leaves. On some level, everyone knows this already, but perhaps it is easy to forget this point. My proposed solution to this problem: Create motivational posters starring Hyman’s Maxim. Put them up in every psychological scientist’s office and bathroom.

*********************************************
Edit (29/3/15): I created some motivational posters starring Hyman's Maxim. Sorry for my awful photoshop skills; feel free to improve or make your own!



No comments:

Post a Comment