I came across a tweet last week by an academic at a conference (I can’t remember who). They were indignant that presenters were using the word ‘predict’ to describe a correlations. My first reaction was to sigh. Prediction has no causal connotation. When you go to the fair and some huckster offers to guess your age for a price, they are making a prediction based on your physical appearance. This prediction does not require a belief that your physical appearance caused your age. Such a belief is absurd. Yet prediction is still the right word.

This was my first reaction. My second, was to reflect on my use of prediction in reporting research results. While I believe ‘to predict’ requires no causal beliefs, it implies a certain level of accuracy. I have used the word predict to describe a correlation of .20. On reflection this seems wrong. Not because I am implying causation. I am not. But because I am implying a level of accuracy in predicting x from y. The implication is that by knowing a person’s value on x I can make a good guess at their value of y. But a correlation of .20 or even .40 as a basis of such a prediction would be atrociously inaccurate. Reporting such weak results using the word predict leads the public, who rightly read ‘predict’ as ‘predict accurately’, to overstate vastly the significance of the finding.

The social sciences are well known for being terrible at prediction.

Predictive accuracy is often woeful, even on the data that the researchers used to build the statistical model (see here for example). The social sciences often seem to not know about the importance of, let alone test, the predictive accuracy of a model on unseen data. Really, the only metric of predictive accuracy that matters.

In his fantastic paper, Shmueli argues that the social sciences have neglected the prediction component of science in favor of a complete focus on explanation. Mostly this is because of the mistaken belief that explanation is synonymous with prediction. 

And here lies the problem. The social sciences are scathing on anyone who uses the word prediction outside of RCT research. But this fit of pique is misdirected. The willing skeptic at the fair may say “I have $10 with your name on it, if you can guess my age to within a year”. So too we should call authors on their use of “predict “ when their models are scarcely better than chance.

I hate academic conferences. What seems like a chance for free travel to an exotic location turns out to be an endless bore in a stuffy room. For an introvert, the need to be constantly ‘on’ when talking to students, peers, and that big name you are desperate to collaborate with is tiring. The point being, I am not usually in the best of moods when at conferences. Which is probably why I found a particular presentation so irksome.

Why so much Person-centered Seems so Hollow to Me

The presenter at this hot and stuffy conference gets up and smugly states that previous, crappy, social science has used a variable-centered approach to research. He, however, would use a person-centered approach. The motivation was, I confess, solid.

Person-centered analysis starts with the assumption that within any large group of people there are likely smaller distinct groups (within a school there are jocks, goths, nerds, etc). Too much research treats humans as a bunch of mini clones that are driven by the same processes and differ only in degree. I can get behind this sentiment.

I was surprised then to not hear mention of a single person for the rest of the presentation. No explanation was given of how people in the different groups think, believe, feel, or act differently from each other. Nor was there discussion about whether people chose to be a member of their group or whether they were forced into it. Did they jump or were they pushed? Instead, the entire presentation focused on various configurations of variables. This was not, to me at least, person-centered.

This is a disturbing trend in person-centered research. The almost total absence of people.

The overall impression I get from most person-centered analysis is that people believe human diversity has been ill-treated by regression like approaches. But many researchers assume that by applying things like cluster analysis or similar they will magically fix this problem. In my experience, researchers don’t seem to put in a lot of thought into how these approaches better represent real people or what the results are really saying about them. Researchers tend not to describe a prototypical human from each of the researcher’s groups. Researchers also apply little imagination to what drives people in different groups.

Greater attention to this could truly transform the social sciences. A truly person-centered ontology and epistemology could serve disadvantaged groups better. Researchers could better acknowledge that the experience of, say, an Indigenous girl is qualitatively different from a South East Asian Australian boy. But to do this, person-centeredness needs to be about more than methods. And it needs to be motivated less by an appeal to what it isn’t (e.g., “Unlike previous research we use person-centered approaches” is not a convincing rationale).

Give me person-centered person-centered analysis

A move in the right direction would be to consider what Rob Brockman and I have recently called the four S’s of person-centered analysis:

1. Specificity. Once you have your groups, can you describe what makes these groups distinct? By this I don’t mean a profile graph of variables used to create the groups. I mean a deeper insight to what these groups of people are like. What do their members do? How do they think? What do they want?
2. Selectivity. How do people end up in these groups? By what process does a person end up in group A and not group B? Were people born into different groups? Did some person, institution, or cultural practice push them into their group? Or is their group membership their choice?
3. Sensitivity. Do these same groups occur in different samples? If not, why not? Do differences in groupings across—for example—countries illuminate how people’s context shapes grouping, or do differences just reflect unreliable research findings?
4. Superiority. The beauty of cluster analysis is that it will always return the number of groups you asked for. And like a Rorschach test, it is easy to make something out of whatever the computer gives you. Researchers should attempt to show that their groups tell us something we did not already know. And researchers need to show us that groups really differ from each other qualitatively rather than merely quantitatively.