Sizeless statistics

We were shown that the log-odds-ratio was 0.101 and the number of sigificance stars attached to it was three. But we were not told the subject-matter implication of this finding, other than that the “log odds” of migrating was larger in one group than in the other.

But why is sizeless statistics so common today? This is an increasing problem in applied statistics, especially in the social sciences, since a couple of decades ago. It has been addressed by several researchers lately, among them Ziliak and McCloskey: The Cult of Statistical Significance: How the Standard Error Costs Jobs, Justice, and Lives (see a review of the book). The most extreme step was taken by Basic and Applied Social Psychology, who banned p-values and hypothesis testing in published articles from 2015. This is of course a too extreme measure to take, though.

One major reason for the deterioration of statistical practice in the social sciences, especially economy and psychology, is the “doityourself” mentality that is common in these circles: Statistical science, model building, reasoning, and statisticians, are replaced by quick fixes from Stata manuals (Stata is one of the top destructors today: We heard for instance about discrete time logit modeling as a recipe). This wouldn’t work, of course, if not many editors and referees of scientific journals played the same game. But many do. One reason is the growing commersialism (authors have to pay the journal), which triggers partiality and tendencies to publish papers with deficient treatment of statistical modeling.

OK, we got a number, the log-odds-ratio was 0.101, fine. But I doubt that many non-statisticians really understand what it means. And the question is if it really means anything to anybody without knowledge of the log-odds of the reference category. So this is really sizeless statistics.

Ignored nonresponse

We were also told that in a survey the response rate was 72%, an excellent result, on the ground that it in any other respect was very representative. And because many similar surveys have worse response rates! I commented that a dropout of 28 per cent is a statistical disaster, and it was agreed upon. The same story here: Noone cares. Publish or perish. Enough said.

Copyright © 2016 Göran Broström and CEDAR. All rigths reserved.