Medical researchers care about main effects, psychologists care about interactions. In psychology, the main effects are typically obvious, and it’s only the interactions that are worth studying.
Medical researchers care about main effects, psychologists care about interactions. In psychology, the main effects are typically obvious, and it’s only the interactions that are worth studying.
I think you give us psychologists too much credit. I suspect it is not because the main effects are obvious (because they not sometimes) but because we are trained in the ANOVA model, as opposed to the regression model. In the ANOVA that is taught from undergrad to PhD and beyond when you have two effects you must enter the interaction. We do it without thinking, it is the way it has to be done.
This difference may not be as black-and-white as your statement suggests (in fact, it may be simply wrong). For example, H. Kraemer has carved out a very visible set of reports calling for interactive models in controlled trials, see (a) (b) and (c) . These papers have been cited a collective ~1200 times (Google Scholar) in less than 10 years. This tells me there is plenty of effect modification work in medicine.
doesn't appear to be true For example, most of the
Jwa:
I'm glad that these papers have been cited 1200 times. Overall, though, I think medical research is focused on main effects, and you often hear advice of the sort that main effects are what's important and that inference for interactions is just exploratory and not the most serious thing.
Longtime lurker here… I laughed when I saw this because it's so true. I was trained in cognitive psychology but I work now as a data analyst in a medical research setting. The worldview difference took a few months to get used to.
I have a lot of psychologists as clients and many do not look at interactions; many do not understand what an interaction is. Until I explain it, of course :-)
I do think, though, that psychologist are less likely to arbitrarily dichotomize or otherwise categorize their variables.
Actually, psychologists are very likely to arbitrarily categorize their variables. Try to find one APA journal article that contains a model with categorical and quantitative predictors and their interaction. It will probably take a while. Indeed, interactions are often more interesting then main effects in psychology, but they are almost exclusively treated in ANOVA style, even though it is often the case that the predictors are quantitative in nature.
Dr. Flom,
I really like your statistical consulting website.
One cannot study the interactions without studying the main effects. And as for orientation, both ANOVA and regression are statistically the same thing, but expressed differently.
-Ralph Winters
Medical trials are usually big and paid for by other people's money. That means medical researchers have to say up front what they are testing for, how many people it will take and so how much money they need.
There is also implicit weeding out of some interactions by specifying the study population e.g. if they think women aren't going to benefit from the drug (compared to standard treatment) but men are then they'll only have men in the trial (and so omit a gender by treatment interaction).
Medical trials (phase 3 anyway) tend to be huge while pshychology trials tend to be small and spurious interactions are more likely to be picked up in smaller trials. Spurious interactions have a tendancy to be novel (because noone else has found it before/thought of it before) which makes them interesting and highly publishable.
(I might be slighly unfair to psychologists because I don't read their literature but what tends to be newsworthy are these really odd results based on 40 college students.)
In psychology it is incredibly difficult to discover a new main effect. So practically all of the experimental work is about (a) new variations of an already-known main effect and (b) interactions.
Thanks HA!
I wouldlike to know more about the way you propose hanlding interaction-effects:
1. I am trying to gather some opinions on the right way to deal with interaction in logistic regression models. Since Ai & Norton's contribution to the analysis of interaction effects in logistic regression models (or earlier), one can not interpret estimated coefficients in a usual manner (OLS etc.).
With their stata command inteff, they allow you to derive unit-specific interaction effects and unit specific levels of significance.
I do not know whether you are familiar with the possibilities these modeling procedure povides. In my opionion, this methodology can contribute bridging gap between deductivists (let's call them that way, you know, the critical rationalists that derive hypotheses etc) the bayesian side as is enable deductivists to differentiate between groups of respondents when the interaction effect is non-sig, positively sign. and neg. sign. So the interaction effect coefficients tell you nothing, you need to derive the unit-specific significance levels and effect magnitudes.
2. Recently, William Greene from NYU in has criticized the approach of Ai and Norton in EconomicsLetters (2010) 107:297-296. I guess the thinking is right in your line, arguing for a more profound graphical display of interaction effect. It would be great to get you opinion on both modeling alternatives, as Greene's emphasizes conditional plots instead of unit specific interpretations.
Coming from economics to psychology, the thing I find most maddening is how many psychology talks will present a series studies that show interactions; then you ask them what their theory is and they rarely have one. A lot of basic Economic theory may be mostly wrong, but at least it's wrong.
Maybe it sounded as if I thought that ANOVA and MR were different animals… they are taught as different processes from intro to stats to grad courses (at least my experience in psychology). And it has been ingrained in the psychology research ethos to treat them as different things. If I had it my way t-tests, anovas, and regressions should be taught as one common topic, as they actually are.