Svar til Paul Barrett
Av Ronny Klæboe Dato: 19.5.2009
I’d like to underline a comment I made earlier so that it sinks in.
SEM has had a tremendous impact on the behavioral and social sciences. 1972 was a watershed year that led to a sea change in the way behavioral and social scientists used quantitative methods in their research. Prior to that time psychology and the behavioral sciences were making little progress because it was bogged down with trying to reject the nil hypothesis. Very little theoretical effort went into developing testable theoretical hypotheses, especially among those wedded to correlational methods, which essentially looked for significant correlations, ignored insignificant zero correlations and took the view that correlation does not imply causation, so we should not study causes with correlations.
Peter Bentler once said in a lecture I heard given by him, that “While correlation does not imply causation, causation does imply correlation”. That’s the key idea of SEM. If we can, through what we observe in nature and in behavior, gain some idea of what may cause what, we can then formulate models that seek to test such hypotheses by the manner in which such causal relations will manifest themselves in patterns of correlations among many variables. We specify a model that determines these patterns and see how well they reproduce the observed correlations (covariances).
Sure, to use SEM well requires that we learn to do things a bit differently. We need to understand how to formulate our hypotheses in terms of variables, quantities.
I am continually frustrated by attending thesis presentations where social psychologists have taught their students to think of constructs as nouns, e.g., “intelligence”, “leadership”, “anxiety”, “socialization”, “affiliation”, ‘internal locus of control”, and so the students come up with regression or even SEM models in which they draw arrows between these things and then try to relate them to some indices, as if these things are unidimensional. If taught to think of constructs as something like “degree to which…” “extent to which…”, “frequency with which…” etc., then the student/researcher will focus on specific unidimensional variables and think then of what sort of things cause variation in such variables, and be farther along in formulating meaningful causal hypotheses. So, SEM doesn’t get well implemented in these situations, and that perhaps leads people like Paul Barrett, who hungers for major breakthroughs in psychology to condemn SEM as just another useless methodology. It is not useless. It is just misused.
We have to pay more attention to ‘experimental design’
in our SEM studies, in finding instrumental variables to help establish causality and causal direction. Often the instrumental variables can be manipulated by the researcher, perhaps randomly, thus breaking any relations with other causes of the variables in question, so we can get a clean pathway from our manipulated cause through the endogenous cause to the downstream endogenous effect.
I mentioned 1972 above. That was when Karl Jöreskog began to make available his programs for doing confirmatory studies, leading to LISREL, and from there others have followed.
But SEM has had a major impact. But researchers have to think of the right kinds of observations and causal theories needed to implement it.