Donald F. Klein’s answers to Martin M. Katz’s questions
Marty Katz sensibly raises a central problem in scientific discussion. A word may derive its precise meaning from a particular mathematical or well defined psychological context. However in verbal discussion there can be semantic slippage so that terms are misused, because in a different context they are now inappropriate.
Katz gives examples, “One problem in regard to discussing the mixture issue may be the several meanings we encounter in psychometrics for factor analysis. When using Hotelling's principal components, I would restate that of the factor analytic techniques involved, principal components is characterized as a strictly mathematical approach, based on deriving dimensions generated by the inter correlations of the factored variables, inappropriately requires no maintenance investigators confined to minimal interpretation, i.e., interpreting the meaning underlying the most highly "loaded" variables of an extracted component.”
However each loaded variable is a composite of correlated variables, each with a somewhat ambiguous label. Labeling the composite is not due to “minimal interpretation”. Ratherit affords ample grounds for disagreement and misunderstanding.
Katz continues, ”Factor analysis, in general, in psychometrics can, as we know however, take several forms, several of the techniques relying more heavily on the investigator's choice of the form and on his interpretations at several stages of the procedure”.
“I find it unclear about what meaning of "mixture" you are using in this context. Are you asking, e.g., whether the wide range of symptoms that we observe in depression, is the result of a mixture of the underlying syndromes of major depressive and generalized anxiety disorders, as against in the other case, the results of the interaction of independent dimensions, uncovered through factor analysis?”
Also a good example of communication difficulty, Katz clearly raises the mixture issue, “whether the wide range of symptoms ….are due to a mixture of the underlying syndromes … as compared to …the results of the interaction of independent dimensions, uncovered through factor analysis?”
I do not understand this last clause. Can dimensions be independent, but nevertheless have interactions? How can we resolve this? My general conclusion is that a complex verbal statement is best illuminated by a simple concrete example. I believe Katz is arguing that some form of factor analysis would produce results equivalent to a model of latent categories. An example would help.
Katz asks what is meant by inclusive design. This fits very well with the mixture model discussion. The term “Mixture” is well defined within modern statistical analysis. Muthen, in his online notes states: “M plus Class Notes
Analyzing Data: Latent Class Other Mixture Models. Mixture models are measurement models that use observed variables as indicators of one or more latent categorical (diagnostic) variables. One way to think about mixture models is that one is attempting to identify subsets or ‘classes’ of observations within the observed data. The latent variable (classes) is categorical, but the indicators may be either categorical or continuous”.
It is often unclear how to model the relationship of outcome to baseline data. For instance, in the 50s, NIMH and the VA hoped that multiple regression analysis might finddifferent treatment relevant diagnoses within an overall diagnosis, by using outcome as a validity criterion.Unfortunately these promising investigations failed on replication and the approach was abandoned.
Perhaps this was due to the heterogeneity of treatment outcome. This remained unclear in such studies. For instance, a study might find that 60% of medication treated patients remitted whereas only 30% of those on placebo did so. Given statistical significance, this was sharp evidence, sufficient for the FDA, that the medication was causally effective. However, identifying the responders who required medication for benefit had not been solved.
In 1967, J. B. Chassan extensively discussed the issue of how to identify drug responders in “Research Design in Clinical Psychology and Psychiatry” (The Century Psychology Series). However, this concern fell out of fashion, probably because the FDA sufficient successes of the parallel group extensive model design made it seem trivial.
Chassan’s ideas were revived and extended in a chapter; Klein D.F.: Causal Thinking for Objective PsychiatricDiagnostic Criteria: A Programmatic Approach in Therapeutic Context, that appeared within the monograph “Causality and Psychopathology: Finding the Determinants of Disorders and their Cures“, Eds. Patrick Shrout , Katherine Keyes, Katherine Ornstein (American Psychopathological Association 2010).
Chassan recommended “intensive design”, that is repeated periods of intervening and non-intervening, judging whether benefit synchronized with intervention. This concept suggests a different clinical trials design. Openly treat all relevant patients with the study medication program, titrating for optimal dose. Patients, who clearly did not respond to treatment are set aside. Responders would be divided randomly into two double blind groups; either to be weaned onto placebo or to remain on medication. All would be closely followed, double blind, for defined signs of worsening. Sufficient worsening would restart medication. Those who both worsened on placebo substitution and then improved on blind medication retreatment are very likely specific drug responders. In contrast, those switched to placebo, who continued to do well, wouldprobably not be specific medication responders.
A higher worsening rate among those switched to placebo than those maintained on medication would be clear evidence of medication efficacy, quite comparable to the inference established by the parallel groups, extensive design.
But better, the intensive design dissects the initial latent mixture into three response specific categories: likely medication specific responders, likely non-specific responders and non-responders. Each group’s meaningful outcome homogeneity,as well as increased heterogeneity between groups, may illuminate the drug’s specific benefit on pathophysiology.
Donald F. Klein
April 24, 2014