Scott, I believe that his conclusion is overstated, based on the summation he himself presents. Similarly, Hróbjartsson and Gøtzsche’s review, which he quotes, does not lend itself to as blanket a condemnation as a non-technical interpretation of the words might suggest.
Here are the issues:
When performing a meta-analysis, an analysis spanning numerous studies, one basically excludes studies (albeit not necessarily without acknowledging the fact) that are too flawed for inclusion in quantitative comparisons. The piece reported that the placebo effects found ranged from negligible to clinically significant. This must mean that some were indeed found to affect physiological markers as well as subjective reports. (N.B. My impression is that if there is any modification of subjective report that appears to be unaccompanied by physiological change, it is because the physiological changes accompanying it were not deemed to be of interest–perhaps a valid perception.)
When I say in casual speech, “She’s showing bias,” I mean she’s a prejudiced judge in some realm, unwilling to look at what’s in front of her. When I say that a scientific study or series of studies betray bias, I mean something far more specific. One of two things, actually. Sometimes I may say there’s systematic bias, another particular factor that should have been accounted for. For instance, classically, I might find that certain heritable racial characteristics correlate with early death. If I don’t account for the fact that in a particular country race is accompanied by many other characteristics, my data are a hell of a lot less useful than if I account for confounding elements that would bias the findings, such as social class and income.
Other times, I may say that study outcomes appear biased, and what I mean is that the outcomes are not a nice clean demonstration, showing the same consistent pattern of effect(s). There are many possible reasons for this, but one of the most common ones is that I don’t have a good way to either control the factors that affect the outcomes, or measure the pesky ones and thus control their effects in my data analysis.
An intuitive example is psychotherapy. A single talk therapy has been manualized well enough that it could be formally studied and proven to be effective, as effective in treating many depressions as medication regimes. This is cognitive-behavioral work, which I think of as straightforward elicitation and reframing of internal verbal patterns. Formalized and manualized or not, its effectiveness still varies with the practitioner. However, efforts to specify the differences between more- and less-effective cognitive-behavioral therapists have been of limited value. non-specific bias, or noise in the data set, from one perspective.
But even beyond this data set, what does it mean that this is the only approach studied well enough that its manualization could be demonstrated to work? It means that other approaches are messier, not that all the seeming results from their use in some of the more-rigorous peer-reviewed journals are actually artifacts of the passage of time (and money).