1 Boring Old Man

Now back to the last of the three studies analyzing and reanalyzing the children of depressed mothers [Children of depressed mothers 1 year after remission of maternal depression: findings from the STAR*D-Child study]. In that study, the mothers were divided into three groups: Early Remission [HAM-D < 7 by 3 months]; Late Remission [HAM-D < 7 between 3 months and 1 year]; and Non-Remission [HAM-D < 7 at one year]. The children’s course was followed at 3 month intervals for 1 year after the mother’s remission or 1 year – whichever came first. They followed: both child an mother reported symptom lists; a Problem scale with Total, Externalizing , and Internalizing Problems; and a Child Global Functioning Scale. Because this paper is not online, I’ve included the Total Problem and Global Functioning Data [as exemplary] below. I’ve also added their explanation of the statistical analysis.

^{Data Analysis: The relation of maternal remission status to child outcome in the year following remission was assessed with Poisson regression analysis when the outcome was the number of childsymptoms at each time point, and was assessed with longitudinal mixed-effects regression when the child outcome was a continuous variable, such as the Children’s Global Assessment Scale or Child Behavior Checklist score. Maternal remission status (early, late, or nonremitting) was included in the models as a three-level categorical predictor variable. All regressions were conducted within a generalized estimating equations framework to adjust for correlations between repeated observations over time. The Poisson regressions used an exchangeable correlation matrix, and the longitudinal mixed-effects regressions used an unstructured covariance matrix. We performed analyses using PROC GENMOD (for Poisson regression) and PROC MIXED (for linear regressions using longitudinal mixed-effects analysis) in SAS, version 9 (SAS Institute, Cary, N.C.). For all of the models described above, we included the following potential confounding variables in the analyses: child’s age and gender; household income; mother’s baseline HAM-D score, presence of an anxiety disorder or substance use disorder at baseline, and marital status; child’s treatment status over the course of the study (any versus no treatment); and child’s baseline score for the outcome under consideration. An interaction term representing time by maternal remission status was included in the regression analyses to formally test whether the rate of change in child symptom outcomes varied according to maternal remission status. Missing outcome data were handled by initially assuming that the data were missing at random for the continuous variables analyzed using mixed-effects models, and missing completely at random when using Poisson regression analysis using generalized estimating equations. Under this assumption, valid analyses can be conducted using models such as the mixed-effects mod-els based on maximum likelihood inference. Because of the large number of dropouts at the 12-month assessment in the late remission category, we performed a sensitivity analysis by rerunning all models after deleting the 12-month assessment for all participants and comparing patterns of change as represented by the beta coefficients for the 9-month period after maternal remission, and comparing results. Data on household income were missing for three mothers, and the average household income of all mothers in each maternal remission category was imputed.}

In the earlier paper, the data was displayed in graphs of the usual type. In this one, we got a table [partially reproduced above]. I’m suspicious when long tables are used when a graphic format would be easier, and since I was already put off by the erudite but impenetrable data analysis write-up, I thought I’d just make a few of the missing graphs to see what they looked like [I was suspicious of another significant but trivial example]:

This is the longitudinal data on the kids beginning with either the time of maternal remission or 1 year of treatment without remission [mean and one standard deviation above and below]. We are asked to believe that the improvement in the early remission group [left graph, blue] is important. I’m underwhelmed. Frankly, I’d throw out the 12 month late remission value [so many drop-outs, right graph, red]. So the early/late responses then seem the same. And when I graph the Children’s Global Improvement, I’m beyond underwhelmed:

If the point of the multiple publication of this same study was to get me or other clinicians to be aware that maternal depression has an impact on their children, they didn’t need to bother. No person who has done any psychotherapy with adults would doubt for a second that maternal depression takes its toll – our patients have told us about it for years. But my suspicion is that these papers have another point. By the time STAR*D came along, the newer antidepressants had been around for a while and the disappointment in the results was growing. Drs. Rush, Trivedi, and others were out to convince us that it wasn’t the drugs that were not as effective as we wished, it was that we just weren’t pushing them hard enough, or dosing them high enough, or treating people long enough, or weren’t changing drugs often enough, or weren’t adding other drugs enough. That was the whole agenda of STAR*D, CO-MED, TMAP, or the other algorithmic strategies. So I think this study was designed to prove that treating the mothers helped the kids, emphasizing their "vigorous treatment" meme.

If you look at the increasingly indefinite conclusions from the papers, they started with a full court press and ended in the range of wimpy. The data didn’t support their conclusion [which kind of surprised me]. But one point I want to make is that STAR*D, CO-MED, TMAP, and the other algorithmic strategies were based on a pharmaceutical industry agenda, "Use more of our great antidepressants!" Their antidepressants are useful, but nowhere near the wonder drugs of the ads [or our literature]. There was another agenda – resume inflation [this graph was previously compiled from pubMed]:

All four of these authors are on all three of these articles. And all three of these articles are about the same patients and could’ve easily been covered in a single paper. Instead we have three different articles in primo journals over a five year span.