Placebo Response in Trials of Negative Symptoms in Schizophrenia: A Critical Reassessment of the Evidence

Abstract Background Summarizing evidence from clinical trials of patients with schizophrenia with predominant or prominent negative symptoms (NS), a prior meta-analysis reported a large placebo effect in negative symptoms (Cohen’s d = 2.909). Assuming that such an effect was clinically not plausible, we performed a critical re-assessment and an update of the previous results with newly available data from add-on and monotherapy studies. Study Design Random-effect meta/regression analysis of trials that focused on predominant or prominent NS; and adopted a double-blind, randomized, placebo-controlled design. The final pooled meta-analytic database, based on the available add-on and monotherapy studies combined, included 24 publications containing data on a total of 25 studies (21 add-on, 4 monotherapy). Study Results The pooled overall estimate for the placebo effect from the primary analysis for all included studies had a medium effect size, with a Cohen’s d value of 0.6444 (SE = 0.091). The estimates were similar in the add-on and monotherapy studies. Meta-regression indicated that the high placebo response was significantly associated with clinical trial characteristics, including the high ratio of patients assigned to active vs. placebo treatment and short trial duration. Conclusions These results represent a major downward correction for a current effect size estimate of the placebo response in the negative symptoms of schizophrenia. Our findings also pinpoint certain clinical trial characteristics, which may serve as important predictors of the placebo response. The knowledge of these factors can have important implications for drug development and trial design for new drugs for negative symptoms of schizophrenia.


Online Supplement, Part 1: Details of the Search Stategy
The current meta-analysis focused on studies that were published in English. We formulated three query expressions (Q1-Q3) to select studies from the Pubmed and Web of Science databases according to the above criteria. Our search targeted all available fields in the bibliography database, including (but not limited to) the key words, the title and the abstract. The first two query expressions (Q1, Q2) were the same for the add-on and the monotherapy studies, whereas the third one differed depending on whether add-on (Q3a) or monotherapy studies Q3m) were selected for the analysis.
Specifically, the first query expression (Q1) focused on the selection of studies with predominant or prominent negative symptoms. It included a number of potentially relevant search phrases (see in quotation marks below) which were connected with logical operators (AND, OR) in the following form: Q1 = ("negative symptoms" AND (predom* OR "severe negative" OR "persistent negative" OR "prominent negative" OR prominent* OR "persisting negative" OR "enduring negative" OR "dominant negative" OR ("deficit syndrome")) AND (2002:2022[pdat])) (Please note that the character * was used as a wildcard.) The second query expression (Q2) that we used for the selection of randomized, double-blind placebo controlled (DBRPCT) studies was the following: Q2 = ("placebo" AND ("double-blind placebo-controlled" OR "double-blind controlled" OR ("double-blind")) AND (2002:2022[pdat])) The third query expression (Q3) was applied to select, respectively, studies with either an add-on (Q3a) or a monotherapy design (Q3m).
For the add-on studies, we applied the following expression: Q3a = ("add-on" OR "augmentation" OR "adjunctive" OR "supplement" OR "supplementation" OR "adjuvant" AND (2002:2022[pdat])) For the monotherapy studies, the search expression was the following: Q3m = NOT "add-on" NOT "augmentation" NOT "adjunctive" NOT "supplement" NOT "supplementation" NOT "adjuvant" Because all three conditions (negative symptoms, DBRPCT, and add-on or monotherapy setting) had to be met in order to be included in the analysis, for our final selection of the publications the three query conditions were connected by the logical operator AND using the following form.
For the add-on studies: Qadd-on = Q1 AND Q2 AND Q3a.
On the basis of the results of the automatic search, two of the authors (PC, BK) identified potentially relevant publications, and obtained the published reports for further evaluation. The same two authors independently evaluated the pre-selected articles for their eligibility for inclusion according to the previously defined inclusion criteria. In case any doubt arose, the uncertainties were resolved through a consensus discussion with the third member (IB) of the research group. In addition to the automatic search, we also checked the references of relevant publications for additional reports not identified by the database search. At the conclusion of the search process, the search results from the add-on and monotherapy queries were pooled into a single database. From the final pooled set of selected publications, we extracted the data onto standard, simple forms, and converted them into a Statistical Analysis System (SAS) 9.4 database.
Online Supplement, Part 2. Technical Details of the Statistical Analyses.

Statistical Analyses
The pooled effect size for the placebo effect (i.e., change from baseline to endpoint in the placebo arms) was estimated using the random-effect model for meta-analysis (Van Houwelingen, Arends, & Stijnen, 2002). The random-effect model was chosen because heterogeneity, regardless of statistical significance, is likely to be present in trials that have marked differences in design and study populations. The analysis under heterogeneity applied the DerSimonian and Laird (DerSimonian & Laird, 1986) normal mixture model, which considers each sample as an independent random sample from a normal population. First, the pooled effect size for the add-on and the monotherapy trials was estimated in the meta-analysis, which included trial design (add-on vs. monotherapy) analysis as a factor. Subsequently, the data from the two types of trials were pooled and examined jointly since the estimates from the add-on and monotherapy trials were essentially the same. Trial type was added as covariate in the pooled analyses. The meta-analytic effect-size estimates can be distorted by a tendency of not publishing negative or inconclusive studies in the literature. The possibility of publication bias in our meta-analysis was investigated on the basis of funnel plots using Begg & Mazumdar's rank correlation approach (Begg & Mazumdar, 1994); and the Egger test (Egger, Davey, Schneider, & Minder, 1997). In the Egger test, the standard normal ES is regressed on its precision, defined as the inverse of the standard error(Egger et al., 1997).
The meta-regression analysis was based on van Houwelingen et al.'s general linear mixed model (GLMM) approach using the approximate likelihood method (Van Houwelingen et al., 2002). In the GLMM, the effect size from each study was regressed on an intercept and on the study-level moderator variables from the individual studies (described in the previous subsection). The GLMM model incorporated both fixed and random effects. Based on the random effect part of the model, a common weighted statistical meta-regression effect-size estimate was calculated through the DerSimonian-Laird estimator (DerSimonian & Laird, 1986). The meta-regressions examined the moderators both in univariate and multivariate analyses. The meta-regression analysis was performed for all studies (i.e., for the add-on and monotherapy trials combined). A separate meta-regression by trial type was not performed since the number of monotherapy trials was too low (n=4). All analyses were conducted using the SAS statistical software With respect to the add-on studies, the detailed evaluation of the full reports resulted in the exclusion of 21 papers. The exclusions occurred due to ineligibility of the study sample (n=10); updated reports becoming available for the same study (n=5); subgroup analysis (n=1); issues with the study design (n=3); and other reasons (not a pharmacological trial, n=1; the negative symptoms [NG] measure was not used as a primary end point, n=1). These exclusions led to the use of 15 papers from the automatic search. Additionally, five more papers were added to this set. Specifically, one paper was added on the basis of manual search, while 4 others were included as a result of reclassification from the monotherapy set to the add-on set. In these cases of reclassification, none of the reports' respective key fields (titles, key words, abstract) explicitly indicated that the test treatment was applied as add-on medication (i.e., none of the relevant descriptors such as addon/augmentation/adjunctive/supplement/supplementation/adjuvant were included in the key fields); however, the reading of the text revealed this information. As a result of the above described selection process, we identified 20 papers which reported with data for a total of 21 add-on studies (i.e., 21 samples receiving placebo).
With respect to monotherapy studies, the detailed evaluation of the full text reports resulted in the inclusion of only 1 paper from the automatic selection (n=27). The reasons for not qualifying for inclusion were that the publication was a review paper/meta-analytic summary using aggregate data (n=6); the report focused on theoretical/methodological issues on negative symptoms (n=2); the study sample did not represent the target population of the current investigation (n=6); or improvement on negative symptoms was not used as a primary endpoint (n=8). Additionally, 4 of the 26 studies from this set were reclassified as an add-on trial (see above). Three additional articles were added to the monotherapy set as a result of manual search, hence the monotherapy set comprised 4 publications. Finally, it should also be noted that 2 of the 4 monotherapy studies ( ; however, the effect size for the placebo response was not determined in that meta-analysis for these two studies. The list below provides the set of publications that were considered for potential inclusion in the metaanalysis for the add-on and monotherapy trial set, respectively, based on the automatic (query) and manual search.

Final disposition status (Included/Not Included in the metaanalysis)for the MONOTHERAPY trials. In case of exclusion, the reason of exclusion is indicated
Online Supplement, Part 5. Description of the negative symptom measures used in each study & the computational details for the severity of negative symptoms at baseline.
Online Table 1  . For example, for the PANSS negative subscale, which has 7 items with an item-score range of 1 through 7, the complete absence of symptoms at baseline would be associated with a subscale score of 7. Thus, although the true baseline severity score fraction should be at 0% in this case, the uncorrected formula, used in the prior meta-analysis, yields 14.3% (=100*(7/49)), as it expresses baseline severity as a fraction of the maximum without adjusting for the scale's actual zero-point. The table below indicates (in column 9) whether the computation of baseline in the current meta-analysis was different from that of Fraguas et al.'s meta-analysis due to the aforementioned problem, or for any other reason.
Online Table 1. Measures of negative symptoms and computation of baseline severity for the studies included in the current meta-analysis 40.5 n/a n/a n/a Subtraction of the min. scale value of 7 needed, original paper used uncorrected PANSS subscale.
Notes: was needed to correct for the condition that the subscale's minimum value is non-zero (i.e., 7); this was not done in prior meta-analysis.
- : The SANS rating did not include inappropriate affect and subjective items while ATTN was included; adjustment was done only for one of these two changes in prior meta-analysis.
-Shoemaker, 2014: the composite score was computed based on items 1 through 22 as indicated in the original paper; this was overlooked in prior meta-analysis. -Usall, 2011: prior meta-analysis indicated that the baseline value was missing; it was, however, available in the original paper. -Weiser, 2012: the modified SANS was used, which did not include ATTN; this was overlooked in prior meta-analysis. : prior meta-analysis indicated that the baseline value was missing; for the current meta-analysis the value was determined from the pertinent figure in the original paper.
-Additional note for Umbricht, 2014: PANSS negative factor (i.e., Marder factor) score was used instead of the PANSS negative symptom subscale (which was indicated in prior meta-analysis).
Online Supplement, Part 6. Basic Demographical and Clinical Characteristics of the Study Sample Included in the Publications.
The patient population in the placebo arms of the included studies was chronically ill, with an average age of 44.0 years (SD=6.9) and a mean duration of illness of 20.1 years (SD=8.1). Females were underrepresented in the samples (mean female %=39.6, SD=26.2), especially if we consider the median value (Q50=34.2% females). In terms of baseline severity, the study samples were located in average somewhat higher than the midpoint in the range between the minimum and maximum of the negative symptom scale used in the study (mean=53.4%, SD=9.5%). The SANS scale for the negative symptom rating was applied 10 of 25 studies, while in 12 and 3 of the included studies the PANSS and NSA-16 scales were used, respectively. Of the 12 studies that applied the PANSS, 7 used the negative symptom subscale, while 5 used a negative symptom factor (4 employed the Marder factor; and 1 employed the negative factor from the PANSS's pentagonal factor structure). Details regarding the computation of the baseline severity of the negative symptoms for each study and the further description of the negative symptom measures used in the individual studies included in the meta-analysis are summarized above in Online Supplementary Table 1.