Randomized clinical trials are the principal method by which researchers assess treatment efficacy. Although clinical trials can provide valid evidence of a treatment's average effect (relative to some control), they often fail to demonstrate meaningful drug effects relative to placebo. Some researchers have cited high “placebo response” as the main cause of these “failures” (1–4), suggesting it may be prudent to exclude “high placebo responders” prior to trial commencement.

Researchers have sought out correlates to predict pain relief in the placebo-treated group, which could then be used to exclude patients who would contribute to clinical trial “failure” via their “placebo response.” One of these identified correlates is pain variability, which has been shown to negatively correlate with subsequent pain relief in the placebo-treated group across several chronic pain conditions (5–7). In other words, patients with the greatest pain variability at baseline tend to have the greatest decreases in pain following placebo administration. This relationship is specific to pain relief in the placebo group in some (6) but not all studies (7), and even absent in others (8).

Although previous work demonstrates a relationship between baseline pain variability and pain relief, other factors such as regression to the mean and natural history can contribute to improvements in pain reports (6, 9). Indeed, previous studies have acknowledged, but have not accounted for, the influence of these factors on decreases in pain reports. In this work, we aim to improve our understanding of the prognostic value of baseline pain variability by adjusting for baseline pain, natural history, and regression to the mean. Since baseline variability is simple to collect and calculate [cf. neuroimaging and genetic traits that are also correlated with greater pain relief following placebo; e.g., (10, 11)], its prognostic value and (placebo-) specificity could be easily exploited in both trials and practice.

In total, 139 subjects were examined (63 subjects in Placebo I; 76 subjects in Placebo II). Figure 1 depicts the independent relationship between baseline pain variability and relief for each group after adjusting for pre-intervention pain, allowing each group to have a different effect of baseline pain variability. Model parameters and semi-partial correlations associated with adjusted group effects can be found in Table 2. Including SD_baseline in the models as a linear (not interactive) predictor increased Placebo I and Placebo II model R²'s by 0.01. Our observed effects were not strongly influenced by any individual patient (Supplementary Figure 1).

The purpose of this study was to quantify and isolate the relationship between baseline pain variability and post-intervention pain by group in two randomized, placebo-controlled clinical trials. Our work extends that of previous research by adjusting for the effects of regression to the mean and natural history via a no treatment control group (15–17). By assuming that pain relief is a linear combination of natural history, regression to the mean, the placebo effect, and the drug effect,¹ we estimated the placebo- and drug-specific effects of SD_baseline on post-intervention pain.

Contrary to previous work, we observed negligible correlations in our primary model, with SD_baseline capturing ≲ 4% of the variance in post-intervention pain across groups in both studies. Of principal interest was the placebo-specific effect, which previous studies suggest is on the order of r ≈ −0.3 (6, 7). After adjusting for the no treatment group and pre-intervention pain, our placebo-specific estimates were incompatible with these previous estimates (Table 1). However, our results are consistent with the recent findings of Gillving et al. (8), who observed negligible correlations between variability and improvements in patients who received placebo. Together, these results suggest that SD_baseline may not be a strong, consistent, and “placebo-specific” predictor of pain relief across populations.

The magnitude and precision of our estimates were sensitive to the modeling strategy. When modeling the groups separately, our effect estimates were larger and had greater variance, especially in Placebo II (Supplementary Table 1; Supplementary Figure 2). Thus, modeling the groups separately produced CIs that encompass previously reported estimates, but our point estimates were still relatively small and did not favor the placebo group. These results are suggestive that modeling differences may partly explain the discrepancy between studies. Similarly, differences in populations and sample sizes are important factors to consider (5–8).

Studies validating a SD_baseline-based prediction model are lacking. Nevertheless, the utility of SD_baseline for trial exclusion is dubious. Even if SD_baseline or some other variable was strongly predictive of pain relief following placebo, the removal of so-called “placebo responders” would also affect the active treatment group, especially since “placebo effects” are thought to be one component of active treatment effects. Finally, although removing “placebo responders” would theoretically improve treatment effect estimates, the observed treatment effect for such a study would answer a different question since the sample is conditioned on SD_baseline. This may result in an optimistic, ecologically questionable estimate that may be unlikely to translate to the clinic.

Rather than trying to optimize treatment effect estimates in trials using peculiar exclusion criteria, researchers should optimize treatments and thus their effect estimates for the ecological patient population. After all, the goal of research is not to find large effects—it is to find large effects that will successfully translate and improve lives. Notwithstanding the limitations of conditioning on SD_baseline for trial inclusion, there are two salient issues to note concerning SD_baseline as a prognostic variable. First, since we and others (7, 8) did not observe consistent or placebo-specific effects, the strength and consistency of previously reported placebo-specific results deserves greater scrutiny and, hopefully, reconciliation (5, 6). Second, although multiple studies have observed relationships between baseline pain variability and pain relief, its prognostic value has only been estimated, not validated (19). Neither of these two points preclude SD_baseline from having value, however. SD_baseline's ability to capture variance in trial endpoints (5–7) indicates it may be useful to include as a covariate in trial analyses, which may serve to improve statistical efficiency and estimates of treatment effects (20).

The original contributions presented in the study are included in the article/Supplementary Materials, further inquiries can be directed to the corresponding author.

The studies involving human participants were reviewed and approved by Northwestern University IRB. The patients/participants provided their written informed consent to participate in this study.

ST, AV, and AA conceived the research question. AA supplied data and resources. ST and AV performed analyses, drafted the manuscript, and created figures and tables. All authors read, edited, and approved the final version of the manuscript before submitting.

This work was funded by the National Institutes of Health (P50DA044121-01A1). This material is based upon work supported by the National Science Foundation Graduate Research Fellowship Under Grant No. DGE-1324585.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.