What clinicians should know about differential placebo effects

In order to appreciate the clinical impact of placebos, it is essential to understand the concepts of overall treatment effect and specific treatment effect. Overall treatment effect refers to the total benefit a patient experiences from a treatment, including associated placebo effects. Specific effect refers to the treatment benefit that extends beyond placebo effects. The specific effect is essentially the overall effect minus placebo effects, and it is representative of the treatment's intended therapeutic action. For example, if we have a treatment with an effect size of 5 points and a placebo with an effect size of 2 points, the specific effect of this treatment is 5 minus 2, which is 3 points, whereas the overall effect of the same treatment is 5 points (Figure 1). In real-world clinical practice, all observed effects are overall treatment effects.

Participant expectation and regression to the mean are methodological concerns that may impact the magnitude of the placebo effects observed in individual trials. Participant expectation, a key mechanism of placebo effect, is a psychological phenomenon which relates to a trial participant's assumptions of a treatment's effect [4]. Participants enrolled in clinical trials tend to disproportionately recall positive results as treatment progresses due to their expectations of the treatment's effect, regardless of receiving active treatment or placebo, while simultaneously downplaying negative effects or not acknowledging a complete lack of effect [5]. Often described as Hawthorne effects, the overestimation of intervention effects due to clinical trial participation is one of many manifestations of patient expectation [6]. The manner in which investigators introduce treatments to participants at the initiation of a clinical trial, the characteristics of active treatments, and the measures taken to blind patients to treatment allocation heavily influence participant expectation; these qualities can vary widely between trials. Expectation on the part of participants can ultimately obscure or amplify the specific effects of active treatments and the effects of placebo interventions [7]. When synthesizing the results of multiple clinical trials to ascertain a comprehensive treatment effect, it is important for researchers to bear the impact of patient's expectation in mind. Similarly, it is essential for clinicians to be aware of the inherent patient expectation associated with clinical trial participation and its resulting impact on observed treatment effects when applying clinical trial data to real-world treatment decisions. Oftentimes, the efficacy of treatments does not entirely translate into clinical practice, as the expectation of a new drug effect which is observed in a clinical trial is no longer at play.

Regression to the mean refers to a phenomenon in which outcomes that produce extreme values upon the baseline measurement are statistically more likely to produce values which fall closer to the mean of the sample upon the final measurement [8]. Sample and outcome characteristics can influence the extent to which values will regress to the mean, although control over these characteristics is limited. Regression to the mean has been erroneously attributed, according to some researchers, to the placebo effect, particularly when measuring subjective pain [9]. Under this assumption, the placebo effect observed in clinical trials may actually be significantly smaller than it appears to be due to regression to the mean and other nonspecific effects. The ‘true placebo effect’ is defined as the effect that remains after nonspecific effects have been subtracted from the ‘perceived placebo effect’ (Figure 2) [10].

In light of the methodological challenges described, it would be ideal to examine the effects of different placebo interventions in the most direct way possible, either in comparison with each other or against an untreated study arm in a randomized controlled trial (Figure 3A). However, due to ethical concerns, these types of trials are extremely rare. Consequently, conducting a traditional meta-analysis using direct comparisons is not an option, and indirect comparison methods such as network meta-analysis must be undertaken. In a traditional meta-analysis, two treatments are compared against each other. Network meta-analysis is an extension of traditional meta-analysis that compares multiple interventions, even in the absence of randomized trials which directly compare these interventions with each other (Figure 3B). In doing so, network meta-analysis allows researchers to estimate relative effects of all available treatments and to develop a hierarchy of treatments based on their effect sizes [11,12]. Our own research in this field explored the strength of the placebo effect in a network of clinical trials in patients with knee osteoarthritis (OA) [13]. Knee OA was ideal for our study due to its high prevalence, well-defined diagnostic criteria, validated outcome measures and extensive clinical research involving a variety of therapeutic approaches. One of the key assumptions of a network meta-analysis is that all included trials should share similar characteristics, and the characteristics of knee OA trials in particular provided us with an excellent opportunity for studying the placebo effect and assessing the different impacts various types of placebos may have on pain [11,12].

In a network meta-analysis comparing many interventions involving different types of placebo comparators, the assumption is often made that various placebos can be categorized as a single group on the basis of having the same null effects. Combining placebos that differ in route of administration or in other characteristics could be problematic if they are not completely ineffectual and if variations in efficacy do exist between the different types. The differential placebo effect leads to a phenomenon called ‘efficacy paradox’, a concept that has the potential to influence clinicians’ decision-making in real-world clinical practice. The efficacy paradox refers to a scenario in which a treatment whose overall effect appears moderately superior to a placebo counterpart may be more efficacious than a different treatment which shows a greater superiority over its placebo, if the placebo effects of the two treatments vary (Figure 4) [14]. This scenario can be seen when treatments with different routes of administration are compared. In clinical practice, choosing a more effective oral treatment by examining randomized trials that compare separate oral treatments against a similar oral placebo is reasonable. However, if choosing between a topical and an oral treatment, it would be erroneous to base the clinical decision on separate trials comparing these treatments against their respective placebos. For example, an oral treatment with an overall effect size of 5 points compared against an oral placebo with an effect size of 2 points will have a specific effect of 3 points. A topical treatment with an overall effect size of 6 points compared against a topical placebo with an effect size of 4 points will have a specific effect of 2 points. Thus, even though the specific effect of the topical treatment is inferior to that of the oral treatment, a patient would perceive a greater overall benefit (6 vs 5 points) using the topical preparation (Figure 4).

Most of the published meta-analytic studies that measure the strength of placebo effects do so by employing a single-group analytic design. Single-group analysis breaks randomization by selecting only treatment arms of interest from each trial, making it difficult to control for participant expectation and/or regression to the mean and to segregate them from the perceived placebo effect. In utilizing network meta-analysis methodology, we are able to preserve randomization by analyzing both active and placebo groups in all of the included studies. Maintaining randomization allows us to account for regression to the mean and other nonspecific effects and to negate them from the overall placebo effect, which in turn allows us to measure the ‘true placebo effect’.

In this study, we compared placebo treatments of varying routes of administration among clinical trials specifically focused on knee OA patients [13]. We included only randomized controlled trials of adults with knee osteoarthritis which compared leading pharmaceutical products against oral, intra-articular or topical placebos. We applied limitations to both the population and to the placebo administration types included in order to reduce heterogeneity in our sample. We included 149 trials including 39,814 participants and performed network meta-analysis using a Bayesian random effects model. Additionally, we performed another network meta-analysis in which we assumed equivalency of all placebos and compared the results of both network meta-analyses in order to assess the impact of ignoring differential placebo effects. We calculated the standardized mean differences between placebo types using the changes in pain scores from baseline to 3 months. Our analysis provided strong evidence toward the existence of differential placebo effects attributable to treatment administration route. It was shown that intra-articular and topical placebos were more effective when compared with oral placebo, with effect sizes of 0.29 (95% CI: 0.09–0.49) and 0.20 (95% CI: 0.20–0.38), respectively.

The comparison between intra-articular corticosteroids (IACS) and various placebo administration types clearly illustrated differential placebo effects. Against their typical intra-articular placebo comparator, IACS demonstrated an effect size of 0.29 (95% CI: 0.15–0.44). When compared with topical placebo, the effect size of IACS increased to 0.38 (95% CI: 0.09–0.67), and in comparison with oral placebo, the original effect size was nearly doubled (0.58 [95% CI: 0.34–0.82]). Conversely, we found that comparison of active treatments with matching placebos can produce positive effects for some active treatments which are diminished when compared with a placebo that has a different route of administration. When compared with oral placebo, oral celecoxib (COX-2 inhibitor) showed a statistically significant benefit of 0.34 (95% CI: 0.27–0.42). The effect size decreased dramatically, however, and was no longer statistically significant, when celecoxib was compared with topical placebo (0.14 [95% CI: -0.04–0.32]) or intra-articular placebo (0.05 [95% CI: -0.15–0.32]). These results provide an example of the efficacy paradox, demonstrating that comparing active treatments with placebo treatments can potentially lead to a misinterpretation of a treatment's specific effect, particularly if one or both of the treatments involves a route of administration associated with strong placebo effects (Figure 5). In addition to the evidence from this network meta-analysis, differential efficacy of placebo interventions has also been demonstrated in previously published meta-analyses, as well as in direct comparisons in prior clinical trials [15–18]. These trials spanned a variety of patient populations and included an assortment of placebo procedures.

One of the limitations in employing a network meta-analysis method for studying placebo effects is the scarcity of direct evidence available for comparisons between different placebo interventions. Our study, for example, had only one randomized trial of oral versus topical placebo among 149 randomized controlled trials [13,19]. The results from network meta-analyses should be interpreted with caution, and the amount of indirect evidence contributing to the analysis should be carefully considered. Despite the limitations, network meta-analysis continues to be the best possible method for differentiating between placebo interventions in the absence of direct comparisons by a randomized trial. Though indirect treatment comparisons are subject to bias, carefully conducted indirect treatment comparisons can, in tandem with the existing body of evidence provided by randomized controlled trials, provide a more detailed and informative profile of treatment effects and their relationship to placebo effects [12].