Badly chosen outcomes

Surrogate outcomes Surrogate markers are often used to infer or predict a more direct patient-oriented outcome, such as death or functional capacity. Such outcomes are popular because they are often cheaper to measure and because changes may emerge faster than the real clinical outcome of interest. This can be a valid approach when the surrogate marker has a strong association with the real outcome of interest. For example, intra-ocular pressure in glaucoma and blood pressure in cardiovascular disease are well-established markers. However, for many surrogates, such as glycated haemoglobin, bone mineral density and prostate-specific antigen, there are considerable doubts about their correlation with disease [4]. Caution is therefore required in their interpretation [5]. Authors of an analysis of 626 randomised controlled trials (RCTs) reported that 17% of trials used a surrogate primary outcome, but only one-third discussed their validity [6]. Surrogates generally provide less direct relevant evidence than studies using patient-relevant outcomes [5, 7], and over-interpretation runs the risk of incorrect interpretations because changes may not reflect important changes in outcomes [8]. As an example, researchers in a well-conducted clinical trial of the diabetes drug rosiglitazone reported that it effectively lowered blood glucose (a surrogate) [9]; however, the drug was subsequently withdrawn in the European Union because of increased cardiovascular events, the patient-relevant outcome [10].

Composite outcomes The use of combination measures is highly prevalent in, for example, cardiovascular research. However, their use can often lead to exaggerated estimates of treatment effects or render a trial report uninterpretable. Authors of an analysis of 242 cardiovascular RCTs, published in six high-impact medical journals, found that in 47% of the trials, researchers reported a composite outcome [11]. Authors of a further review of 40 trials, published in 2008, found that composites often had little justification for their choice [12], were inconsistently defined, and often the outcome combinations did not make clinical sense [13]. Individual outcomes within a composite can vary in the severity of their effects, which may be misleading when the most important outcomes, such as death, make relatively little contribution to the overall outcome measure [14]. Having more event data by using a composite does allow more precise outcome estimation. Interpretation, however, is particularly problematic when data are missing. Authors of an analysis of 51 rheumatoid arthritis RCTs reported >20% data was missing for the composite primary outcomes in 39% of the trials [15]. Missing data often requires imputation; however, the optimal method to address this remains unknown [15].

Subjective outcomes Where an observer exercises judgment while assessing an event, or where the outcome is self-reported, the outcome is considered subjective [16]. In trials with such outcomes, effects are often exaggerated, particularly when methodological biases occur (i.e., when outcome assessors are not blinded) [17, 18]. In a systematic review of observer bias, non-blinded outcome assessors exaggerated ORs in RCTs by 36% compared with blinded assessors [19]. In addition, trials with inadequate or unclear sequence generation also biased estimates when outcomes were subjective [20]. Yet, despite these shortcomings, subjective outcomes are highly prevalent in trials as well as systematic reviews: In a study of 43 systematic reviews of drug interventions, researchers reported the primary outcome was objective in only 38% of the pooled analyses [21].

Complex scales Combinations of symptoms and signs can be used to form outcome scales, which can also prove to be problematic. A review of 300 trials from the Cochrane Schizophrenia Group’s register revealed that trials were more likely to be positive when unpublished and unreliable and non-validated scales were used [22]. Furthermore, changes to the measurement scale used during the trial (a form of outcome switching) was one of the possible causes for the high number of results favouring new rheumatoid arthritis drugs [23]. Clinical trials require rating scales that are rigorous, but this is difficult to achieve [24]. Moreover, patients want to know the extent to which they are free of a symptom or a sign, more so than the mean change in a score.