Statistics from Altmetric.com
Lack of generalisability limits use.
From a GP’s perspective, the accompanying article by Janine Janosky on the single subject design is both interesting and stimulating.1 This type of design, as Dr Janosky highlights, is infrequently used in research and has some potential advantages. Most notably, it is the only type of design that can provide information about effects at an individual level. There are obvious benefits in formalising what all GPs do on a day to day basis, namely observing the effects of individual treatments on individual patients. However, the article suggests a scope and potential for the n = 1 trial that I would take issue with, and the author fails to adequately describe the limits and disadvantages of this type of design.
While single subject designs have the potential of examining effects at an individual level, they do not provide data that can readily be applied to others. The author does mention that the generalisability of results from this type of study is limited, but goes on to suggest that if a subject that is “representative of the general type of patients for which this intervention would be used” then the results become more generalisable. A person can be chosen that has a certain disease at a certain stage and with certain sociodemographic characteristics. But is this person really representative? How do we know exactly which variables are relevant to the effect being shown? And how can we judge “biological representation”. For example, are there aspects of a person’s metabolism that influences the way they respond to a drug? Furthermore, demonstrating an effect in a person, even if the person is similar to a population, provides little evidence about the probable effects of an intervention on a population, or the effects on another, inevitably unique, person.
Other potential problems with single subject trials are problems of bias and the determination of statistical significance. Effects that are likely to lead to bias in this type of trial include regression to the mean and “carry-over” effects. Values towards the extremes are likely to normalise on repetition for statistical reasons, an effect that is described as regression to the mean. This will occur without any clinical change in the subject, and in a single subject trial is likely to lead to a false impression of treatment effect when none may be present. Secondly, the author describes using a washout period between different phases of treatment. However, treatments can have lasting effects and it can be difficult to distinguish whether any prior treatments or indeed the rotation of treatments plays a part in any observed effect. Furthermore, once a subject has “changed”, how can you ever really know what you are comparing to? Can change revert back to its original state in every way? After all, you can only ever be a virgin once!
Ways of attempting to minimise bias in n = 1 trials, as the author points out, include randomisation, blinding, and multiple treatment phases. An effect that is observed by a blinded observer, in a blinded patient, each time a treatment is randomly introduced, and that disappears each time the treatment is withdrawn, is more likely to represent a true treatment effect. However, trials that entail randomisation, blinding, and multiple phases are likely to be more difficult to implement in primary care, and more costly. Larson et al estimate the costs of formal n = 1 trials at $450 to $500 per patient.2 This is much less than large scale RCTs, but still a significant cost for a primary care clinician interested in evaluating a treatment on a patient. Furthermore, if the single subject trial is repeated in a number of patients to try and prove generalisability, the costs could easily add up to the sort of amount that a small scale RCT would cost. A series of n = 1 trials combined could thus be seen as a long, drawn out, “sequential RCT”. Some comments from the author on the feasibility and methods for combining the results of multiple single subject trials, and the applicable lessons that have been learnt from cross over trials, would prove very useful for those interested in conducting studies of this kind.
The working example given by the author shows fasting blood glucose values before and after a comprehensive intervention for diabetes management. This is a two phase, A-B, design that shows fasting blood glucose values after the intervention as being lower than those before the intervention. The author concludes that, “it seems that the intervention was effective in lowering the measured fasting blood glucose in this subject.” Is this a fair conclusion from this type of study? Is it not possible that some other, unmeasured variable changed during the course of this study, and that this was the cause of the change in measured fasting blood glucose? For example, the subject may have independently taken up a new form of exercise.
In conclusion, Janosky’s paper on single subject trials has relevance to both GP clinicians and researchers. In primary care, making treatment decisions with patients and then monitoring their response to those treatments is a daily occurrence. This, on occasion, entails a number of phases, where either different treatments or no treatment are tried. Formalising this process, and consideration of the use of blinding or placebo treatments, provides opportunities for primary care clinicians to assess the effectiveness of treatments on a person in a more comprehensive and rigorous fashion. Conducting formal n = 1 studies, compared with the informal treatment trial, can however, have significant time, cost, and ethical considerations. Given these constraints, it is unlikely that this type of trial will have wide applicability in everyday general practice. Furthermore, it must be emphasised that the results from this type of study can generally not be applied to anyone other than the person that has been studied, and that this design is particularly vulnerable to certain types of bias as described above.
I am pleased to read the commentaries that accompany this article. Each of the commentaries raises relevant issues that serve to illuminate the strengths and limitations of single subject research designs. As with all methodological designs, each affords strengths and limitations to answer tailored research questions. The research questions best answered, through the application of a single subject research design, are questions that garner their interest in the potential to have high internal validity while acknowledging that the external validity is most probably weak. The single subject design can be easily contrasted with a clinical trial or true randomised experimental study that aims to examine effectiveness where the strength of the design is its strong external validity. The results of these clinical trials provide conclusions regarding effectiveness on the average treatment effects for only the studied populations; however at times, the results from these clinical trials might not always be applicable in determining the most effective treatment for an individual patient. A single subject research design, although limited for external validity, provides an opportunity to examine the applicability of the study findings to a specific patient. As methodologists, having an arsenal of research designs with strengths and limitations identified for each affords us a comprehensive means to answer diverse research questions.
Lack of generalisability limits use.