|Home | About | Journals | Submit | Contact Us | Français|
There is considerable clinician and researcher interest in whether the outcomes for patients with low back pain, and the efficiency of the health systems that treat them, can be improved by 'subgrouping research'. Subgrouping research seeks to identify subgroups of people who have clinically important distinctions in their treatment needs or prognoses. Due to a proliferation of research methods and variability in how subgrouping results are interpreted, it is timely to open discussion regarding a conceptual framework for the research designs and statistical methods available for subgrouping studies (a method framework). The aims of this debate article are: (1) to present a method framework to inform the design and evaluation of subgrouping research in low back pain, (2) to describe method options when investigating prognostic effects or subgroup treatment effects, and (3) to discuss the strengths and limitations of research methods suitable for the hypothesis-setting phase of subgroup studies.
The proposed method framework proposes six phases for studies of subgroups: studies of assessment methods, hypothesis-setting studies, hypothesis-testing studies, narrow validation studies, broad validation studies, and impact analysis studies. This framework extends and relabels a classification system previously proposed by McGinn et al (2000) as suitable for studies of clinical prediction rules. This extended classification, and its descriptive terms, explicitly anchor research findings to the type of evidence each provides. The inclusive nature of the framework invites appropriate consideration of the results of diverse research designs. Method pathways are described for studies designed to test and quantify prognostic effects or subgroup treatment effects, and examples are discussed. The proposed method framework is presented as a roadmap for conversation amongst researchers and clinicians who plan, stage and perform subgrouping research.
This article proposes a research method framework for studies of subgroups in low back pain. Research designs and statistical methods appropriate for sequential phases in this research are discussed, with an emphasis on those suitable for hypothesis-setting studies of subgroups of people seeking care.
Several authors [1,2] have argued that low back pain is most accurately classified as pain associated with serious pathology, pain associated with nerve compression, or non-specific low back pain (NSLBP). Under this approach, approximately 80% of low back pain in primary care is classified as NSLBP  and investigations into treatment efficacy for this condition have identified only moderate treatment effects.
However, most clinicians [3,4] and researchers  believe NSLBP to be a number of conditions, and subgrouping NSLBP is currently of clinical and research interest [5-10]. This interest is premised on the notion that patient outcomes might be improved with more precise targeting of treatment, and health system efficiency might be improved with more effective triage of patients.
Many NSLBP subgrouping systems have been proposed. Some aim to identify people whose pain is associated with a particular pathoanatomical condition, based on their presenting symptoms and signs (diagnostic subgroups) [11,12]. Other systems aim to identify people likely to respond favourably to particular treatment regimens (treatment effect modifier subgroups) [13-15], while other systems aim to identify people with particular prognoses (prognostic factor subgroups) - such as those at risk of chronicity . While there is no shortage of opinions about the composition of clinically important NSLBP subgroups, there is very little consensus regarding the symptoms and signs that identify these subgroups .
Subgrouping studies have previously been classified into three broad stages of research: exploratory studies that seek to identify subgroups, studies that attempt to validate subgroups and studies that test the capacity of subgrouping to positively influence routine clinical care . The research designs and statistical methods appropriate for subgrouping studies vary depending on whether the aim of the subgrouping is prognostic, therapeutic or diagnostic, and also vary depending on the stage of the research.
Subgrouping research is fraught with methodological pitfalls and many authors have described reasons for caution in the conduct, interpretation and reporting of such studies [18-22]. In this context, there has been a proliferation of subgrouping studies in NSLBP, most of which have been hypothesis-setting and they report highly variable methods. Even among studies that report similar methods, their authors may have different opinions about the level of evidence these studies are capable of providing. Therefore, in subgrouping research not only is methodological rigor very important but there is also a need for an accepted method framework in which to classify, evaluate and discuss this research with a common vocabulary.
The aims of this debate article are to present a method framework for conducting and evaluating subgrouping research in low back pain, and to discuss the strengths and limitations of research methods suitable for hypothesis-setting studies. The focus of the article is on research method and where appropriate, examples of studies are used to illustrate concepts. However, this article is not a review of the findings of subgrouping research and other examples of studies may have been equally appropriate.
In this proposal we adopt earlier recommendations in distinguishing between prognostic factors and treatment effect modifiers [23,24]. Prognostic factors are symptoms, signs or other characteristics that indicate likely outcomes regardless of treatment. Treatment effect modifiers are symptoms, signs or characteristics that indicate likely response to a specific treatment (a subgroup treatment effect). This distinction has important implications for the methods suitable for research of subgroups. Patient outcomes are usually the product of a combination of treatment effects and prognostic factor effects, unless the treatment is completely ineffective. Therefore, studies need to use particular designs if these effects are to be teased apart. It has been reported that this distinction is commonly misunderstood .
The effect of prognostic factors can be studied in data from cohort studies ('single-group' designs) of usual care. Usual care implies that treatments are various, uncontrolled and reflective of common practice. Under these circumstances, it is assumed that the heterogeneity of treatments washes out specific treatment modifier effects. In contrast and at the other extreme, when predictors of outcome are investigated in cohort studies in which all participants receive only one treatment, it is not possible to differentiate between which factors predictive of outcome are prognostic factors and which are effect modifiers specific to that treatment .
Prognostic factors have also been studied using data from randomised controlled trials ('two-group' designs). One approach is to study predictive factors only in a control group that received either placebo care or usual care. Conceptually, this is similar to a prospective cohort study. Another approach is to study (as a single group) the whole cohort from a trial that showed no differences in outcome for the experimental and the control treatments. Where the control treatment was not placebo or no treatment, this latter approach is problematic if the treatments had a clinical effect, as the predictive factors may contain treatment effect modifiers common to both treatments. The generalisability of findings from studies of prognostic factors are always limited by the selection criteria of the study, and clinical trials tend to have more restrictive inclusion criteria than cohort studies. Contemporary summaries are available of key methodological issues for cohort studies of prognostic factors [18,20,25,26].
In contrast, the precise measurement of treatment effect modifiers requires data from randomised controlled trials . The appropriate trial design varies depending on the type of research question being investigated. As precise identification of the presence of treatment effect modification requires a test of subgroup/treatment effect interaction, currently only two designs for controlled trials are well suited for measuring treatment modifier effects [23,28,29]. Examples of such studies are those by Childs et al (2004)  and Brennan et al (2006) . There are a number of concise summaries available for readers who are seeking greater detail on methodological issues for randomised controlled trials in which treatment effect subgroup analysis is planned [19,21-24,30].
Whether used in the investigation of prognostic factors or treatment effect modifiers, many statistical techniques produce measures of association that can be difficult for clinicians to apply to individual patients. In response to this, clinical prediction rules are increasingly being used as a means to express the likely response of a subgroup in clinically interpretable ways. They also allow the accuracy of this predictive capacity to be described . A method for forming a prediction rule is detailed in Additional File 1.
The proposed method framework classifies subgrouping studies into six phases of research: studies of assessment methods, hypothesis-setting studies, hypothesis-testing studies, narrow validation studies, broad validation studies, and impact analysis studies. These are defined in Figure Figure1.1. This framework extends descriptive terms previously suggested by McGinn et al (2000)  as suitable for classifying studies of clinical prediction rules. It does so by adding an initial phase of 'studies of assessment methods', and by splitting the process described by McGinn as 'derivation studies' into two phases: hypothesis-setting studies and hypothesis-testing studies.
Extending and modifying these phases in the proposed framework was undertaken to achieve a number of purposes. The first purpose was to allow categorisation of an increasing number of studies that seek to devise measures of subgroup-specific characteristics, especially measures of physical impairment. The second purpose was to better describe the quality of evidence provided by studies in the pre-validation phases, in recognition that authors were interpreting that quality in different and contradictory ways. The third purpose was to broaden the framework to include subgrouping studies that do not express findings using clinical prediction rules.
Within the proposed method framework, the first phase in subgrouping research comprises studies that attempt to create or improve tools that assess clinical characteristics potentially indicative of subgroup membership, or to determine the measurement properties (clinimetrics) of those tools. Guidance is available on suitable research designs and statistical methods to perform such studies [32-36]. An example of a study of a novel tool for assessing potential subgroup membership is Ferreira et al (2004), which investigated an ultrasound test to measure, in clinical settings, the automatic recruitment of trunk muscles in people with low back pain.
The second phase (hypothesis-setting) is represented by studies that attempt to determine which characteristics identify people in clinically important subgroups, and the magnitude of any prognostic effects or treatment effect modification attributable to these subgroups. Other authors have argued that treatment effect modifiers can only be determined in randomised controlled trials and that the probability (p value) of treatment responses associated with specific subgroups should be adjusted to reflect multiple statistical comparisons. We suggest that within a hypothesis-generating phase, these criteria can be relaxed and instead applied later during rigorous hypothesis-testing studies. For example, we suggest that during this exploratory phase, data from cohort studies may generate useful hypotheses about potential treatment effect modifiers (for example Flynn et al 2004), and that it is permissible to perform post-hoc multiple comparisons without Bonferroni-type corrections.
The third phase (hypothesis-testing) in subgrouping research comprises studies that test pre-specified (a priori) hypotheses about subgrouping effects in samples of people independent from but similar to those people who participated in the hypothesis-setting phase. During this confirmatory phase there is a need for the rigorous testing of only pre-specified hypotheses and for appropriate statistical adjustment for multiple comparisons. We believe that this distinction between hypothesis-setting and hypothesis-testing studies would reconcile differences in interpretation as to the quality of evidence of subgroup effect that particular studies provide.
Replication of prognostic effects or treatment effect modifion in an independent sample under stringent research conditions is the central aim of hypothesis-testing studies and is a method of external validation. The chance of spurious, sample-specific effects or associations in hypothesis-setting studies is so high in subgrouping research, that Rothwell (2005)  suggests that the best test of the validity of subgroups is not significance testing but replication in an independent sample. Quasi-replication within the hypothesis-setting stage by use of iterative statistical techniques (such as boot-strapping) is an inadequate substitute for replication in an independent sample [17,38]. This is because the repeated testing of findings on sub-samples of the original data only partially counters problems associated with sample-specific relationships between predictors and outcomes, as they are test a relationship in the cohort in which the relationship was first established.
The fourth phase (narrow validation) comprises studies that attempt to validate the findings of hypothesis-testing studies in samples of people who are independent from, but similar to, those who participated in the hypothesis-testing phase. Such studies provide insight into the variability of a subgroup effect in the target population .
The fifth phase (broad validation) comprises studies that test the findings of hypothesis-testing studies in samples of people who differ from those who previously participated. The clinical characteristics of these samples of patients may differ on dimensions such as the spectrum of the disorder, demographic and psychosocial profile, culture and language, co-morbidities and care settings (primary/secondary/tertiary care). Similarly, the experience, training and professional discipline of the clinicians may vary from those providing care in earlier studies. Broad validation establishes the generalisability of the subgroup findings beyond the clinical and professional profile of the people originally studied .
The sixth and last phase (impact analysis) comprises studies that seek to establish the feasibility of uptake of the subgrouping scheme in practice and the capacity of subgrouping to improve outcomes in routine clinical care. Subgroups that have been shown to have predictive capacity in hypothesis-testing and validation studies may still not be effectively implemented in routine care, due to issues such as perceived importance by clinicians, patient-perceived acceptability and the practicality of assessing predictor variables [26,39].
Most of the subgrouping studies in low back pain have been hypothesis-setting and they have used highly variable research designs and statistical methods. We have classified these designs and methods into three categories. The purpose of the categories is to clarify suitable method pathways in hypothesis-setting studies, although it is possible that studies may exist that contain elements from more than one pathway.
The first of these categories is opinion-based subgroups that originate from clinical observation. An example is the McKenzie subgroup of patients who display a directional preference , which was initially based on an astute clinical observation that some people display pain that responds to particular movements.
The second category is subgroups based on physiological/psychosocial constructs that are derived from experimental observation. An example is O'Sullivan's 'Mechanism-based classification' .
We argue that within the hypothesis-setting phase, opinion-based subgroups and subgroups based on experimental constructs need to be formally tested for treatment modification effects using appropriately designed randomised controlled trials, and/or tested for prognostic effects using a cohort study design. Method pathways for opinion-based subgroups and subgroups based on physiological/psychosocial experimental constructs are shown in Figure Figure22.
The third category of designs and methods used in hypothesis-setting studies is 'data-driven' subgroup analysis, where data from cohort studies or randomised controlled trials are investigated using cross-sectional statistical analysis, or investigated using longitudinal statistical analysis. These forms of analysis are called 'data-driven' because a subgroup is being formed retrospectively (post-hoc) from the characteristics of the sample data , rather than on clinical observation or a physiological/psychosocial experimental model. Method pathways for data-driven subgroups are shown in Figure Figure33.
Data-driven subgroups can be identified in two ways: either in relation to an outcome or by identifying variables that are associated with each other without regard to an outcome. Statisticians call these two main classes of statistical approaches 'supervised' techniques and 'unsupervised' techniques respectively. Both have methodological advantages and disadvantages. The first main class of these statistical techniques are called 'supervised' because these techniques work backwards from an outcome in longitudinal data, such as people classified as responders or non-responders to a treatment regimen. Examples of supervised statistical techniques include regression analysis, discriminant function analysis, recursive partitioning analysis, and classification and regression trees. The other main class of statistical techniques used in data-driven subgrouping is called 'unsupervised' because these techniques do not work backwards from an outcome but instead look for relationships between measurable characteristics inherent in cross-sectional data. Examples of unsupervised statistical techniques include cluster analysis, data-mining and neural networks.
A sub-class of statistical techniques are known by statisticians as 'clinimetric' techniques. These techniques include sensitivity, specificity, likelihood ratios, odds ratios, risk ratios and pre- and post-test probability. In hypothesis-setting studies of subgroups of low back pain, one set of circumstances where these statistical techniques have been used is with cross-sectional data. An example is when seeking to identify the clinical characteristics of people who respond to 'diagnostic' injections ('diagnostic' subgroups). Another circumstance where these statistical techniques have been used is within the formation of clinical prediction rules. For example, where subgroup characteristics have been identified using other methods, they have been used to determine the optimal combination of predictor variables that provides the greatest classification accuracy.
In hypothesis-setting subgrouping analysis of longitudinal data, the most common data-driven approach is the use of 'supervised' statistical techniques. An example of such a study is the formation of the Flynn manipulation prediction rule . In this cohort study, a group of people who all received the same treatment (spinal manipulation and range-of-motion exercises) was investigated with the aim of constructing a clinical prediction rule capable of identifying people likely to improve with this treatment. Logistic regression was used to determine which symptoms and signs were predictive of people who improved, and clinimetric statistics were used to determine what combination of those symptoms and signs provided the greatest predictive capacity (Figure (Figure4).4). Being a cohort study, hypotheses formed about treatment effect modifiers could only be tentative, as such a study design cannot clearly differentiate between treatment effect modifiers and prognostic factors.
However, subsequent to this study, Childs et al (2004)  performed a randomised controlled trial, in which a treatment modifier effect of the Flynn manipulation prediction rule was demonstrated using a test of subgroup/treatment/time interaction. In this example, the tentative hypothesis regarding treatment effect modifiers was formed in a cohort study and then demonstrated in a subsequent randomised controlled trial, using a test of interaction. In the proposed method framework, these two studies would be categorised as sequential steps in the hypothesis-setting phase. Had the initial study been a randomised controlled trial, the hypothesis formation could have occurred in a single study. The method for such a single study is shown in Figure Figure55.
The Childs (2004) study examined the treatment modification effect of the entire set of five symptoms and signs in the Flynn manipulation prediction rule. Due to the possibility that a prediction rule derived from a cohort study may contain some predictors that are treatment effect modifiers and some that are prognostic factors, it would be ideal for there to be a mechanism to tease these apart. One way is to perform retrospective (post-hoc) exploratory analysis on the individual symptoms or signs that were included in a randomised controlled trial. Using the Childs (2004) data , Fritz et al (2005)  did for this one item in the Flynn manipulation prediction rule and showed, using a test of subgroup/treatment interaction, a treatment modifier effect of lumbar spine segmental hypomobility. Theoretically, this post-hoc analysis could be performed on all the prediction rule items to identify which are treatment effect modifiers and further refine the prediction rule.
An important advantage of subgrouping studies that use supervised statistical techniques to analyse longitudinal data, is that the subgroup has immediate face validity. This is because the subgroup is formed using a clinically relevant dependent (outcome) variable and therefore the clinical utility of the subgroup is readily apparent. A characteristic of supervised techniques is that clinical prediction rules based on the results of such research are usually dependent on a single outcome and therefore this type of research may lead to a proliferation of competing prediction rules. For example, a clinical prediction rule for the outcome of 'return-to-work' may be quite different from a clinical prediction rule for the outcome of 'moderate or more pain', even if the cohort of people, the treatment and the time period of interest are all the same. This is because outcomes such as 'return-to-work' are influenced by other factors than those that are associated with pain reduction, such as the availability of alternative duties or workplace support structures. Similarly, a prediction rule formed when comparing two treatments may not be the same when the comparison treatment is different. In addition, the prediction rule for a monotherapy (such as manipulation) may not hold when that therapy is applied in combination with other treatment (such as manipulation and exercise). Moreover, treatment effects can be time-dependent, and so prediction rules for the same treatment may vary depending on the time period over which participants are studied. Therefore, supervised analysis is likely to result in multiple clinical prediction rules for the same cohort of people and rules that also vary across cohorts of people.
A number of subgrouping studies have used supervised analysis techniques, such as logistic regression, in forms that can only model two subgroups. This is appropriate for modelling dichotomous subgroups, such as responders and non-responders. However, in circumstances where more than two subgroups are to be modelled, other techniques, such as polytomous or multinomial forms of logistic regression, could be used. For example, where a therapy was expensive and had significant side effects, it might be desirable to identify people very likely to respond, people less likely to respond and those very unlikely to respond. In this case, these subgroups could be used to triage people into 'good candidate for this treatment', possible candidate under particular circumstances' and 'poor candidate for this treatment'. Hypothetically, such a therapy in low back pain might be the use of TNF-inhibitor medication for anklyosing spondylitis.
One data-driven approach to analysing cross-sectional data for subgroups is the use of 'unsupervised' statistical techniques. As seen earlier, unsupervised techniques do not work backwards from an outcome but are instead used to look for inherent relationships between measurable characteristics in cross-sectional data.
An example of a subgrouping hypothesis-setting study that used a data-driven unsupervised statistical technique on cross-sectional data is that by Scholtz et al 2009 . In this study of a mixed cohort of people experiencing pain but not necessarily low back pain, hierarchical cluster analysis was used to identify six subgroups of people with neuropathic pain and two subgroups of people with non-neuropathic pain (Figure (Figure6).6). In a second step, classification tree analysis was used to isolate which symptoms and signs had the greatest discriminatory capacity to classify people into these subgroups. Though not undertaken in this study, the next step within the hypothesis-setting phase of our proposed method framework would be to test these subgroups for treatment modifier or prognostic effects using longitudinal data.
Unsupervised techniques have some advantages: subgroups detected in this way can later be studied against a range of treatments and outcomes, subgroup formation is not dependent on only one outcome, subgroup formation is not dependent on the efficacy of current treatments, more than two subgroups can be detected in a single analysis, and some unsupervised techniques, such as forms of data-mining, also perform well in the presence of missing data. However, unsupervised techniques are more exploratory than supervised techniques and subgroups are not modelled using a clinical outcome. Therefore, they always require, still within the hypothesis-setting phase, subsequent testing against clinically important outcomes to determine if they are clinically relevant. The major disadvantage of this method is that it is possible, maybe probable, that many subgroups derived using unsupervised techniques have no clinical relevance.
Another data-driven method of analysing cross-sectional data for subgroups is the use of 'clinimetric' statistical techniques. In this particular context, this method has been used when researchers seek to detect symptoms and signs that indicate an increased probability that a patient's pain is associated with the presence of a particular pathoanatomic structure (diagnostic subgroup).
An example of a subgrouping hypothesis-setting study that used data-driven 'clinimetric' statistical techniques on cross-sectional data is Laslett et al 2005 . In this study, people with chronic low back pain seeking a diagnostic evaluation in a radiology clinic, were evaluated using provocative discography and a clinical examination by a skilled physiotherapist. The radiologist and physiotherapist were blind to each other's results and the physiotherapist was blind to previous imaging and injection results. Clinimetric statistics were used to determine the strength of association (diagnostic accuracy) between a positive result on the provocative discography and each of the other symptoms or signs (Figure (Figure7).7). A clinical decision rule was then formed, consisting of the optimal combination of those symptoms and signs that provided the greatest predictive capacity. Again, though not undertaken in this study, the next step within the hypothesis-setting phase of our proposed method framework would be to test these subgroups for treatment modifier or prognostic effects.
It is understandable why this 'diagnostic' approach has appeal, as it mimics the Medical Model that has been useful across broad areas of health care. Typically in NSLBP the imaging and clinical findings that have been associated with pathoanatomic structures capable of generating back pain have such a high prevalence in the asymptomatic population that there is considerable uncertainty as to whether they indicate the source of pain in an individual symptomatic patient. Therefore, this 'diagnostic' approach has used the pain response to invasive tests, such as controlled facet injection or provocative discography, to determine the certainty with which a set of clinical symptoms and signs indicate that an individual patient's pain arises from a particular pathoanatomic structure [46-51]. The approach of using pain response to injections in individuals to generalise to populations has been criticised for a number of reasons, including the validity of using response to injection as a reference standard, somatotopic imprecision in the low back, selection bias, spectrum bias, and a lack of concordance in replication studies. However, these criticisms relate more to the construct, design and interpretation of these 'diagnostic' studies than to the clinimetric statistics used.
The findings of subgroup research can be erroneous if overfitting or data dredging have occurred. Overfitting is present when the statistical analysis contains too many predictor variables for the size of the dataset. Though differences of opinion exist, some authors have argued that multivariable analysis requires at least 10 outcome events per independent variable to avoid overfitting [52,53]. The number of outcome events is the sum of the occurrences of the outcome of interest in the data. For example, if the outcome of interest were the people who had a very good recovery, and 35% of a cohort of 300 people did recover well, the number of outcome events is 105 (35% of 300) and therefore, approximately 10 independent variables could be simultaneously entered into a multivariable analysis. The presence of overfitting will markedly weaken the probability that the original findings are reproduced in an independent sample.
Data dredging is the search in large data sets for chance findings that are statistically significant and their reporting without testing if they are spurious associations through replication in an independent data set. Overfitting and data dredging reinforce the desirability of independent sample replication before subgroup predictors are given credence. This may prevent the clinical application or further fruitless testing of chance findings. Despite the importance of this step, a recent systematic review of predictors of chronicity in NSLBP found that, depending on the outcome measure used, only 1 in 12 to 1 in 30 included studies had tested their multivariable findings in an independent sample .
All statistical methods for subgroup research require larger sample sizes than studies that are powered to detect effects observable in a whole group. For example, two-group randomised controlled trials designed to quantify the impact of treatment effect modifiers, require approximately four times the sample size of a conventional controlled trial powered to detect a main effect of the same size . Similarly, the variability present in NSLBP is likely to warrant larger sample sizes in cohort studies and 'diagnostic' studies, than in conditions where the link between pain and pathology is stronger. Hancock et al (2009)  have argued that estimates of treatment effect modification require narrow confidence intervals to be convincing. Narrower confidence intervals, whether around point estimates of treatment effect modification, prognostic risk or diagnostic accuracy, are measures of increased certainty. Narrower confidence intervals allow increased confidence in inferences about clinically important subgroups but do not preclude the need for validation studies.
All the research designs and statistical techniques shown in the method pathways in Figures Figures22 and and33 have their advocates and detractors. The perfect study has not been conducted, as methods are constantly evolving. However, if we are to determine whether subgroup-tailored treatment or generic treatment is better clinical practice, imperfect hypothesis-setting studies will need to be tolerated in the knowledge that further testing is required in a rigorous hypothesis-testing phase and subsequent validation phases. Regardless of the methods used to form subgroup hypotheses, whether these hypotheses concern prognostic effects or treatment effects, there is a need to continue through the other phases of subgrouping research to determine whether these effects are reproducible, generalisable and of clinical importance. Erroneous subgroup findings will not survive the challenge of these later phases of investigation.
We propose that a commitment to the rigor implicit in the method framework is a standard that all proponents of subgroups should meet if subgroup hypotheses are to have scientific credibility. Similarly, it could be argued that subgroup findings that are still in the hypothesis-setting stage are premature to market to clinicians, due to the high probability of spurious findings.
Even where subgrouping findings have shown acceptable reproducibility, generalisability and important effect size, it will take appropriately-designed impact studies to demonstrate whether subgrouping does change practice and improve outcomes in routine care settings. Only where all these criteria are satisfied can evidence-based clinical guidelines confidently recommend subgrouping for routine care.
There is a need for a method framework for subgrouping studies in low back pain due to considerable interest in subgrouping and clinical prediction rules, a proliferation of research methods, and variability in how subgroup results are interpreted. The method framework presented in this article is not prescriptive but is presented to further conversation amongst researchers and clinicians about a suitable roadmap with which to coherently plan, stage, perform and evaluate subgrouping research. The studies used as examples in this method framework were from low back pain research but the methods are equally applicable to neck pain and may also be applicable to other musculoskeletal conditions. The suggested framework provides a platform for modification and extension by the research community, and needs to be regularly updated as method evolves.
The manuscript submitted does not contain information about medical devices or drugs. No benefits in any form have been, or will be, received from a commercial party related directly or indirectly to the subject of this manuscript.
The conception and design of the paper was by PMK. All authors (PMK, JLK, CLY) were involved in the analysis and interpretation of data, drafting and revision of the manuscript, and gave final approval of the manuscript.
The pre-publication history for this paper can be accessed here:
A method for constructing a clinical prediction rule.
The authors are grateful to Per Kjaer for insightful comments on draft manuscripts.