Identifying back pain subgroups: developing and applying approaches using individual patient data collected within clinical trials

Article history

The research reported in this issue of the journal was funded by PGfAR as project number RP-PG-0608-10076. The contractual start date was in October 2010. The final report began editorial review in October 2014 and was accepted for publication in September 2015. As the funder, the PGfAR programme agreed the research questions and study designs in advance with the investigators. The authors have been wholly responsible for all data collection, analysis and interpretation, and for writing up their work. The PGfAR editors and production house have tried to ensure the accuracy of the authors’ report and would like to thank the reviewers for their constructive comments on the final report document. However, they do not accept liability for damages or losses arising from material published in this report.

Declared competing interests of authors

Sarah E Lamb is chairperson of the Health Technology Assessment Clinical Evaluation and Trials (HTA CET) Board. Martin Underwood is a member of the National Institute for Health Research Journals Library Editorial Group.

Copyright statement

© Queen’s Printer and Controller of HMSO 2016. This work was produced by Patel et al. under the terms of a commissioning contract issued by the Secretary of State for Health. This issue may be freely reproduced for the purposes of private research and study and extracts (or indeed, the full report) may be included in professional journals provided that suitable acknowledgement is made and the reproduction is not associated with any form of advertising. Applications for commercial reproduction should be addressed to: NIHR Journals Library, National Institute for Health Research, Evaluation, Trials and Studies Coordinating Centre, Alpha House, University of Southampton Science Park, Southampton SO16 7NS, UK.

Background

Chronic non-specific LBP (CNSLBP) is a common problem affecting a large proportion of the population. 1–4 In the UK, around 70–80% of adults will experience back pain at some point in their life. 5 Some argue that episodic LBP is a universal part of human experience. 6,7 Half of the adult population in the UK (49%) report LBP lasting at least 24 hours in a 1-year period. 5 The 2010 Global Burden of Disease study8 identified LBP as the leading cause of years lived with disability internationally. LBP affects around one-third of the world’s population. 8

Most episodes of back pain are short lived, resolving without the need for any specific treatment. It is the minority of episodes that develop into CNSLBP which create the greatest health need. The natural history of LBP is untidy; around 70% of those affected will experience at least one recurrent episode within a 12-month period. 9

The true prevalence of CNSLBP is difficult to estimate, as definitions and populations vary between studies and countries. However, a review of prevalence studies, reported, between 1966 and 1998, a 12–33% point prevalence; 22–65% 1-year prevalence and up to 84% lifetime prevalence. 10

Since this review, further reviews on the prevalence, focusing on older people and adolescents, have been published. 3,11 A 2012 systematic review synthesised the global prevalence of LBP in studies published between 1980 and 2009. The greatest prevalence was in females aged 40–80 years. After adjusting for methodological variations the point prevalence of back pain lasting for > 1 day was 11.9% [95% confidence interval (CI) 7.98% to 15.82%] and 1-month period prevalence was estimated at 23.2% (95% CI 17.52% to 28.88%). 12

Defining low back pain

The International Association of the Study of Pain defines pain as ‘an unpleasant sensory and emotional experience associated with actual or potential tissue damage, or described in terms of such damage’. 13 The British Pain Society defines acute pain as ‘short term lasting less than 12 weeks’ duration’, whereas chronic pain is defined as ‘long-term pain of more than 12 weeks or after the time that healing would have been thought to have occurred in pain after trauma or surgery’. 14

Low back pain is diagnosed based on the presence of pain and discomfort in the lumbosacral area. 15 Some people also experience pain in the upper leg as a result of LBP. In the majority of cases it is difficult to identify a single cause for back pain. A 2013 systematic review16 of studies of new presentations of LBP found a combined prevalence of 1.5% for fracture and malignancy in primary care; in secondary and tertiary care, prevalence was 6.5%. Once specific causes for LBP have been excluded [malignancy, fracture, infection, inflammatory disorders (such as ankylosing spondylitis)] then a diagnosis of non-specific LBP (NSLBP) is made. This recognises the difficulty in producing robust classification criteria to identify different populations of people affected by chronic LBP.

There is no evidence for a reduction in the population burden of LBP over time. Between 1990 and 2010, in the UK, the number of disability-adjusted life-years attributable to LBP increased by 3.7% from 2231 (95% CI 1555 to 3015) of 100,000 to 2313 (95% CI 1574 to 3113) of 100,000 of the age-standardised population. 17

Economic burden of low back pain

Low back pain is a costly condition to society, health care and the individual. It is the leading cause of sickness absence and health-care use. 18–21 In the UK, the direct health-care cost of back pain in 1998 was £163M. However, the larger burden is that of the indirect costs related to lost production and informal care, which were estimated to be at least £5018M. 22 More up-to-date UK estimates are not available. The current cost is likely to be substantially larger. It is difficult to make direct comparisons of the cost of LBP internationally because of varying health and social care systems. 23

Low back pain results in approximately 4% of the UK population taking time off work. This translates to around 90 million working days lost and between 8 and 12 million general practitioner (GP) consultations per year. 22,24 In 2013 the Office for National Statistics reported 131 million lost working days due to sickness absences in that year in the UK; 30.6 million of these (23%) were lost because of musculoskeletal conditions including back and neck pain. 25

Treatment options for low back pain

People experiencing LBP will often seek medical and drug therapies, as well as therapist-delivered complementary therapies, such as acupuncture, chiropractic or osteopathy, to help relieve pain. 26 Until comparatively recently there were few robust trials of treatments for LBP, and no convincing evidence for the effectiveness of any back-pain treatments. Guidance on the management of LBP was based largely on expert opinion, custom and practice. Since the mid-1990s, there has been a substantial investment in high-quality randomised controlled trials (RCTs) of different treatments for NSLBP. We now have good evidence to show that several therapist-delivered treatment approaches are effective, and for some of these there is also evidence that they are cost-effective. 15,27 By ‘therapist-delivered interventions’ we mean non-drug, non-surgical approaches to the treatment of LBP. Typically, these are delivered by physiotherapists or health/clinical psychologists, but they may be delivered by doctors, health trainers, statutorily regulated complementary practitioners (such as osteopaths or chiropractors), or independently registered professionals providing treatments such as acupuncture or the Alexander technique. The types of interventions offered include acupuncture, manual treatments, exercise regimens, cognitive behavioural approaches or combinations of these.

A number of therapist-delivered interventions are superior to ‘treatment as usual’ (GP care) for participants with chronic LBP. There are numerous treatment options for LBP and several guidelines recommending treatment, including the National Institute for Health and Care Excellence (NICE), the European Corporation in Science and Technology, and the American College of Physicians and American Pain Society guidelines. Such guidance is typically framed as examining independent treatment modalities. Any recommendation for a treatment modality is, inevitably, recommending a package of care including both the non-specific effects of the therapist encounter and the specific effects of the treatment modality in question.

In 2009, NICE guidance15 advised that all people with persistent LBP should be given advice and encouraged to self-manage. As part of this advice, people are encouraged to remain physically active and to engage in daily activity. Subsequently, those affected should be offered a course of acupuncture, exercise or manual therapy. 15 The decision on which treatment to select should be a collaborative decision, taking into account the patient’s treatment preferences. If the selected treatment option is not effective then the patient should be offered another option from the remaining recommended treatments. If the patient is still troubled by back pain then he/she should be considered for an intense physical and psychological intervention. NICE is currently revising its LBP guidelines.

Effectiveness and cost-effectiveness of treatments for low back pain

Although the effectiveness of adding a range of therapist-delivered interventions to best usual care or to no treatment has been well established, the typical mean effect sizes are, at best, modest. By way of illustration, the minimally important (within-person) change in the Roland–Morris Disability Questionnaire (RMDQ) score,28 the most commonly used outcome measure in back pain trials, has been established as 5 points. 29,30 Typical between-group differences in high-quality RCTs are in the order of 1–2 points on the RMDQ, although a few studies have found larger effect sizes (Table 1). These modest mean differences probably translate into ‘numbers needed to treat’ in the order of 5–10. 29,33 These are similar to the numbers needed to treat that are found with antidepressant or antiepileptic drugs which are used to treat chronic painful disorders. 36

The cost per quality-adjusted life-year (QALY) for some of these treatments is well within cost-effectiveness thresholds that are usually used by NICE. Despite this, evidence of access to such treatments within the UK NHS remains patchy. The guideline-endorsed treatments of interdisciplinary rehabilitation, exercise, acupuncture, spinal manipulation and cognitive–behavioural therapy for subacute or chronic LBP have been shown to be cost-effective, but evidence for other endorsed treatments for NSLBP do not yield conclusive or consistent evidence about their relative cost-effectiveness. 37 The scarcity of economic evaluations for some guideline-endorsed treatments means that well-conducted economic evaluations are required to strengthen the evidence base of treatments for LBP.

Subgrouping

Identifying which participants are likely to gain the greatest benefit from different treatments for LBP is an identified high research priority internationally and was one of the key recommendations for future research in the 2009 NICE guidelines for the management of persistent LBP. Current research does not provide any robust data on how to match back pain treatments to participants to maximise effects on outcomes relevant to the participant and cost-effectiveness for the health service.

As different treatment options are argued to work in very different ways, it is a reasonable hypothesis that matching people with LBP to those treatments that are more likely to be effective for their back pain will be a more efficient use of health-care resources and will improve patient outcomes. One might expect that people with high levels of psychological distress that is related to their back pain may gain greater benefit from a psychologically orientated intervention, such as cognitive–behavioural therapy; those with marked loss of physical fitness to benefit most from an exercise intervention; or those with poor back function to benefit most from manual therapy interventions. Developing an evidence base to inform the development of such a stratified care approach has great potential to improve outcomes for people with LBP.

We are aware of one trial of a stratified care approach, published after this programme of work started. The STarT Back trial38 successfully demonstrated that a combination of using a stratification tool and enhanced physiotherapy packages for selected participants improves outcomes and reduces costs when compared with usual physiotherapy care. This study38 does not, however, allow the performance of the stratification tool to identify subgroups to be assessed.

There is a myriad of RCTs that could be designed to address individual components of this problem. High-quality trials in this area are very costly and time-consuming, and can address only one small part of this complex problem. Alternative approaches, which make the best possible use of existing data, can produce timely answers to a range of important research questions and provide substantial added value to the money that is already invested in this area.

We present a programme of work – using systematic reviews, methodological development and secondary analyses of existing data sets – to identify strategies to improve outcomes for people seeking treatment for back pain, by improving how participants, clinicians and purchasers choose treatments. Our programme of work ensures that the maximum information is gleaned from existing substantial trial data sets. The analysis plan for these data and modelling of clinical effectiveness and cost-effectiveness are informed by our literature reviews.

Aim and objectives

The overall aim was to improve the clinical effectiveness and cost-effectiveness of LBP treatment by providing participants, their clinical advisors and health service purchasers with better information about which participants are most likely to benefit from which treatment choices. To achieve this, our objectives were to:

1. synthesise what is already known about the validity, reliability and predictive value of possible treatment moderators

2. develop a repository of individual participant data from RCTs testing therapist-delivered interventions for LBP

3. determine which participant characteristics, if any, predict clinical response to different treatments for LBP

4. determine which participant characteristics, if any, predict the most cost-effective treatments for LBP.

We have defined a therapist as a person trained in administering any of the available recommended treatments, excluding drug interventions and surgical interventions, for the management of LBP.

Structure of this report

This report has been structured as shown in Figure 1. In this report we use some specific terminology that needs additional definition to aid understanding. We have defined these in the Glossary at the start of this report and in more detail at relevant points in the report.

FIGURE 1.

The structure of the current report.

Systematic review 1: identification of potential moderators

This review has been published in Physiotherapy under the terms of the Creative Commons Attribution – NonCommercial – NoDerivs (CC BY-NC-ND 4.0) Licence (http://creativecommons.org/licenses/by-nc-nd/4.0/). Here we present a summary of the paper. 39

Results

Our initial electronic searches generated 7208 hits; 6294 were removed based on title, abstract and duplicates. We obtained 64 papers for detailed review; of these, 60 papers were excluded (Figure 2). Four studies45–48 were included in this review (Table 2). All four trials45–48 were RCTs, constituting a total sample of n = 5514.

FIGURE 2.

Review 1: Quorum statement flow diagram.

TABLE 2 - Review 1: included studies

BeST, Back Skills Training Trial; UK BEAM, UK Back pain Exercise And Manipulation.

Study	Country	Sample	Interventions
UK BEAM45	UK	1334	Group exercise, manual therapy and combination therapy
BeST46	UK	701	Group cognitive–behavioural approach
Witt47	Germany	2841	Acupuncture
Cherkin49	USA	638	Acupuncture

Once we had identified these papers we revisited our search results to include any studies with a sample size of ≥ 300 in a two-group comparison because the trial by Cherkin et al. 49 was a four-arm trial with a sample of n = 638, whereas our sample size calculation of ≥ 500 was based on a two-arm trial. As this paper48 generated some useful moderators for our exploratory work we decided to include it. We did not identify any additional relevant studies with between 300 and 499 participants.

Although the Witt et al. 47 paper provided insufficient data to judge the quality of its exploratory analysis, it did include a specific test for interaction. The data presented did not allow for any pooling of moderator analyses across studies testing similar interventions.

Risk of bias and methodological quality for subgroups

To assess risk of bias and quality of subgroups we used both the original main trial papers and the associated secondary papers where appropriate (Tables 3 and 4).

TABLE 3 - Review 1: results of the risk of bias assessment

BeST, Back Skills Training Trial; H, high risk of bias; L, low risk of bias; MRC, Medical Research Council; NIHR HTA, National Institute for Health Research Health Technology Assessment; U, unclear; UK BEAM, UK Back pain Exercise And Manipulation.

Quality of the study based on main trial paper(s)	UK BEAM31	BeST33,34	Witt47	Cherkin48,49
Random sequence generation	L	L	L	L
Allocation concealment	L	L	L	L
Blinding of participants and personnel	H	H	H	H
Blinding of outcome assessment	L	L	H	L
Incomplete outcome data	L	L	U	L
Selective reporting	L	L	U	L
Generalisability	L	L	L	L
Sample size calculation	L	L	U	L
Conflict of interest	L	L	H	L
Source of funding	MRC	NIHR HTA	Social Health Fund Providers	National Institutes of Health

TABLE 4 - Review 1: results of methodological quality assessments

BeST, Back Skills Training Trial; CE, confirmatory evidence – fulfils all five criteria for moderator studies; EE, exploratory evidence – fulfils three, four or five criteria for moderator studies; IE, insufficient evidence to judge quality; N, no; U, unclear; UK BEAM, UK Back pain Exercise And Manipulation; Y, yes.

Quality of the moderator analyses based on subgroup paper(s)	UK BEAM31	BeST33,34	Witt47	Cherkin48,49
Was the subgroup analysis specified a priori?	N	Y	N	N
Was the selection of subgroup factors for analysis theory/evidence driven?	N	Y	N	N
Were subgroup factors measured prior to randomisation?	Y	Y	U	Y
Was measurement of subgroup factors measured by adequate (reliable and valid) measurements, appropriate for the target population?	Y	Y	N	Y
Does the analysis contain an explicit test of the interaction between moderator and treatment?	Y	Y	U	Y
Strength of evidence	EE	CE for two potential moderators	IE	EE

Table 5 presents the potential moderators with strong and/or weak evidence from the four included trials. 45–48 The many interactions tested that were not statistically significant are not reported here.

TABLE 5 - Mean difference (95% CI) of potential moderators with strong evidence (p < 0.05) and weak evidence (p < 0.20 ≥ 0.05)

BeST, Back Skills Training Trial; B/L, baseline; FFbHR, Hannover Functional Ability Questionnaire for measuring back pain-related functional limitations (Funktionsbeeintrachtigung durch Ruckenschmerzen); HADS, Hospital Anxiety and Depression Scale; IA, individualised acupuncture; MVK, Modified von Korff; NS, no significant interaction found; SiA, simulation acupuncture; StA, standardised acupuncture; UK BEAM, UK Back pain Exercise And Manipulation.

Study ID	Potential moderators		Significant interaction on selected outcomes (12 months)
			RMDQ		MVK pain		MVK disability
BeST34,46	Troublesomeness (very/extremely – moderately)		p = 0.190; –1.01 (–2.52 to 0.50)		p = 0.184; –5.04 (–12.47 to 2.40)		NS
	Age (≥ 54 years – < 54 years)		p = 0.035; –1.58 (–3.05 to –0.12)		NS		NS
	Female – male		p = 0.102; –1.27 (–2.79 to 0.25)		NS		NS
	Left FT education (> 16 years of age – ≤ 16 years of age)		p = 0.098; 1.29 (–0.24 to 2.82)		NS		NS
	Employed – not employed		p = 0.011; 1.89 (0.43 to 3.35)		p = 0.181; 5.01 (–2.33 to 12.34)		NS
	HADS – anxiety (≥ 11 – < 11)		p = 0.195; –1.12 (–2.83 to 0.58)		NS		NS
	HADS – depression (≥ 11 – < 11)		p = 0.135; –2.07 (–4.79 to 0.65)		NS		p = 0.051; –14.58 (–29.19 to 0.03)
Study ID	Potential moderators	Significant interactions; outcome, RMDQ
		8 weeks			52 weeks
		IAd	StAe	SiAf	IA	StA	SiA
Cherkin48,49	Age	NS	p = 0.08; 0.08 (–0.02 to 0.18)	NS	NS	p = 0.15; 0.07 (–0.03 to 0.17)	NS
	Self-efficacy	p = 0.04; –6.17 (–12.01 to –0.33)	NS	NS	NS	NS	NS
	RMDQ (B/L)	p < 0.0001; –0.48 (–0.72 to –0.24)	p = 0.004; –0.37 (–0.62 to –0.12)	p = 0.001; –0.41 (–0.66 to –0.16)	p = 0.07; –0.23 (–0.48 to 0.02)	p = 0.07; –0.24 (–0.49 to 0.01)	NS
	Bothersomeness score (B/L)	NS	p = 0.10; 0.47 (–0.10 to –1.04)	NS	NS	NS	NS
	Heavy lifting	p = 0.03; 4.29 (0.43 to 8.15)	p = 0.13; 3.00 (–0.86 to 6.86)	p = 0.18; 2.73 (–1.27 to 6.73)	p = 0.01; 5.19 (1.17 to 9.21)	p = 0.15; 3.03 (–1.05 to 7.11)	p = 0.04; 4.45 (0.28 to 8.62)
	Sedentary	NS	NS	NS	p = 0.12; 2.73 (–0.72 to 6.18)	p = 0.15; 2.47 (–0.90 to 5.84)	NS
	Use of narcotic medication	p = 0.08; 3.52 (–0.38 to 7.42)	NS	p = 0.01; 4.81 (0.97 to 8.65)	NS	p = 0.04; 4.06 (0.18, 7.94)	p = 0.19; 2.71 (–1.31 to 6.73)
	Acupuncture expectation (top tertile)	p = 0.05; –2.65 (–5.28 to –0.02)	NS	NS	NS	p = 0.17; –1.9 (–4.60 to 0.80)	p = 0.03; –2.91 (–5.56 to –0.26)
Study ID	Potential moderators	Significant interactions; outcome, bothersomeness score
		8 weeks			52 weeks
Cherkin48,49	Age	NS	p = 0.09; 0.04 (0.001 to 0.08)	p = 0.07; 0.04 (0.001 to 0.08)	NS	p = 0.15; 0.04 (–0.02 to 0.10)	p = 0.08; 0.05 (–0.01 to 0.11)
	Self-efficacy	p = 0.14; –2.21 (–5.13 to 0.71)	NS	NS	NS	NS	NS
	Baseline RMDQ score	p = 0.01; –0.15 (–0.27 to –0.03)	NS	p = 0.0005; –0.22 (–0.34 to –0.10)	p = 0.16; –0.09 (–0.23 to 0.05)	NS	NS
	Heavy lifting	p = 0.05; 1.97 (0.03 to 3.91)	NS	p = 0.04; 2.10 (0.10 to 4.10)	p = 0.02; 2.51 (0.43 to 4.59)	NS	NS
	Light/medium lifting	NS	p = 0.12; –1.28 (–2.87 to 0.31)	NS	p = 0.12; 1.35 (–0.36 to 3.06)	NS	NS
	Sedentary	NS	NS	NS	p = 0.19; 1.20 (–0.58 to 2.98)	NS	NS
	Acupuncture expectation (top tertile)	p = 0.10; –1.10 (–2.41 to 0.21)	NS	NS	p = 0.051; –1.44 (–2.87 to –0.01)	NS	p = 0.06; –1.29 (–2.64 to 0.06)
Study ID	Potential moderators	3 months for RMDQ outcome, combined treatment					12 months for RMDQ outcome, combined treatment
UK BEAM31,45	Quality of life	p = 0.174; –0.1 (–0.26 to 1.43)					NS
	Treatment expectation (helpful)	p = 0.073; –3.2 (–6.74 to 0.30)					p = 0.038; –3.8 (–7.39 to –0.20)
	Treatment expectation (very helpful)	p = 0.192; –2.2 (–5.49 to 1.11)					p = 0.019; –4.0 (–7.38 to –0.67)
	Manipulation
	Beliefs	p = 0.07; –0.8 (–1.62 to 0.06)					NS
	Quality of life	p = 0.118; 1.4 (–0.35 to 3.07)					NS
	Pain/disability	p = 0.176; –1.9 (–4.61 to 0.85)					p = 0.143; –2.2 (–5.16 to 0.75)
	Treatment expectation (helpful)	NS					p = 0.083; –0.1 (–0.16 to 0.01)
	Treatment expectation (very helpful)	p = 0.113; 1.6 (–0.38 to 3.60)					NS
Study ID	Potential moderators	Outcome, FFbHR
Witt47	Worse initial back function	p < 0.001 Back function and pain improvement at 3 months with acupuncture treatment
	Younger	p < 0.001
	> 10 years of schooling	p = 0.01

Moderator variables identified

Potential moderators with strong evidence (p < 0.05) in one or more studies include age (younger participants may gain more benefit), employment status and type (those employed or in sedentary occupations may gain greater benefit), back pain status (those who are worse may gain greater benefit), narcotic medication use (users may benefit less), treatment expectations (those with a greater positive expectation gained more benefit) and education (those with > 10 years of schooling gained a greater benefit). Potential moderators with weaker evidence (0.05 < p ≤ 0.20) include gender (female participants may gain greater benefit), psychological distress (those with anxiety and depressive symptoms may benefit more), pain/disability (those with greater pain/disability at baseline may benefit more) and quality of life (those with a better quality of life may benefit more). It should be noted that these findings might just be a chance finding, particularly as these conclusions come from different studies.

Age: the BeST (Back Skills Training Trial), Cherkin and Witt trials46,49,50 found an interaction with age. In the BeST trial,46 younger participants gained more benefit from cognitive behavioural therapy than older participants on the RMDQ score. The treatment difference was –1.58 (p = 0.035; 95% CI –3.05 to –0.12). As the p-value was < 0.05, the interactions provided strong evidence. Witt et al. 50 found a statistically significant additional benefit from acupuncture treatment in younger participants (p < 0.001).

Gender: the BeST trial46 found that gender had a moderating effect on treatment. In this trial, females had comparatively greater improvement following group cognitive behavioural therapy than males. The treatment difference between male and female was –1.27 (p = 0.102; 95% CI –2.79 to 0.25) for the RMDQ score. As the p-value was 0.05 < p ≤ 0.20, the interaction provides weak evidence.

Employment status: employment was found to be one of the positive moderating factors. In the BeST trial,46 the authors found that employed participants gained additional benefit from a cognitive behavioural approach compared with those who were unemployed. The treatment difference between employed and unemployed was 1.89 (p = 0.011; 95% CI 0.43 to 3.35) and 5.01 (p = 0.181; 95% CI –2.33 to 12.34) for the RMDQ and Modified von Korff (MVK) pain scores, respectively. The interaction effect in the analysis of the MVK pain score was weak. 46 The Cherkin trial48,49 found some moderating effect according to types of employment status. The participants in this trial48 received acupuncture therapy. Those participants whose job involved heavy lifting showed positive moderating effect against back-related dysfunction score at 8 weeks (p = 0.03 to 0.18) and 52 weeks (p = 0.01 to 0.04). Those participants doing medium/light lifting at work showed positive moderating effect in terms of the bothersomeness score (p = 0.12) at 8 and 52 weeks; however, the interaction was weak. Finally, those participants with sedentary work showed positive moderating effect at 52 weeks (p = 0.12 to 0.19). The interaction was generally weak.

Education: the BeST trial46 found that participants who had left full-time education after the age of 16 years had better improvement from cognitive behavioural therapy than participants who left full-time education aged ≤ 16 years. The treatment difference was 1.29 (p = 0.098; 95% CI –0.24 to 2.82) for the RMDQ score. The interaction effect was > 0.05 and, therefore, this provides weak evidence. Witt et al. 50 found that those participants who have had > 10 years of schooling gained a greater benefit from acupuncture (p = 0.01).

Back pain status: In the Cherkin and Witt trials48–50 participants with a worse initial back pain status (baseline RMDQ score) gained an increased benefit from acupuncture compared with those with a better back pain status at baseline (p-values ranged from < 0.001 to 0.16). The extent to which LBP inconveniences participants – how troublesome or bothersome it is – was found to be a moderator in two trials, with a greater benefit from treatment in those with a more troublesome/bothersome condition. The interaction was weak, with the p-values being > 0.05. In the Cherkin trial,48,49 the p-value was 0.10, whereas in the BeST trial46 the treatment difference for the RMDQ score was –1.01 (p = 0.190; 95% CI –2.52 to 0.50) and –5.04 (p = 1.184; 95% CI –12.47 to 2.40) for MVK pain score.

Pain/disability: similarly, those participants with greater pain/disability at baseline seemed to benefit more at 3 months (p = 0.176) and 12 months (p = 0.143) for the RMDQ score with manipulation treatment [UK Back pain Exercise And Manipulation (UK BEAM45)] (see Table 5). The p-values are > 0.05 and < 0.2, therefore providing weak evidence. 45

Narcotic: Cherkin et al. 48,49 found that use of medication such as narcotics had a negative moderating effect in those receiving acupuncture. The p-value for this interaction ranged from 0.01 to 0.19, demonstrating a spectrum of strong to weak evidence.

Treatment expectations: having better expectations about the treatment was found to be a moderating factor in two trials. 45,48,49 The p-values ranged between 0.03 and 0.192, demonstrating a spectrum of strong to weak evidence for the interactions. 48,49 Cherkin et al. 48,49 found that participants with higher expectation of acupuncture treatment helpfulness gained more benefit in the back-related dysfunction score (p = 0.03–0.17) and bothersomeness score (p = 0.05–0.10). 48,49 In the UK BEAM trial,45 manipulation at 3 months (p = 0.113) and 12 months (p = 0.083), or a combined treatment of manipulation and exercise (p = 0.03 to 0.192) at both 3 and 12 months, showed positive moderating effect, as was demonstrated by the RMDQ score. Overall, the interactions were found to range between a spectrum of strong to weak evidence.

Quality of life: good quality of life showed weak evidence for a moderating effect on treatment outcome for both manipulation treatment (p = 0.118) and a combined manipulation and exercise treatment (p = 0.174). 45

Psychosocial status: in the BeST trial,46 psychosocial status moderated treatment effect. The trial46 investigated whether psychological status moderated better outcome from a cognitive behavioural therapy. Participants with higher levels of anxiety at baseline gained more benefit from treatment in terms of the RMDQ score. The treatment difference was found to be –1.12 (p = 0.195; 95% CI –2.83 to 0.58), demonstrating a weak interaction. Similarly, those participants who were depressed considerably gained more benefit from the treatment than those who were less depressed as was found in the RMDQ and MVK disability scores. The treatment difference was found to be –2.07 (p = 0.135; 95% CI –4.79 to 0.65) and –14.58 (p = 0.051; 95% CI –29.19 to 0.03) for the RMDQ and MVK disability scores, respectively.

Systematic review 2: quality of subgroup analyses in low back pain trials

This review has been published in Spine. 51 Here we present a summary of the paper.

Results

Our initial search identified 5581 papers. All titles and abstracts were screened to identify potential papers reporting results of RCTs of therapist-delivered interventions for LBP. We excluded 5521 papers during the screening process. The full text for the remaining 60 papers was then thoroughly examined to look for subgroup analyses, of which 21 were excluded as they either did not meet the inclusion criteria or they met one or more of the exclusion criteria. We included 39 papers in the final review (Figure 3).

FIGURE 3.

Review 2: Quorum statement flow diagram. Reproduced from Mistry D, Patel S, Hee SW, Stallard N, Underwood M. Evaluating the quality of subgroup analyses in randomized controlled trials of therapist-delivered interventions for nonspecific low back pain: a systematic review. Spine 2014;39:618–29; with permission from Lippincott Williams & Wilkins.

A summary of the included studies is given in Table 6 and a summary of excluded studies can be found in Appendix 1. A total of 63% of the included papers were from the Netherlands, the UK or the USA. The median study size was 223, ranging from 100 to 3093.

TABLE 6 - Summary of included papers in descending order by subgroup quality assessment

AIP, active individual therapy; APT, active physical therapy; ATEP, active trunk exercise protocol; CBT, cognitive–behavioural treatment; CTrt, combination treatment; FD, flexion distraction; Hannover Functional Ability Questionnaire for measuring back pain-related functional limitations; FRP, functional restoration programme; GAP, graded activity with problem solving training; GI, guideline implementation; GPE, global perceived effect; GPP, general practitioner programme; ITP, intensive training programme; MC, motivational counselling; MIS, minimal intervention strategy; MTP, manual therapy programme; ODI, Oswestry Disability Index; OPCO, operant behavioural treatment with cognitive coping skills training; OPDI, operant behavioural treatment with group discussion; RRP, Roessing back rehabilitation; SF-36, Short Form questionnaire-36 items; SiA, simulation acupuncture; SMT, spinal manipulative therapy; StA, standardised acupuncture; VAS, visual analogue scale; WL, waiting list; WLC, waiting list control.

Reproduced from Mistry D, Patel S, Hee SW, Stallard N, Underwood M. Evaluating the quality of subgroup analyses in randomized controlled trials of therapist-delivered interventions for nonspecific low back pain: a systematic review. Spine 2014;39:618–29; with permission from Lippincott Williams & Wilkins.

Subgroup quality assessment	Author	Date of publication	Country	Study size	Interventions compared	Outcome measure and follow-up	Subgroups identified (interaction test only)
Confirmatory findings	Sheets58	2012	Australia	148	First-line care group vs. McKenzie group	Pain measured at 1 week and 3 weeks GPE at 3 weeks	None
	Smeets59	2009	Australia and New Zealand	259	Exercise and advice vs. exercise and sham advice vs. sham exercise and advice vs. sham exercise and sham advice	Pain intensity (11-point scale) and patient-specific function scale (0–10 scale) measured at baseline 6 weeks and 52 weeks	None
	Underwood46	2011	UK	701	Advice plus cognitive–behavioural intervention vs. advice only	RMDQ and MVK scores measured at baseline and 3, 6 and 12 months	Age and employment
Exploratory findings	Becker60	2008	Germany	1378	Multifaceted GI vs. GI plus MC vs. postal dissemination of guideline (control)	FFbHR measured at baseline and 6 months	None
	Cecchi61	2012	Italy	210	Back school vs. individual physiotherapy vs. spinal manipulation	RMDQ score measured at baseline and 3, 6 and 12 months	None
	Cherkin62	1998	USA	321	Physical therapy vs. chiropractic manipulation vs. educational booklet	Bothersomeness of symptoms and RMDQ score measured at baseline, 4 weeks and 12 weeks	Mental health
	Cherkin63	2001	USA	262	Chinese acupuncture vs. therapeutic massage vs. self-care education	Bothersomeness of symptoms and RMDQ score measured at baseline, 4 weeks, 10 weeks and 1 year	None
	Cherkin49	2009	USA	638	Individualised acupuncture vs. StA vs. SiA vs. usual care	Bothersomeness of symptoms and RMDQ score measured at baseline, 8 weeks, 26 weeks and 1 year	None
	Hansen64	1993	Denmark	180	Intensive dynamic back-muscle exercise vs. conventional physiotherapy vs. placebo control (semi-hot packs and light traction)	Pain level (10-point scale) measured at baseline, 4 weeks, 6 weeks and 1 year	None
	Hay65	2005	UK	402	Brief pain management vs. manual physiotherapy	RMDQ score measured at baseline, 3 months and 12 months	None
	Juni66	2009	Switzerland	104	Standard care alone vs. standard care plus SMT	Pain intensity (11-point scale) and analgesic use measured at baseline, days 1–14 and 6 months	None
	Karjalainen67	2004	Finland	170	Mini-intervention group vs. worksite visit group vs. usual care group	Pain intensity (11-point scale) measured at baseline, 3 months, 6 months, 1 year and 2 years	Perceived risk for not recovering and type of occupation (comparing Mini-intervention vs. usual care and worksite visit vs. usual care)
	Kole-Snijders68	1999	Netherlands	159	OPCO vs. OPDI vs. WLC	Main outcome unclear Outcomes measured post treatment and at 6 months and 1 year	None
	Roche69	2007	France	132	AIP vs. FRP	Main outcome unclear Outcomes measured at baseline and 5 weeks	Sorenson score
	Sherman48	2009	USA	638	Individualised acupuncture vs. StA vs. SiA vs. usual care	Bothersomeness of symptoms and RMDQ score measured at baseline, 8 weeks, 26 weeks and 1 year	Baseline RMDQ score
	Smeets70	2006	Netherlands	223	APT vs. CBT vs. Combined APT and CBT (CTrt) vs. WL	RMDQ score measured at baseline, 10 weeks, 6 months and 12 months	Baseline RMDQ
	Smeets71	2008	Netherlands	223	ATP vs. GAP vs. CTrt vs. WL	RMDQ score measured at baseline, 10 weeks, 6 months and 12 months	None
	Tilbrook35	2011	UK	313	Yoga vs. usual care	RMDQ score measured at baseline and 3, 6 and 12 months	None
	Underwood45	2007	UK	1334	Control (best care in general practice) vs. exercise programme vs. spinal manipulation vs. combined treatment (manipulation and exercise)	RMDQ score measured at baseline, 3 months and 1 year	Expectation
	Van der Hulst72	2008	Netherlands	163	RRP vs. usual care	RMDQ score measured at baseline, 1 week after treatment and 4 months after treatment	Pain intensity and depression
	Witt50	2006	Germany	3093	Acupuncture vs. control (delayed acupuncture treatment 3 months later)	FFbHR (0–100 scale) measured at baseline and 3 and 6 months	Initial back pain, age and years of schooling
Insufficient findings	Bendix73	1998	Denmark	816	FRP programme vs. outpatients programme (control)	Main outcome unclear Outcomes measured at baseline and 1 year
	Beurskens74	1995	Netherlands	151	Traction vs. sham traction	GPE and severity measured on VAS at baseline and 5 weeks
	Bishop75	2011	USA	112	Supine thrust technique vs. side-lying thrust vs. non-thrust technique	ODI measured at 1 week, 4 weeks and 6 months	None
	Carr76	2005	UK	237	Group exercise programme vs. individual physiotherapy	RMDQ score measured at baseline, 3 months and 6 months
	Ferreira77	2009	Australia	191	General exercise vs. motor control exercise vs. SMT	GPE (11-point scale), Patient specific functional status, RMDQ score, Pain intensity (10-point scale) and spinal stiffness measured at baseline and 8 weeks	None
	Glazov78	2009	Australia	100	Laser acupuncture vs. sham acupuncture (control)	Pain (VAS) measured at baseline, immediately after treatment, 6 weeks and 6 months
	Gudavalli79	2006	USA	235	FD vs. ATEP	Perceived pain (VAS), RMDQ score and SF-36 measured at baseline, 4 weeks, 3 months, 6 months and 1 year
	Hsieh80	2004	China	146	Acupressure vs. physical therapy	Short-form pain questionnaire measured at baseline, 4 weeks and 6 months
	Jellema81	2005	Netherlands	314	MIS vs. usual care	RMDQ score, perceived recovery (7-point scale) and sick leave measured at baseline; 6, 13 and 26 weeks; and 1 year
	Johnson82	2007	UK	234	Group exercise and education using a cognitive behavioural approach vs. usual care	Pain (VAS) and RMDQ score measured at baseline and 3, 9 and 15 months	Patient preference
	Kalauokalani83	2001	USA	166	Acupuncture vs. massage (subanalysis of Cherkin 2001 paper)	RMDQ score measured at baseline, 4 weeks, 10 weeks and 1 year	Patient expectations
	Mellin84	1989	Finland	456	Inpatient treatment vs. outpatient treatment vs. control (advice)	LBP disability index (scale 0–45) measured at baseline and 3 months
	Klaber Moffett85	2004	UK	187	Exercise vs. usual care	RMDQ score measured at baseline, 6 weeks, 6 months and 1 year
	Myers86	2008	USA	444	Usual care vs. usual care plus patient choice of acupuncture, chiropractic or massage	RMDQ score measured at baseline, 5 weeks and 12 weeks	None
	Seferlis87	1998	Sweden	180	MTP vs. ITP vs. GPP	Main outcome unclear Outcomes measured at baseline and 1, 3 and 12 months
	Thomas88	2006	UK	241	Traditional acupuncture vs. usual care	Bodily pain dimension of the SF-36 (0–100 scale) measured at baseline and 3, 12 and 24 months	Expectation
	Van der Roer89	2008	Netherlands	114	Intensive group training protocol vs. guideline group	RMDQ score measured at baseline and 6, 13, 26 and 52 weeks
	Vollenbroek-Hutten90	2004	Netherlands	163	RRP vs. usual care	RMDQ score measured at baseline, 1 week after treatment and 4 months after treatment

Methodological quality of subgroup analyses

The methodological quality of the subgroup analyses performed in the identified papers was assessed to determine the strength of evidence that they provide. Of the 39 papers:35,45,46,48–50,58–90

• Three (8%) papers46,53,54,58,59 met all five criteria and therefore provided confirmatory evidence. Two of these papers58,59 were too small to anticipate finding any important interaction if it were present (n = 148 and 259).

• Eighteen (46%) papers provided exploratory evidence, that is, they met criteria 3, 4 and 5 (see Table 6).

• Eighteen (46%) papers provided insufficient evidence (see Table 6).

Assessment of conduct and reporting of subgroups

We examined the conduct and reporting of subgroups in terms of design and methods and found that:

• One study50 had sufficient power to detect an interaction; however, subgroups of interest were not prespecified a priori.

• Thirty-one (79%) studies35,45,48–50,60–63,66–74,76–78,80,81,83–90 did not prespecify subgroups of interest.

• Eight studies46,58,59,64,65,75,79,82 reported prespecified subgroups for confirmatory analyses; six of these studies also carried out exploratory analyses without clear distinction between analysis types.

• Sometimes it was not clear from the methods that subgroup analyses were going to be performed; they were just presented in the results. 62,69,74,80

• All papers measured subgroups of interest prior to randomisation, with most using adequate measurements.

• Prior to performing analyses, only one paper58 reported the expected size and direction of the subgroup effect. A further three papers46,59,85 predicted the direction of the subgroup effect.

• One-third (13/39) of the papers45,46,48,58,59,72,77,79,83,85–87,90 provided some justification regarding the choice of subgroups to be analysed.

• In two papers45,59 around 60 interaction tests were conducted, substantially increasing the chances of detecting false-positive findings. Of the three papers46,58,59 that provided confirmatory findings, only one of them46 adjusted for multiplicity. The authors applied a Bonferroni correction to their confirmatory subgroup analyses.

• Twelve (31%) of the papers64,73,74,76,79–81,84,85,87,89,90 did not use a statistical test for interaction to assess for treatment effect modification. Of these, two of the papers74,87 did not give any indication as to what statistical method they used. Two papers73,84 looked at correlations between individual subgroups and outcomes within each treatment arm separately. Two papers79,80 used t-tests between treatment groups within individual subgroups. Five papers76,81,85,89,90 used either multiple linear regression or multiple logistic regression for each individual subgroup. One paper64 compared the medians across three trial arms within individual subgroups using Kruskal–Wallis tests.

We examined the conduct and reporting of subgroups in terms of reporting of results and found that:

• A statistical test for interaction was reported to have been used in 27 (69%) of the papers. 35,45,46,48–50,58–72,75,77,82,83,86,88

• Six studies45,48,61,72,75,77 reported both the interaction effect sizes with CIs and the corresponding p-values.

• Four studies46,58,59,82 reported only the interaction effect sizes with CIs.

• Eight studies35,50,66,67,69,83,86,88 reported only the p-values.

• Nine papers49,60,62–65,68,70,71 did not report the interaction effect sizes, CIs or p-values.

• Four studies60,66,70,88 reported subgroup analyses within individual subgroups rather than between-group interaction.

We examined the conduct and reporting of subgroups in terms of reporting of interpretation and discussion and found that:

• Four60,66,70,88 out of 27 papers that performed interaction tests reported subgroup analyses within individual subgroups and thus based the interpretations and discussion on this as well.

• Reference to other relevant studies (supporting or contradicting) were made in around one-third of the papers.

• The limitations of subgroup analyses were reported in 12 papers. 45,46,48,58–61,65,76,79,86,90

Summary of reviews

Both reviews conducted during this programme of work have been informative in developing our understanding of subgrouping in LBP.

Review 1 looked at identifying potential moderators to be tested within the back pain repository. The literature on moderators is weak and, subsequently, lacking in rigour to inform clinical practice. Despite this, the review has helped us to identify some potential moderators of treatment effect, including age, educational attainment, employment status, symptoms of anxiety or depression, longer history of back pain and treatment expectations in at least one trial. We used these variables in our later analyses within our repository of data.

Review 2 looked at the quality of subgroup analyses conducted in the LBP literature. This review concluded that the overall quality was poor. A trial that is sufficiently powered to detect subgroups would need to be approximately four times larger than a traditional trial powered to detect a main effect of the same magnitude. 92 This would be a timely and costly undertaking, for which care would also need to be taken to select moderators that were clinically relevant and applicable.

In addition to these reviews we have previously published a systematic review93 that summarised findings from RCTs testing the effects of a clinical prediction rule for NSLBP. Clinical prediction rules have been developed and are being used in clinical practice to help clinicians to make decisions on treatment; however, the overall effect of such tools is unclear. Multicomponent clinical prediction rules have the potential to be much more powerful tools for targeting treatments than single-component measures. We identified 1821 potential citations after all duplications had been removed. Two reviewers independently screened the titles and abstracts, and consensus was reached on obtaining 35 papers for full detailed evaluation. Of these, only three papers94–96 were included in the review. The results from the available trials do not convincingly support the use of clinical prediction rules in the management of NSLBP. We concluded that the existing RCTs looking to validate clinical prediction rules in LBP are limited. Methodologies for the validation of these rules lack clarity and, subsequently, the evidence for, and development of, the existing prediction rules in LBP is generally weak.

Current approaches have failed to provide the data needed to target treatments for LBP. There is therefore a need to look at alternative methods to address this problem. We propose three recommendations:

1. To develop new and novel methods to identify multiple participant characteristics or clusters of moderators that would identify who is most or least likely to benefit. 97–99

2. To apply individual participant data meta-analysis to homogeneous pooled data sets, as this would improve statistical power.

3. To develop subgroups, and suggested interventions, based on clinical reasoning, and test these within trials to determine if the targeted intervention produces a larger average effect size than existing non-specific interventions.

In this programme we address points 1 and 2, leaving point 3 for others within the back pain research community to consider and address.

Identification of potential trials

We used the search results generated from review 1 (Identification of potential moderators, described in Chapter 2) as a starting point for identifying trials of interest. In the first instance we were interested in only:

• RCTs

• trials of therapist-delivered interventions

• trials with a sample size of > 179 participants.

Based on these criteria we filtered the original search output to identify 658 citations. These were systematically screened by two members of the team independently (see Figure 2). Additionally, we also obtained further data through snowballing; essentially, we were offered data (from researchers aware of the project) from trials that were not on our original list. Although some of the trials obtained through the snowballing process are smaller in sample size than our target studies, we decided to include these to add power to our analysis.

Justification of sample size

We started with an original lower limit of 200 for the sample size. Allowing for some loss to follow-up, a trial of 200 participants would have 90% statistical power to identify a SMD of 0.5 between two treatment groups. Any individual trials smaller than this are likely to be seriously underpowered for their primary outcome. Upon screening the trials there were many that obtained a final sample size of just fewer than 200 participants; typically these were studies aiming for around 200 participants, which fell short of the final target. We therefore revised our inclusion to more than 179 participants. From a practical perspective of approaching trial investigators, this yielded a manageable number of trials to approach; large trials (those with thousands of participants) and small trials (fewer than 100 participants) each create a similar amount of work to collate.

Process for approaching investigators

We identified 42 trials33,49,50,62,63,65,70,76,82,84,94,100–130 that fitted our inclusion criteria. For these trials we identified the chief investigator and the best e-mail contacts for them. Between 2011 and 2012 each investigator was sent an e-mail to invite him/her to participate in the repository. Each e-mail included the following attachments:

• formal invitation letter (see Appendix 2)

• information sheet (see Appendix 3)

• sample data sharing agreement (see Appendix 4).

If a response was not received within a 6- to 8-week period, a reminder e-mail was then sent. If a response was received indicating an interest in sharing data then the data sharing agreement was personalised and sent back to the investigator for review and signature. Once the signed document was received by the university, the investigator was provided with details on how to securely send the data to us. We used the University of Warwick secure file transfer service.

Secure data transfer

We requested all data from a trial. Investigators were advised that any data sets being sent to us needed to be anonymised and encrypted using an open-source compression software programme such as 7-Zip 9.20 © Igor Pavlov (www.7-zip.org/). Investigators were then provided with details on how to securely transfer this data to the University of Warwick (see Appendix 5) using an upload system that was set up for the project (available at https://files.warwick.ac.uk/repositorylbpdata/sendto).

Once these data were received it was the responsibility of the team’s statisticians and/or health economists to transform the original data to the repository standard. To aid this process we requested all trial-specific information, including the protocol and questionnaires if they were available.

Final data set obtained

We obtained 14 (33%) trial data sets31,33,50,65,70,76,101–107,131 from the original 42 trials33,49,50,62,63,65,70,76,82,84,94,100–130 we approached. A further five trials132–136 were obtained through snowballing, resulting in a total of 19 data sets (Figure 4). We were unsuccessful in getting a response from 15 (36%) investigators and a further six (14%) data sets were not available for data sharing. We still have seven (17%) data sets in negotiation, for which we were unable to agree on the data sharing before starting our formal analysis; therefore, these trials have not been included in this report.

FIGURE 4.

Quorum statement flow diagram for database identification.

Through the process of snowballing, further smaller data sets were offered to be included in the repository. The offer of these trials was carefully considered by the research team, and it was decided that any additional data would be helpful in increasing power. Therefore, three (16%) of the 19 trials obtained have a sample size of < 179 participants.

Table 7 shows the trials that were excluded and the reason for the exclusion. Details of papers excluded as a result of multiple publications can be found in Appendix 6. A list of trials that were unavailable because of a lack of response from the investigator, data sets not available and those still under negotiation are documented in Appendix 7. A final table of included trials and associated papers is presented in Table 8.

Summary of the included trials in the repository

The agreed and included trials in this repository are detailed in Table 9.

Grouping of interventions

Initial examination of the data showed that no two trials studied identical interventions. Even the usual-care arms of included studies are likely to differ according to jurisdiction, site of recruitment and age of the study. Even with our initial large sample size it was clear that, to be able to make meaningful comparisons, we would need to pool interventions into broad groups for our analyses. As a first stage we identified the control interventions and classified these as either usual care or a sham control. There is, for example, evidence from the acupuncture literature that the difference between sham acupuncture and usual care is greater than any difference between sham and verum acupuncture. 145 We therefore opted to separate the sham interventions from the usual care control in our analyses comparing different treatments with control or with each other.

There may be qualitative differences between sham treatments. For example, sham acupuncture, through which the participant has had the sensation of being needled, might have a different effect from a sham educational intervention. In some analyses we have included sham interventions, typically sham acupuncture as a separate category. For this reason we have, where appropriate, specified the nature of the sham intervention considered.

We used the following approach to develop our final grouping of interventions:

1. Careful reading of each trial intervention to decide on core groups [individual physiotherapy, exercise, manipulation, advice/education, psychological therapy, graded activity, acupuncture, combination therapy, mock transcutaneous electrical nerve stimulation (TENS), sham acupuncture and control]. We listed all of the trials contributing to each of the core groups together with the number of participants. Subsequently, links were made between core groups to indicate potential direct and indirect comparisons (Figure 5).

2. To explore further the potential direct and indirect comparisons, a second figure was constructed (Figure 6). This shows the same groups presented in the first step with the additional information on the number of trials and total number of participants contributing to each of the comparisons.

3. Finally, to allow for any meaningful comparisons, we split the groups mentioned in steps 1 and 2 into three broad categories, namely active physical (exercise and graded activity), passive physical (individual physiotherapy, manipulation and acupuncture) and psychological (advice/education and psychological therapy) (Table 10).

FIGURE 5.

Step 1: classification of trials into core groups. a, Plus sham advice/education; b, plus sham electrotherapy.

FIGURE 6.

Step 2: classification of trials with indication of number of trials and participants for direct and indirect comparisons. m, Number of trials; n, total number of participants.

In this programme of work we were not seeking to estimate the true effect size of any individual intervention. Rather, we were seeking to identify predictors of treatment response. These analyses were constrained by the availability of data on potential moderators that could be pooled across trials. Considering the potential mechanisms through which the potential moderators might affect outcome, the study team concluded that it was reasonable to pool interventions that might under other circumstances appear rather heterogeneous. In particular, the decision to include several superficially different interventions as passive physiotherapy might surprise some readers. Our view, however, is that these are very distinctly different from active exercise-based interventions, or those working through a psychological approach. Essentially, they all consist of an assessment, whatever reassurance and education is provided as part of the treatment session, plus whatever modality is being offered, be it massage/mobilisation/manipulation or needling. We consider these to be conceptually sufficiently close in their mode of action that it is unlikely there will be distinctions in how the potential moderators included in our analyses might affect outcomes. They are, however, distinctly different from their active physical or psychological interventions in how treatment moderation might operate.

In organising the data we also identified combined interventions but there were too few data points for it to be worthwhile pursuing these analyses. For this reason these were excluded from our final analyses.

Typographical conventions

This chapter presents the methods we used to create the repository database. To distinguish database vocabulary and commands from regular texts, different typographical fonts are used. Database object-class vocabulary is printed in sans-serif font [like this] and the command for mapping and transformation procedures is printed in monospaced typewriter font . In addition, coloured command fonts in the text are for ease of referencing between program commands shown in figures and text explanations.

Background

Clinical trial data sets can be stored in a tabular format, for example Microsoft Excel^® or SPSS (Statistical Package for the Social Sciences). A tabular format typically uses each row to represent data from a participant and each column to represent an item from a case report form (CRF).

Tabular formats have the advantage of being intuitive, relatively simple to create and machine readable. However, this format can be susceptible to excessive growth, especially when clinical and non-clinical items are measured across multiple time points. Data collected for withdrawn participants or non-responders would still require columns for all variables irrespective of whether or not they were used. Repeating questions pose a similar problem, whereby storage space must be allocated across the whole domain to accommodate all responses. For example, asking for a participant’s medical history of prescribed drugs would require a new column to be added for every drug listed. If only one participant documented a long list of drugs then many columns would have to be created for all participants.

Tabular formats are effective for only the smallest of trials and quickly become inefficient and difficult to maintain when the range of data collected increases. For larger trials, a more robust solution is to use a relational database. The relational database model allows individual tables to be created for each CRF and for repeating sets of questions. Normalisation rules are often applied to define the columns for each table and the logical relationships are used to create table joins. 146

Figure 7 shows sample data in a tabular format and the normalised equivalent in a relational database. The sample data consist of the subject identification, recruitment date, demographic data and the RMDQ scores taken at baseline and at 3-month follow-up. The data are normalised into four tables, namely SUBJECT, DEMOGRAPHICS, RMDQ (for the RMDQ measurement) and FU (follow-up); the last is used to store the time points for each follow-up visit.

FIGURE 7.

(a) A sample of original tabular format data; and (b) normalised relational interpretation of the original tabular data.

Each table has a primary key (PKey) column for storing a unique record identifier that is used as the basis for creating relationships between tables (see Figure 7b). The relationship between SUBJECT and DEMOGRAPHICS is one-to-zero-or-one, that is a subject can have zero or one demographic record. The PKey from the SUBJECT table is copied to the DEMOGRAPHICS tables, thereby creating a join using a shared value.

The relationship between SUBJECT and RMDQ is one-to-zero-or-many, that is, a subject can have zero or many RMDQ completed questionnaires. The FU table is joined to the RMDQ table using a one-to-zero-or-many relationship. This join allows a RMDQ score to be associated with either a baseline or a 3-month follow-up time point.

To create the relationships to the RMDQ table, the PKeys from both the SUBJECT and FU tables are added as foreign keys (FKeys). This has the result of allowing a subject to have either zero or many RMDQ scores at all time points. A composite unique constraint is applied to the Subject FKey and FU FKey columns to prevent a subject from having duplicate RMDQ scores for the same time point.

The repository differs from a typical clinical trial database in that it is not possible to predetermine requirements by using annotated CRFs. The repository relies on data from multiple trials to be periodically reviewed and classified, and must be frequently altered to accommodate new discoveries. The relational database is not a suitable model for such a scenario because modifications to the schema can be time-consuming and complex, often requiring the expertise of information technology specialists. Thus, the database for this project needs to be flexible so that the end users, namely, statisticians and health economists, can carry out modifications without having to change the database schema.

Our solution is to create a hybrid database that is a cross between an entity-attribute-value (EAV) open schema model and a relational database. This hybrid database has the flexibility of storing sparse heterogeneous data, which allows dynamic changes while enforcing data integrity.

The next section describes the architecture of the hybrid database. The rules used to map and transform the original source data to the repository standard are described below in Mapping and transformation. Using entity-attribute-value data shows how the repository database is manipulated, such that the data can be viewed in an analysis-friendly format from any statistical program that supports Open Database Connectivity (ODBC). Extract, transform and load describes how data from multiple RCTs were extracted, transformed and harmonised to the repository standard and, finally, loaded to the repository database.

System architecture

Tables and columns in a relational database can be represented as classes and attributes in an EAV model. 147 In the subsequent text the terms ‘class’ and ‘attribute’ will be used to conform to the EAV vocabulary. The term ‘entity’ is interchangeable with the term ‘object’ and can be thought of as providing a similar role to a table row but with the significant difference of storing only a pointer to the data and not the actual data itself. The entity–relationship diagram for the hybrid database is shown in Figure 8.

FIGURE 8.

The entity–relationship diagram for the hybrid repository database depicting the fixed schema with the subschema EAV tables. Bold text represents required parameters.

We anticipated that there would be some consistent data present in all of the RCTs for describing the trial and for identifying the trial’s subjects. The two tables Primary Source and Subject were created with fixed schemas to store this data (see Figure 8). The Primary Source table stores the name of the RCT (prms_TrialName), a brief description of the trial (prms_Description) and the date on which the data were imported into the repository (prms_ImportDate). The Subject table stores the original identifier assigned to the trial participant (subj_OriginalID), the date the participant enrolled into the trial (subj_EDate), the date the participant was randomised (subj_RDate) and a unique identifier generated by the system (subj_ID). A foreign key relationship is created to link each subject to the Primary Source.

The EAV model uses a subschema consisting of tables for classes, attributes, objects and the EAV data. The Class table is used to hold a list of all the identified domains, for example RMDQ and demographics. These domains generally map to a CRF but can also be used to describe a subset of repeating questions, for example repeated medical prescriptions.

The Attribute table is used to hold a list of all identified variables that typically map to a CRF question. The Attribute table has columns for storing a short name, a verbose name, a reference to the containing class and data type details. The short name is used to store a standardised version of the original CRF question.

The Object table stores a unique identifier for each instance of a class and a reference to the class itself. A foreign key relationship is created to link each Object to a Subject. This relationship essentially makes the EAV model subject centric, that is, all of the data stored in the Object and EAV tables must be directly related to an imported subject. Relationship between objects is possible by using an ‘ancestor column’ to store the unique identifier of a related object. For example, an object used for repeated medical prescriptions will store the unique identifier of the related follow-up object in the ‘ancestor column’.

The EAV data table has three columns and is used to store all of the repository’s RCT data. Two columns hold references to the related objects and attributes, with the other column used for storing the actual value of each object–attribute combination. The references to the objects and attributes take the form of foreign keys to the object and attribute tables. The format of the value is coerced into a string regardless of the intended data type. The intended data type – for example binary data, small integers or strings – details are stored in the related attribute table.

A simplification of how tabular data are represented in an EAV table is shown in Figure 9. In this example, the tabular data have one row for each subject (see Figure 9a). When the data are shown in the EAV table there are four rows for subject #1000, three rows for subject #1001 and three rows for subject #1002. For each populated cell in the tabular data a row is created in the EAV table. Subject #1000 has all cells populated and, therefore, has a row for each entry. Only three rows are entered for the other subjects because there was no RMDQ baseline score for #1001 and age was not recorded for #1002 (see Figure 9c).

FIGURE 9.

(a) A sample of original tabular format clinical data; and (b) the XML mapping and transformation instructions; and (c) the sample data represented as EAV. XML, extensible mark-up language.

In reality the EAV table will use the column Attribute ID to store the unique attribute identifier and not the text value as shown in Figure 9c. In addition, the column Object ID stores a reference to the object and not the subject ID. It is the related object that links back to the subject and to the class.

Mapping and transformation

Early evaluation of data sets from various RCTs in the project identified large variations between variable naming and coding conventions. For example, the RMDQ was used to measure back pain disability and the participant would tick all of the items that were applicable to him/her on that day. There are 24 items in the questionnaire and the score is the sum of all of the ticked items. One trial might name each column ‘rm1’, ‘rm2’ and so on until ‘rm24’ for all 24 individual items and ‘rmscore’ as the RMDQ score measured at baseline, ‘rm1_3mo’, ‘rm2_3mo’, . . ., ‘rm24_3mo’ and ‘rmscore_3mo’, for the 3-month follow-up data, and so on. Another trial might name them ‘rdq1’, ‘rdq2’, . . ., ‘rdq24’ and ‘rdq’ for items measured at baseline, ‘rdq11fu’, ‘rdq21fu’, . . ., ‘rdq241fu’ and ‘rdq1fu’ for items measured at the first follow-up, which could have been 1 or 3 months post randomisation, depending on the protocol. In addition, some trials might use the numerical value ‘1’ to represent a tick for that item and ‘0’ if it was not ticked. Other trials might use ‘1’ as ticked and ‘2’ as not.

Mapping and transforming health-care resource-use data

Mapping health-care resource-use variables was more challenging because the different types of resources used across all RCTs do not conform to any standard and are completely variable. However, each question and answer in a typical health-care resource-use questionnaire can be broken down to the recall period, the type of resource, the reason for using the resource, the location of the resource, the unit of measurement, the quantity, the cost or expenses incurred and the payer.

Figure 10 shows a simplified version of a typical health-care resource-use questionnaire. In this example, participants were asked to record all of the health-care resources that they used at the 3-month follow-up time point (see Figure 10a). The answers provided by the participants were stored in a tabular format, which used 12 columns to capture all of the responses to the five questions (see Figure 10b). By using this format, the number of required columns to accommodate the data would grow in line with the maximum number of responses provided by any one individual. For example, if only one participant listed three items he/she bought over the counter to treat his/her LBP, the number of columns required would have to be increased from 12 to 13.

FIGURE 10.

(a) A sample of questions in a CRF at 3-month follow-up; (b) a sample of original tabular format health-care resource-use data; and (c) a sample of how the health-care resource-use data populate the repository standard. CC, community clinic; GP, general practice; IND, individual; N/A, not applicable; PHS, public health service; Pri, primary care; Physio, physiotherapist; Priv, private clinic; Px, prescription; RP, recall period.

Figure 10c shows a view of the repository health-care resources data, generated from the EAV tables. This view displays the eight standard repository health-care resource-use attributes (table columns) and an additional attribute called ‘Text’, which is used to store all of the characters that are captured as comments in the CRF.

The process for creating the transformed health-care resource-use data involves splitting the original questions into a number of derived parts that will map to the standard attributes. For example, question 1 asked how many times the participant had consulted his/her doctor or any primary care doctor, for any reason, in the last 3 months. Using the information contained in the question, the recall period is set to ‘3’, the type of resource is ‘GP’, the reason for using the resource is ‘Any condition’, the location of the resource is ‘Primary Care Setting’, the unit of measurement is ‘Visit’ and the payer is ‘Public Health Service’. All of these values are derived solely from the information contained in the original question as opposed to the value of the variable. Only the attribute ‘Quantity’ is directly mapped to the original variable’s value.

The health-care resource-use data are stored in the EAV tables by creating relationships between objects. For each time point, one or many resource-use objects can be created. The HE class is used to define the time points for collecting only the health-care resource-use data. The actual resource-use data are defined in the HE-DATA class and the time point value is used to link an HE-DATA object to an HE object. The XML schema was modified to allow related classes to be described, which, in turn, gets interpreted by the system to create the relationships in the Object table.

Figure 11 shows the HE-DATA class being used as a child class, that is, it has the HE class as its parent. Creating child classes signifies to the system that a relationship exists between two classes. The linkedValue attribute is used to specify a shared value between the parent and child classes. In a relational database, this shared value would be created as a foreign key constraint. In the example shown in Figure 11, an HE class has been defined for the 3-month follow-up time point using the attribute . A child HE-DATA class has been defined and linked to the parent HE class by specifying the value for the linkedValue: . This corresponds with the 3-month follow-up time point specified in the HE class.

FIGURE 11.

The XML mapping and transformation instructions for the sample data in Figure 10.

Child classes in the XML use elements to signify the number of objects that need to be created. In a relational database, this would result in adding a new element for every table row to be inserted. The value for the element has no significance except that it must be unique. In the example shown in Figure 11, six groups have been created for the 3-month resource-use data, namely , , , , and . These groups represent each question in the CRF shown in Figure 10a and the data shown in Figure 10b.

The original tabular data required 13 columns across three rows to store all of the data for the three participants. Instead of creating a new column for every resource, the repository creates a new object. The seven groups are used to create objects for GP visit (), NHS physiotherapist visit (), private physiotherapist visit (), two instances of prescribed medicine (, ) and two instances of aids or medications bought over the counter (, ). Although seven groups have been defined in this example, the ETL system will create objects only where data exist. For example, subject #1000 will create only four objects for GP visit (), NHS physiotherapist visit (), private physiotherapist visit () and medicine prescribed by GP ().

Once all resources have been identified and a group has been defined, the mapping rules are used to populate the repository’s standard resource-use attributes. Within the structure, the is used to allow the system to locate the correct object to process and the is used to store the name of the original variable. The element stores the name of the mapped repository attribute.

The original variable stores the quantity of doctor visits and hence is mapped to the repository attribute for the group . The other information required to make sense of this value are hard coded to the repository standard within the structure, which is within the structure. For example, the recall period (), the type (), the reason (), the location (), the unit () and the payer () of the resource allocated in group is hard coded to , , , , and , respectively (see Table 11 for list of values and corresponding labels). These values can be hard coded in the XML because they are known to be based on the CRF and do not affect the original data. When the system processes this mapping instruction, subject #1000 would have a health-care resource-use object that shows that there was one GP visit made during the 3-month follow-up time point (see Figure 10b).

Other rules can be applied to manipulate the original health-care resource-use data. For example, the original medicines prescribed have to be transformed to the repository standard to the standardised drug coding. Figure 11 shows a transformation for the attribute that uses a match rule to check for the value . If matched, the rule has been set to update the attribute’s value to .

The XML mapping and transformation instructions shown in Figure 11 were based on only one follow-up time point. For mapping data from more than one follow-up time point, simply create more HE objects, and map and transform health-care resource data within the child class HE-DATA that is linked to that follow-up time point, for example:

Using entity-attribute-value data

Using the EAV with classes and relationships data in its raw state for any kind of analysis work would be extremely difficult because of the fragmented nature of the EAV schema. For analysis purposes, it is therefore necessary to piece together the data to form complete data sets that are comparable to the data sets outputted from relational or tabular data sources. This task is achieved by processing the EAV table to derive a table for each class, a column for each attribute and a row for every object. An excerpt of the SQL statement to join the various data to extract the required data items for class RMDQ (whose identifier is in this example) is shown below:

The statement produces a table in a long format, which was subsequently pivoted to produce a row for each object and a column for every attribute. The outcome of this query is a data set that resembles a tabular structure that can easily be processed for further analysis.

Although this solution provides a means for generating a usable tabular format, the scalability is severely limited. The server performance was found to decrease as the volume of data increases, and multiple pivot operations were used for transforming object relationships. Querying the derived data sets directly was also impractical because of the huge numbers of data that can be generated in the server’s temporary database, causing the server to be unstable.

An initial solution used to overcome these issues was to disconnect from the actual query by using the in-built functionality of the statistical analysis software to create a copy of the query results. A more permanent solution, which is the current practice, is to periodically create a copy of the query results into actual tables within the database.

Extract, transform and load

The bespoke ETL application was required to read the original source data, automatically apply mapping and transformation rules from an XML document, and load the processed data into the repository. In addition to these basic functions, the ETL application was also required to permit end users to set up new RCTs for import, create new classes and attributes and make changes to existing ones, and to switch between a testing and live environment.

The bespoke ETL application was distributed as a Microsoft Windows^® desktop application. It works by first uploading the original data set and the XML mapping and transformation rules. The instructions defined in the XML file are applied to the original data set and the transformed data are loaded into the repository database. The ETL application allows the statistician and health economist to execute these steps from their desktop computers. The ability to switch between a test and live environment gives the users the flexibility and convenience of checking whether or not the instructions that they have delineated in the XML file are correct before loading the data sets into the live database.

Data validation

Data integrity is vital throughout the repository ETL process. To check that the mapping and transformation procedures were carried out correctly, the repository data were routinely checked against the original data sets. To achieve this, at each time point (baseline and all follow-ups), a random sample of data was extracted and manually cross-checked against the source data. Any inconsistencies were flagged and, if required, the XML instructions were amended. This process was repeated until the data were deemed to have been transformed correctly.

Storage

In condition of our data sharing agreements to hold the RCT data sets and to meet local governance and standard operating procedures, the repository database server is held in a secure data centre, with robust disaster recovery policies in place.

The appeal of having this hybrid system architecture is that the structure takes up very little space in the server, and the time needed to query and retrieve data is very little, too. Naturally, the disk space needed to store the data in this repository will grow in proportion in accordance to the number of data points.

Future data sharing

At the end of this programme of work we would like to make the pooled data available for future analyses. We will go back to all of the principal investigators (PIs)/data custodians with a new data sharing agreement to enable us to share their pooled data. Once these agreements have been signed, we will set up a website with details of how to apply for the data. All requests will be:

1. forwarded to the study statistician who will carry out internal checks to ensure that the data being requested can be provided; the response from the study statistician will be supplied with the original request for the independent committee consideration

2. sent via e-mail to an independent committee, who will review the application and make a final decision on data sharing; for the data requested, if a PI/data custodian has:

   1. agreed to sharing the data but has asked to see a copy of the request, a copy will be sent to them via e-mail for information purposes only

   2. not agreed to sharing their data, this data set will be removed from the pooled data before providing the requested data to the applicant.

Background

There are six PROMs that have been used in one or more studies within the repository that aim to measure back pain-related disability, namely the Chronic Pain Grade Scale (CPG) disability score (CPG-DS), which is one of the two domains in the CPG that aims to grade chronic pain status,150 Hannover Functional Ability Questionnaire for measuring back pain-related functional limitations (Funktionsbeeintrachtigung durch Ruckenschmerzen) (FFbHR),151 Oswestry Disability Index (ODI),152 Pain Disability Index (PDI),153 Patient-Specific Functional Scale (PSFS)154 and RMDQ. 28 Some trials also included generic health-related quality-of-life instruments, such as the Short Form questionnaire-12 items (SF-12)155 or the Short Form questionnaire-36 items (SF-36),156 for which the physical component score (PCS) measures the physical functioning. As mentioned later in Chapter 6 (see Outcome variables), no common instrument was used by the trials that were included in the repository. We sought to assess the agreement of these instruments by determining their correlation and responsiveness at a trial level, in order to decide whether or not data pooling was feasible. After we had completed this work, a National Institutes of Health task force identified developing crosswalking values for ‘legacy’ measures of back pain outcome as a key priority for back pain research. 157

Data

We used data from 11 trials that had used at least two of the following measurements: CPG, FFbHR, PCS, PSFS, PDI, ODI and RMDQ. For all of these analyses we used the short-term change score, as this is where any treatment effects are likely to be greatest. For the purposes of this report we have defined a short-term follow-up as a measurement taken at between 2 and 3 months post randomisation or entry to the trial. The short-term change score is the difference between the baseline and the short-term follow-up (see Chapter 6, Follow-up time point). In each case we have standardised the reporting so that a positive change score is interpreted as an improvement. Where appropriate, we used the standardised response; change score divided by the SD of the change. We used this in preference to the standardised effect size (change score divided by the SD of the measure at baseline), so that all of the standardised scores had a SD of one. This enables visual comparisons to be made between all of the scatterplots.

Outcome conversion

All comparisons between instruments were carried out at an individual trial level. Each pair of outcome measures was fitted with simple linear regression models. Denoting the change scores for the two outcome measures by x and y, the simple linear model was:

where the intercept, α, and the coefficient, β, are parameters to be estimated and ϵ is the error term. For the conversion to be meaningful, the standardised change scores have to be correlated and have similar responsiveness; the latter is explained below. 158

Correlation

Correlation was assessed by scatterplots and Pearson’s correlation coefficient, with a correlation coefficient considered to be at least moderately high if it was > 0.5.

Responsiveness

Responsiveness is the ability to detect a change in condition; if a participant’s condition improves or worsens over time then this should be reflected by a change in the participant’s score. If two outcome measures do not have similar responsiveness then combining them in a meta-analysis may introduce heterogeneity that could be falsely attributed to other sources, such as the treatment effect.

Similarity of responsiveness of two outcome measures was examined by categorising the change scores as negative change (change score of < 0), no change (change score = 0) or positive change (change score of > 0), and applying Cohen’s kappa (κ) to these categorisations. 159 We considered κ > 0.4 to indicate sufficiently similar responsiveness. 160 These broad categories were chosen to demonstrate whether or not the outcome measures had similar responsiveness in the most basic sense (improved, worsened or no change). We also planned to examine narrower categorisations in the event that the agreements within these three categories were good (κ > 0.4). However, as there was no standard on the levels of categorisations, a few would be examined.

For it to be acceptable to pool two measures, they needed to meet two criteria; to be at least moderately correlated (correlation > 0.5) and to have at least moderately similar responsiveness (κ > 0.4).

Results

Eleven trials31,33,50,76,101,103,104,107,131,132,134 (n = 6089) and seven instruments were included in these analyses (Table 12). There was a total of 21 within-trial pairwise comparisons between two outcomes. Figures 12–16 show scatterplots of standardised change scores for each such pair of outcome measures. See Appendix 8 for scatterplots between raw change. It is clear from these plots that the outcomes were positively correlated. Note also that the standardised change scores were widely scattered around the reference line, suggesting that there was a lack of agreement between the outcomes.

FIGURE 12.

Scatterplots of standardised change scores for (a) PCS vs. CPG (n = 2451); and (b) PCS vs. FFbHR (n = 3620) outcome measures. PCS of SF-12/36.

FIGURE 13.

Scatterplots of standardised change scores for (a) PCS vs. RMDQ (n = 1694); and (b) PCS vs. ODI (n = 206) outcome measures. PCS of SF-12/36.

FIGURE 14.

Scatterplots of standardised change scores for (a) PCS vs. PSFS (n = 158); and (b) CPG vs. FFbHR (n = 1110) outcome measures. PCS of SF-12/36.

FIGURE 15.

Scatterplots of standardised change scores for (a) CPG vs. RMDQ (n = 1661); and (b) PSFS vs. RMDQ (n = 625) outcome measures.

FIGURE 16.

Scatterplots of standardised change scores for (a) PDI vs. PCS (n = 281), and (b) FFbHR vs. and PDI (n = 284) outcome measures. PCS, physical component scale of SF-12/36.

The correlations between outcomes ranged from 0.21 to 0.70; implying that the linear associations between them range from weak to moderately strong (Table 13). Three trials50,101,132 had both SF-12/36 PCS and FFbHR data, and their correlations were very similar, about 0.58. Another three trials33,101,132 had both SF-12/36 PCS and CPG, and the correlations were reasonably similar, ranging from 0.41 to 0.56, and four trials31,33,76,134 had both a SF-12/36 PCS and a RMDQ score with range 0.38–0.52, again similar. However, correlations between other outcomes were quite wide ranging: between CPG and RMDQ scores (m = 3 trials;31,33,104 range 0.21–0.47) and between PSFS and RMDQ scores (m = 3;103,131,134 range 0.40–0.70).

Cohen’s kappa was < 0.4 for all 21 comparisons. Some were similar between trials, namely for PCS and FFbHR (range 0.27–0.30) and for PCS and CPG (range 0.27–0.31). However, the level of agreement was never more than fair. 160 As the Cohen’s kappa agreement was not > 0.4 narrower categorisations were not investigated.

There were no pairs of outcomes that satisfied both criteria of at least moderately correlated (correlation > 0.5) and at least moderately similar responsive (Cohen’s kappa > 0.4). Therefore, it was not meaningful to convert any outcome to another one.

Conclusion

In view of the lack of correlation and responsiveness, it is not recommended to map any physical disability outcome measures to another considered in this investigation.

For each of our subsequent analyses we have pooled data only where the same participant-reported outcomes are available from multiple trials. The one exception is that the SF-12 and SF-36 are explicitly designed to have similar measurement properties when converted into their physical and mental component scores. We have therefore pooled the mental component score (MCS) and PCS from studies using SF-12 or SF-36.

Background

In this chapter we present the results of preliminary statistical analyses performed on the individual participant data, specifically the analysis of covariance (ANCOVA) comparing all treatments with all controls (usual care plus sham) to identify individual potential moderators to take forward into our main analyses. The methodological development work to identify multiple covariates’ baseline characteristics that moderate treatment effect is presented in later chapters (see Chapters 7–10). We do not, in this preliminary analysis, seek to define subgroups using multiple parameters.

Statistical analysis plan

In accordance with the standard operating procedure in the Warwick Clinical Trials Unit, a detailed statistical analysis plan was written by the study’s statistician (SWH) and health economist (JJ). The plan was subsequently reviewed and approved by the study team and members of the repository oversight committee (see Appendix 9), whereas the overview of the plan is described in the following sections.

Definitions

Data sets

Individual participant data without treatment assignment were excluded from the repository. This exclusion criterion applies to individual participants whose data were included in the data set but the treatment allocation was not available in the data set. We were not able to allocate these participants to a treatment group and they were thus excluded.

Methods

Results

One-step meta-analysis

Box plots of change of outcome measures from baseline to short-, mid- and long-term follow-up by treatment arms show that participants in all groups are behaving as expected, with all groups improving over time (data not shown). This observation was examined further in the one-step meta-analysis (adjusting for study effects) and the results are shown in Figures 17–19 and Table 16. There was a statistically significant difference between control and intervention for all outcomes at the short-term follow-up.

FIGURE 17.

The estimated efficacy between control (non-active usual care and sham) and intervention treatments from one-step meta-analysis for (a) FFbHR; and (b) RMDQ score. AT, number of participants in the intervention arm; Est (95% CI), estimated treatment efficacy and 95% CI; m, number of trials; UC, number of participants in the control arm.

FIGURE 18.

The estimated efficacy between control (non-active usual care and sham) and intervention treatments from one-step meta-analysis for (a) average pain (based on VAS); and (b) PCS of SF-12/36. AT, number of participants in the intervention arm; Est (95% CI), estimated treatment efficacy and 95% CI; m, number of trials; UC, number of participants in the control arm.

FIGURE 19.

The estimated efficacy between control (non-active usual care and sham) and intervention treatments from one-step meta-analysis for (a) MCS of SF-12/36; and (b) EQ-5D. AT, number of participants in the intervention arm; Est (95% CI), estimated treatment efficacy and 95% CI; m, number of trials; UC, number of participants in the control arm.

TABLE 16 - One-step meta-analysis: estimated mean change from baseline to short-term follow-up by treatment arms and the estimated difference between treatment arms (95% CI)a

Adjusted by random intercept, trial and interaction between treatment and trial effects.

Control, usual care/GP and sham control.

Difference, intervention−control (thus, positive = favours intervention arm).

Obtained from either VAS or CPG-PS (see Selection of instrument, above).

PCS of SF-12/36.

MCS of SF-12/36.

Outcomes	Number of trials, m	Intervention	Controlb	Differencec	p-value
FFbHR	3	n = 1841	n = 2118		0.0165
		13.88	5.80	8.08
		1.24 to 26.51	–6.93 to 18.53	3.46 to 12.69
RMDQ	8	n = 1778	n = 897		< 0.0001
		4.43	2.97	1.46
		1.56 to 7.29	0.10 to 5.84	1.10 to 1.81
Average paind	10	n = 2061	n = 1546		< 0.0001
		18.03	11.57	6.46
		8.65 to 27.41	2.18 to 20.97	4.86 to 8.06
PCSe	6	n = 2793	n = 2415		0.0006
		6.86	3.72	3.15
		4.90 to 8.83	1.75 to 5.68	1.99 to 4.30
MCSf	6	n = 2793	n = 2415		0.0044
		2.69	0.62	2.07
		1.54 to 3.84	–0.55 to 1.79	0.93 to 3.20
EQ-5D	4	n = 1271	n = 503		< 0.0001
		0.1065	0.03422	0.072
		0.008 to 0.205	–0.059 to 0.127	0.04538 to 0.099