Moderators of the Effect of Spinal Manipulative Therapy on Pain Relief and Function in Patients with Chronic Low Back Pain: An Individual Participant Data Meta-analysis.

STUDY
DESIGN: Individual participant data (IPD) meta-analysis.
OBJECTIVE: The aim of this study was to identify which participant characteristics moderate the effect of spinal manipulative therapy (SMT) on pain and functioning in chronic LBP. SUMMARY OF
BACKGROUND: The effects of SMT are comparable to other interventions recommended in guidelines for chronic low back pain (LBP); however, it is unclear which patients are more likely to benefit from SMT compared to other therapies.
METHODS: IPD were requested from randomized controlled trials (RCTs) examining the effect of SMT in adults with chronic LBP for pain and function compared to various other therapies (stratified by comparison). Potential patient moderators (n = 23) were a priori based on their clinical relevance. We investigated each moderator using a one-stage approach with IPD and investigated this interaction with the intervention for each time point (1, 3, 6, and 12 months).
RESULTS: We received IPD from 21 of 46 RCTs (n = 4223). The majority (12 RCTs, n = 2249) compared SMT to recommended interventions. The duration of LBP, baseline pain (confirmatory), smoking, and previous exposure to SMT (exploratory) had a small moderating effect across outcomes and follow-up points; these estimates did not represent minimally relevant differences in effects; for example, patients with <1 year of LBP demonstrated more positive point estimates for SMT versus recommended therapy for the outcome pain (mean differences ranged from 4.97 (95% confidence interval, CI: -3.20 to 13.13) at 3 months, 10.76 (95% CI: 1.06 to 20.47) at 6 months to 5.26 (95% CI: -2.92 to 13.44) at 12 months in patients with over a year LBP. No other moderators demonstrated a consistent pattern across time and outcomes. Few moderator analyses were conducted for the other comparisons because of too few data.
CONCLUSION: We did not identify any moderators that enable clinicians to identify which patients are likely to benefit more from SMT compared to other treatments.Level of Evidence: 2.

Low back pain (LBP) is the world's leading cause of disability.[1] Nonpharmacological approaches are the first choice of treatment.[2] The treatment options include spinal manipulation and mobilization which are used by a variety of heath care providers such as osteopaths, chiropractors, and physiotherapists. These approaches can be used together or alone to treat patients with chronic LBP, and collectively defined as spinal manipulative therapy (SMT).

Many systematic reviews and meta-analyses have found that SMT is an effective treatment for patients with chronic LBP with a modest mean effect compared to other interventions.[3-8] Although SMT can relieve LBP in some patients, it is not effective for everyone; the number needed to treat is in the range of five to ten.[9] One potential explanation is that patients with “nonspecific” chronic LBP have different characteristics that influence the intervention effect, whereas another explanation can be the variation in duration, number, and type of SMT. Relevant subgroups of patients with chronic LBP may exist that might benefit more or less from SMT.[10] The first step in identifying these subgroups is to examine which participant or treatment characteristics moderate the treatment effect (e.g., age, duration of LBP).[10-12] These moderators are typically not presented in the traditional meta-analyses[13] because aggregate data on relevant patient characteristics are often not available, are poorly reported, or derived and presented differently across studies. More importantly, if the results of subgroup analyses are reported, group averages or proportions are presented, which can result in ecological bias. The patient-level intervention-covariate interactions are usually not examined or reported, even though they have the potential to better target the intervention.[14] Although some authors have presented appropriate analyses of treatment moderation, few trials are large enough to exclude important moderator effects.[10]

One way to test interactions of these characteristics with the intervention is to use individual participant data (IPD). IPD provide much increased statistical power and allow for standardized analyses across studies, using direct derivation of information desired on an individual level, independent of whether and how it was reported in original publications. Therefore, IPD potentially allows identification of clinical characteristics of patients with chronic LBP that may moderate treatment effects.

The specific objective of this IPD meta-analysis is to identify individual participant characteristics measured at baseline that moderate the effect of SMT for pain and function at 1, 3, 6, and 12 months in adults with chronic LBP versus recommended interventions, nonrecommended interventions, sham SMT, SMT + other intervention versus SMT only, and mobilization versus manipulation.

METHODS

This study was conducted and reported according to the Preferred Reporting Items of Systematic Reviews and Meta-Analyses for IPD (PRISMA-IPD) guidelines[15] (appendix eTable1). The protocol was registered with PROSPERO (https://www.crd.york.ac.uk/prospero/display_record.php?RecordID=25714) and approved by the Scientific Review Board of the Vrije Universiteit Amsterdam and by the Ethical Committee of the VU University Medical Centre Amsterdam. (Projectnr. 2015.544).

A detailed protocol has been published previously.[16] The methodology presented in the present article gives an overview of the moderator selection and analysis, whereas the eligibility criteria, search methods for identification of new trials (appendix eTable 11), risk of bias assessment, funnelplots (appendix eFig 1 and 2), collection, and extraction of IPD are fully described in the published protocol.[16]

Types of Outcome Measures

Primary outcomes were pain (reported on a 0–10 or 0–100 NRS or VAS scale) and back-specific function on any scale, for example, Roland Morris Disability Questionnaire, Oswestry disability Index. All outcomes were self-reported and converted following decision rules (appendix eTable 4).

Moderators of Treatment Effect

Candidate moderators of treatment response were identified by the research team a priori based on consensus (appendix eTable 5). In short, the selection of patient moderators was based on a specific rationale (e.g., understanding behavioral and sociocultural mechanisms by which response is modified or from prognostic research [treatment effect modification studies or prognostic factor research])[10-12] (see protocol).[16] Of 23 potential moderators identified, six were not analyzed because data were insufficient, unavailable, or only available at study level (patient preference/expectancy, comorbidities, alcohol use, income, nonspecific LBP, and for all treatment characteristics). The number and frequency of SMT treatments were measured at study level and not at patient level in the vast majority of the trials. The same accounts for type of SMT technique used. Therefore, contrary to the description in our protocol, we could not analyze moderating effects by these types of variables.

For psychological factors, analyses were only performed for combined depression scales. For other psychological scales, there were insufficient data.

Moderator analyses were classified into confirmatory or exploratory. Moderators in confirmatory analyses are those related to specific theory or evidence, whereas moderators in exploratory analyses relate to moderators for which no empirical evidence exists or for which a specific theory or mechanism is lacking. Our potential confirmatory moderators were age, sex, duration of LBP, psychological factors, treatment preference/expectation and baseline pain, function, and quality of life. Other moderator analyses were exploratory (appendix eTable 5).[17] In both cases, the analytical technique was the same, but for the interpretation of those confirmatory moderators the evidence was considered to be stronger.

Preparing Data for Moderator Analyses

If data on a variable of interest were not available, we attempted to extract this information based on other data in the trial (e.g., information about employment was missing, but there was a variable on sick leave). Whenever possible, we used continuous data as presented, unless dichotomizing facilitated the translation of findings to clinical practice or was needed to meaningfully combine data across trials. The cutoff points were determined by consensus of the steering committee (ADZ, MRdB, SMR, MWT, and RO) (Appendix eTable 5). For age, we used 65 years as a cutoff point. Additionally, duration of LBP was dichotomized into <1 year versus >1 year. Similarly, physical activity was categorized into low (one or less exercise sessions per week), medium (two to three exercise sessions per week), and high (more than exercise sessions per week) and subsequently dichotomized into low-medium or high. The choice of cut-off for physical activity was further evaluated in a sensitivity analysis (low vs. medium-high).

For the outcome pain, all pain scores were converted to a 0 to 100 points scale following a decision rule (appendix eTable 4). To allow pooling of different functional status measures, we recoded the individual scores into z scores. For each separate time point using pooled standard deviations as nominator . Analyzing these z scores resulted in standardized mean differences (SMDs). To ease interpretation of SMDs, we converted these to a mean difference (MD) for the 24-point Roland Morris Disability Questionnaire, by multiplying the SMD with the population standard deviation (SD) of the studies measuring Roland Morris Disability Questionnaire ( ni = sample size for each trial; S = standard deviation for each trial).

Data Analysis

We studied moderators of intervention effects when three or more trials within a comparison had data on the moderator and the outcome at a specific time point. We used the following comparisons: 1) SMT versus recommended interventions including nonpharmacological treatment (e.g., exercise) and pharmacological treatment (e.g., NSAIDs, analgesics); SMT versus nonrecommended interventions (e.g., light massage, diathermy, ultrasound; sham “placebo”SMT; SMT + intervention versus intervention alone; high-velocity low-amplitude SMT versus low-velocity low-amplitude SMT (i.e., manipulation vs. mobilization).[16]

Potential moderators were analyzed using a one-stage random-effect IPD meta-analysis. The baseline outcome, treatment, potential moderator, and interaction between treatment and moderator were included as fixed effects. Study-specific intercepts were also included as fixed effects. Random treatment and interaction effects were added to the model (see protocol[16] and Eq. (2) in appendix eTable 6).[18] We performed these analyses for each time point and each moderator separately to facilitate convergence of models. Centering the patient-level covariates about their study-specific means enabled us to separate the within- and across-study interactions.[19] The within-study interaction explained the patient-level variation in treatment response, while the across-study interaction represented the moderator effect on study level. We present the within-study interactions. A negative interaction coefficient indicates a more positive or less negative estimate of the intervention effect of SMT vs comparison for the index group compared to a/the reference group (e.g., females compared to males).

We refrained from presenting stratified results for subgroups of moderator variables, because these include a combination of within- and across-study information because of differences in proportions of persons within the separate subgroups between studies.

Synthesis of Evidence

Assessment of clinical relevance for the main effects analyses was defined as a small, medium, or large difference and based upon the recommendations of the Cochrane Back and Neck group:[20]

In a consensus meeting with the project group, we discussed our results to determine whether a moderating effect was present. We considered a moderator effect to be present if the magnitude of the effect was at least half of our pre-specified clinically relevant main effects; that is, more than 5-points (on a 100-point pain scale) or more than 0.25 for SMD on function and there was consistency in the direction of the moderators across three consecutive follow-up intervals for both pain and function. The arbitrary cut-offs of 5-points or 0.25 SMD were used to detect small differences within a moderator as SMT is a low intensity, low cost, intervention.

As a crude method to guide interpretation and further synthesize evidence, we combined interaction effects with the main treatment effects, in case a moderator fulfilled the criteria for described above. Based on this, we assessed whether we could identify a clinically relevant treatment effect for potentially relevant moderators (e.g., interaction-effects around main effects near zero usually do not imply clinically meaningful effects within subgroups, whereas interaction effect reaching 10 points on a 100-point scale does). We assumed that subgroup effects based on these interaction effects lie symmetrically around the main effects. For example, consider a moderator effect of “sex” of –six points on a 0 to 100scale, and a main effect of SMT of -8 points.

This would result in an approximated estimate of main treatment effect for men of -5 points (–8+3) and for women of 11 points (–8–3). These subgroup effects might indicate minimally relevant treatment effect for women, but not for men.

RESULTS

Characteristics of trials are presented in Table 1, risk of bias criteria and assessment are presented in the appendix eTable 2 and 3. For more details, see appendices and protocol.[16]

TABLE 1

Descriptives of Studies Evaluating the Effects of SMT on Outcomes Included in the Database (n = 21) in Alphabetical Order of First Author

Author (Year) Acronym	Country	N	Interventions	Duration of LBP According to Inclusion Criteria	Type of Manipulator	Type of Manipulation	Max No. Treatments Allowed and Duration of Treatment
Balthazard et al (2012)[21]	Switzerland	42	1. Spinal manipulation therapy plus active exercise (n = 22)2. Detuned ultrasound plus active exercise (n = 20)	>12 and <26 wks	Physiotherapist (n = 1)	Manipulation and mobilization	Eight over 4–8 wks
Bronfort et al (2011)[23]	United States	301	1. Supervised exercise (n = 101)2. Spinal manipulative therapy (n = 100)3. Home exercise and advice (n = 100);	>6 wks	Chiropractor (n = 9)	Manipulation	Participants were discharged from care if the treating clinician felt that maximum clinical benefit was obtained. 12 wks of care
Bronfort et al (2014)[22]	United States	192	1. Spinal manipulative therapy plus home exercise and advice (n = 96)2. Home exercise and advice (n = 96)	>4 wks	Chiropractor (n = 11)	Manipulation and mobilization	As many as 20 over 12 wks
Cecchi et al (2010)[24]	Italy	210	1. Back school (n = 70); 2. Individualized physiotherapy (n = 70); 3. Spinal manipulative therapy (n = 70)	>6 mo	Physician (n = 2)	Manipulation and mobilization	Four to six sessions per wk for 4–6 wks
Cook et al (2013)[25]	United States	154	1. Thrust manipulation (n = 77)2. Nonthrust manipulation (n = 77)	No restriction	Physiotherapist (n = 17)	Manipulation or mobilization (depending upon grp. assignment)	First two visits only afterwards clinician was allowed to choose technique they felt most beneficial for the patient
Ferreira et al (2007)[26]	Australia	240	1. General exercise (n = 80)2. Motor control exercise3. Spinal manipulative therapy	>3 mo	Physical therapist (n = ?)	Mobilization or manipulation; Maitland	12 Over 8 wks
Gudavelli et al (2006)[27]	USA	235	1. Flexion distraction mobilization (n = 123)2. Exercise therapy (n = 112)	>3 mo	Chiropractor (n = ?)	Mobilization (flexion-distraction)	16 Over 4 wks
Haas et al (2014)[28]	United States	400	1. 0 SMT + 18 LM (n = 100)2. Six SMT + 12 LM (n = 100)3. 12 SMT + six LM (n = 100)4. 18 SMT + 0 LM (n = 100)	>3 mo	Chiropractor (n = 12)	Manipulation or mobilization	18 Over 6 wks
Hidalgo et al (2015)[41]	Belgium	32	1. Spinal manipulative therapy (n = 16)2. Sham spinal manipulative therapy (n = 16)	No restriction	Physiotherapist (n = 1)	Mobilization	One over 2 wks
Hondras et al (2009)[29]	United States	240	1. High-velocity low-amplitude spinal manipulative therapy (n = 96)2. Low-velocity variable amplitude spinal mobilization (n = 95)3. Medical care (n = 49)	>4 wks	Chiropractor (n = 4)	Manipulation or mobilization (flexion-distraction) (depending upon group assignment)	12 Over 6 wks
Hsieh et al (2002)[30]	United States	206	1. Back school (n = 48) 2. Myofascial therapy (n = 51)3. Joint manipulation (n = 49)4. Combination of treatments 2 and 3 (n = 52)	>3 wks to <6 mo	Chiropractor (n = ?)	Manipulation	Nine over 3 wks
Petersen et al (2011)[39]	Denmark	350	1. McKenzie therapy (n = 175)2. Spinal manipulative therapy (n = 175)	>6 wks	Chiropractor (n = 3)	Manipulation or mobilization	Max 15 over 12 wks
Rasmussen-Barr et al (2003)[38]	Sweden	47	1. Stabilizing training group (n = 24)2. Manual therapy group (n = 23)	>6 wks	Manual therapist (n = ?)	Mobilization	Six over 6 wks
Skillgate et al (2007)[31]	Sweden	409	1. Naprapathy (n = 206)2. Standard care or “evidence-based” care (provided by physician) (n = 203)	>2 wks	Naprapath (n = 8)	Manipulation or mobilization	Six over 6 wks
UK Beam Trail Team (2004)[40]	UK	1334	1. Best care in general practice (n = 338)2. Best care plus exercise alone (n = 310)3. Best care plus private manipulation alone (n = 180)4. Best care plus NHS manipulation alone (n = 173)5. Best care plus private manipulation plus exercise (n = 172)6. Best care plus NHS manipulation plus exercise (n = 161)	(Essentially) >3 wks	Chiropractor, osteopath or physiotherapist (n = 84)	Manipulation or mobilization	Eight over 12 wks
Verma et al (2013)[32]	India	30	1. Exercise (n = 15)2. Lumbar mobilisation and exercise (n = 15)	>3 mo	Physiotherapist (n = ?)	Mobilization	Eight over 4 wks
Vismara et al (2012)[33]	Italy	21	1. Osteopathic manipulation and Specific exercise (n = 10)2. Specific exercise (n = 11)	>6 mo	Osteopath (n = 1)	Manipulation or mobilization	10 Over 10 wks
Walker et al (2013)[34]	Australia	183	1. Sham group (n = 91); 2. Usual chiropractic care group (n = 92)	>1 wk	Chiropractor (n = 8)	Manipulation or mobilization	Two over 2 wks
Wilkey et al (2008)[35]	UK	63	1) Hospital pain clinic (n = 33)2) Chiropractic treatment (n = 30)∗	>3 mo	Chiropractor (n = ?)	Manipulation	16 Over 8 wks
Xia et al (2015)[36]	United States	192	1. Thrust spinal manipulation (n = 72)2. Nonthrust spinal manipulation (n = 72)3. Control (n = 48)	>4 wks	Chiropractor (n = 4)	Manipulation or mobilization	Four over 2 weeks
Zaproudina et al (2009)[37]	Finland	73	1. Traditional bone setting (n = 36)2. Physical therapy (n = 37)†	>3 mo	Bone-setter (n = 8)	Mobilization	Five over 10 wks

? indicates unclear/unknown; LM, light massage; LBP, low back pain; NHS, National Health Service; SMT, spinal manipulative therapy.

More patients’ data provided than published.

Only patient data used if patient consented to be included in our database, and therefore less patients than published.

Identification of Trials

In total, 43 RCTs met the inclusion criteria, of which 21 (50%) provided data[21-41] (Figure 1) from 4223 participants. Baseline characteristics were compared to the published results of the individual trials. In two trials, the results differed from the published results: one trial provided only data from participants who gave consent to share their data,[37] whereas for the second trial, all relevant baseline moderator data of the participants were lost.[35]

Figure 1

Flow diagram of study inclusion.

Characteristics of Study Participants

Participant characteristics were fairly similar across all comparisons (Table 2 and appendix eTable 7). All trials except one[35] provided data on sex and age. The average age of the participants was 46.1 (SD 13.78) years, 54.4% were women.

TABLE 2

Patient Characteristics at Baseline for Groups Receiving SMT vs. Groups Receiving Recommended Interventions

	SMT vs. Recommended Interventions (m = 12; n = 2475)
Demographic Data	SMT	Recommended Interventions
Age, mean (SD) y (m = 11, n = 2409)	47.13 (13.63)	47.18 (13.99)
Sex, n (%) female (m = 11, n = 2412)	667 (56.9)	684 (55.2)
BMI, mean (SD) (m = 8, n = 1434)	26.85 (5.12)	26.79 (5.10)
Ethnicity, n (%) white (m = 5, n = 861)	409 (90.9)	388 (88.1)
Lifestyle factors
Physical activity, n (%) (m = 6, n = 824)
Low (one or less than once a wk)	115 (31.9)	166 (35.9)
Medium (two to three times a wk)	146 (40.4)	166 (35.9)
High (more than three times a wk)	100 (27.7)	131 (28.3)
Smoker, n (% nonsmokers) (m = 6, n = 1173)	451 (79.5)	453 (74.8)
Alcohol use (%)	∗	∗
Sociodemographics
Marital status, n (%) married; living with a partner (m = 6, n = 1173)	397 (69.0)	404 (67.6)
Level of education, n (%) low/middle (m = 7, n = 1672)	600 (68.0)	534 (67.6)
Income, n (%)	∗	∗
Employment status, n (%) at work (m = 9, n = 2126)	818 (77.9)	770 (71.6)
Nature and severity of LBP
Duration of LBP, n (%) <12 mo (m = 7, n = 1252)	121 (20.3)	149 (22.9)
Leg pain, n (%) (m = 5, n = 1038)	320 (59.0)	281 (56.7)
Previous LBP treatment received, n (%) (m = 5, n = 930)	258 (27.7)	218 (23.4)
Previous physiotherapy for low back pain received, n (%) (m = 5, n = 771)	64 (8.3)	72 (9.3)
Previous SMT for low back pain received, n (%) (m = 6, n = 988)	209 (21.2)	111 (11.2)
Used medication for low back, n (%) (m = 6, n = 1018)	200 (19.6)	269 (26.4)
Nonspecific, n (%)	∗	∗
Comorbidities	∗	∗
Type of treatment	∗	∗
Psychological factors	SMT	Control
Depression, n (%) (m = 5, n = 1297)	43 (6.2)	75 (12.5)
Treatment preference/expectations	∗	∗
Primary outcomes
Pain	SMT	Recommended interventions
Combined pain score at baseline, mean (SD), (m = 12, n = 2441)	49.47 (22.27)	49.75 (21.59)
Combined pain score at 1 mo, mean (SD), (m = 10, n = 1948)	34.19 (22.95)	35.81 (23.91)
Combined pain score at 3 mo, mean (SD), (m = 9, n = 1673)	27.92 (23.03)	32.12 (24.25)
Combined pain score at 6 mo, mean (SD), (m = 8, n = 1321)	27.35 (23.12)	32.31 (23.90)
Combined pain score at 12 mo, mean (SD), (m = 10, n = 1816)	31.80 (25.81)	33.32 (25.38)
Function
RMDQ sum score at baseline, mean (SD), (m = 9, n = 2174)	8.99 (4.96)	10.07 (5.44)
RMDQ sum score at 1 mo, mean (SD), (m = 8, n = 1760)	5.62 (5.02)	6.65 (5.37)
RMDQ sum score at three months, mean (SD), (m = 8, n = 1648)	4.81 (5.14)	5.52 (5.34)
RMDQ sum score at six months, mean (SD), (m = 8, n = 1348)	4.99 (5.44)	6.26 (5.95)
RMDQ sum score at 12 mo, mean (SD), (m = 7, n = 1575)	5.44 (5.67)	6.16 (5.92)
Secondary outcomes
SF36 Physical Component Scale of SF36 at baseline, mean (SD), (m = 5, n = 1362)	40.69 (7.15)	41.06 (7.59)
SF36 Mental Component Scale of SF36 at baseline, mean (SD), (m = 5, n = 1362)	43.83 (9.05)	45.08 (9.60)
Medication use at baseline, n (% medication use) (m = 3, n = 668)	145 (21.7)	216 (32.3)

BMI indicates body mass index LBP, low back pain; m, number of studies; n, number of participants; SMT, spinal manipulative therapy; SD, standard deviation.

Less than three studies or combining categories was not meaningful.

For employment and BMI, moderator data were missing in 9.6% and 5.9% of the participants respectively, whereas for all other moderators, data were missing in <2% of the participants.

Moderators of SMT for Primary Outcomes: Pain and Function

SMT Versus Recommended Interventions

For most moderators, no moderating effects were identified except for the moderators described below (Tables 3 and 4).

TABLE 3

Moderator Effects of SMT vs. Recommended Interventions for Pain. Within-study Interaction (β) and 95% CI With the Intervention Effects of Random-effect Models Adjusted for Baseline Using REML Separating, Between-study and Within-study Variation Are Presented

Combined Pain scale	Follow-up 1 mo β (95% CI) m; n	3 mo β (95% CI) m; n	6 mo β (95% CI) m; n	12 mo β (95% CI) m; n
Demographic moderators
Sex: (reference: male)	–1.63 (–5.02 to 1.75) 9; 1859	–2.76 (–6.81 to 1.28) 8; 1592	–5.11 (–9.69 to –0.54) 7; 1270	–6.69 (–10.71 to –2.67) 9; 1740
Age∗: (reference: <65 years’ old)	–1.74 (–6.03 to 2.55) 9; 1859	–0.56 (–7.96 to 6.84) 8; 1590	–9.42 (–17.65 to –1.20) 7; 1268	2.60 (–5.27 to 10.47) 6; 1042
Body mass index: (reference: <30)	–1.96 (–7.30 to 3.38) 6; 1060	3.78 (–3.10 to 10.65) 5; 924	4.36 (–2.28 to 11.00) 6; 1056	6.01 (–0.84 to 12.86) 6; 1042
Ethnicity: (reference: other than white)	–1.96 (–10.08 to 6.15) 4; 699	5.12 (–7.37 to 17.60) 3; 492	–0.08 (–13.39 to 13.22) 3; 465	2.77 (–9.48 to 15.02) 3; 469
Lifestyle factors
Physical activity: (reference: three or less a wk)(reference: one or less a wk)	–0.46 (–7.45 to 6.54) 4; 533-3.96 (–10.07 to 2.16) 4; 533	−4.84 (–11.59 to 1.92) 7; 721-2.06 (–8.31 to 4.18) 7; 721	−4.67 (–11.47 to 2.14) 4; 6811.00 (–5.89 to 7.60) 4; 681	−0.02 (–7.58 to 7.54) 5; 661-3.22 (–10.35 to 3.90) 5; 661
Smoker: smoker (reference: nonsmoker)	3.19 (–3.20 to 9.58) 5; 886	2.45 (–3.68 to 8.58) 4; 866	6.02 (0.12 to 11.92) 5; 891	4.85 (–1.33 to 11.03) 4; 806
Alcohol use	†	†	†	†
Sociodemographics
Marital status: (reference: not involved in relation)	2.60 (–3.11 to 8.31) 6; 1076	1.39 (–4.47 to 7.35) 4; 802	1.84 (–4.19 to 7.87) 5; 834	–0.28 (–6.11 to 5.57) 4; 764
Employment status: (reference: not employed)	5.69 (0.32 to 11.06) 7; 1622	0.88 (–5.72 to 7.49) 7; 1440	5.12 (–2.94 to 13.19) 6; 1191	0.96 (–4.93 to 6.86) 8; 1572
Level of education: (reference: low or middle)	–0.42 (–4.62 to 3.77) 6; 1377	–1.52 (–7.51 to 4.48) 5; 1117	–3.31 (–14.50 to 7.88) 4; 625	–2.54 (–9.27 to 4.19) 5; 1066
Income	†	†	†	†
Nature and severity of LBP
Duration of LBP∗: (reference: <1 y)	–1.69 (–10.37 to 7.00) 5; 876	4.97 (–3.20 to 13.13) 4; 700	10.76 (1.06 to 20.47) 5; 875	5.26 (–2.92 to 13.44) 5; 854
Radiation: (reference: no leg pain)	–4.13 (–9.68 to 1.43) 3; 649	–2.56 (–9.18 to 4.06) 4; 716	0.43 (–6.56 to 7.41) 3; 636	–4.41 (–11.33 to 2.52) 4; 682
Previous LBP treatment received: (reference: no)	0.23 (–5.06 to 5.53) 5; 878	2.42 (–9.09 to 4.25) 4; 675	–0.36 (–7.18 to 6.45) 3; 626	–0.95 (–8.15 to 6.25) 3; 622
Previous Physio for LBP: (reference: no)	–10.83 (–19.42 to –2.25) 3; 448	––3.38 (–11.51 to 4.74) 4; 679	2.41 (–6.64 to 11.46) 4; 639	–1.55 (–10.13 to 7.02) 4; 624
Previous SMT for LBP: (reference: no)	–3.97 (–13.12 to 5.19) 4; 629	5.34 (–3.25 to 13.95) 4; 676	6.86 (–1.63 to 15.35) 4; 636	15.59 (6.18 to 24.99) 4; 621
Previous medication for LBP: (reference: no)	1.32 (–3.47 to 6.13) 5; 887	–2.99 (–8.44 to 2.46) 6; 902	–2.43 (–8.22 to 3.37) 4; 790	–1.69 (–11.33 to 7.96) 6; 862
Comorbidities	†	†	†	†
Psychological factors
Depression∗: (reference: no depression)	1.45 (–6.78 to 9.68) 4; 1053	–1.06 (–11.23 to 9.11) 4; 860	†	–1.20 (–13.23 to 10.83) 4; 817
Treatment preference/expectations∗	†	†	†	†
Primary/secondary outcomes at baseline as moderator
Baseline pain score per 10 points change∗	–0.20 (–1.2 to 0.80) 10; 1922)	–0.70 (–2.10 to 0.80) 9; 1922)	–0.40 (–0.18 to 1.10) 8; 1922)	–1.1 (–2.90 to 0.70) 9; 1791)
Baseline function scales combined (z score)∗	–1.25 (–3.90 to 1.39) 10; 1914	–0.77 (–3.43 to 1.87) 9; 1641	–1.73 (–4.27 to 0.82) 8; 1313	–0.90 (–4.07 to 2.26) 10; 1783
MCS	–0.09 (–0.40 to 0.21) 5; 1190	–0.01 (–0.28 to 0.26) 4; 1121	0.17 (–0.16 to 0.49) 4; 681	–0.19 (–0.48 to 0.11) 4; 1046
PCS	0.19 (–0.36 to 0.73) 5; 1190	–0.07 (–0.52 to 0.38) 4; 1121	–0.40 (–1.07 to 0.28) 4; 681	0.09 (–0.50 to 0.69) 4; 1046

A negative interaction coefficient indicates a more positive/less negative effect of SMT vs. recommended therapies for the index group (e.g., females) as compared to the reference group (e.g., males). CI indicates confidence interval; LBP, low back pain; m, number of studies; MCS, mental component summary of SF 36; MD, mean difference; n, number of participants; PCS, physical component summary of SF 36; REML, restricted maximum likelihood; RMDQ, Roland Morris Disability questionnaire; SMT, spinal manipulative therapy.

β (95% CI), within-study interaction and confidence interval: the mean difference in pain score for the specific moderator for SMT vs. recommended therapies on scale from 0–100.

Confirmatory moderator analysis.

Not enough data.

TABLE 4

Moderator Effects of SMT vs. Recommended Interventions for Function. Within-study Interaction (β) and 95% CI With the Intervention Effects of Random-effect Models Adjusted for Baseline Using REML, Separating Between-study and Within-study Variation Are Presented

Standardized Mean Difference of Combined Function Scales	Follow-up 1 mo β (95% CI) m; n	3 mo β (95% CI) m; n	6 mo β (95% CI) m; n	12 mo β (95% CI) m; n
Demographic moderators
Sex:(reference: male)	–0.03 (–0.16 to 0.10) 9; 1876	0.06 (–0.09 to 0.21) 10; 1837	(–0.15 to 0.18) 8; 1439	–0.02 (–0.17 to 0.14) 9; 1775
Age∗: (reference: <65 years’ old)	–0.03 (–0.17 to 0.12) 9; 1876	–0.06 (–0.25 to 0.13) 10; 1835	–0.23 (–0.45 to –0.01) 8; 1437	–0.32 (–0.57 to –0.07) 9; 1773
Body mass index: (reference: <30)	–0.07 (–0.26 to 0.12) 6; 1047	–0.18 (–0.42 to 0.07) 7; 1139	–0.06 (–0.30 to 0.17) 7; 1225	0.15 (–0.14 to 0.44) 6; 1055
Ethnicity: white vs. other (reference: other than white)	–0.19 (–0.55 to 0.17) 4; 691	0.07 (–0.39 to 0.53) 5; 707	–0.23 (–0.65 to 0.19) 4; 630	0.03 (–0.38 to 0.44) 3; 469
Lifestyle factors
Physical activity: (reference: three or less a week) (reference: one or less a week)	–0.002 (–0.008 to 0.002) 4; 511–0.11 (–0.33 to 0.11) 4; 511	–0.002 (–0.009 to 0.004) 6; 739–0.05 (–0.27 to 0.18) 6; 739	–0.13 (–0.40 to 0.14) 4; 659–0.13 (–0.34 to 0.09) 4; 659	0.00 (–0.01 to 0.01) 5; 676–0.10 (–0.33 to 0.14) 5; 676
Smoker: smoker (reference: nonsmoker)	0.24 (0.00 to 0.48) 5; 873	0.22 (–0.02 to 0.47) 6; 1075	0.14 (–0.11 to 0.38) 6; 1050	0.29 (0.02 to 0.56) 4; 801
DZ converted to a MD on the 24-point RMDQ scale	1.08	1.11	0.72	1.54
Alcohol use	†	†	†	†
Sociodemographics
Marital status: (reference: not involved in relation)	–0.12 (–0.29 to 0.05) 6; 1066	–0.09 (–0.29 to 0.12) 6; 1016	–0.08 (–0.26 to 0.10) 6; 1005	–0.12 (–0.33 to 0.10) 4; 777
Employment status: (reference: not employed)	0.10 (–0.08 to 0.27) 7; 1657	–0.17 (–0.39 to 0.04) 8; 1665	0.06 (–0.16 to 0.27) 7; 1377	–0.06 (–0.36 to 0.23) 8; 1606
Level of education: (reference: low or middle)	–0.17 (–0.37 to 0.01) 6; 1398	–0.11 (–0.29 to 0.08) 7; 1363	–0.07 (–0.27 to 0.13) 5; 796	–0.14 (–0.35 to 0.05) 5; 1106
Income	†	†	†	†
Nature and severity of LBP
Duration of LBP∗: >1 y (reference: <1 y)	0.07 (–0.29 to 0.43) 5; 861	0.02 (–0.30 to 0.34) 6; 910	0.19 (–0.02 to 0.39) 6; 1031	0.13 (–0.25 to 0.52) 5; 848
DZ converted to a MD on the 24-point RMDQ scale	0.35	0.10	1.06	0.78
Radiation: leg pain (reference: no leg pain)	–0.16 (–0.40 to 0.80) 3; 660	0.04 (–0.18 to 0.27) 5; 911	–0.01 (–0.22 to 0.21) 4; 820	–0.08 (–0.33 to 0.17) 4; 681
Previous LBP treatment received: (reference: no)	0.09 (–0.11 to 0.29) 5; 884	0.11 (–0.12 to 0.34) 5; 871	0.02 (–0.21 to 0.15) 5; 812	0.12 (–0.14 to 0.39) 4; 637
Previous Physio for LBP: (reference: no)	–0.08 (–0.63 to 0.48) 3; 426	0.04 (–0.33 to 0.40) 5; 696	0.18 (–0.19 to 0.56) 4; 615	0.11 (–0.27 to 0.50) 4; 634
Previous SMT for LBP: (reference: no)	–0.11 (–0.49 to 0.26) 4; 616	–0.07 (–0.46 to 0.32) 6; 885	0.14 (–0.20 to 0.49) 5; 795	0.52 (0.09 to 0.95) 4; 631
DZ converted to a MD on the 24 point RMDQ scale	–0.49	–0.36	0.74	2.93
Previous medication for LBP: (reference: no)	–0.02 (–0.22 to 0.19) 5; 887	–0.11 (–0.33 to 0.09) 6; 903	–0.20 (–0.41 to 0.004) 4; 790	–0.11 (–0.39 to 0.17) 6; 877
Comorbidities	†	†	†
Psychological factors
Depression∗: (reference: no depression)	0.28 (–0.03 to 0.58) 4; 1075	0.33 (–0.02 to 0.69) 5; 1066	0.13 (–0.26 to 0.51) 3; 505	0.02 (–0.39 to 0.43) 4; 825
Treatment preference/expectations∗	†	†	†	†
Primary/secondary outcomes at baseline as moderator
Baseline function scales combined (z score)	–0.04 (–0.17 to 0.09) 10; 1939	+0.01 (–0.10 to 0.13) 11; 1892	–0.04 (–0.17 to 0.08) 9; 1490	–0.05 (–0.17 to 0.07) 10; 18,266
Baseline pain dichotomized∗: (reference: baseline pain <50)	–0.20 (–0.36 to –0.04) 10; 1932	–0.20 (–0.37 to –0.03) 11; 1882	–0.22 (–0.39 to –0.06) 9; 1506	–0.14 (0.33 to 0.04) 10; 1806
	–0.90	–0.92	–1.00	–0.64
MCS per 10 points change	–0.00 (–0.10 to 0.10) 5; 1204	–0.03 (–0.16 to 0.08) 6; 1358	0.09 (–0.07 to 0.25) 5; 864	0.00 (–0.12 to 0.12) 4; 1069
PCS per 10 points change	0.03 (–0.17 to 0.23) 5; 1204	–0.06 (–0.18 to 0.06) 6; 1358	–0. 13 (–0.32 to 0.07) 5; 864	0.00 (–0.22 to 0.21) 4; 1069

CI indicates confidence interval; DZ, difference in Z-score; LBP, low back pain; m, number of studies; MCS, mental component summary of SF 36; MD, mean difference; n, number of participants; PCS, physical component summary of SF 36; REML, restricted maximum likelihood; RMDQ, Roland Morris Disability questionnaire; SMT, spinal manipulative therapy. A negative interaction coefficient indicates a more positive/less negative effect of SMT vs. recommended therapies for the index group (e.g., females) as compared to the reference group (e.g., males).

β (95% CI), within-study interaction and confidence interval of the interaction term: the difference in z score for the specific moderator for function for SMT vs. recommended therapies.

Confirmatory moderator analysis.

Not enough data.

Confirmatory Moderator Analysis

For pain and/or function, we found a consistent moderator effect for duration of LBP. Patients with <1 year LBP showed more positive/less negative point estimates for SMT versus recommended therapy on pain, with MD of 4.97 (95% CI: −3.20 to 13.13) at 3 months, 10.76 (95% CI: 1.06 to 20.47) at 6 months ,and 5.26 (CI −2.92 to 13.44) at 12 months; for function: SMD were 0.07 (–0.29; 0.43) at one month, 0.02 (–0.30; 0.34) at three months; 0.19 (–0.02; 0.15) at six months to 0.13 (–0.25; 0.52) at twelve months) (Tables 3 and 4). These effects were small, except for pain at 6 months, which showed a moderate effect. Converted to a MD for the 24-point Roland Morris Disability Questionnaire, these moderating effects amount to 0.35 at one month, 0.10 at 3 months, 1.06 at 6 months, and 0.78 at 12 months (Tables 3 and 4).

The direction of the main treatment effect of SMT versus recommended interventions for the outcome pain was in favor of SMT (e.g., 6 months: MD −5.56, 95% CI: −9.63 to −1.50) (see appendix eTable 10). When adding the moderator effect of duration of LBP to the main treatment effect, the results may indicate minimally relevant effects, meaning that patients with shorter duration of LBP may benefit from SMT. For those with longer duration, SMT has similar benefit compared to recommended interventions.

For function, patients with pain score >50 showed more positive/less negative point estimates for SMT vs recommended therapy on pain, with SMD of −0.20 (–0.36; −0.04) at 1 month, −0.20 (–0.37; −0.03) at 3 months, −0.22 (–0.39; −0.06) at 6 months, and −0.14 (–0.33; 0.04) at 12 months. These effects were small.

Exploratory Moderator Analysis

For pain and/or function, we found a consistent moderator effect for smoking. Nonsmokers showed more positive/less negative point estimates for SMT versus recommended therapy on pain, with MD of 3.19 (–3.20; 9.58) at 1 month, 2.45 (–3.68; 8.58) at 3 months, 6.02 (0.12; 11.92) at 6 months to 4.85 (–1.33; 11.03) at 12twelve months; for function: SMD were 0.24 (0.00; 0.48) at 1 month, 0.22 (–0.02; 0.47) at 3 months; 0.14 (–0.11; 0.38) at 6 months to 0.29 (0.02; 0.56) at 12 months) (see for conversion to the 24-point Roland Morris Disability Questionnaire Tables 3 and 4). These effects were small.

For pain and/or function, we found a consistent moderator effect for previous SMT for LBP. Patients that had no previous SMT showed more positive/less negative point estimates for SMT versus recommended therapy on pain, with MD of −3.97 (–13.12; 5.19) at 1 month, 5.34 (–3.25; 13.95) at 3 months, 6.86 (–1.63; 15.35) at 6 months to 15.59 (6.18; 24.99) at 12 months; for function, SMD were −0.11 (–0.49; 0.26) at 1 month, −0.07 (–0.46; 0.32) at 3 months; 0.14 (–0.20; 0.49) at 6 months to 0.52 (0.09; 0.84) at 12 months (Tables 3 and 4). This effect (i.e., patients that had no previous SMT improved more than patients that had previous SMT for the outcomes pain and function with SMT compared to other recommended treatments) was small, except for pain and functional status at twelve months, which showed a moderate effect (Tables 3 and 4).

SMT Versus Nonrecommended Interventions, SMT as Adjuvant Therapy, and Manipulation Versus Mobilization

Ninety percent of the moderator analyses for these comparisons were not performed due to too few data. In the analyses performed (mainly age, sex and BMI), we found no consistent effect for any moderator (Appendix eTables 8–9)

DISCUSSION

This study is the first large-scale IPD meta-analysis which attempted to identify potential moderators for those undergoing SMT for chronic LBP. In short, the results suggest no substantive moderation in the effect of SMT compared to other interventions. We did, however, identify (possible) small moderation effects for the following confirmatory moderators: duration of LBP and greater pain at baseline, and for the exploratory moderators smoking and previous exposure to SMT. However, these effects were too small to be clinically relevant. This suggests that targeting SMT, based upon individual characteristics examined in this study, is not warranted at this time.

Analyses of moderator effects of SMT have rarely been performed and have largely been restricted to aggregate meta-analytic approaches. These different approaches make it difficult to compare those results to ours. Results from an earlier systematic review[10] indicated moderating effects of psychosocial and belief factors, expectations, and baseline pain and disability.

Two earlier IPD studies evaluated moderators for other types of treatment for LBP and identified small effects for the following moderators: age, sex, BMI, no heavy physical demands, psychosocial factors, back pain disability, pain severity, and medication use.[42,43] However, our analyses suggest only a weak moderating effect of baseline pain for functional status in our confirmatory analysis. There are a number of reasons why the results of our moderator analyses might differ from the other studies.[10,43-45] Most importantly, these earlier studies examined various types of conservative treatments for LBP (e.g., cognitive behavioral therapy) and the comparisons were chosen differently than in our study.

An important difference of our IPD analysis compared to traditional aggregate meta-analyses is that we could adjust for covariates and were not dependent upon how these data were reported in the study publications. IPD allowed investigation for moderators in a more sophisticated and valid way. In the IPD analysis one can separate the between-study and the within-study interaction. The between-study interaction describes the moderation effects at study level. This is what is analyzed in a meta-regression or subgroup analyses in traditional aggregate meta-analysis. Results of these analyses can be severely affected by ecological biases.[19] The real interest lies in the within-study interaction, which describes the effects of covariates on the treatment effectiveness at the patient level.

Strengths and limitations: The most important strength is our large data set (i.e., 21 RCTs) from various countries resulting in a dataset, which included many different moderators and thousands of patients of which >2000 were in the SMT versus recommended intervention comparison. Most moderator analyses in this comparison included >500 patients provided from at least three trials. It has been suggested that this might be robust.[10] Far fewer patients were included in the moderator analyses for the other comparisons. Therefore, our (exploratory) moderator results should serve as a guide for future research only.

We collected a wide variety of moderators, but many moderators were measured differently across trials or were not measured at all. For example, duration of LBP was measured as a continuous variable in some trials and as a categorical variable. Only age, sex, and BMI were measured similarly. This meant that in many instances we had to compromise our best detailed measures by categorizing the data, which led to loss of information. Importantly, there was a large diversity in frequency (one to six times a week), duration (2–12 weeks), and number of treatments (two to 36 [average of 8]) in included trials and these characteristics were measured at study-level in most trials. Therefore, the moderator analyses with treatment characteristics were not possible in contrast to what we planned in our protocol. A better understanding of the etiology of chronic LBP and key mechanisms involved in the effects of SMT would help to identify moderators.

We did not assess the effects of imputing missing data on outcomes and moderators. Methodology for imputing missing values in IPD meta-analysis is still in the developmental phase.[46,47] To our knowledge standard imputation methods for IPD meta-analysis of moderator effects have not been described in the literature. These models are especially challenging as they should result in valid estimates for the one-stage models we used that distinguish within-study and between-study interaction effects.

Additionally, we did not investigate multiple moderators in the same analysis as no evident clinically important moderators were found, although others[43] looked at multiple moderators at a lower level of statistical significance. At this moment, our study clearly presents exploratory results to inform future studies.

Another challenge we encountered is the definition of clinical relevance of the treatment moderator effects. For main effects, three levels of clinical relevance (small, medium, and large) are broadly used across-systematic reviews, and are recommended by the Cochrane Back and Neck Review Group.[20,48] However, for moderator analysis, we think clinical relevance should not be defined by the same criteria as for the main treatment effects. Importantly, we interpreted the moderator effects considering hypothesized mechanisms and consistency of results across time and outcome measures. In summary, we used a consensus approach for the arbitrary cutoff points for drawing our conclusions to detect small difference within a moderator for low-intensity, low-cost intervention comparisons. This is subjective, but in our view, the best method currently available.

Another potential limitation is selection bias. We included only 50% of the eligible trials, which is comparable to other IPD studies.[49,50] However, the effect sizes, methodological quality, and range of publication dates of studies where IPD was collected was comparable with the studies where no IPD was present. We also missed the data of the most recent trials as we only included trials until 2016, because collection of data for an IPD is time-consuming as also seen in other IPD studies.[42] It took 4 years to collect and analyze the data, which is comparable to IPD meta-analyses in other fields.[43,51] When we updated our search May 4th, 2018, we found that the most recent trials were small in size, had few data on patient characteristics, and were considered to have a high risk of bias.[52-56] Therefore, it is not likely that these most recent trials or the studies where IPD was not provided, would materially change our results.

The clinical implication of this IPD study is that based on the evidence to date there is no justification for using specific patient characteristics to target SMT for chronic LBP patients.

In addition to more detailed study of the etiology of chronic LBP and mechanism(s) of SMT, future initiatives should focus on standardizing the manner in which inclusion and exclusion criteria, outcomes and moderators are measured and reported.[54,57-59] This will facilitate an effective comparison of interventions across trials. Additionally, our wish is to form an international IPD repository of RCTs which have examined the effect of conservative treatment for LBP. This will provide an excellent resource for researchers with advantages such as the potential for future network meta-analysis and to standardize, safeguard, and store data centrally. To facilitate this, we encourage researchers in future grant applications to obtain permission to share their data and to include costs of uploading their final data into a repository as well as permission from Research Ethical Committees and participants for sharing these data. However, three large IPD meta-analyses of nonpharmacological treatments for LBP have failed to find any consistent and clinically important moderation effects indicates that this line of research is very unlikely to generate important finding to improve patient care.

CONCLUSION

Based on the current IPD analyses, there is no evidence for moderating effects of specific patient characteristics that enable clinicians to identify which patients are likely to benefit more from SMT compared to other treatments. Future research dealing with the effectiveness of SMT would benefit from shared procedures for including important treatment effect modifiers.

The effects of SMT are comparable to other interventions recommended in guidelines for chronic LBP; however, it is unclear which patients are more likely to benefit from SMT compared to other therapies.

Based on this review, there is no evidence to suggest that specific patients or treatment characteristics are associated with clinically better response to SMT as compared to other (recommended) treatments for chronic LBP.

This may well be a result of the great variation in reporting of potential treatment and patient moderators.

Future initiatives should also focus on standardizing the manner in which inclusion and exclusion criteria, outcomes, and moderators are defined, measured, and reported.

This will facilitate an effective comparison of interventions across trials.