A matched cohort study design was performed using the Danish Chiropractic low back pain cohort (ChiCo) study, a prospective longitudinal observational study with one year follow-up performed between November 2016 and December 2019 [28]. As the ChiCo study followed ethical regulations [28] and only anonymised data was provided for use in this study, the Macquarie University Human Research Ethics Committee provided a waiver to use the ChiCo data without further ethical review.
Cohort
The ChiCo cohort includes 2818 patients who presented to Danish chiropractors with a new or recurrent episode of low back pain, with or without leg pain, for which they were not in an ongoing course of treatment or long-term management. To be included, patients needed to be 18 years old or over and able to complete electronic questionnaires in Danish. Patients with a confirmed diagnosis of fracture, infection, cancer or other serious pathology were excluded from the ChiCo study. Baseline questionnaires were completed by both participants and chiropractors, and participants underwent a baseline clinical assessment with the treating chiropractor. Follow-up questionnaires were completed by participants at two weeks, three months, and one year. Treatment was provided as needed and not impacted by participation in the study. No limitation was placed on access to other healthcare services. For the current study, participants identified by the chiropractor at baseline to have previous imaging relevant to their current clinical presentation were excluded (N = 656). All other participants in the ChiCo study were eligible for inclusion (N = 2162).
Data collection
Baseline assessment for each participant included two patient questionnaires and a clinical assessment questionnaire completed by the treating chiropractor. The first patient baseline questionnaire was performed at the clinic prior to assessment with the chiropractor. This baseline questionnaire included: age; sex; low back and leg pain intensity, measured on a 0–10 numerical rating scale; low back disability, measured using the Roland Morris Disability Questionnaire (RMDQ), with results proportionally recalculated to a score from 0 to 100; duration of current episode of low back pain (1–2 days; 3–7 days; 1–2 weeks; 2–4 weeks; 1–3 months; 3–12 months; more than 1 year); risk of poor prognosis, measured using the STarT back tool; previous imaging for back pain (yes/no); and expectations about the use of imaging within the initial assessment (yes/no). The second patient baseline questionnaire was performed immediately after the initial assessment, and included: previous treatment for back pain (surgical or non-surgical); and existing co-morbidities. The clinical assessment questionnaire completed by the chiropractor at the first visit included: suspicion of serious pathology as the cause of low back pain (yes/no); low back pain diagnosis (non-specific low back pain, spine-related leg pain with nerve root involvement, spine-related leg pain without nerve root involvement, or suspected fracture, infection, cancer, cauda equina syndrome, or inflammatory arthritis); pre-existing and relevant imaging for the presenting low back pain (yes/no); intention to treat the patient with high velocity, low amplitude manipulation (yes/no); and whether they referred the participant for imaging (X-ray, CT, or MRI).
Follow-up questionnaires were completed online via REDCap by participants at two weeks, three months, and one year. All follow-up questionnaires included the low back pain numerical rating scale and RMDQ. In addition, the two-week questionnaire included items on satisfaction with care and global perceived effect on low back pain.
Diagnostic imaging
Participants who were referred for diagnostic imaging (X-ray, CT, or MRI) by the chiropractor at the first visit were defined as the exposure group. Participants who were not referred for imaging at the first visit were defined as the non-exposure group.
Matching variables
Matching was performed using baseline variables which were considered by the research team to potentially affect the decision to refer for imaging and were also associated with outcome measures on regression analysis. Baseline variable selection was informed by the literature on reasons chiropractors refer for imaging [20, 25, 29,30,31], and the expert opinion and clinical experience of the research team. The baseline variables that were assessed for association with outcome measures were: age; sex; low back pain intensity; duration of low back pain; leg pain intensity; low back disability; risk of poor prognosis (STarT back tool); previous imaging; previous surgery; previous low back pain treatment (surgical or non-surgical); chiropractor suspicion of serious pathology; chiropractor intention to treat with high velocity low amplitude spinal manipulation; participant expectations of imaging; and the presence of co-morbidities.
Outcome measures
The primary outcomes were participant reported low back pain intensity and disability at three months.
Secondary outcomes were participant reported low back pain intensity and disability at two weeks and one year, and global perceived effect on low back pain and satisfaction with care at two weeks. Global perceived effect was measured on a seven-point Likert scale from ‘much better’ (response 1) to ‘much worse’ (response 7). Responses were dichotomised for analysis to ‘improved’ (responses 1 or 2, ‘much better’ or ‘better’) or ‘not improved’ (responses 3–7, ‘slightly better’ to ‘much worse’). Satisfaction with care was measured on a five-point Likert scale from ‘to a very high degree’ (response 1) to ‘not at all’ (response 5). Responses were dichotomised for analysis to ‘satisfied’ (responses 1 or 2, ‘to a very high degree’ or ‘to a high degree’) or ‘not satisfied’ (responses 3–5. ‘to some degree’ to ‘not at all’).
Missing data
Missing baseline data were imputed using multiple imputation in SPSS (IBM Corp. Released 2017. IBM SPSS Statistics for Windows, Version 25.0. Armonk, NY: IBM Corp.), depending on the type and reason for the missing data. Twenty-seven percent of ChiCo participants did not receive the second baseline questionnaire to complete. Therefore, data related to previous treatment and co-morbidities were imputed for these participants. Missing data were also imputed for baseline leg pain intensity (2.3% missing), low back pain intensity (2.0% missing), and duration of back pain (0.7% missing). Participants with missing outcome data were removed from the analyses.
Data analysis
To account for potential differences between the groups that may have impacted on outcomes, the exposure and non-exposure groups were matched on baseline variables using coarsened exact matching (CEM) [32, 33]. In CEM, exact matches of participants in the exposure (imaging) and non-exposure (no imaging) groups were compared. However, as it was not realistic to match exactly on continuous variables, the data were ‘coarsened’ to allow exact matching in pre-determined categories. Cut-points to define the strata for matching were selected for age (18–50, 51–70, greater than 70 years), low back pain intensity (0–3, 4–7, 8–10 on the numerical rating scale); low back disability (0–20, 21–40, 41–60, 61–100 on the RMDQ); and leg pain intensity (0–3, 4–7, 8–10 on the numerical rating scale). CEM assigns individuals to strata of baseline characteristics and discards any strata that do not include at least one exposed and one non-exposed participant. Each stratum may contain different numbers of exposed and non-exposed participants. This matching was used in the subsequent effect estimation while applying weights to compensate for differences in strata size.
Baseline variables in the exposure and non-exposure groups were presented descriptively. Observed and CEM-weighted standardised differences in means between the exposed and non-exposed groups were calculated for baseline variables used in the matching process.
The effect of imaging was estimated using mixed models with random effects of chiropractors and CEM-weightings. The primary model was adjusted for the outcome measure at baseline (low back pain intensity or disability) as well as for duration of pain. Sensitivity analyses were performed with additional adjustment for covariates that did not achieve almost perfect balance with CEM matching (standardised difference in means of zero). Mixed linear models were used for continuous outcomes (back pain intensity and disability), and mixed effect logistic regression was performed for dichotomous outcomes (global perceived effect and satisfaction with care). Separate models were conducted for each time point (two weeks, three months, and one year). All data analyses were performed using STATA (StataCorp. 2019. Stata Statistical Software: Release 16. College Station, TX: StataCorp LLC).