Association of Lumbar Spine Radiographic Changes With Severity of Back Pain–Related Disability Among Middle-aged, Community-Dwelling Women

Key Points Question Are lumbar spine radiographic changes associated with severity of back pain–related disability among middle-aged, community-dwelling women? Findings In this population-based cohort study of women from the UK, there was no evidence to support an association between a higher number of lumbar segments with radiographic changes (Kellgren-Lawrence grade, osteophytes, and disc space narrowing) and more severe back pain–related disability in cross-sectional (650 women) or longitudinal (443 women) analyses. Meaning The findings suggest that the changes detected on lumbar radiographs provide limited value for decision-making regarding back pain management in this population.


Exploratory analyses
We examined whether the association between lumbar spine radiographic changes and the severity of back pain-related disability differed by age, BMI or smoking status through testing of multiplicative interactions using WALD statistics.

Sensitivity analyses
1) To assess the validity of our cut-off points for defining lumbar spine radiographic changes, we changed the cut-off points from 'no change vs any change' to 'no or mild change vs moderate-to-severe change'.
2) We changed the composite score by using the total original score, which is a sum score from the original grade score at each segment. For example, the K/L grade based on total original score ranged from 0-16 (at each segment 0-4).
3) We built a new composite score based on the disease severity: for disc space narrowing and osteophyte based score, grade 0 vs grade 2 and 3; for K/L grade based score, grade 0 and 1 vs grade 3 and 4. 4) Considering the potential heterogeneity of the population, we restricted our analyses to women with back pain. 5) Considering the potential model misspecification issue, we changed it to linear model and ordinal logistic regression with cloglog link function. 6) To explore the potential influence from unmeasured confounding, E-value was calculated. 7) As one reviewer suggested, we additionally adjusted pain medication and depression considering they might affect the results as potential strong prognostic factors.

eAppendix 2. Details of Statistical Methods
Ordinal logistic regression: Liu et al. indicated that ordinal regression models are robust for continuous outcomes, especially when the distributions of continuous responses are skewed (Stat Med. 2017 Nov 30;36(27):4316-4335). Our continuous outcomes were skewed so that we chose ordinal logistic regression.
Stepped modelling framework: If all covariates were measured at the same time, two models were sufficient: step 1: unadjusted analyses; step 2: analyses adjusted for age, BMI, back pain status, physical activity, bisphosphonates usage status, and smoking status (additionally adjusted for Year 9 back pain-related disability for the longitudinal analysis). The reason for the inclusion of Year 9 back pain-related disability for the longitudinal analysis is that it is a strong prognostic factor for Year 15 back pain-related disability. As VanderWeele et al.
indicated, it is often important to control, whenever possible, for the outcome at or prior to the time of the baseline exposure assessment so that confounding control assumption is as plausible as possible (Statistical Science 35.3 (2020): 437-466.). In our study, among all selected covariates, physical activity was measured at Year 6 and others were measured at Year 9 (baseline at our study). Considering physical activity could change between Year 6 and Year 9, we built three models: step 1: unadjusted analyses; step 2: analyses adjusted for age, BMI, back pain status, bisphosphonates usage status, and smoking status (additionally adjusted for Year 9 back pain-related disability for the longitudinal analysis); step 3: analyses further adjusted for physical activity.
Exposure modelling: For each type of radiographic changes, our exposure has five values: 0, 1, 2, 3, and 4. The value equals the number of lumbar spine segments affected by radiographic changes. We could consider the exposure as the unordered categorical variable. In this case, we set 0 as the reference level and obtained the estimate by comparing other values with 0. We could also consider 0-4 as the continuous variable. As a continuous variable, we tested the linear trend by modelling the exposure as continuous variable and reported the P-value (whether the regression coefficient of the exposure variable equalled zero); we also tested the non-linear trend by adding the quartic term to the previous model and reported the P-value (whether the added quartic term could improve performance of the previous model through analysis of variance).

Multiple imputation:
It is a general approach to handle missing data in epidemiological and clinical research (BMJ 2009;338:b2393 (1) We calculated E-value through Online Calculator (https://mmathur.shinyapps.io/evalue/) based on results from step 2 of the stepped modelling framework.
Explanation: for an unmeasured confounder to explain the OR estimate of 1.22, the unmeasured confounder would have to be associated with both the exposure and the outcome by 1.44-fold above and beyond the measured confounders.