Do medical conditions predispose to the development of chronic back pain? A longitudinal co-twin control study of middle-aged males with 11-year follow-up

A major challenge for observational research in CBP is how to isolate the effects of specific risk factors by controlling for relevant confounding factors. Most commonly this is done by multivariate statistical adjustment, yet even with the most rigorous methods there remains the potential for residual confounding due to unknown or unmeasured factors. An alternative approach is to design studies that may account for some of the unknown or unmeasured confounding factors. One such approach is to use the co-twin control design to examine the relationship between a putative risk factor and CBP. Contemporary approaches to the co-twin control design permit comparing the association of a risk factor and CBP using both ‘individual-level’ analyses (such as those used in conventional studies of unrelated individuals), with ‘within-pair’ analyses comparing twins to their co-twins [30]. Within-pair analyses account for ‘familial factors’, including both genetic factors and early life experiences (also called ‘shared’ environmental factors), which are common to both twins. Such within-pair analyses are matched for factors such as age and ethnicity, as well as other factors which are not directly measured, but which twins share. Since genetic predispositions and early life experiences which influence different health conditions are accounted for in the within-pair analyses, associations between putative risk factors and CBP which persist in within-pair analyses are less likely to be explained by confounding due to genetic or shared environmental factors, and might be possibly modifiable and therefore targets for intervention [31].

This longitudinal co-twin control study was a secondary analysis of existing data from the Vietnam Era Twin (VET) Registry. The VET Registry was constructed from military discharge records, and is a national sample of male twin pairs from all United States (US) military branches who were born between 1939 and 1957 and who served on active duty during the Vietnam era (1965–1975) [32]. The VET Registry was not compiled based on specific diagnoses, health behaviors, or military service characteristics aside from military discharge; details of the Registry construction and zygosity ascertainment process have been described previously [33‐35]. VET Registry members reside in all 50 US states, and are comparable to older US males from the general population with respect to income and education; most obtain healthcare outside the Veterans Affairs (VA) system [36].

In 1999–2000, VET Registry members were invited to participate in an observational study of men’s health, the ‘Men’s Health Study’, which included a mailed survey obtaining information regarding self-reported medical conditions and health behaviors. Responses from this survey formed the baseline assessment for the current analysis. Between 2010 and 2012, VET Registry members were invited to participate in an observational study of PTSD among Veterans (VA Cooperative Studies Program #569: The Course and Consequences of PTSD in Vietnam Era Twins, or ‘CSP #569’). This included a mailed survey obtaining information about mental and physical health conditions. The current analysis included VET Registry participants who participated in both the Men’s Health Study and CSP#569, 11 years later (Fig. 1). Written informed consent was obtained from all VET Registry members, and study procedures were approved by the VA Puget Sound Institutional Review Board.

The main medical conditions of interest for the current study measured at study baseline during the Men’s Health Study were arthritis, diabetes, hypertension, and CAD (which was considered a proxy for general atherosclerosis [19‐22]). At the baseline assessment, participants reported whether they had ever previously been diagnosed with specific medical conditions, choosing from a list of common conditions, including “arthritis of any kind, or rheumatism”, “diabetes”, “hypertension or high blood pressure”, “coronary heart disease”, “asthma”, “chronic bronchitis”, “emphysema or chronic obstructive pulmonary disease”, “gastroesophageal reflux disease or reflux esophagitis”, “kidney disease”, or “liver disease”. For each condition, participants also reported whether they had received medical treatment for the condition in the past year. Reported medical conditions were not independently corroborated by other means. For the purposes of the current study, the definitions of the main medical conditions of interest were: arthritis defined as self-report of prior diagnosis of “arthritis of any kind, or rheumatism”, without specification as to the type of arthritis, or the joints involved; diabetes defined as self-report of prior diagnosis of “diabetes”, without distinguishing type 1 diabetes from type 2 diabetes; hypertension defined as self-report of “hypertension or high blood pressure”; and CAD defined as self-report of “coronary heart disease”.

Data on self-reported medical conditions were used to generate an overall score reflecting the burden of major medical conditions (the ‘comorbidity score) in a manner analogous to that used in the Self-Administered Comorbidity Questionnaire (SACQ), a validated self-report measure for comorbidity burden that is commonly used in orthopedic research [37]. The comorbidity score used these medical diagnosis groups included in the SACQ: heart disease, hypertension, lung disease, diabetes, ulcer/stomach disease, kidney disease, liver disease, and arthritis. For calculation of the comorbidity score, lung disease was defined as a self-report of “asthma”, “chronic bronchitis”, or “emphysema or chronic obstructive pulmonary disease”, and ulcer/stomach disease was defined as “gastroesophageal reflux disease or reflux esophagitis”. Although typically part of the SACQ, back pain was not included in our comorbidity score since it was the outcome of interest, and information for three other diagnosis groups from the SACQ (anemia, cancer, and depression) were also not available as part of the baseline assessment in the Men’s Health Study. Scoring followed the same general practice as used in the SACQ: individuals could receive up to a maximum of 2 comorbidity ‘points’ for each medical condition: 1 point for the presence of the condition, and another point for having received treatment for it in the past year. The number of points were summed across all 8 conditions, for a possible range between 0 (lowest possible comorbidity burden) to 16 (highest possible burden). This comorbidity score differed from the SACQ in that the SACQ also includes a 3rd item inquiring whether each condition results in functional limitations, permitting up to 3 comorbidity points for each condition [37].

At baseline, respondents reported whether or not they had ever previously been told by a doctor that they had ‘back problems’, slipped or ruptured disk, or sciatica. Individuals who reported having no prior back problems at baseline constituted the study sample that was followed longitudinally. At the follow-up assessment 11 years later, respondents reported whether they had ever had chronic back pain, without specification as to the location or duration of their back pain. Those participants without back problems at the study baseline, who went on to report chronic back pain 11 years later, were classified as having developed CBP (‘incident CBP’).

Adjustment variables and potential confounders included participant age, race, and educational attainment. Data obtained from military records classified race as white vs. non-white. Participants reported educational attainment as the highest grade/year of school completed, and the highest degrees obtained.

We used statistical methods for co-twin control studies [31] to examine associations between baseline medical conditions and incident CBP in longitudinal analyses restricted to individuals who did not report back problems at baseline. First, we conducted multivariable-adjusted individual-level analyses equivalent to that used in conventional studies of unrelated individuals, using generalized estimating equations (GEE) to account for clustering by twin pair when calculating odds ratios (ORs) and 95% confidence intervals (CIs). In these analyses we adjusted for age, race, and education, yielding effect estimates similar to those yielded by any conventional epidemiology study. Due to the interrelatedness of the different cardiovascular factors examined, and to avoid conditioning on intermediates along theoretical causal pathways, only one medical condition at a time was included in the multivariate models [38]. In a separate analytic step of the individual-level analysis, we also included the comorbidity score as an additional adjustment variable. In each such model, the main medical condition of interest was excluded from calculation of the comorbidity score so that medical conditions were not ‘counted’ twice; for instance, the model for the arthritis-incident back pain association adjusted for the comorbidity score, but for the purpose of that specific model, arthritis was not included in calculation of the comorbidity score. Next, we conducted within-pair analyses comparing twins to their co-twins using conditional logistic regression analyses, in which only twin pairs who are discordant for the CBP outcome are informative. These within-pair analyses account for familial factors (which include both genetic and early ‘shared’ family environmental factors) as well as unmeasured confounders due to the similarities within a twin pair. The degree to which within-pair estimates differ from individual-level estimates of association can be used to infer whether familial factors are a source of confounding [31]. If individual-level and within-pair associations are of similar magnitude, this implies no confounding or minimal confounding due to familial factors (Fig. 2, Scenario A). If, however, familial factors confound the association between a health-related factor and CBP then the magnitude of association in within-pair analyses will typically be closer to the null (ie, an OR of 1.0) than that observed in the individual-level analyses (Fig. 2, Scenario B) [39]. The analytic approach involved complete-case analysis. Small sample sizes are common in longitudinal within-twin-pair analyses using categorical outcomes, since only twin pairs where both twins lack the outcome at baseline, and where twins are discordant for the outcome at follow-up, are informative. Thus, inferences regarding the strength of associations must not only refer to statistical significance, but also to the magnitude of point estimates and width of confidence intervals (CIs). Due to sample size concerns, the within-pair analyses were not stratified by twin zygosity (monozygotic vs. dizygotic); however, for completeness, zygosity-stratified results are provided in Additional file 1.

Self-reported arthritis was significantly associated with incident CBP in multivariable-adjusted individual-level analyses (OR 1.8 [95% CI 1.4–2.3]; p < 0.001), and further adjustment for comorbidity score resulted in a slight decrease (OR 1.7) in the magnitude of this association (Table 2). However, the association was notably attenuated and no longer significant in the within-pair analysis (OR 0.9 [95% CI 0.4–1.8]; p = 0.72).

Diabetes was not significantly associated with incident CBP in individual-level nor within-pair analyses (Table 2). Hypertension was significantly associated with incident CBP in multivariable-adjusted individual-level analyses (OR 1.3 [95% CI 1.0–1.5]; p = 0.04); the magnitude of this association was slightly smaller and not statistically significant after further adjustment for comorbidity score (OR 1.2 [95% CI 0.9–1.5]; p = 0.14). Although the magnitude of the hypertension-CBP association was slightly larger in within-pair analysis, the confidence intervals were wider and the results not statistically significant (OR 1.3 [95% CI 0.6–2.6]; p = 0.48). CAD was significantly associated with incident CBP in multivariate-adjusted individual-level analyses (OR 1.6 [95% CI 1.0–2.3]; p = 0.05)); the magnitude of this association was slightly smaller and not statistically significant after further adjustment for comorbidity score (OR 1.5 [95% CI 0.9–2.3]; p = 0.09). Within-pair analyses yielded wide confidence intervals and associations that were not statistically significant, although associations were of comparable magnitude (OR 2.0 [95% CI 0.5–8.0]; p = 0.33) to the individual-level analyses. Further adjustment of the within-pair analyses in Table 2 of specific medical conditions for comorbidity score showed no material differences (data not shown).

The medical comorbidity score was significantly associated with incident CBP in multivariate-adjusted individual-level analyses (Table 2) (OR 1.2 per comorbidity point [95% CI 1.1–2.3]; p < 0.001). The magnitude of this association was slightly smaller in within-pair analyses (OR 1.1 per comorbidity point [95% CI 0.9–1.4]; p = 0.32) and not statistically significant.

Within-pair analyses stratified by zygosity were not materially different from the main within-pair analyses, although confidence intervals generally were wider due to the smaller sample sizes involved (see Additional file 1).