As depicted in Fig. 1 we analyzed the relationship between CSM and CeAD with three different types of controls: (1) a case-control design consisting of cases with CeAD and controls from the general population of Medicare beneficiaries matched by sex, age (in years) and calendar year of the CeAD; (2), a case-control design with the same cases and controls with ischemic stroke from the population of Medicare beneficiaries; and (3) a case-crossover design, in which exposures prior to the CeAD are compared to exposures in the time period 6 months earlier in the same patient.
The study subjects included fee-for-service Medicare beneficiaries using 100% 2007–2015 Medicare Part A (covering hospitalizations) and B (covering outpatient encounters and physician services) files. We included beneficiaries aged 65 and older with at least one Part B claim in a calendar year with the following annual exclusions: (a) any Medicare Advantage; (b) less than full Part B enrollment for the entire calendar year (or from the month turning 65 to month of death); and (c) residence outside the 50 United States or Washington, DC. All subjects were concurrently and continuously enrolled in Medicare Parts A and B for at least two consecutive years.
The primary outcome was the occurrence of a CeAD which was sub-divided into (1) VAD and (2) CAD. The cases were identified as beneficiaries with a new (not recorded in the prior year) diagnosis of International Classification of Disease (ICD-9) code 443.24 (VAD) or 443.21 (CAD) in the primary diagnosis field on at least one inpatient hospital claim or primary/secondary diagnosis for outpatient hospital and Part B claims on at least two separate days. The majority (73%) of CeAD cases had a diagnosis of stroke within 30 days of the CeAD diagnosis.
As discussed above, there were three types of controls. For population controls, we matched Medicare beneficiaries without CeAD, which we refer to as population controls, to the CeAD cases in a 10:1 ratio. Controls were matched for age (in years), sex and having at least one claim on the same day (+/- 1 week) as the case. Controls were excluded if they ever had a diagnosis of CeAD. For the ischemic stroke controls, we identified beneficiaries with a diagnosis code for non-CeAD-associated ischemic stroke (ICD9 codes 431, 432, 434, 433.10, or 433.11) in the primary diagnosis field on at least one inpatient hospital claim or primary/secondary diagnosis for outpatient hospital and Part B claims on at least two separate days. For the case-crossover study we evaluated claims for the CeAD cases in the corresponding time period 6–7 months prior to their CeAD.
The index date is defined as the date of diagnosis of CeAD in the cases, as the date of the corresponding claim in the population controls, and as the date of diagnosis of ischemic stroke in the stroke controls. In the case-crossover analysis, the index date for the control period is the date 180 days prior to the occurrence of CeAD.
Primary and secondary exposures
The primary exposure was CSM, as identified by Current Procedural Terminology (CPT) codes 98,940–98,942 (indicating spinal manipulation by a doctor of chiropractic) associated with a primary diagnosis of headache (ICD-9 code 339.xx) or neck pain (ICD-9 codes 721.0, 721.1, 722.0, 722.4, 722.71, 722.81, 722.91, 723.1-723.8, 739.1, 756.16, 839.0x, 847.0, 953.0, or 953.4) or other disorders of the head or neck pain that are commonly treated by spinal manipulation (ICD-9 codes 739.1, 723.1, 739.0, 722.4, 839.xx, 723.3, 847.0, or 839.00) in order to try to localize the manipulation to the cervical region. The secondary exposure was the occurrence of an encounter for Evaluation and Management (E&M) as indicated by ICD-9 codes 99,201–99,205 and 99,211–99,215 with the same associated diagnoses discussed above. Using the primary and secondary exposure we created a 3-level categorical variable, (i) CSM, (ii) E&M but no CSM and (iii) neither CSM nor E&M. The E&M only category was selected to the referent group. Individuals with both a CSM and an E&M visit in the requisite time period were analyzed in the CSM group.
Timeframes for the exposure
We created the 3-level exposure described above for each of the following time frames, up to 7, 14 and 30 days prior to the index event. For instance, for the 7-day time window, the 3-level categorical exposure is (i) CSM in the 7 days before index, (ii) E&M but no CSM in the 7 days before index, and (iii) neither CSM nor E&M in the 7 days before index.
Covariates included demographics age, sex, race (categorized as White, Black, Asian, Hispanic, North American Native, and Other) and calendar year. In addition, to control for comorbidities we considered all diagnoses 14 to 365 days preceding the index date grouped using the Multi-level Clinical Classification Software (CCS) of the Healthcare Cost and Utilization Project . ICD-9 codes were grouped into categories using the third level of the multi-level classification.
We used odds ratios to characterize the association between CeAD and the 3-level exposure, CSM vs. E&M vs. neither, with E&M set as the referent group. To control for the covariates described above we applied multivariable logistic regression to the dataset consisting of CeAD cases and ischemic stroke controls. To control for covariates in the analysis comparing CeAD cases to matched population controls we used multivariable conditional logistic regression. The conditional logistic regression estimates the odds ratios conditional on the sex-age-year matches, in addition to race and diagnostic covariates. Sex, age and calendar year therefore have null coefficients due to matching on them.
Due to large number of diagnostic covariates (over 430) we used variable selection methods. To select predictors of CeAD cases versus ischemic stroke controls we employed Least Absolute Shrinkage and Selection Operator (LASSO) for logistic regression to select covariates . In particular, we selected those covariates that had a nonzero LASSO coefficient. The penalty parameter in the LASSO was determined using 10-fold cross-validation with optimization of the binomial deviance (analogous to log-likelihood). To select predictors of CeAD cases versus population controls that accounts for matching by sex, age, and year, we employed stepwise conditional logistic regression (we are not aware of a LASSO adaptation to conditional logistic regression). This approach to covariate selection served to identify any comorbidities that predict diagnosis of CeAD and may act as confounders.
We tested if there was an association of the exposure (CSM vs. E&M vs. neither) with CeAD in the case-crossover analysis using conditional logistic regression conditioning on subject (e.g., the pair of observations from CeAD and 6 months earlier).
All analyses above were repeated for each of the three time frames for exposure to CSM and E&M (7, 14 and 30 days). That is, we report odds ratios comparing CSM to E&M and neither CSM nor E&M to E&M exposure in each of these time periods.
As yet another perspective, we employed a propensity score approach to estimate the odds ratio relating CeAD to CSM. This consisted of the following steps. Using data from the population controls we modeled the occurrence of CSM in the 7 days before index as a function of demographics and comorbidities using logistic regression. The predicted probabilities (fitted values) from this logistic regression were used to calculate the inverse weighted propensities. The final step was to employ a weighted logistic regression to estimate the odds ratio relating VAD (or CAD) to CSM in the previous 7 days.
Our study is powered (at 80%) to detect odds ratios of VAD with CSM in the previous week (relative to E&M) for the population of 2.0. The corresponding detectable odds ratio for CAD with CSM is 1.8. The detectable odds ratios using the Ischemic stroke controls are slightly smaller as we had more than 10 of those controls per CAD case. Statistical software employed was SAS 9.4, and R (including libraries, tidyverse, & glmnet).