Our analyses were performed using two linked data sources: (1) the baseline (1993-1994) interview responses of self-respondents in the Survey on Assets and Health Dynamics Among the Oldest Old (AHEAD); and, (2) the Medicare carrier claims for those respondents from 1993 to 2007. The design and sampling approach in the AHEAD have been well described elsewhere [16–19]. All analyses are weighted to adjust for the over-sampling of African-Americans, Hispanics, and residents of Florida.
There were 7,447 older adults who completed baseline AHEAD interviews in 1993-1994. A total of 1,937 people were excluded from the analytic sample due to (a) the inability to link their Medicare claims (N = 802), (b) being in a managed care Medicare plan at baseline (N = 605), or (c) not being a self-respondent at baseline (N = 530). The final analytic sample consisted of 5,510 individuals, some of whom were censored post-baseline due to death (N = 3,369) or subsequent enrollment in managed care (N = 988). In previous work propensity score re-weighting was used to address the potential sample selection bias introduced by these exclusion criteria; however, such adjustments did not meaningfully alter the results and thus were not used here .
Measuring Chiropractic Use
Chiropractic visits were identified by using the Health Care and Financing Administration's (HCFA) specialty provider code for chiropractors in the Medicare carrier claims file. Claims were aggregated to the individual level in each calendar year of service, as well as across the entire period for which the user was in the sample. Users were partitioned into two groups: those exceeding the 12 chiropractic visit "soft-cap" in any calendar year (high volume users) and those with 12 or fewer annual chiropractic visits in any calendar year (lower volume users).
Framework for selection of covariates
The AHEAD survey data contains a litany of information on individuals, the totality of which has not been available in previous studies of chiropractic use. In order to bring structure to our inclusion of covariates, we selected them based on Andersen's Behavioral Model of Health Services Use . This model highlights predisposing characteristics that play a role in predicting and explaining health services use in general, and for our analysis, chiropractic use. Variability in demographic factors, such as age, gender, and race could be expected to play a role in explaining chiropractic utilization variation. Social structure variables like education, marital status, and income might also influence why services are sought. Personal enabling resources like having a job, having supplemental insurance, and being able to drive a car would predictably improve access to chiropractic. Need for services can be differentiated by "evaluated" need, such as that identified by a health provider (i.e. arthiritis), or "perceived" need, such as how people view their own health status, functional limitations, psycho-social state, or experience symptoms of pain and illness. Health behavior and lifestyle choices, such as smoking, alcohol consumption, and weight also arguably reveal individual preferences for health that may affect demand for chiropractic. Prior health services utilization measures indicate individual propensity to use health services and prior access to these services. Measures of physician supply, rural-urban characteristics, and distance to a chiropractic college are also included in our model to provide indicators of access and possible familiarity with chiropractic as a health profession. Geographic location measures may reflect differences in regional preferences for chiropractic. All covariates were obtained from baseline interview responses to the AHEAD survey, with the exception of the distance to chiropractic college measure, which was calculated as the distance between a subject's baseline census tract and that of the nearest chiropractic college.
Demographic, socioeconomic and geographic variables
Demographic covariates are age at baseline, sex, race, and marital status. Socioeconomic measures included educational attainment, income distribution (quintiles), the number of supplemental health insurance policies (zero vs. one or more), whether the respondent was working for pay at baseline, and whether the subject was able to drive a car or not. We included a set of indicators measuring geographic location based on the Health Resources and Services Administration's (HRSA) ten region definition , with the Midwest region as the reference group. A measure of rurality was also included, defined by whether a person lived in a non-metropolitan (rural) or metropolitan (non-rural) area. The final measure of geographic interest was distance to nearest chiropractic college. This was included because relative nearness to a chiropractic college might influence local demand for chiropractic care, and chiropractic college graduates may be more likely to locate closer to these institutions thereby increasing supply of the service. Distance was re-coded into two categorical levels: near (under 150 miles to the nearest chiropractic college) and far (greater than 150 miles to the nearest chiropractic college).
Health and health services use measures
Disease history and comorbidity were measured by participants' responses to survey questions about whether they were ever told by a medical doctor they had a specific health condition. The health conditions included were arthritis, cancer, any heart condition, diabetes, lung disease, hip fracture, or hypertension. In order to reflect the extent of a respondent's comorbidity, we re-coded the count of comorbid conditions into four categories: zero, one (reference category), two, or three or more comorbid conditions. Self-rated health measures at baseline assessed each respondent's view of their own health in terms of "excellent", "very good", "good", "fair", or "poor".
Functional health status was measured in multiple ways. The first was how the respondent answered the question "Are you often bothered by pain?"(yes/no). In addition to the standard activities of daily living (ADLs) and instrumental activities of daily living (IADLs), the AHEAD respondents were asked about five additional measures of upper and lower body limitations to further assess physical impairment. These measures were 'difficulty picking up a dime', 'difficulty lifting ten pounds', 'difficulty pushing or pulling large objects', 'difficulty climbing a flight of stairs', and 'difficulty walking several blocks'. We included pain, ADLs, IADLs, and the additional measures of physical function in our analysis because these are conditions likely to be associated with seeking chiropractic care.
Because pain and lack of physical function are also associated with depression and cognition in older adults, we also included measures of depressive symptoms based on a respondent's score on the Centers for Epidemiologic Studies Depression (CES-D)  and measures of cognition based on their scores from the Telephone Interview for Cognitive Status (TICS-7) . In assessing depressive symptoms we classified individuals into three categories: zero depressive symptoms (reference group), one or two depressive symptoms, or three or more depressive symptoms. Similarly we classified cognitive function into three discrete groups: zero-to-ten on the TICS-7 (low cognitive functioning), eleven-to-thirteen on the TICS-7 (normal functioning, and the reference category), and fourteen-to-fifteen on the TICS-7 (high cognitive functioning).
Health lifestyle related factors included cigarette smoking, alcohol consumption, and body mass. Cigarette smoking and alcohol consumption could be related to a person's way of coping with physical pain . Respondents were asked to describe themselves as a current smoker, former smoker, or someone who has never smoked. We re-coded these responses into a single indicator of 'never smoked' versus 'current and/or former smoker' in order to discern between those who might have other underlying health conditions from smoking from those who never smoked. Regarding alcohol consumption, respondents were asked if they ever drank beer, wine or liquor, and if they did how many drinks they averaged per week. In this analysis we used the 'ever drink' variable indicating that a respondent has at least one alcoholic drink during a week vs. never drinking alcohol. BMI measures (kg/m2) were included to reflect a potential association between carrying excess weight and back pain. In our study we have four BMI categories: obese ≥ 30 BMI, overweight 25 ≤ BMI < 30, normal weight 18.5 ≤ BMI < 25, and underweight < 18.5 BMI. The normal weight group was the reference category.
Two measures of health services use in the twelve months prior to baseline interview were included: hospital stays (raw count) and the number of physician visits. With regard to the measure of physician visits, the AHEAD survey asks how many times the respondent talked to a medical doctor about their health in the last 12 months, so this measure does not capture visits to non-MD health providers. These are included as baseline indicators of access to health services as well as health status markers. To capture non-linearities in outpatient services use this variable was re-coded into four levels based on the number of physician visits: one or fewer physician visits, two to three physician visits, four to six physician visits, and seven or more physician visits. The reference category was two to three physician visits. A third variable was included to represent the supply of physicians at the local level, as measured by the number of active, nonfederal MDs per 1,000 people in the respondent's county of residence.
Two separate binomial logistic regression was used to identify factors associated with (1) chiropractic use, and (2) conditional upon any use, to identify factors associated with high levels of annual chiropractic use (i.e., those exceeding a "soft cap" of 12 visits in a calendar year versus lower levels of annual use). The logistic regression models used forced entry that included all covariates described above to determine the odds of using chiropractic versus not using. The odds ratios for each of the covariates in the second regression are of being in the high user group versus not being in the high user group. No interaction terms were hypothesized or included in these analyses. In each logistic regression we followed standard procedures for model development and evaluation [26–28].