Sample and Setting
Modeled after the National Health and Nutrition Examination Survey, the NYC HANES is a population-based, cross-sectional survey of non-institutionalized NYC residents ≥ 20 years old. The NYC HANES was designed to provide citywide prevalence information on conditions identifiable only through a physical examination or biologic specimen testing (i.e., hypertension or diabetes) and conditions that are not easily ascertained by a telephone survey (i.e., drug use, incarceration, and domestic violence). The study design is described in detail in previous publications [20]. Briefly, NYC HANES measured key health indicators in a sample of randomly selected community residents. A 3-stage, cluster sampling plan was used to recruit participants between June and December of 2004. Sample selection included random selection of census blocks or groups of blocks, followed by households within the selected segment, and finally study participants within selected households. The survey included a face-to-face computer assisted personal interview, private audio computer assisted self-interview for questions about drug use and incarceration, and a physical examination and laboratory testing. Of the 3047 selected eligible survey participants, 1999 individuals completed the interview for an overall response rate of 55%. All participants gave informed consent, and the study received approval from the NYC Department of Health and Mental Hygiene Institutional Review Board. This study focuses on the role of incarceration in the health of this population of 1999 individuals, among whom 160 (8%) reported a history of incarceration.
Analytic Variables
History of incarceration
To assess history of incarceration, participants were asked: "Have you ever spent any time in a correctional facility, jail, prison, or detention center as an adult, that is, 18 years old or older?" A positive response to this item was categorized as having been incarcerated. Individuals who did not respond to this question were excluded from the analysis (N = 9).
Chronic Disease Measures
To assess asthma, we used a series of questions that captured an individual's history of asthma and measures of asthma severity. Participants who responded yes to both: "Has a doctor or other health professional ever told you that you have asthma?" and do you "still have asthma?" were categorized as currently suffering from asthma. Participants who reported currently having asthma were then asked the following measures of disease severity: "During the past 12 months have you had an episode of asthma or an asthma attack?" and "During the past 12 months have you had to visit an emergency room or urgent care clinic because of your asthma?" Diabetes was defined by self-reported disease, fasting glucose ≥ 126 mg/dL, or use of anti-diabetic medication. Hypertension was defined as a systolic blood pressure ≥ 140 mmHg, diastolic blood pressure ≥ 90 mmHg, or current use of prescribed anti-hypertension medication.
Potential confounders
The NYC HANES contains a variety of potential confounders allowing us to better assess the independent contribution of having a history of incarceration on the prevalence of asthma, diabetes, and hypertension. Socioeconomic status of each participant was approximated using measures of self-reported educational attainment and family income. Participants were asked: "Which of these categories best describes your total combined family income for the past 12 months?" Those who responded that their family income was less than $20,000 were defined as meeting 100% of the federal poverty line and being of low socioeconomic status [21].
For cigarette smoking, participants were categorized as current smokers if they indicated that they "smoke every day or some days." Former smokers were defined as those who reported having smoked at least 100 cigarettes in their entire life, but did not currently smoke. Use of cocaine, intravenous drugs, and excessive alcohol was ascertained with private audio computer assisted self-interview. For cocaine and intravenous drug use, participants were categorized as current users if they had used in the past 12 months, former users if they had ever used but not in the past 12 month, or never users. For alcohol use, participants were categorized as excessive alcohol consumers based on the at-risk consensus thresholds of the National Institute on Alcohol Abuse and Alcoholism, which were met if a man consumed equal to or greater than 14 drinks per week or a woman more than 7 drinks per week, where a drink was counted as 12 ounces of beer, 5 ounces of wine or 1-1/2 ounces of spirits [22]. Binge drinking was categorized as drinking more than 5 drinks on one day in the past 12 months. Intimate partner violence was measured by a positive response to the following question: "In the past 12 months have you been frightened for the safety of yourself, your children, or friends because of anger or threats of an intimate partner?"
Health care access
To assess health care access, participants were asked if they had a primary care provider with a series of questions, beginning with: "Is there a place that you usually go when you are sick or you need advice about your health?" Respondents who answered "yes" were then asked, "What kind of place do you go to most often: is it a clinic, doctor's office, emergency room or some other place?" Those who responded "clinic or health center" or "doctor's office or HMO" were coded as having a regular primary care provider. Participants were also asked whether they had "health insurance or some other kind of health care plan" to define insurance status.
Analytic Methods
A challenge to estimating the true impact of incarceration on health outcomes is that studies which randomly assign individuals to incarceration are unethical and unfeasible. Observational studies are complicated by the fact that individuals who have been incarcerated are inevitably different from individuals who have never been, which may bias the estimated impact of incarceration on health outcomes. As a result, researchers have devised strategies for observational studies that approximate the design features of a randomized experiment and reduce confounding. These include statistical control with multiple regression and propensity score matching. Because multiple regression may lead to overparameterization if the number of potential confounders is large relative to the number of study units, we chose to use both methods to estimate the influence of having a history of incarceration on the prevalence of chronic diseases.
Propensity Score Matching
Propensity score methods use measured participant demographics and clinical and behavioral characteristics to match individuals on the basis of their likelihood to experience a treatment, in this case, their likelihood of having been incarcerated [23]. By matching on background attributes of individuals that might otherwise confound the estimated impact of incarceration on asthma, diabetes, and hypertension prevalence, we can draw inference from the comparison of similar "treated" and "non-treated" individuals. The identifying assumption is that, conditional on measured characteristics, individuals who have been incarcerated would have the same health outcomes as those who have not been incarcerated, were it not for having been incarcerated:
where Y is a health outcome of interest, D is incarceration status, 1 denotes individuals who have been incarcerated and 0 denotes those individuals who have not been incarcerated, and S is a vector of observable characteristics of individuals to which the matching estimator is applied. The propensity score is an estimate of the probability that an individual has previously been incarcerated:
with D = 1 denoting having been incarcerated and the Xs are a series of observable background attributes of the individuals. Assuming incarceration is properly modeled, using the propensity score allows us to treat incarceration as random and thus helps alleviate concerns about sample heterogeneity. In contrast to traditional regression, propensity scores often produce better adjustment of baseline differences than simply including potential confounders in a regression model for two primary reasons. First, the analyst does not have to make assumptions regarding linearity and additivity in the relationship between the individual confounders and the outcomes. Second, propensity scores create a homogenous sample, given the covariates in the model, where individuals differ only in whether they have experienced incarceration [24].
We estimated average treatment on treated (ATT) effects by first calculating the difference in mean prevalence of asthma, diabetes, and hypertension between those individuals who have and have not been incarcerated in the unmatched sample. In order to further demonstrate the robustness of our results, we matched on propensity scores using four different matching procedures (with and without common support and with and without replacement) to assess the effect of incarceration on prevalent asthma, diabetes, and hypertension [25, 26]. Common support matching begins with defining a common support region, which excludes "treated" units whose propensity score is higher than the highest of the control units and control units whose propensity scores are lower than the lowest of the "treated" units. This process is followed by randomly selecting one control that matches on the propensity score with a "treated" participant. The advantage of this matching procedure is that it reduces the probability of a bad match, though some "treated" participants may not be matched [27]. Alternatively, matching without common support enables a match with the control with the closest propensity score, guaranteeing that a match is always found for all the "treated" units. Replacement means that once a control has been matched, it can no longer be matched. As each matching procedure has potential strengths and weaknesses, the results for all four matching estimators are presented.
Logistic Regressions
For medical conditions found to be significantly associated with history of incarceration, we used logistic regressions excluding and including incarceration in order to examine whether having been incarcerated mediates racial and ethnic disparities, or to ascertain the extent to which incarceration is a social pathway through which Black and Latino individuals are placed at higher odds of having that chronic condition [28]. We used a 3-step Baron and Kenny process to assess mediation: demonstrate that (1) race-ethnicity is associated with the medical condition, (2) race-ethnicity is associated with incarceration, and (3) the association of race-ethnicity is attenuated, after adjustment for incarceration [29]. As a final specification, we included the estimated propensity score of each individual - their predicted probability of having been incarcerated, in the regression model. Parameter estimates and standard errors were estimated using probability weights provided by NYC HANES to adjust for complex sampling design, non-response, and post-stratification.