Methods
Description of the South Texas HIV Cohort
The South Texas HIV Cohort includes patients receiving care from January 1, 2007 to December 31, 2010 in the Family Focused AIDS Clinical Treatment and Services clinic. This clinic is the largest HIV treatment center in south-central Texas and located in a publicly funded county hospital affiliated with an academic medical center in San Antonio, Texas. Study data were obtained from an electronic medical record (EMR) system and included demographics, health insurance information, physiologic measures, clinical diagnoses, laboratory data, visit data, and prescribed medications. Cohort patients were ≥18 years old at their first clinic encounter and not known to be incarcerated per Institutional Review Board specifications. If patients under 18 years of age were seen in the clinic, EMR data were censored before their 18th birthday. HIV diagnosis was validated by an ICD-9-CM code for HIV infection (v08 or 042.xx) in the EMR or a visit to the HIV clinic and confirmatory laboratory results (positive HIV ELISA and Western blot, or plasma HIV-1 RNA level >1000 copies/mL). The 19 questionable cases were resolved by chart review. The resultant cohort totaled 1890 individuals who received longitudinal care in the HIV clinic.
For our analysis of weight change, we selected patients with: (1) at least 2 HIV clinic visits during the observation period from January 1, 2007 to December 31, 2010, (2) at least 6 months between consecutive BMIs, (3) no pregnancy in this time frame, and (4) white, black, or Hispanic race–ethnicity because other racial-ethnic groups comprised only 1.5% of the cohort. We excluded patients with missing plasma HIV-1 RNA or initial CD4 count data. We also excluded patients who were underweight (BMI, <18.5 kg/m at baseline measurement) because weight gain for underweight individuals is often a treatment goal. Our final cohort totaled 1214 individuals (Fig. 1). Of the 959 excluded patients, 20% were women; median age was 39 years (intraquartile range, 30–47); and race–ethnicity was 34% white, 42% Hispanic, 17% black, and 7% other.
(Enlarge Image)
Figure 1.
Cohort sample selection flow diagram.
Outcome Variables
The primary study outcome was significant weight gain specified dichotomously as ≥3% annual increase in BMI, calculated as the difference between the first (baseline) and last BMI in the observation period divided by the observation time in years. We selected this outcome because a 3% annual weight gain leads to obesity in only 3 years among persons at the midpoint of the overweight range (eg, 27.5 kg/m), and it is 5–10 times greater than the average increase in BMI for the general US population, which is 3.4% in men and 5.2% in women over a 10-year period. To calculate change in BMI, we excluded heights that differed by >10 cm for an individual (5.4% of all measurements) and used each patient's average height to calculate BMI using the formula: weight/(average height during observation period). All BMIs <15 kg/m or >50 kg/m and any BMI that changed more than ±5% from the previous recorded BMI were manually checked, and apparent data entry errors were excluded (406 of 16,451 or 2.5%).
Independent Variables
Race–Ethnicity/Insurance Status. Self-reported race–ethnicity was categorized as Hispanic, non-Hispanic black (black), or non-Hispanic white (white). Health insurance status, based on the most common type for all visits, was classified as: (1) insured (ie, Medicare, Medicaid, and all private insurance programs) or (2) uninsured (ie, medical care provided through a county or federal financial assistance plan, available on a sliding scale for those with incomes <300% of the Federal Poverty Level). Insurance status serves as an indicator of SES as in other studies. We specified the hypothesized interaction between insurance and race/ethnicity in a 4 category variable: (1) insured white, (2) uninsured white, (3) insured Hispanic or black (minority), or (4) uninsured minority.
Demographic Variables: These characteristics included age at first clinic visit in the observation period, gender, and mean household income per year for the patient's residential zip code as an alternative measure of SES. Self-reported HIV transmission categories were heterosexual sex, men who have sex with men, injection drug use (IDU), and other, including unknown or missing. The IDU category includes those men who have sex with men reporting IDU due to small sample size of that dual risk category (n = 13).
Clinical Variables: BMI is measured in the clinic before each provider visit using one of 2 regularly calibrated standing scales to calculate weight. Height is calculated using a stadiometer, and BMI is calculated as weight in kg/(height in meters)2." The first or baseline BMI in the observation period was categorized as normal (18.5 to <25 kg/m), overweight (25 to <30 kg/m), or obese (≥30 kg/m). Immunologic status, based on the first CD4 cell count in the observation period (hereafter initial CD4 cell count), was categorized as <200, 200–349, or ≥350 cells per microliter. Because of changes in plasma HIV-1 RNA assay threshold during the observation period, virologic failure was defined as any 2 plasma HIV-1 RNA levels >1000 copies per milliliter at least 24 weeks after starting ART. Diagnosis of diabetes mellitus was based on ICD-9-CM diagnosis codes 250.XX at ≥2 visits or hemoglobin A1c ≥6.5%; and hypertension from ICD-9-CM codes (401.xx at ≥2 visits) or 2 blood pressure measurements ≥140 mm Hg systolic or ≥90 mm Hg diastolic. Hepatitis C Virus (HCV) infection was determined by positive HCV antibody test or 2 ICD-9-CM codes for chronic active HCV. ART type was based on EMR prescription records during the observation period: (1) receipt of ART without a protease inhibitor (PI) prescription, (2) receipt of ART with a PI prescription, and (3) no prescriptions for ART, because of the known association of PIs with metabolic syndrome, obesity, and weight change in previous studies.
Health Care Utilization Variables: The per-patient observation period was the time between the first and last available BMI measurement within the 2007 to 2010 observation period. The number of HIV clinic visits with documented weights during the observation period was used as a metric of care intensity, with those patients in the lowest quartile (<7 visits) characterized as receiving infrequent HIV clinic follow-up.
Statistical Analyses
We first examined the interaction between race–ethnicity and health insurance status with the primary outcome, significant annual weight gain. We then estimated multivariate logistic regression models to examine the association of race–ethnicity/insurance status with significant weight gain after adjustment for demographic, clinical, and health care utilization variables. We excluded HIV transmission risk from the final model because it was not significant in ours or other studies. Because of collinearity with insurance type, we also excluded mean household income in residential zip code from the final model. We excluded indicators for diabetes or hypertension because these conditions may result from weight gain instead of being in the causal pathway. HCV status was excluded because it did not contribute specificity to any of the models. Interactions between individual predictors were examined using second-order interaction terms in the logistic regression model.
In sensitivity analyses, we excluded the few patients who were not treated with ART (4%) and conducted a comparison between only Hispanic and non-Hispanic white patients. We also examined an alternative specification of the outcome as absolute change in BMI per year over the observation period. These additional analyses did not change our conclusions so they not reported.
We examined the longitudinal trajectory of BMI change using all BMI measurements for each patient over the observation period in a mixed-effects model, including both a random intercept and a random slope and using continuous BMI values as the response variable. The fully adjusted model included: the 4-level race/ethnicity-insurance variable, years since baseline BMI, age at the first visit, length of observation, gender, virologic failure, initial CD4 cell count, initial BMI, ART type, and infrequent HIV clinic follow-up. Second-order interactions between explanatory variables were also examined.
For all analyses, predictor variables were considered statistically significant if associated with a P value of <0.05% and 95% confidence intervals (CIs) were used. Pearson χ, Kruskal–Wallis H test, and logistic regression analyses were conducted using PASW Statistics (Version 17.0, Armonk, New York), and mixed effects models were generated using SAS (Version 9.2, Cary, North Carolina), GLIMMIX procedure.