Study sample and data collection
San Francisco residents ages 13 years or older diagnosed and reported to the San Francisco DPH with HIV or AIDS, and who died from January 1, 2002 through December 31, 2016 were included. Socio-demographic and HIV transmission characteristics, AIDS-related clinical data, and vital status were collected as part of routine HIV surveillance on all study participants. Multiple coded causes of death data were obtained from matches with the National Death Index Plus. HIV surveillance data collection is required by state law (California Health and Safety Code (HSC) 121022 and HSC 120130), therefore Institutional Review Board (IRB) approval and consent was deemed unnecessary according to state regulations. Data used in this study is protected by state law and is not publicly available.
The cause of death information from death certificates was coded using the International Classification of Diseases 10th edition16. All coded causes and conditions contributing to death (including the underlying cause of death), as listed on the death certificate, were included in our multiple cause of death category. For our dependent variable, cause of death, we included the most frequently occurring causes, both HIV/AIDS-related (presented as all HIV/AIDS-related deaths and separately as AIDS-related malignancies and AIDS-related opportunistic infections) and non-HIV-related. In addition, we included the following causes of death regardless of the frequency of their occurrence because they are more likely to occur in people who are homeless: non-drug-related accident, assault, suicide, drug-related overdose, alcohol-related liver disease, and mental disorders. Composite categories were made for AIDS-defining opportunistic infections17, any HIV/AIDS cause of death, AIDS-defining and non-AIDS-defining cancers, heart disease (including cardiomyopathy, cerebrovascular, and ischemic disease), and mental disorders (including mental disorders due to substance abuse).
Most socio-demographic characteristics in our analyses included those obtained at the time of HIV diagnosis, except for age at death, county of death, and housing status. For the purpose of HIV surveillance, the Centers for Disease Control and Prevention (CDC) uses the federal definition of homelessness. Thus, a person is defined as homeless if their medical record states that the patient is homeless or not housed at the time of HIV or AIDS diagnosis, or the patient’s address at diagnosis is a known homeless shelter or free postal address not connected to a residence (i.e., general delivery). This definition does not include those with marginal (living in a single room occupancy units) or unstable (living with friends) housing. For our analyses, a person was defined as homeless if periodic review of the medical record stated that the patient was homeless at the time of HIV diagnosis, AIDS diagnosis, at follow up, or at death1.
In contingency table analyses, age at death was categorically defined by decades: 20-29, 30-39, 40-49, 50-59, 60-69, and 70-79. Race was categorized as non-Hispanic White, Hispanic, African-American, or other, including multi-race/ethnicity. HIV transmission risk category was categorized as men who have sex with men (MSM), person who injects drugs (PWID), men who have sex with men and who also injects drugs (MSM-PWID), heterosexual, or other. Gender was classified as male, female, or transgender. We defined an individual as living below the federal poverty level at diagnosis if they lived in a census tract where more than 20% of persons aged 18 years or older had a median annual household income that was below the United States poverty level18. Country of birth was dichotomized as USA/US Dependency or other and county of residence at death was dichotomized as San Francisco or other. Prescription of ART was dichotomized as yes or no.
Statistical analyses
The primary independent variable was housing status dichotomized as homeless at or after HIV diagnosis or not. The primary dependent variable was multiple condition or disease-specific cause of death.
We compared the distribution of case characteristics by housing status using contingency tables, and p values were calculated using chi-square and Cochran-Armitage trend tests. This method was also used to explore the differences in causes of death among PLWH stratified by county of residency at time of death (San Francisco vs. non-San Francisco residents), which serves as a surrogate measure for access to San Francisco specific interventions.
To compare differences in cause-specific deaths between the housed and the homeless populations, unadjusted and adjusted Poisson regression models for binary outcomes were performed to calculate the prevalence ratios (PR’s), p values, and 95% confidence intervals (CI’s). The binary explanatory variable for the unadjusted model was housing status, with housed individuals being the reference group. We considered a p value of less than 0.05 significant. Adjusted models were constructed for each cause of death if the housing variable resulted in a statistically significant PR in the unadjusted model.
Adjusted regression models controlled for the following factors with their respective reference group noted in parentheses: gender (male), race (non-Hispanic white), a concurrent initial diagnosis of HIV and AIDS (yes or no), HIV transmission category (MSM), low income (yes or no), and county of residence at death (San Francisco). In addition, the regression models were adjusted for age, which was continuous per decade, and year of death, which was continuous per year. All statistical analyses were performed using SAS® software version 9.419.