All adult and adolescent (aged ≥13 years) San Francisco residents who were diagnosed with AIDS from January 1, 1996 through December 31, 2006 and reported to the San Francisco Department of Public Health (SFDPH) by November 30, 2007 were included in the study. The AIDS surveillance system is evaluated annually and consistently found to be over 95% complete [26].

More than 90% of AIDS cases undergo a complete medical chart review at the time of report with records re-reviewed and updated every 18–24 months. Data collected include demographic and risk information, insurance status, AIDS-defining illnesses, results of HIV, CD4, and viral load tests, date of initiation and type of antiretroviral therapies, and of prophylaxis against Pneumocystis jirovecii pneumonia (PCP) and Mycobacterium avium complex (MAC).

Housing status is collected at diagnosis. Cases were considered to be homeless if the medical record noted that the patient was homeless or if the address recorded was a known homeless shelter, a health care clinic, or a free postal address not connected to a residence (e.g., general delivery). Persons with missing addresses in the medical record were considered to be housed.

Documentation of deaths was obtained through weekly review of local death certificates, reports from other health departments, and annual matches with the National Death Index, which includes deaths through 2005. Underlying and contributory causes of death were coded according to the International Classification of Diseases – 9^th and 10^th revisions.

Differences in the characteristics of homeless and housed persons were assessed using the chi square test for differences in proportions and the t test for differences in means. The Kaplan-Meier method was used to estimate the distribution of the time from AIDS diagnosis to death among homeless and housed persons; cases not known to have died were censored at the more recent of either the date of their last follow-up or December 31, 2005. The log-rank test was used to assess differences in survival. As with previous studies of AIDS survival, we used all-cause mortality [27-29].

A Cox proportional hazards model was used to estimate the independent associations of housing status with mortality. For the multivariable analysis, we included demographic and risk characteristics, insurance status at diagnosis, the AIDS-defining condition (low CD4 count versus an opportunistic illness), the CD4 count at diagnosis, use of antiretroviral therapy (as a time-dependent variable), and prophylaxis against PCP and MAC.

To assess the proportionality assumption, we checked for interaction between each risk factor and time since AIDS diagnosis; for highly active antiretroviral therapy, a time-dependent covariate, we assessed interaction with time since start of treatment. Because the risk associated with homelessness increased during the study period (p < 0.05), we calculated the estimated relative hazard (RH) at one, three, and five years after diagnosis. Results did not suggest qualitative changes from harm to protection or from protection to harm; therefore, we also calculated a summary RH for this predictor, averaged over the entire follow-up period.

To examine in more detail the factors related to high mortality in homeless people, we compared the five most frequent contributory causes of deaths among homeless and housed AIDS cases, as obtained from the National Death Index.

The SFDPH provides supportive housing services to homeless persons who have the most severe or greatest number of chronic medical or psychiatric needs, or both, through the Direct Access to Housing (DAH) program. The program houses people directly from shelters, street living situations, or institutions. Residents must pay rent on a sliding scale of 30% to 50% of their income; those who are ineligible for public entitlements are housed for free. All DAH sites have dedicated case managers and provide medical services that range from an on-site, full-time nurse and part-time, mid-level clinician (e.g. nurse practitioner) to obtaining care at designated health care facilities located near the DAH residences. Supportive housing services were initiated in 1999, although the DAH program database includes move-in dates as early as 1992. The database contains the names, dates of birth, and move-in dates of persons who are housed by the program.

To identify homeless AIDS cases who obtained DAH housing, we computer matched the AIDS case registry to the DAH database using the patient name and date of birth as the matching variables. Cases who did not match using these criteria were matched using the soundex code (an alpha numeric code that allows for minor differences in the spelling of names) of the last name, first initial of the first name, and date of birth. All matches were manually reviewed. When available, matched cases were further compared using sex and race. To be classified as a homeless AIDS patient who obtained DAH housing, the move-in date must have been after diagnosis. By matching all AIDS cases to the DAH database, we identified persons who were not originally reported as homeless but received supportive housing. Because the DAH program offers housing only to persons who have been identified by their case manager as homeless, AIDS cases who were housed through DAH were included in the analysis of the impact of housing on survival and considered to have been homeless at the time of diagnosis as long as housing was obtained after the AIDS diagnosis. Such individuals were likely misclassified as housed at AIDS diagnosis or they may have become homeless sometime later.

Differences in the characteristics of homeless persons who received housing and those who did not were assessed using the chi square test for categorical variables and the t test for continuous variables. A multivariable proportional hazards model was performed to estimate the independent association of supportive housing with mortality. Obtaining public housing was included as a time-dependent variable to avoid creating a survival bias among those surviving long enough to obtain housing. Included in the multivariable model were demographic and risk characteristics, the AIDS-defining diagnosis (low CD4 count or opportunistic illness), the CD4 count at diagnosis, use of antiretroviral therapy (as a time-dependent variable), and prophylaxis against PCP and MAC. Because there were so few cases with private insurance, this variable was excluded from the model. The proportionality assumption was assessed using the methods previously described. No statistically significant violations were found.

To estimate the increase in average survival time over the first five years in supportive housing, we matched each homeless AIDS case who obtained housing to one control who did not. The cases and controls were matched by date of AIDS diagnosis (plus or minus six months), CD4 count at diagnosis (≤100, 101–200, >200), age at diagnosis (<40 years to ≥40 years), and whether they had injected drugs. Some cases matched with more than one control; in this situation, we selected the control whose date of diagnosis and CD4 cell count most closely matched. To avoid both a survival bias and a potential bias from controls being denied housing because of imminent death, controls were required to survive at least the same number of months from date of diagnosis as the case at the time that the case obtained housing and at least three months beyond the date that the case obtained housing. Follow-up time for cases and controls began the date the case obtained housing.

Using the matched sample, we estimated the increase in mean survival time over the first five years after receipt of housing for cases and their matched controls by calculating the difference in areas under the Kaplan-Meier survival curves [30]. The survival time for cases and controls still alive and in follow-up five years or more after receiving supportive housing was considered to be five years. Housing costs were limited to those incurred by the health department: the lease, property management, and support services. We estimated the cost of housing as the restricted mean survival time in the housed group multiplied by the average monthly cost of housing. The cost per year of life saved then was estimated as the annualized average cost of housing divided by the increase in the restricted mean survival gain associated with housing. We used bootstrap methods to compute 95% bias-corrected percentile confidence limits (CL) for restricted mean survival, survival gain from housing, and the cost per year of life saved [31].

Figure 1 displays the derivation and composition of the samples used in each of the analyses. In the proportional hazards models, records with missing values were excluded. Analyses were done using SAS (SAS Institute, Cary, North Carolina) and Stata software (StataCorp, College Station, Texas). The study protocol was approved by the Institutional Review Board of the University of California, San Francisco.

Of the 6,558 AIDS cases included in this analysis, 641 (9.8%) were homeless at diagnosis. Homeless cases were more likely to be women or transgender, less than 30 years old, African American, and injection drug users, to have public health insurance or to be uninsured, and to be less likely to be aged 50 years or older, or to be receiving antiretroviral therapy than were the housed cases (Table 1).

Survival was significantly worse for homeless persons. Sixty-seven percent of persons who were homeless at the time of AIDS diagnosis survived five years compared with 81% of housed persons (p < 0.0001, Figure 2).

After adjustment for potential confounders, homelessness was significantly associated with increased mortality (RH 1.20; 95% CL 1.03, 1.41; Table 2). The analysis of the changes in the RH over time showed that the adjusted increase in the risk of mortality associated with homelessness increased from 1.00 (95% CL 0.82, 1.22) at one year, to 1.22 (95% CL 1.04, 1.42) at three years, and to 1.48 (95% CL 1.22, 1.79) at five years.

Other significant predictors of mortality were older age at diagnosis (RH 1.67; 95% CL 1.50, 1.86), injection drug use among persons other than men who have sex with men (MSM) (RH 1.88; 95% CL 1.59, 2.23), injection drug use among MSM (RH 1.52; 95% CL 1.32, 1.76), heterosexual contact (RH 1.48; 95% CL 1.14, 1.93), having public or no health insurance at diagnosis (RH 2.12; 95% CL 1.82, 2.48 and RH 1.60; 95% CL 1.39, 1.84, respectively), and having an opportunistic illness as the AIDS defining diagnosis (RH 1.41; 95% CL 1.24, 1.60). Use of antiretroviral therapy, prophylaxis against PCP, and higher CD4 counts at diagnosis were all associated with decreased risk of death (Table 2).

Although HIV/AIDS was the most frequent cause of death for all homeless persons with AIDS, the proportion of homeless persons who died from HIV/AIDS was significantly lower than for housed cases (Table 3). Causes of death that are associated with substance use (e.g., hepatitis, other causes of liver disease, septicemia) and mental illness occurred more frequently among the homeless. In contrast, chronic diseases were the most frequent non-HIV/AIDS-related causes of death among the housed cases.

Seventy AIDS cases matched with the DAH database and had move-in dates that were later than the AIDS diagnosis date, 35 of these cases were listed as homeless in the AIDS registry at the time of diagnosis and the other 35 were added after database matching. The result was a total of 676 homeless persons for this analysis (70 housed and 606 non-housed). The characteristics of persons who received housing and those who did not were similar except those who received housing were older (Table 4).

Of the 676 persons characterized in Table 4, 652 had complete data and were included in the proportional hazards model. Of these, there were two deaths among persons who received supportive housing, 219 deaths among those who were not housed, and 431 censored cases. After adjusting for potentially confounding variables, obtaining supportive housing was independently associated with an 80% reduction in mortality (RH 0.20; 95% CL 0.05, 0.81; Table 5). Mortality also was significantly lower for African Americans and persons with higher CD4 counts at diagnosis and significantly higher for persons 40 years or older at diagnosis and injection drug users.

To assess the possible effect of misclassification due to some housed patients not being initially coded as homeless in the AIDS registry, we repeated the multivariable Cox proportional hazards regression excluding these 35 cases. The RH for receiving housing (0.29) was similar to that found in our main analysis (0.20), although with this smaller sample the 95% CL included one (0.07, 1.20).

It is possible that unmeasured characteristics could explain the 80% reduction in mortality from supportive housing, and we determined the conditions that would be needed for this to occur (see Appendix). Specifically, we considered binary confounders of prevalence from 10% to 90% and correlations with not receiving housing that ranged from 0.1 to 0.5. For each combination, we calculated the strength of the association between the omitted confounder and mortality that would completely explain the reduction. In one scenario, we found that an unmeasured confounder affecting 70% of the population and correlated at 0.5 with not receiving supportive housing would be needed to cause a nine-fold increase in mortality to explain the apparent protective effect of receiving housing.

We identified 49 case-control pairs. The restricted mean survival time during the first five years of follow-up was 57.2 months (95% bootstrap CL, 51.7, 60.0) for the housed cases and 42.5 months (95% bootstrap CL, 33.4, 51.2) for the matched controls, resulting in an average gain of 14.7 months. At an approximate cost of $1,000 per month, the average cost of housing per year of life saved was $ 46,800 (95% bootstrap CL, $28,946, $171,746).