The study population is comprised of adults who had a record of first entering a homeless shelter run by the NYC Department of Homeless Services (DHS) from 1990 through 2002, as captured by DHS administrative records, and whose identifying information was validated through records from the Social Security Administration (SSA). Among single adults, all persons meeting these criteria were selected; among families, all adults were selected who were designated by DHS as heads of households. Thus, one adult per family household was selected.
DHS operates or funds separate shelter networks for unaccompanied (i.e., single) adults and families. Combined, these two shelter networks include approximately 85% of all general homeless shelter beds in NYC. As part of administering this system, DHS maintains databases for each of the two shelter networks, which have compiled information on persons staying in DHS shelters, including their shelter utilization, since 1986. While there is broad availability of shelter in NYC, there is a minority of persons who, when homeless, will stay in places not meant for habitation (outdoors, abandoned buildings, etc.). This homelessness cannot be captured with administrative data.44
For this study, 175,524 unduplicated DHS records with complete identifying information (name, gender, date of birth, and social security number [SSN]) were submitted to SSA, which was able to validate 160,525 unique records (91%) where identifying information from DHS matched that of SSA records using SSA’s standard matching protocol.45
The unvalidated records were discarded. The unvalidated records do not include an unknown number of records with missing or partial SSNs, which were never included in the original dataset.
Among the unvalidated records, 90% were for single adults, 13% of all records for single adults that were submitted. The single adults’ records considered invalid were more likely to be for persons who were male; under age 50; not of Hispanic ethnicity or white race (mutually exclusive groupings); had stayed in shelters less often and for shorter durations; and who first entered shelter earlier in the time period covered by this study (early 1990s). For adults sheltered with families, the rate of invalid records was minimal (2% of all records submitted), with these records being more likely among persons of Hispanic, other, or unknown race/ethnicity, over age 30, and who first entered shelter in the early 1990s.
Records from single adult and family shelters were each grouped into three categories based on shelter use patterns. Total number of days and stays spent in DHS shelters in the 3-year period following initial shelter entry were the criteria for a cluster analysis, which was applied to data in a manner consistent with procedures previously used on DHS shelter data.46,47
As a result, each record in this study was assigned to one of three distinct groups by virtue of shelter use patterns, with each of these groups representing different homeless trajectories. These groups included transitional
users, where persons used shelter for one or two stays and for a limited number of days; long-term
users, where persons used shelter for a limited number of extended stays; and episodic
users, where persons tallied multiple shelter stays of relatively brief durations.
Records in the study group were matched (by SSN) to the Social Security Death Index (SSDI). The SSDI has been compared with a national death registry in that it collects records of all deaths in the USA. With a correct SSN, using the SSDI has been determined to be an accurate means by which to identify mortality in research studies.48
However, SSDI records do not include an undetermined number of decedents who did not have SSNs or whose SSNs were not available for their death records. Upon a match with the SSDI, records for persons in the study group were appended with the appropriate date of death. No other information related to the circumstances of death was available. Data on deaths were current through June 2008.
A series of analyses provided multiple perspectives by which to view mortality among this homeless population. Due to differences in the family and single adult shelter populations, separate analyses were performed and reported on each of these subgroups. Approximately 10% to 15% of households in the family shelters were childless couples and pregnant unaccompanied women. These “adult families” were more likely to be older, white, and, on average, stayed in shelter longer.49
Other unpublished DHS analyses have found that adult families were more likely to enter shelter self-reporting physical health problems, substance abuse disorders, and a history of incarceration. According to personal communication with Joanna Weissman of DHS (March 2011), adult families were also more likely to have received mental health treatment prior to shelter entry.
Deaths were summarized by calculating mortality rates for persons by sex and age groupings. These mortality rates were expressed per 100,000 person-years of observation. Person-years were derived from the time period between the date of each individual’s initial shelter entry to either the date of the individual’s death or the last date of the study period (June 30, 2008). These values for sheltered single adults were compared with the mortality rates for sheltered adults in families through age-adjusted and sex-adjusted standardized mortality ratios (SMR). Ninety-five percent confidence intervals were also calculated for each SMR.50
In addition to mortality rates, life tables were computed to estimate life expectancy at a given age for each sex and shelter type, based on age at initial shelter entry. As it was not possible to calculate life expectancy from age 0 for the sheltered population, an alternate estimate of life expectancy was calculated based on a weighted average of age at shelter entry.
Finally, Cox regression analysis was used to assess the associations of various factors on the hazard of death. Of particular interest is whether time actually spent in a homeless shelter, shelter use pattern (long-term, episodic, or transitional), and placement from shelter to housing had lasting associations with the hazard of dying. The measure used for housing here indicates whether or not the person was known to have exited shelter to a stable housing arrangement. No further specifics about these housing arrangements, such as whether or not they were subsidized or whether they had accompanying support services, were available. Models were fitted from two time points: one in which the risk period began at point of shelter entry, and the other model at point of last shelter exit in the database.
Using the former start point allowed for a prospective look at mortality from the point of initial shelter entry. Here, the measures of shelter status (in shelter or out of shelter) and shelter exit type (to housing or to other living arrangement) were considered time-dependent covariates, as each person was in shelter for only part of the risk period and everyone would have spent time in the risk period before exiting to housing (and many would not exit shelter to housing at all). This model precluded shelter use pattern from being in the model, as this measure could only have been determined retrospectively at the point of shelter exit. In its use of a Cox regression model where shelter status was a time-dependent covariate, this model resembled the method used by Hwang in the previously mentioned shelter study.31
The second model, using shelter exit as the start point, precluded having a measure for shelter status in the model, given that the time spent in shelters was no longer part of the risk period. Exit to housing and shelter use pattern were included in this model as fixed covariates. Persons who did not have a record of death before July 1, 2008 were considered censored observations.51