background image
ties, and housing, that might distinguish families on welfare who requested shelter
from other families in the public assistance caseload. (Families who had used shel-
ter previously were excluded from both groups.) Although 18 factors were related
to homelessness, taken one at a time, the "best" multivariate model included 10
predictors that reliably contributed to the prediction of homelessness in the context
of the other variables in the model.
The model yielded a summary score of "homeless risk" for each family. Deter-
mining who should be eligible for a prevention program corresponds to choosing
some cutoff for risk scores. A liberal cutoff score selected to deliver prevention
services to a large portion of those who would otherwise become homeless also tar-
gets many families who would not become homeless in the absence of services
("false alarms"). A conservative cutoff yields fewer false alarms but also has a
lower "hit rate"; that is, it reaches fewer of those who would become homeless
without preventive efforts. Thus, a plot of hit rates versus false-alarm rates for dif-
ferent predictive models is a useful policy tool (Camasso & Jagannathan, 1995;
Swets, 1973; Swets, Dawes, & Monahan, 2000). Shinn et al. (1998) found that the
best model was able to correctly "hit" 66% of welfare families who requested shel-
ter with a false-alarm rate of 10%.
Although this ratio of hits to false alarms may sound good, the population to
which the false-alarm rate refers is far larger than the group who will end up in
shelter. At the time the data were collected, there were about 270,000 families on
welfare in New York City, over the course of a year, excluding families with previ-
ous shelter stays, and about 90% of the approximately 10,000 families who first
entered shelter over the course of the year came from the welfare caseload. Thus, to
correctly reach 6,000 families (90% of 66% of 10,000), a primary prevention pro-
gram would have to offer services to 27,000 families (10% of 270,000) who would
not become homeless. With respect to preventing shelter entry, over 80% of the
services would be wasted (although such help might be valuable to families for
other reasons). A more narrowly targeted prevention program that confined false
alarms to 2% of the public assistance caseload and reached only 36% of those
Shinn, Baumohl, and Hopper
The 10 were race/ethnicity (African Americans were at greater risk than Latinos or others), being
pregnant or having an infant under the age of 1 year, childhood poverty, being married or living with a
partner (surprisingly, marriage increased risk for homelessness), domestic violence in adulthood, fam-
ily disruption in childhood (a scale that included foster care or other types of separation from the family
in childhood or childhood abuse), and four housing factors (doubling up with others, lack of subsidized
housing, frequent moves, and overcrowding). Unrelated, in the context of other variables, were youth,
education, work history, having been a teen mother, positive social ties, mental illness, substance abuse,
health problems, imprisonment, and building problems. (Note that youth was related to homelessness
taken alone, but not after housing factors were entered in the equation, suggesting that youth affected
homelessness primarily via access to the housing market.) At the univariate level, homeless mothers ac-
tually had stronger networks than housed mothers (80% had stayed with network members before re-
questing shelter). Building problems were severe for both groups. Mental illness, substance abuse, and
imprisonment were relatively rare for both.