background image
applying for shelter would still "waste" three-fifths of its services (correctly identi-
fying 3,600 families against 5,400 false alarms). In addition to the problem of
wasting services on those who will not become homeless, there is the problem of
failing to serve those who will become homeless. Even a targeting cutoff that
wastes 80% of services misses 34% of families who in fact become homeless.
("Waste" here simply means that families would have avoided homelessness in
any case, not that their circumstances are untroubled.)
The best predictive model includes some risk factors (such as childhood
abuse) that might prove hard to verify. If access to an attractive prevention program
(such as subsidized housing or valued social services) depended on such risk
factors, and the prediction formula became even roughly known (as it inevitably
would), the targeting effort would create incentives for people to dissemble in
order to obtain services and could create an adversarial relationship between
clients and service providers charged with certifying eligibility. In that event,
reports of the key risk factors would increase, more people would be deemed eligi-
ble for services, and the predictive power of the model would decline.
5
Most targeting programs use a single criterion, such as eviction. From the New
York city data, we estimate that a program that targeted welfare families facing
eviction would serve four families who would not in fact enter shelter in the
absence of the program while reaching only one-fifth of the shelter population.
6
This one- predictor model correctly identified less than one-third as many families
as the multivariate model, at a constant false alarm rate (80%).
The salient lesson is that a prevention program aimed at people with any single
characteristic, such as those being evicted, is likely to target only a small portion of
all who become homeless. Even sophisticated multivariate models with very nar-
row targeting (which therefore reach a very small proportion of those who become
homeless) are likely to have far more false alarms than hits.
7
If the outcome criterion to be predicted were months in shelter (which is more
closely associated with costs than is simple shelter entry), it might be possible to
develop more efficient predictive models. Culhane and Kuhn (1998) showed that
in New York city, 18% of single-adult, first-time shelter users accounted for 53%
of the total days in shelter for first-time users in their first year; in Philadelphia,
10% accounted for 35% of these days. The authors described several individual
The Prevention of Homelessness Revisited
101
5
Interestingly, a model with only seven easily verified predictors did almost as well as the full
model at intermediate levels of risk (65% versus 66% hits at 10% false alarms, among families on public
assistance). The model included race/ethnicity, pregnancy/newborn and all five housing variables.
However, this model did less well for narrow targeting and includes one factor (race) on which it would
be illegal to base access to services.
6
For this calculation and an explanation of why an alternative calculation by the New York State
Department of Social Services (1990) is erroneous, see Shinn and Baumohl (1999).
7
A second lesson, perhaps less general, is that in the case of New York city families, targeting
based primarily on housing variables did about as well as models that took into account less verifiable
indicators of individual risk.