2.4 Preventing Family Homelessness
Speaker: Andrew Greer and Marybeth Shinn
One of the keys to ending homelessness is to prevent it from happening in the first place. This workshop will examine the most effective strategies to prevent family homelessness, including using homelessness data to target interventions and partnering with providers serving high-risk families. Presenters will cover a wide array of services and solutions.
2. Background & Rationale
Targeting services to prevent homelessness is
difficult:
Numbers of shelter entrants are small and people
with many risk factors for shelter entry avoid
shelter
Prevention should be aimed at those most at-
risk of becoming homeless
3. Study Questions
Question 1: What was the pattern of shelter
entry over time among families who applied for
Homebase services?
Question 2: What families were at highest risk
of entering shelter?
Question 3: Is it possible to develop a short
screening instrument to target services?
Question 4: If Homebase adopted better
targeting, how much more effective might it
be?
4. Data base
City provided a database of 11,105 families
who applied for services between Oct 1, 2004
and June 30, 2008
Intake workers interviewed families about
program eligibility and risk factors for
homelessness
The City provided administrative data on
shelter entry over the next 3 years
5. Risk Factor Domains
Demographics
Human capital and poverty
Housing
Disability
Interpersonal discord
Childhood experiences
Previous Shelter
Dependent Variable: Time to Shelter Entry
6. Methods: Question 1
What was the pattern of shelter entry?
Survival Analysis
Technique borrowed from medicine where
“survival” is how long a patient lived after
treatment
Forus, the end point was not mortality, but shelter
entry
Questions:
“how long did people stay out of shelter?” (Survival
Curve)
“which periods of time were applicants at greatest risk
of shelter entry?” (Hazard Estimate)
7. Survival and Hazard Curves
Survival and Hazard Curves
Usedto illustrate survival and hazard rates for
subjects over time
8. Results: Question 1
What was the pattern of shelter entry over time
among families who applied for Homebase
services?
12.8% entered shelter within three years of
applying
Most families who entered shelter did so shortly
after applying for services
9. Methods: Question 2:
What families were at highest risk of entering
shelter?
Survival Analysis
Included predictors of shelter entry to see
which families were most at risk of entering
shelter
10. Results: Questions 2
Coefficient Haz Ratio Risk Conf Interval
direction
Demographics
Female 1.28 + 1.01-1.63
Black 1.35 .90-2.04
Hispanic 1.07 .71-1.62
Age .98 - .98-.99
Child under 2 yrs old 1.14 + 1.01-1.29
# of Children 1.04 1.00-1.09
Pregnant 1.24 + 1.08-1.43
Married 1.09 .906-1.31
Veteran 1.119 .54-2.34
11. Results: Question 2
Coefficient Haz Ratio Risk Direction Conf Interval
Human Capital/ Poverty
High School / GED .85 - .75-.96
Currently Employed .81 - .71-.93
Public Assistance History 1.30 + 1.13-1.49
Lost benefits in past year 1.14 .96-1.35
Housing
Name on lease .816 - .75-.96
Overcrowding or Discord 1.02 .87-1.20
Doubled up 1.14 .93-1.38
Threatened with eviction 1.20 + 1.04-1.38
Rent > 50% Income .93 .79-1.08
Arrears 1.00 1.00-1.00
Level of disrepair 1.02 .99-1.05
Number of times moved in past 1.16 + 1.08-1.24
yr
Current subsidy .85 .68-1.07
12. Results: Question 2
Coefficient Haz Ratio Direction Conf Interval
Disability
Chronic health probs or 1.10 .96-1.26
hosp
Mental illness or hosp .82 .67-1.02
Substance abuse 1.22 .95-1.56
Criminal justice 1.11 .92-1.33
Interpersonal Discord
Domestic violence .87 .73-1.04
History with protective 1.37 + 1.13-1.66
services
Legal involvement .98 .75-1.28
Av Discord with 1.09 + 1.05-1.13
landlord/household
13. Results: Question 2
Coefficient Haz Ratio Risk Direction Conf Interval
Childhood Experiences
Teen mother .95 .81-1.10
Childhood Disruption index 1.15 + 1.08-1.22
Shelter
Shelter as an adult (self 1.43 + 1.22-1.66
report)
Applied for shelter in last 3 1.63 + 1.31-2.02
mos
Seeking to reintegrate into 1.29 + 1.06-1.59
community
14. Results: Question 2
Coefficient Haz Ratio Risk Direction Conf Interval
Administrative
Variables
Previous Shelter 1.15 .89-1.50
# Prior shelter 1.18 + 1.08-1.30
applications
Previously found 1.10 .85-1.43
eligible for shelter
Exited shelter to a .96 .73-1.24
subsidy
16. Methods: Question 3
Is it possible to develop a short screening
instrument?
Eliminated location and administrative variables
Eliminated racial categories
Omitted variables that didn’t contribute reliably to
prediction of shelter entry
Examined hazard ratios to assign 1-3 points for
each predictor
For continuous variables like age, examined
patterns of shelter entry at different ages to
decide on cut points
17. Results Question 3: Screener
1 point adult
Pregnancy Age
Child under 2 1 pt: 23 - 28;
No high school/GED 2 pts: ≤22
Not currently employed Moves last year
Not leaseholder 1 pt: 1-3 moves;
Reintegrating into community 2 pts: 4+ moves
2 points Disruptive experiences in
Receiving public assistance (PA) childhood
Protective services 1 pt: 1-2 experiences;
Evicted or asked to leave by 2 pts: 3+ experiences
landlord or leaseholder Discord (landlord, leaseholder, or
Applying for shelter in last 3 household)
months 1 pt: Moderate (4 – 5.59);
3 points 2 pts: Severe (5.6 – 9)
Reports previous shelter as an
18. Methods: Question 4
If Homebase adopted better targeting, how much
more effective might it be?
Compare decisions based on our screening model
to:
1. Administrative data only
2. Current Decisions
3. Our full model
Consider the percentage of shelter entrants
targeted at different levels of risk
19. Results: Question 4 Accurate
Model
TargetingRisk Criterion % % Shelter
Applicants Entrants
Served Targeted
Current Approach Judged eligible 62.4% 69.1%
• The intake worker assessment approach gives services to 62% of
applicants and correctly targets 69% of shelter entrants.
20. Results: Question 4 Accurate
Model
TargetingRisk Criterion % % Shelter
Applicants Entrants
Served Targeted
Admin Data Any admin data 13.0% 25.7%
Current Approach Judged eligible 62.4% 69.1%
• People with past contact with the shelter system are at very high risk, but
only 13% of HomeBase applicants have any past contact
• Giving services to them would reach only 26% of shelter entrants
21. Results: Question 4 Accurate
Model
TargetingRisk Criterion % % Shelter
Applicants Entrants
Served Targeted
Admin Data Any admin data 13.0% 25.7%
Current Approach Judged eligible 62.4% 69.1%
Full Model Cutoff based on % 62.5% 89.6%
of Applicants
• If we use the full model to target the same proportion of HomeBase
applicants who currently get services, we do a much better job of
reaching those families who enter shelter
• We would reach almost 90% of shelter entrants, while the current system
reaches 69%
22. Results: Question 4 Accurate
Model
TargetingRisk Criterion % % Shelter
Applicants Entrants
Served Targeted
Admin Data Any admin data 13.0% 25.7%
Current Approach Judged eligible 62.4% 69.1%
Full Model Cutoff based on % 62.5% 89.6%
of Applicants
Screener 62.3% 88.9%
• A quick screener does almost as well as the full model
• Is this the right proportion? That’s a hard question that depends on lots
of factors: How much do prevention or shelter stays cost? What are
some of the other financial and moral costs of homelessness? How
effective are services?
• Our data don’t answer these questions. But we can say what proportion
of shelter entrants are reached at different proportions of applicants
23. Results: Question 4 Accurate
Model
TargetingRisk Criterion % % Shelter
Applicants Entrants
Served Targeted
Admin Data Any admin data 13.0% 25.7%
Current Approach Judged eligible 62.4% 69.1%
Full Model Cutoff based on % 62.5% 89.6%
of Applicants
Screener 62.3% 88.9%
Screener 5 or more points 67.8% 91.9%
Screener 6 or more points 53.6% 84.4%
Screener 7 or more points 41.6% 73.8%
Screener 8 or more points 30.5% 61.0%
• The last lines show what happens when we target people by their risk
24. Conclusions
Our short screener can predict likelihood
of shelter entry more accurately than
current decisions
Prediction is hard: Even at the highest
levels of risk, most families avoid shelter.
Determination of the proportion of families
to serve is a question of available funds
and costs, both to the homeless service
systems and to society.
25. Recommendations
Workers should be able to override the
recommendation of the model with written
explanations
Although this exact screener may not work
as well in other locations, the methods can
be shared
Any model should be tested periodically to
see if it misses recently vulnerable
populations