1. An Apple A Day Can Keep the Doctor Away, But
Does SNAP Improve Your Health?
Christian A. Gregory*1 , Partha Deb2 ,
Geetha Waehrer3
1 Economic
Research Service, USDA
College
3 Pacific Institute for Research and Evaluation
2 Hunter
The analysis and views expressed are the authors’ and do not represent the
views of the Economic Research Service or USDA.
Southern Economic Association
Tampa, Florida November 25, 2013
2. Background and Introduction
Background and Motivation
•
SNAP largest food assistance program of USDA
•
2012: $80 billion, 48 million participants
•
participation has doubled since 2007
•
policy concerns:
– does it reduce food security?
– does it support healthy diets?
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013
3. Background and Introduction
Background and Motivation
• Empirical Work
–
–
food security: Nord and Prell (2011), DePolt et al. (2009), Yen
et al. (2008), Shaefer and Gutierrez (2012), Ratcliffe et al. (2011),
Cole and Fox (2008), Gregory et al. (2013a), Mabli et al. (2013);
encouraging but mixed findings (natural experiments ⇑, binormal
error structure ⇑, ⇓ cross sectional data ⇑, ⇓)
nutrition: Fox et al. (2004), Yen (2010), Waehrer and Deb
(2012),Gregory et al. (2013b)
• Why would SNAP have any effect on health?
–
–
⇓ food insecurity = ⇑ health
• but Bhattacharya et al. (2004)
• obesity ?
SNAP as income transfer: ⇑ income ⇑ health (Deaton and Paxson
(2001))
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013
4. Background and Introduction
Background and Motivation
• What are possible other avenues for SNAP’s effect?
–
–
–
by relaxing income constraint, SNAP makes resources and time
available for activities that are conducive to well-being but not
necessarily related to diet
relaxation of budget constraint relieves stress that includes but
goes beyond food hardship
as in Oregon Medicaid experiment–increase in income ⇒
improvement in self-regard, feelings of well-being
• where do we look for evidence of SNAP’s effects?
self assessed health (SAH)
• has strong objective validity
• contains “private” information about well-being not captured in
other measured outcomes
– healthy time
• Grossman (1972, 2000): principle measure of health is healthy
time; healthy time is both investment in labor market activities
and home production, and consumption: sick days bring
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013
disutility
–
5. Background and Introduction
Background and Motivation
•
where do we look for evidence of SNAP’s effects?
– healthcare utilization
• Meyerhoefer and Pylypchuk (2008) finds increased
spending as effect of SNAP (pathway through obesity)
• doesn’t control for utilization–ie. services or Rx–or market
heterogeneity
• Grossman (1972, 2000) Ht = Ht−1 (1 − δ) + It−1 :
instantaneous recalibration of health capital through It−1
⇒ ⇑ It−1 (Mt−1 ) ⇑ Ht ; higher utilization = better health
• Galama and Kapteyn (2011) consumers have a threshold
of H: above threshold (healthy state), refrain from I (M);
below, increase I ; better health ⇒ ⇓ utilization
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013
6. Background and Introduction
Other Research, Our Contribution
3 issues: selection, distribution of outcomes, sample (data)
• previous (non-diet-outcome related) research; data/methods
limitations
– Nicholas (2011) SNAP and Medicare expenditures,
diabetics; FE methods, no accounting for skewed
distribution of outcomes (count, expenditure)
– Fey-Yensan et al. (2003) convenience sample of elderly
persons in CT public housing (SAH), descriptive statistics
– Gibson (2001) SNAP, SAH, 4 chronic conditions, single
wave of NLSY97
– Yen et al. (2012) participants in TN welfare program; SAH,
full switching model, copula approach
⇓ Pr (Excellent, VeryGood) health
• we use nationally rep. sample of non-elderly adults, methods
take into account selection and distribution of outcomes
•
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013
7. Background and Introduction
Preview of Results
•
SNAP improves SAH ⇑ Pr (Excellent, VeryGood) health,
⇓ Pr (Good, Fair , Poor ) health
•
SNAP reduces sick days– between 1 and 2 a year
•
SNAP reduces office based visits–between 1 and 2 a year
•
SNAP reduces outpatient visits – a statistically significant
fraction
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013
8. Data: MEPS
Data: MEPS 1999-2008
•
•
•
•
•
•
data from 10 years (1999-2008) of MEPS
rolling panel: 5 interviews over 2 years; demographic, labor
market, health insurance, health condition, health expenditure
and utilization data for all respondents
frequency of info differs: health insurance (monthly), BMI
(yearly), SNAP (yearly), ability status (at interview), priority
conditions (at interviews), SAH (at interviews), expenditure
(yearly), utilization (yearly)
because we use yearly measures of utilization, we use year’s
last recorded SAH response (3rd and 5th interview)
sick days = sum of work days, school days, and days of other
activities lost due to illness, respondent spent at least half of
the day in bed
utilization measures are in consolidated yearly data file
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013
9. Data: MEPS
Data: MEPS 1999-2008
SNAP 1999−2008
.06
Participation Rate
.07
.08
.09
.1
Data: MEPS, FNS
1998
2000
2002
2004
2006
2008
Year
FNS
Gregory, Deb, Waehrer
SNAP Health
MEPS
November 25, 2013
10. Methods
Methods
Treatment Effects Ordered Probit
Si∗ = Xi βS + Zi δ + εi
Hi∗
(1)
= Xi βH + Si ζ + υi .
•
S ∗ and H ∗ latent variables, utility of SNAP, underlying health,
X are factors effecting both SNAP and health, Z are
instruments: simplified reporting, β, ζ parameters
•
S is binary, H ∈ (1, 2, 3, 4, 5)
•
ε and υ ∼ Φ2 , model estimates ρ, correlation of unobservables
•
parameters estimated by maximum likelihood
•
Greene and Hensher (2010): semi-ordered bivariate probit.
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013
11. Methods
Methods, cont’d
• joint normal distribution of errors for count models not available
• use latent factor structure, developed in treatment effects literature
(Heckman and Vytlacil, 2005; Aakvik et al., 2005; Meyerhoefer and
Yang, Autumn 2011)
Treatment Effects: Count Models
To fix ideas, let:
Si∗
E (Ci |Xi , Si , li )
= Xi βS + Zi δ + li λ + ǫi
= g (Xi βC + Si ζ + li λ).
(2)
• S ∗ , S, X , Z , β, δ, and ζ are defined as above.
• Ci count outcome, li latent characteristic underlies correlation b/w
selection and the outcome; g is a negative-binomial 1 density
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013
12. Methods
Methods, cont’d
Treatment Effects: Count Models
• assume that li have a normal distribution
• could get joint distribution (Ci , Si |Xi , Zi ) by integrating over the
distribution of li :
Pr (Ci , Si |Xi , Zi ) =
{f (Xi βC +Si ζ+li λ)×Φ(Xi βS +Zi δ+li λ)φ(li )dli }.
(3)
• no closed form solution; really hard
• we use MSM:
N
lnℓ(Ci , Si |Xi , Zi ) ≈
ln[
i =1
1
S
S
{f (Xi βC +Si ζ+˜is λ)×Φ(Xi βS +Zi δ+˜is λ)}].
l
l
i =1
(4)
• 400 Halton sequence draws–efficiency properties compared to pseudo
random draws
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013
15. Results
Summary Statistics cont’d
Wage Income ($)
Unemployent Income
Other Program Income
SSI Income ($)
Family Size
Excellent Health
Very Good Health
Good Health
Fair Health
Poor Health
Total Sick Days
Office Based Visits
Outpatient Visits
N
Gregory, Deb, Waehrer
Non-SNAP
6199.11
(59.86)
88.29
(5.67)
21.18
(2.14)
362.25
(13.03)
2.86
(0.01)
0.19
(0.00)
0.27
(0.00)
0.32
(0.00)
0.15
(0.00)
0.06
(0.00)
9.80
(0.29)
4.64
(0.10)
0.46
(0.03)
33423
SNAP Health
SNAP
4261.07
(69.83)
123.33
(8.40)
490.94
(15.88)
1016.25
(29.49)
3.44
(0.02)
0.13
(0.00)
0.20
(0.00)
0.33
(0.01)
0.22
(0.00)
0.12
(0.00)
17.90
(0.54)
6.73
(0.17)
0.89
(0.07)
November 25, 2013
16. Results
SAH Results
Table : Marginal Effects of SNAP on SAH, 130% FPL
Parameter (se) : -.446*** (.08)
Excellent Very Good Good Fair Poor
0.11
.04
-.04
-.06 -.05
N
33423
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013
17. Results
Count Outcomes
Predicted Difference in Sick Days
SNAP−Non−SNAP
0
.1
Density
.2
.3
.4
.5
Data: Non−Elderly Adults < 130 % FPL, MEPS
−15
−10
−5
Predicted Difference in Sick Days
0
Median Difference = −1.54
Figure : Distribution of Marginal Effects: Sick Days
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013
18. Results
Count Outcomes
Predicted Difference in Office Visits
SNAP−Non−SNAP
0
.1
Density
.2
.3
.4
.5
Data: Non−Elderly Adults < 130 % FPL, MEPS
−10
−8
−6
−4
−2
Predicted Difference in Office Visits
0
Median Difference = −1.62
Figure : Distribution of Marginal Effects: Office-Based Visits
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013
19. Results
Count Outcomes
Predicted Difference in Outpatient Visits
SNAP−Non−SNAP
0
2
4
Density
6
8
10
Data: Non−Elderly Adults < 130 % FPL, MEPS
−.3
−.2
−.1
Predicted Difference in Outpatient Visits
0
Median Difference = −.08
Figure : Distribution of Marginal Effects: Outpatient Visits
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013
20. Results
Median Effects
•
η.5,days = −1.54
•
η.5,obv = −1.62
η.5,opv = −.08
ˆ
• p-values on β < .001
•
Table : Ancillary Parameters
ρ
χ2
IV
SAH
-0.337***
(0.05 )
9.563***
(.008)
Gregory, Deb, Waehrer
λ
Sick Days
2.113***
(0.021)
7.276**
(.026)
OB Visits
-1.310***
(0.020)
12.202***
(.002)
SNAP Health
OP Visits
0.324***
(0.107)
10.070***
(.007)
November 25, 2013
21. Discussion
Discussion
• find that SNAP has unequivocally positive effect on SAH
• find that SNAP increases “healthy time,” reduces utilization
• might argue that ⇓ utilization index of material hardship
• but all other measures of program participation are positive–public
income, medicaid, SSI etc.
• consistent with Galama and Kapteyn (2011): persons in better
health decrease utilization
• ρsah < 0 implies SNAP participants have better unobserved health
status “before” enrolling
• λdays > 0, λobv < 0, λopv < 0
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013
22. Discussion
Why?
•
as with Medicaid Oregon experiment–subjective states are
affected by changes in income
•
how much is enough?
•
using NHIS (sampling frame for MEPS), we look at the effect
of small changes in income on affective states: feelings of
worthlessness, depression, anxiety–even relatively small
changes (< $500/yr ) in income make a difference in how
people feel; this can account for a lot of what is observed as
improved health
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013
23. Discussion
Next Steps
•
robustness checks: poverty status, gender, instruments
•
interactions between SNAP and Medicaid
•
latent factor in SAH specification–flexible specifications in
SAH
•
what can we learn by linking to NHIS and using panel
component
Gregory, Deb, Waehrer
SNAP Health
November 25, 2013