Measuring the Impact of Hurricane Katrina on Access
to a Personal Healthcare Provider: The Use of the National
Survey of Children’s Health for an External Comparison Group
Tasha Stehling-Ariza • Yoon Soo Park •
Jonathan J. Sury • David Abramson
Published online: 29 March 2012
� Springer Science+Business Media, LLC 2012
Abstract This paper examined the effect of Hurricane
Katrina on children’s access to personal healthcare providers
and evaluated the use of propensity score methods to compare
a nationally representative sample of children, as a proxy for
an unexposed group, with a smaller exposed sample. 2007
data from the Gulf Coast Child and Family Health (G-CAFH)
Study, a longitudinal cohort of households displaced or
greatly impacted by Hurricane Katrina, were matched with
2007 National Survey of Children’s Health (NSCH) data
using propensity score techniques. Propensity scores were
created using poverty level, household educational attain-
ment, and race/ethnicity, with and without the addition of
child age and gender. The outcome was defined as having a
personal healthcare provider. Additional confounders
(household structure, neighborhood safety, health and insur-
ance status) were also examined. All covariates except gender
differed significantly between the exposed (G-CAFH) and
unexposed (NSCH) samples. Fewer G-CAFH children had a
personal healthcare provider (65 %) compared to those from
NSCH (90 %). Adjusting for all covariates, the propensity
score analysis showed exposed children were 20 % less likely
to have a personal healthcare provider compared to unex-
posed children in the US (OR = 0.80, 95 % CI 0.76, 0.84),
whereas the logistic regression analysis estimated a stronger
effect (OR = 0.28, 95 % CI 0.21, 0.39). Two years after
Hurricane Katrina, children exposed to the storm had sig-
nificantly lower odds of having a personal health care pro-
vider compared to unexposed children. Propensity score
matching techniques may be useful for combining separate
data samples when no clear unexposed group exists.
Keywords Hurricane Katrina � National Survey of
Children’s Health � Personal healthcare provider �
Propensity score
Introduction
Medical homes provide high-quality medical care that is
accessible, continuous, comprehensive, coordinated, fam-
ily-centered, compassionate and culturally effective [1–4]
and are endorsed by major primary care organizations [5–7].
The benefits of the medical home concept have been well-
documented [5–18] and are supported even when only some
of the key components are present. Identification with a
specific primary care practitioner, for example, has been
associated with better preventive care, lower healthcare
costs, fewer hospitalizations and fewer emergency depart-
ment visits [8–11].
Not all children have equal access to such health ser-
vices. Children living in poverty, unsafe neighborhoods, or
in households with limited education or a single-parent are
significantly less likely to have a regular source of medical
care or medical home [12–21]. Other factors negatively
associated with having a usual source of care include race/
ethnicity, poor health and being uninsured [13–16, 18].
Children greatly impacted by the 2005 Hurricane Katrina
shared many of these characteristics. Louisiana evacuees
T. Stehling-Ariza (&) � Y. S. Park � J. J. Sury � D. Abramson
National Center for Disaster Preparedness, Mailman School
of Public Health, Columbia University, 215 W. 125th Street,
Suite 303, New York, NY 10027, USA
e-mail: [email protected]
Y. S. Park
Department of Medical Education, College of Medicine,
University of Illinois at Chicago, 808 South Wood (MC 591),
Chicago, IL 60612-7309, USA
e-mail: [email protected]
123
Matern Child Health J (2012) 16:S170–S177
DOI 10.1007/s10995-012-1006-y
were predominantly African-American, poor, and with
limited education [22–24]. In Mississippi, the most seriously
damaged areas had significantly more households in poverty
and with limited education (unpublished analysis, [25, 26]).
In addition to these factors, the overall disruption experi-
enced by households and healthcare systems was likely to
upset pre-existing family-provider relationships [27–29].
One would expect the children exposed to the conse-
quences of Katrina would be less likely to have a personal
healthcare provider. However, responses to the hurricane
by numerous organizations may have mitigated some of the
effect. Programs such as Disaster Relief Medicaid and
longer-term assistance [30] as well as efforts by local and
non-profit organizations to provide health services [29]
may have increased access to personal healthcare provid-
ers. This article tested whether, 2 years after Hurricane
Katrina, exposed children were less likely to have a per-
sonal healthcare provider than similar unexposed children.
Using Multiple Datasets
Epidemiologic research seeks to compare equivalent
groups of people who vary only on exposure status; dif-
ferences in outcomes are then presumed to be due to
exposure. Valid comparison groups, however, are not
always available. Six months after Hurricane Katrina,
residents from southern Louisiana and Mississippi were
recruited into the Gulf Coast Child and Family Health
(G-CAFH) Study in order to assess long-term health,
mental health and service needs [31]. But because the
hurricane affected such a large geographic area [23, 31–
33], finding an unaffected comparison group from the same
communities and with similar characteristics would have
been extremely challenging.
This paper explored the utility of a nationally repre-
sentative sample, the National Survey of Children’s Health
(NSCH), as a proxy for such a comparison group in
investigating the impact of Hurricane Katrina on having a
personal healthcare provider. Furthermore, it examined the
application of the propensity score matching method. This
approach matches respondents on their probability of
exposure so that, within matched groups, the likelihood of
being exposed is essentially random [34–41]. Matched
groups are then used to estimate the effect of exposure
while controlling for confounding. The advantage of the
propensity score approach over traditional methods of
comparison lies in its assessment of the suitability of the
comparison groups [37, 38, 40], reducing the chance of
biased estimates and incorrect inferences regarding the
effect of exposure. This project evaluated results obtained
from both propensity score methods and traditional
approaches and explored the use of secondary data when
no clear control group existed [35, 42, 43].
Methods
Gulf Coast Child and Family Health Study
The exposed group consisted of G-CAFH children greatly
impacted by Hurricane Katrina. Details of the data collection
have been published elsewhere. Briefly, within 12 months of
the hurricane, participants were recruited from randomly-
selected FEMA-subsidized housing sites in Louisiana and
Mississippi; additional respondents were randomly-selected
from moderately- to extensively-damaged census blocks in
Mississippi. Respondents from 1,079 households (response
rate 32.6 %, cooperation rate 67.9 %) agreed to follow-up
and have been reinterviewed annually [44–46].
This analysis focused on 2007 data; 803 households, 320
of which had at least one child under the age of 18 years,
completed telephone surveys (75.2 % follow-up partici-
pation rate). One child from each household was randomly-
selected using a Kish sampling strategy [47]. Data on
selected children were collected from adult respondents,
usually their parents.
National Survey of Children’s Health
The unexposed group was composed of children from the
2007 NSCH sample. US households with children were
selected using random-digit-dial techniques and interviews
were conducted with 91,642 respondents [48]. Additional
details are described in this issue. This analysis used pub-
licly-available NSCH data [49].
Given the scale of Hurricane Katrina, some NSCH
children, especially in Louisiana and Mississippi, may have
had some exposure; however, no exposure data were col-
lected. Complicating matters, Louisiana and Mississippi
differ from other states on important factors, such as pol-
icies supporting the economic security of families [50–52],
suggesting state-level characteristics may introduce
unmeasured confounding. Consequently, three comparison
groups of NSCH children were utilized to address these
issues: (1) a national reference group, excluding Louisiana
and Mississippi, with the least exposure to Hurricane
Katrina; (2) a group from Texas and Alabama, states
characteristically similar to Louisiana and Mississippi [50],
but not as affected by the hurricane; and (3) a group from
Louisiana and Mississippi who may have had some expo-
sure to Katrina, but were not expected to be affected to the
same extent as G-CAFH children.
Personal Healthcare Provider
The outcome was an established relationship with a per-
sonal doctor or nurse. Both samples were presented the
following definition and question: ‘‘A personal doctor or
Matern Child Health J (2012) 16:S170–S177 S171
123
nurse is a health professional who knows your child well
and is familiar with your child’s health history. This can be
a general doctor, a pediatrician, a specialist doctor, a nurse
practitioner, or a physician’s assistant. [Right now,] do you
have one or more persons you think of as [the child’s]
personal doctor or nurse?’’ ‘‘Right now’’ was used in the
G-CAFH survey to focus respondents on post-Katrina
relationships with providers. Positive responses were cat-
egorized as having access to a personal provider.
Covariates
Covariates were selected based on previous research and
availability in the NSCH and G-CAFH datasets. Poverty was
defined as living below the 2007 Federal Poverty Level
(FPL) based on household income and size. Pre-Katrina
income was used for G-CAFH households because 2007
income was potentially affected by exposure. A single
imputation of NSCH poverty data, provided by the National
Center for Health Statistics, was included to account for non-
random missing data [49]. Household educational attain-
ment in both samples was defined as the highest achieve-
ment among all adults in the household and grouped into less
than high school, high school graduate or GED, and more
than high school. Educational attainment of G-CAFH adults
was measured in 2008; 59 households did not participate in
the 2008 interview and were excluded from the adjusted and
propensity score analyses. Household structure was cate-
gorized into two-parent households (biological, adoptive or
step) and other. In both surveys, safe neighborhoods were
those in which respondents reported feeling their children
were ‘‘usually’’ or ‘‘always’’ safe. Child health was cate-
gorized into ‘‘fair’’ or ‘‘poor’’ versus ‘‘excellent,’’ ‘‘very
good,’’ or ‘‘good.’’ These variables were dichotomized in
order to ensure adequate power to detect meaningful dif-
ferences in the traditional logistic regression analysis [53].
Few G-CAFH children were Hispanic (3.2 %) or of
other races (2.2 %), so the race/ethnicity variable was
dichotomized into Black, non-Hispanic versus all others.
Finally, child gender and age (0–5, 6–11, and 12–17 years)
were included in the analyses.
Traditional Analysis
Two types of analyses were conducted using a single dataset
of appended G-CAFH and NSCH data: a traditional
approach and a propensity score matching method. The
traditional approach assessed covariate frequencies using
Pearsons’ Chi-squared tests. Unadjusted and adjusted esti-
mates of effect were calculated using logistic regression;
adjusted estimates controlled for the covariates listed above.
Separate models were run for each comparison group.
NSCH survey weights were used to estimate effect sizes.
Propensity Score Analysis
This analysis was conducted in several steps: (1) propensity
score development, (2) respondent matching, and (3) effect
estimation [35]. Propensity scores are typically calculated
with predictors of exposure. In this analysis, exposure
included living in damaged areas and displacement in
FEMA-subsidized housing for at least 6 months. A priori,
poverty level, household educational attainment, and race/
ethnicity were hypothesized to be associated with expo-
sure. Because household structure, neighborhood safety,
child health and insurance may have been consequences of
exposure to Katrina [54–56], they were excluded from the
propensity scores estimation. Anecdotal evidence, for
example, suggested that some displaced families were sent
to trailer parks in which they did not feel safe; in other
families, post-Katrina stressors led to divorce or separation.
Because these variables were important to control when
estimating Katrina’s effect on healthcare access, yet inap-
propriate to include in the propensity score calculation, we
first generated the propensity scores and then assessed the
effect estimates while controlling for the other variables.
This approach was a mix of propensity score matching
methods and weighted regression, but, for simplicity, it is
referred to as the propensity score approach.
Age and gender were neither potential predictors nor
outcomes of exposure, so two propensity scores were cal-
culated: one without age and gender and one with, referred
to as the 3- and 5-variable analyses, respectively. Proximity
to the coast could not be included due to limitations of the
NSCH public dataset.
The propensity score analysis was conducted with the
pscore and attnd algorithms in Stata [57]. A logistic
regression model calculated the propensity scores using the
specified covariates, then respondents were matched using
nearest neighbor techniques with replacement and without
calipers. The nearest neighbor approach finds the closest
propensity score to match, while ‘‘with replacement’’
allows observations to be matched multiple times. If one-
to-one matching had been used, thousands of NSCH
observations would have been discarded; instead, respon-
dents were returned to the pool of observations available
for re-matching. Calipers set the acceptable range of pro-
pensity scores for matching; smaller values limit the
number of available matches and, theoretically, increase
the exchangeability of those matched. In lieu of setting
calipers, the pscore procedure uses t-tests to assemble
blocks of observations in which the mean propensity scores
of the exposed and unexposed are not significantly differ-
ent, and then tests for covariate balance [57]. The matching
is successful if the covariates are balanced within blocks.
Analyses were restricted to overlapping (matched)
observations.
S172 Matern Child Health J (2012) 16:S170–S177
123
The effect of Hurricane Katrina was estimated using the
average effect of the treatment on the treated (ATT).
Assuming exchangeability across all possible covariates, the
ATT theoretically compares the proportion of exposed chil-
dren (the ‘‘treated’’) with a personal healthcare provider (the
‘‘effect’’) with the proportion that would have had a provider
if they had not been exposed (‘‘untreated’’). In this analysis,
the attnd procedure derived the ATTs using both the esti-
mated propensity scores and weighted regression to adjust for
the additional covariates. The ATTs were converted to odds
ratios for comparison with the traditional results.
The 3-variable propensity score procedure successfully
matched 255 G-CAFH children with children from the US
(87,608), Alabama/Texas (3,552), and Louisiana/Missis-
sippi (3,772) samples, respectively. All blocks of obser-
vations were balanced on poverty level, household
educational attainment, and race/ethnicity. The 5-variable
procedure was less successful in matching 253 G-CAFH
children; several covariates remained unbalanced. In the
national comparison, three of nine blocks were not bal-
anced on education only, education and age, or gender.
Two of the six Alabama/Texas blocks were not balanced
on either age or education. Three of the four Louisiana/
Mississippi groups were not balanced on race/ethnicity.
Sensitivity Analysis
To test the effect of different specifications on the ATT
estimates, the psmatch2 algorithm was utilized [57]. The
procedure allows for specification of calipers and replace-
ment, but does not accommodate the separate control for
additional variables. Since the purpose was to test the
sensitivity of the results, rather than to estimate Katrina’s
effects, the inclusion of potential outcomes of exposure in
the propensity score calculation was considered acceptable.
Analyses were run without replacement (one-to-one
matching), and with caliper widths of 0.05, 0.01 and 0.001.
All analyses were conducted in Stata 10 [58]. Obser-
vations with missing values were excluded from each
analysis. For example, all available observations were
included in the propensity score matching, but the final
adjusted results were limited to 243 G-CAFH and 89,799
NSCH observations with complete data.
Results
Traditional Analysis
As discussed above, the G-CAFH sample was compared
with the national NSCH sample, excluding Louisiana and
Mississippi (‘‘US excl. LA/MS’’ in the tables), Alabama
and Texas (‘‘AL/TX’’), and Louisiana and Mississippi
(‘‘LA/MS’’). Frequency comparisons (Table 1) showed
significant differences in the proportion with personal
healthcare providers between the NSCH (about 90 %) and
G-CAFH groups (65 %). G-CAFH children were signifi-
cantly more likely to be Black, live in poverty, and have
fair/poor physical health, and less likely to live with two
parents or in safe neighborhoods. Adults living in NSCH
households were more highly educated than those in the
G-CAFH households. Insurance status was similar between
Table 1 Frequencies of select
characteristics among the
G-CAFH and NSCH samples
* p 0.05; ** p 0.01;
*** p 0.001 compared to
G-CAFH sample. To conduct
multiple comparisons, a
Bonferroni correction can be
used at a conservative alpha
level of 0.01
G-CAFH NSCH
US excl LA/MS AL/TX LA/MS
Overall sample size (n) 320 87,856 3,566 3,786
Personal healthcare provider (%) 64.67 92.25*** 89.56*** 90.06***
Living below 2007 FPL (%) 68.13 18.28*** 24.27*** 28.06***
Household educational attainment
Less than high school (%) 15.33 9.16*** 14.47*** 11.12***
High school graduate/GED (%) 40.23 23.29 24.54 29.28
More than high school (%) 44.44 67.55 60.98 59.60
Two-parent household (%) 50.00 75.71*** 74.88*** 62.82***
Safe neighborhood (%) 59.55 86.09*** 83.98*** 84.38***
Black, non-Hispanic race/ethnicity (%) 50.00 13.32*** 14.57*** 39.66***
Male (%) 50.00 51.13 50.88 51.07
Child age
0–5 years (%) 29.34 33.19* 34.95* 33.16
6–11 years (%) 29.65 32.42 32.45 32.51
12–17 years (%) 41.01 34.39 32.59 34.33
Fair/poor physical health (%) 16.93 3.44*** 5.50*** 4.83***
Insured (%) 81.73 90.84*** 84.25 93.09***
Matern Child Health J (2012) 16:S170–S177 S173
123
the G-CAFH children and those in Alabama/Texas; how-
ever, the NSCH national and Louisiana/Mississippi chil-
dren were more likely to be insured than the G-CAFH
children. G-CAFH children were slightly older than the
national and Alabama/Texas children, but not significantly
so compared to Louisiana/Mississippi. There were no
gender differences.
The traditional logistic regression estimated the unad-
justed odds of having a personal healthcare provider among
G-CAFH children ranged from 79 % lower than Alabama/
Texas children to 85 % lower than the national group
(Table 2). Living in poverty, limited household education,
not living with two parents, Black non-Hispanic race/eth-
nicity, and being uninsured were significantly associated
with a lower likelihood of having a personal provider
across the three comparisons. Older children were less
likely to have a provider in the national and Alabama/
Texas comparisons. Safe neighborhoods were positively
associated in the national and Louisiana/Mississippi
groups. Child health was associated with having a provider
in the national comparison only. Gender was not associated
with the outcome in any analysis.
The adjusted odds that a G-CAFH child exposed to the
damage and/or displacement of Hurricane Katrina had a
personal doctor or nurse was approximately 70 % lower
than the odds of NSCH children (Table 2). The odds ratios
were as follows: national OR = 0.28 (0.21, 0.39), Ala-
bama/Texas OR = 0.34 (0.20, 0.56), and Louisiana/Mis-
sissippi OR = 0.32 (0.22, 0.47). Poverty, Black race/
ethnicity, and lack of insurance were consistently associ-
ated with a lower likelihood of having a personal provider
after controlling for the other covariates. Educational
attainment was positively associated with the outcome in
the national and Louisiana/Mississippi comparisons, while
age was negatively associated in the national and Alabama/
Texas analyses. Living in a safe neighborhood was
Table 2 Unadjusted and adjusted odds of having a personal healthcare provider: traditional approach
US excl LA/MS AL/TX LA/MS
OR 95 % CI OR 95 % CI OR 95 % CI
Unadjusted estimates
Katrina damage and displacement (G-CAFH vs. NSCH) 0.15*** (0.12, 0.20) 0.21*** (0.15, 0.30) 0.20*** (0.15, 0.27)
Living below 100 % 2007 FPL (vs. C100 % FPL) 0.41*** (0.35, 0.47) 0.40*** (0.25, 0.64) 0.34*** (0.24, 0.48)
Less than high school education (vs. some college) 0.27*** (0.23, 0.33) 0.33** (0.18, 0.59) 0.31*** (0.19, 0.50)
High school graduate/GED (vs. some college) 0.45*** (0.39, 0.53) 0.46** (0.27, 0.78) 0.40*** (0.27, 0.58)
Two-parent household (vs. single parent/other) 1.61*** (1.40, 1.85) 1.66* (1.02, 2.71) 2.14*** (1.53, 3.00)
Safe neighborhood (vs. unsafe neighborhood) 1.86*** (1.58, 2.20) 1.40 (0.78, 2.50) 1.85** (1.19, 2.88)
Black, non-Hispanic (vs. White/Hispanic/other) 0.65*** (0.55, 0.77) 0.43** (0.25, 0.74) 0.35*** (0.25, 0.49)
Male (vs. female) 0.91 (0.80, 1.04) 1.38 (0.88, 2.19) 1.19 (0.85, 1.66)
Age: 6–11 years (vs. 0–5 years) 0.81** (0.68, 0.95) 0.65 (0.35, 1.21) 1.17 (0.74, 1.85)
Age: 12–17 years (vs. 0–5 years) 0.69*** (0.59, 0.80) 0.51* (0.29, 0.92) 0.75 (0.50, 1.11)
Fair/poor physical health (vs. good–excellent health) 0.44*** (0.32, 0.62) 0.75 (0.28, 2.02) 0.75 (0.34, 1.67)
Insured (vs. uninsured) 5.18*** (4.42, 6.06) 6.70*** (4.09, 10.97) 3.83*** (2.48, 5.93)
Adjusted estimates
Katrina damage and displacement (G-CAFH vs. NSCH) 0.28*** (0.21, 0.39) 0.34*** (0.20, 0.56) 0.32*** (0.22, 0.47)
Living below 100 % 2007 FPL (vs. C100 % FPL) 0.67*** (0.56, 0.80) 0.52* (0.31, 0.89) 0.52** (0.33, 0.81)
Less than high school education (vs. some college) 0.48*** (0.38, 0.60) 0.62 (0.31, 1.23) 0.54* (0.31, 0.93)
High school graduate/GED (vs. some college) 0.61*** (0.51, 0.73) 0.63 (0.35, 1.13) 0.56** (0.38, 0.83)
Two-parent household (vs. single parent/other) 1.11 (0.94, 1.32) 0.94 (0.52, 1.72) 0.92 (0.59, 1.43)
Safe neighborhood (vs. unsafe neighborhood) 1.31** (1.08, 1.58) 0.89 (0.45, 1.77) 1.21 (0.77, 1.91)
Black, non-Hispanic (vs. White/Hispanic/other) 0.76** (0.62, 0.93) 0.43* (0.22, 0.84) 0.46*** (0.31, 0.68)
Male (vs. female) 0.93 (0.80, 1.06) 1.39 (0.86, 2.24) 1.11 (0.78, 1.57)
Age: 6–11 years (vs. 0–5 years) 0.83* (0.69, 0.99) 0.67 (0.33, 1.36) 1.35 (0.84, 2.19)
Age: 12–17 years (vs. 0–5 years) 0.72*** (0.61, 0.85) 0.48* (0.26, 0.92) 0.79 (0.53, 1.18)
Fair/poor physical health (vs. good–excellent health) 0.83 (0.58, 1.20) 1.91 (0.71, 5.12) 1.61 (0.68, 3.81)
Insured (vs. uninsured) 4.21*** (3.54, 5.01) 5.94*** (3.45, 10.22) 3.77*** (2.47, 5.74)
* p 0.05; ** p 0.01; *** p 0.001
The traditional approach used logistic regression models to compare children from the G-CAFH study with three NSCH groups
S174 Matern Child Health J (2012) 16:S170–S177
123
significantly associated with having a provider in the
national comparison only. Household structure, health and
gender were not associated with the outcome.
Propensity Score Analysis
The propensity score analysis showed a negative, but
weaker, association between Katrina exposure and having a
personal healthcare provider (Table 3; the adjusted logistic
regression estimates were included for comparison). The
3-variable estimates of Katrina’s effect were 0.80 (0.76,
0.40) in the national comparison, 0.79 (0.75, 0.84) com-
pared to Alabama/Texas, and 0.80 (0.76, 0.85) when lim-
ited to Louisiana/Mississippi, after adjusting for the
covariates. The 5-variable adjusted estimates were com-
parable at 0.81 in all three analyses with similar 95 %
confidence intervals.
Sensitivity Analysis
The psmatch2 algorithm gave slightly higher estimates of
effect compared to the pscore and attnd procedures. When
using the same specifications as the original analyses (with
replacement and without calipers), the estimates were 0.84
(0.77, 0.91) for the national comparison, 0.81 (0.75, 0.88)
for Alabama/Texas, and 0.86 (0.79, 0.93) in Louisiana/
Mississippi. The estimates remained consistent with caliper
widths of 0.05, 0.01 or 0.001, and without replacement.
The effect estimates were 0.83-0.84 in the US comparison,
0.81-0.82 in Alabama/Texas, and 0.84-0.86 in Louisiana/
Mississippi (p 0.05).
Discussion
In the wake of Hurricane Katrina, a number of time-limited
interventions were launched to assure medical care access
for exposed children (e.g., Disaster Relief Medicaid). Both
the traditional and propensity score analyses conducted
here revealed that Katrina-exposed children were less
likely to have access to a personal healthcare provider than
were comparable children. The propensity score analysis
suggested that the exposed children were only 20 % less
likely to have provider access, in contrast to the traditional
regression which estimated that exposed children were
approximately 70 % less likely to have such access. The
more precise matching of the propensity analysis may have
done a better job of estimating the effect of exposure and
addressing potential confounding effects. The propensity
score results were considered superior due to the detailed
assessment of covariate balance within the matched groups,
and the consistency of the estimates. Regardless of the
propensity score used (3- or 5-variable) or the specifica-
tions, the effect estimates showed little variation.
In previous work, the one-to-many matching of the pro-
pensity score approach has led to different effect estimates
Table 3 Adjusted effect of exposure on having a personal healthcare provider: propensity score matching versus logistic regression
US excl LA/MS AL/TX LA/MS
Traditional logistic regression methoda
Adjusted odds ratio (95 % CI) 0.28 (0.21, 0.39)*** 0.34 (0.20, 0.56)*** 0.32 (0.22, 0.47)***
Propensity score method—3 variablesb
Number of matched exposed (G-CAFH) 255 255 255
Number of matched unexposed (NSCH) 87,608 3,552 3,772
Propensity score method
Adjusted odds ratio
c
(95 % CI) 0.80 (0.76, 0.84)*** 0.79 (0.75, 0.84)*** 0.80 (0.76, 0.85)***
Propensity score method—5 variables
d
Number of matched exposed (G-CAFH) 253 253 253
Number of matched unexposed (NSCH) 83,487 3,335 3,544
Propensity score method
Adjusted odds ratio
c
(95 % CI) 0.81 (0.76, 0.85)*** 0.81 (0.77, 0.86)*** 0.81 (0.77, 0.86)***
* p 0.05; ** p 0.01; *** p 0.001
a
OR adjusted for poverty level, household educational attainment, household structure, neighborhood safety, race/ethnicity, gender, age, child
health, and insurance
b
Propensity score calculated with poverty level, household educational attainment, and race/ethnicity. Estimate of effect adjusted for household
structure, neighborhood safety, gender, age, child health, and insurance
c
Derived from the average effect of the treatment on the treated (ATT)
d
Propensity score calculated with poverty level, household educational attainment, race/ethnicity, gender and age. Estimate of effect adjusted
for household structure, neighborhood safety, child health, and insurance
Matern Child Health J (2012) 16:S170–S177 S175
123
than those from regression analyses [38]. In this study,
however, matching without replacement (i.e., one-to-one
matching) yielded comparable results indicating the differ-
ing estimates were more likely due to increased exchange-
ability of the two groups rather than the sample weights.
There were several limitations to this analysis. Propen-
sity score matching can only control for confounding due
to measured covariates. Despite controlling for many
covariates, residual confounding by unmeasured factors
may remain. Misclassification of exposure among the
NSCH children was also a concern; nevertheless, the
findings were consistent across comparison groups making
it unlikely that misclassification affected the results. Some
covariate imbalance was …
