Cunningham_ResearchBrief

Shared by: NiceTime

Tags

linear probability models, household survey, reference group, insurance coverage, ED visits, low income

Stats

views:: 7
posted:: 5/22/2010
language:: English
pages:: 37

Public Domain

Document Sample

Editor’s Note: This online data supplement contains supplemental material that was not included with the published article by Peter
J. Cunningham, “Medicaid/SCHIP Cuts and Hospital Emergency Department Use,” Health Affairs 25, no. 1 (2006), online at
http://content.healthaffairs.org/cgi/content/abstract/25/1/237; 10.1377/hlthaff.25.1.237.

RESEARCH BRIEF

The Effects of Insurance Coverage and Health System Factors
on Hospital Emergency Department Use for Low Income Persons

Peter Cunningham, Ph.D.
Senior Health Researcher
Center for Studying Health System Change
600 Maryland Ave., S.W. Suite 550
Washington, D.C. 20024
202-484-4242
pcunningham@hschange.org

January, 2006

This research was funded by the Kaiser Family Foundation and the Robert Wood Johnson
Foundation through its support of the Center for Studying Health System Change.
2

1.0 Introduction

Understanding how changes in Medicaid and other public programs would affect hospital

emergency department (ED) volumes requires detailed understanding of the effects of insurance

coverage, individual characteristics, health system and policy factors on the ED utilization of

individuals. While there have been numerous studies on health care utilization over the past

three decades, there are comparatively fewer studies that have comprehensively addressed the

effects of insurance coverage and other health system factors on ED use.

Existing studies are focused on specific areas or subpopulations (e.g. based on age,

income, race/ethnicity, or health status), which limits the ability to make inferences for the nation

(Guttman et al. 2003; Piehl et al., 2000; Hurley et al. 1989). Other studies are limited to

individuals who use EDs (e.g. using claims data or medical records), and therefore are unable to

distinguish the characteristics of individuals who use EDs from those who do not use EDs (Oster

and Bindman 2003; Grossman et al. 1998; Powers 2000). Several national studies based on

population surveys address these limitations, but even these studies do not include a broad array

of health system variables that measure the proximity of hospital EDs to individuals, the

availability and capacity of alternative sources of primary care, such as Community Health

Centers (CHCs), and the overall supply of primary care and other office-based physicians in the

area (Cunningham et al., 1995; Zuckerman et al., 2002; McLaughlin and Mortensen 2003).

This paper presents a conceptual framework and analysis that addresses many of the

limitations of previous research. The basic objective is to provide a strong empirical foundation

for estimating the potential effects of public program cost containment efforts on hospital ED

volumes, and the relative proportion of ED visits made by insured vs. uninsured patients. To

this end, this paper provides the background and analytical detail for the findings and
3

conclusions contained in the article, “Do Reductions in Medicaid/SCHIP Enrollment Increase

Hospital ED Use Among Low Income Persons? (Cunningham, 2006)”

2.0 Conceptual framework

The behavioral model by Anderson is one of the most widely used models of health care

utilization and access. The original version of the model posited that health services use is a

function of need (e.g. both perceived and diagnosed health problems), enabling factors (e.g. both

community and personal resources), and characteristics of individuals and their communities that

predispose them to health care use (Andersen, 1968). More recent specifications of the model

have explicitly identified health care system, policy, and other “contextual” characteristics of the

community as influencing both enabling and need factors, in addition to directly affecting health

care use (Andersen, 1995).

The model serves as a useful conceptual framework for estimating the effects of

Medicaid policy changes on ED use. First and foremost, health insurance coverage has been

shown to be the most important “enabling” factor in determining health care use, and a detailed

understanding of insurance coverage differences in ED use can be used to make inferences about

the effects of changes in coverage that may result from changes in eligibility rules, benefits, and

premium requirements. Inclusion of a comprehensive set of need and predisposing factors in the

analysis controls for the fact that individuals with different coverage types also differ on a

number of characteristics that are often correlated with health services utilization.

Other Medicaid policies—such as reimbursement rates—may influence ED use indirectly

because they affect community resources, such as the availability and willingness of office-based

physicians to care for Medicaid patients. Fewer office-based physicians willing to see Medicaid

patients may result in greater use of the ED. Finally, general cutbacks in public programs may
4

negatively affect the capacity of health care providers who are highly dependent on Medicaid

revenue, such as Community Health Centers, which in turn could increase use of the ED.

3.0 Methodology

3.1 Data and sample

The data for this study are based primarily on the 2000-01 and 2003 Community

Tracking Study household surveys. The CTS household survey is designed to produce

representative estimates for the U.S. population and 60 randomly selected communities of

individuals’ health insurance coverage, access to care, use of services, and perceived quality.

The sample for the surveys was obtained primarily through random digit dialing, supplemented

by in-person interviews in order to represent households without telephones. For a more

detailed description of the survey, see Strouse et al., (2003).

The overall sample for the surveys includes about 60,000 persons in the 2000-01 survey,

and about 46,600 persons in the 2003 survey. This study focuses on nonelderly persons (less

than age 65) with family income less than 300 percent of the federal poverty level. The analysis

focuses on low income persons because their use of EDs is higher compared to higher income

groups, and because it can be expected that any changes in Medicaid policy—including spillover

effects for privately insured and uninsured—will have the greatest effects on low income

persons. Although the vast majority of Medicaid and SCHIP enrollees have incomes less than

200% of poverty, low income is defined as less than 300% of the poverty level in the analysis

because a much higher proportion of the uninsured are in the range of 200-300% of poverty.

Nevertheless, sensitivity tests show that the results are similar regardless of whether 200% or

300% is used as the cutoff to define low income. Elderly persons with Medicare coverage are
5

excluded because Medicaid or private coverage is supplemental, and therefore changes in

Medicaid policy are likely to affect their use of EDs very differently. The final sample for the

analysis includes about 18,500 low income nonelderly persons in the 2000-01 sample, and about

16,000 persons in the 2003 sample.

3.2 Model specification

Based on the above conceptual framework, the basic model of ED use in this analysis can

be specified as:

ED use = f(COV, MED, EDDIS, CHC, N, P, HS, COM)

Thus, ED use is a function of type of health insurance coverage (COV), the availability of

office-based physicians in the community willing to accept Medicaid patients (MED), proximity

to hospital emergency departments (EDDIS), proximity to and capacity of Community Health

Centers (CHCs), individual need (N) and predisposing (P) characteristics, other health system

factors (HS), and community characteristics (COM).

One potential source of bias in estimating the effects of insurance coverage on utilization

is that choice of insurance may not be independent of patterns of health care utilization. In other

words, there may be some unobserved factors that influence decisions to use certain types and

quantities of health care that are also correlated with decisions to enroll in public or private

insurance coverage, or remain uninsured. To the extent that these unobserved differences aren’t

accounted for, then estimates of the effects of insurance coverage on health services use could be

biased.

To test for the potential for insurance coverage selection, a standard instrumental variable

analysis was used to compute predicted measures of insurance coverage in a first stage model

that are then included as independent variables in the utilization models (the second stage). The
6

insurance coverage prediction model included a set of identifying variables that are expected to

strongly influence choice of health insurance coverage, but not utilization of EDs. The

identifying variables used in the first stage included measures of eligibility for public coverage,

the cost of private health insurance in the area of residence, and employment characteristics of

individuals and their families that reflect access to employer-sponsored coverage. As is

discussed in greater detail in section 4.5, differences in estimates between the instrumental

variable analysis and OLS regressions were not statistically significant. OLS estimates are used

to report results because of their much higher level of statistical precision.

3.3 Dependent variable specification

The means for the dependent and independent variables for both low income adults and

children are shown in Table 1. For ED visits, survey respondents were asked to report on the

number of visits to hospital EDs in the 12 months prior to the interview. The survey also

distinguished between ED visits that resulted in inpatient hospital stays and other ED visits.

Only a relatively small proportion of ED visits (15 percent) resulted in an inpatient stay, and are

excluded from the study because most are likely to be true emergencies and therefore less

susceptible to changes in Medicaid policy. For visits that did not result in inpatient stays, the

survey did not distinguish between visits for emergent, urgent, and nonurgent problems.

The analysis also includes a measure of non-ED related physician visits as a comparison

to ED use. The survey also asked about the number of visits to physicians in the previous 12

months other than those seen in the hospital ED. These visits include those that took place in a

physician’s office, clinic, hospital outpatient department, or any other ambulatory setting other

than an ED.

3.4 Independent variable specification.
7

• Health insurance coverage. Type of health insurance coverage was ascertained on the day

of the interview. Four types of coverage are distinguished in this study; private insurance

(including employer-sponsored and nongroup policies); Medicaid/SCHIP coverage; other

coverage types (generally Medicare disability, military coverage, Indian Health Service,

other unspecified coverage); and uninsured. About 20 percent of persons had a change in

coverage during the preceding 12 months, and it is possible that some ED use occurred while

they had a coverage type that was different than the day of the interview. However,

restricting the analysis to only those persons who had the same type of coverage throughout

the year does not appreciably change the results or conclusions.

• Medicaid acceptance rates in the community. This measure reflects the predicted percent of

physicians in the community accepting all or most new Medicaid patients. It is based on a

question in the 2000-01 CTS physician survey that asked physicians whether they were

currently accepting all, most, some, or no new Medicaid patients. The CTS physician

survey includes a representative sample of physicians in the same 60 communities as the

CTS household survey (Diaz-Tena et al. 2003).1 The measure in this study is a predicted

measure derived from a comprehensive model that includes Medicaid reimbursement levels

in the state, as well as other physician, health system, and community characteristics

(Cunningham and Nichols 2005).

• CHC availability and capacity. Similar to previous studies, data from the Health Resources

and Services Administration’s Uniform Data System (UDS) were linked to the CTS

household survey (Cunningham and Hadley 2004; Hadley and Cunningham 2004). The

UDS provides extensive information on revenue, staffing, services provided, and patient

characteristics on all CHC federal grantees and their sites (Bureau of Primary Health Care
8

2004). By linking CHC data to the CTS household survey at the zip code level and using

longitude and latitude coordinates to compute distances in miles, measures of total CHC

revenue within 5 miles of the person’s residence were computed. To standardize the

measure based on the potential user population of the local community, the measure was

further refined to reflect total CHC grant revenue per poor person within 5 miles.

• Distance to hospital emergency departments. Using data from the American Hospital

Association annual survey, information on hospital EDs were linked to the CTS survey by

zip code (AHA 2003). As with the measure of CHC availability, distances between each

sample person and every hospital ED within the community were computed. The measure

in this study reflects distance (in miles) to the nearest hospital ED.

• HMO enrollment. All persons who reported health insurance coverage—including those

with Medicaid/SCHIP coverage—were asked if they were enrolled in an HMO. A variable

is included to reflect HMO enrollment for insured persons. In addition, the survey question

was used to compute the percent of persons in the community enrolled in HMOs (both public

and private). The analysis also tested the inclusion of a state-level Medicaid HMO

penetration measure, based on data from CMS. However, this measure was dropped from

the final analysis because it showed no independent effects on ED use, and there were

concerns about collinearity with the other measures.

• Individual need factors. Derived from the CTS household survey, these include self-

reported health status (i.e. excellent, very good, good, fair, or poor) for both adults and

children. In addition, adults in the survey were asked about a selected number of the most

prevalent chronic conditions. The study includes measures of whether adults have 1 chronic

condition or 2 or more chronic conditions.
9

• Other individual characteristics (i.e. other enabling and predisposing factors). Derived

entirely from the CTS survey, these include age, gender, race/ethnicity, family type (i.e.

married with children, single, single-parent), education level, and family income relative to

poverty.

• Other community characteristics. Controls are also added for a variety of community

characteristics, including the number of primary care practitioners per 1,000 persons.

Derived from the CTS household survey, community-level measures are also constructed for

the percent uninsured, average family income, percent black, percent Hispanic, and the

percent in fair or poor health. Indicators for large metro areas (greater than 200,000 persons)

and small metro areas (less than 200,000 persons) are also included. Binary variables for

survey year and each of the states in the CTS survey are included in order to control for state

and year fixed effects.

3.5 Analysis

All regressions were estimated as linear probability models. Separate models are

estimated for adults and children, as the effects of coverage and other factors are likely to differ

for adults and children. Results for the effects of insurance coverage on use are based on pooled

models that combine all insurance coverage groups into a single analysis, with indicators of type

of insurance coverage included in the model. However, since the effects of key health system

variables on ED use may differ depending on type of coverage (e.g. the effects of Medicaid

acceptance rates), separate models are also estimated for Medicaid/SCHIP enrollees, uninsured,

and privately insured persons. All models are estimated using SUDAAN to adjust the standard

errors for the effects of the complex sampling design (Shah et al., 1996).
10

Because measures of ED visits are highly skewed (i.e. a high proportion of the sample

have 0 visits) and because decisions about whether to use the ED at all may differ from decisions

about how often to use the ED, the analysis follows a standard two-step approach for estimating

service use (Duan et al. 1982). The first model estimates the probability of having any ED visit

in the prior year, while a second model estimates the number of ED visits for persons with any

ED use. Predicted estimates (e.g. for insurance coverage) are computed for the probability of

use and number of visits separately, and then the two predictions are multiplied in order to derive

an overall predicted number of visits per person.

4.0 Results

4.1 Effects of insurance coverage on ED use

Table 2 contains the regression results for ED use for low income adults (less than 300%

of the federal poverty level, ages 18-64). The results show that Medicaid/SCHIP enrollees and

persons with “other” coverage are more likely to have an ED visit in the past year relative to

uninsured adults. However, there were no differences in probability of ED use between

privately insured and uninsured adults. For adults with an ED visit, there were few differences

in the number of visits between the coverage groups, with Medicaid/SCHIP enrollees having

only a slightly higher number of visits compared to uninsured adults.

Proximity to key health care providers also has some effect on ED use. Greater CHC

capacity in the local area was associated with a lower probability of ED use, but had no effect on

the number of ED visits. Greater distance to the ED did not affect the probability of use, but

was associated with fewer ED visits.

Characteristics of individuals also have strong effects on ED use. In general, higher ED

use is associated with younger ages, female gender, lower incomes, less than a college education,
11

single adults, poor health, and having chronic conditions. Community characteristics associated

with higher ED use include smaller numbers of primary care providers, a higher African-

American population, small Hispanic population, and being in a small MSA.

Table 3 includes the regression results for ED use by low income children (less than age

18). In contrast to the findings for adults, there were no statistically significant differences in

the probability of ED use between Medicaid/SCHIP, privately insured, and uninsured children.

For ED users, Medicaid/SCHIP and privately insured children had fewer visits compared to

uninsured children. Longer distances to the ED decreased the probability of use for children.

Contrary to expectations, HMO enrollment and higher Medicaid acceptance rates increased the

number of visits for ED users, although these results were statistically significant only at the .10

level.

In terms of individual characteristics, older children have lower levels of ED use

compared to younger children, while children in single-parent families have both a higher

probability of use and a greater number of visits compared to children living with both parents.

As with adults, ED use is highest among those children with fair or poor health. Other

community characteristics have no significant effect on the probability of use, although there

were some effects on number of visits. Higher numbers of ED visits by children was associated

with higher uninsurance rates in the community, a smaller Hispanic population, and residence in

metropolitan vs. nonmetropolitan areas.

4.2 Effects of insurance coverage on physician use.

Table 4 shows the results for the effects of insurance coverage on physician use for low

income adults. Having any type of health insurance greatly increases the probability of a

physician visit relative to being uninsured. For adults with physician visits, Medicaid/SCHIP
12

and “other” coverage increases the number of visits, although there was no statistically

significant difference between privately insured and uninsured adults in the number of physician

visits.

Medicaid acceptance rates had no effect on the probability of physician use, although

higher acceptance rates are associated with a higher number of visits. Longer distances to the

hospital ED increases the probability of a physician visit, but decreases the number of physician

visits. In general, greater use of physicians is associated with younger ages, females, higher

incomes, whites, college graduates, persons with fair or poor health, and persons with chronic

conditions. Higher numbers of physician visits were associated with greater HMO penetration

in the community, greater numbers of blacks and lower numbers of Hispanics, and residence in

small MSAs.

Table 5 shows the results for the effects of insurance coverage on physician use for low

income children. As with adults, having any type of insurance coverage greatly increases the

probability of physician use, and also increases the number of visits for users. Greater CHC

capacity also increases the probability of a physician visit for low income children. The effects

of individual characteristics on physician use for children are similar to the effects for adults.

Community characteristics had virtually no statistically significant effect on physician use for

children.

4.3 Comparison of regression-adjusted utilization with actual utilization.

The above results are used to compute regression-adjusted measures of utilization for

Medicaid/SCHIP, uninsured, and privately insured persons (separately for adults and children).2

The regression-adjusted estimates hold constant all of the individual, community, and health

system factors included in the regression models. Regression-adjusted estimates for
13

Medicaid/SCHIP and privately insured are computed by setting the indicator for that specific

coverage group to 1, setting the indicators for all other coverage groups to 0, and using the actual

values for all other independent variables. Regression-adjusted estimates for uninsured are

similarly computed by setting all insurance coverage indicators to 0. Separate regression-

adjusted estimates are computed for the probability of use and number of visits for users. For

each coverage and age group, the results of these two estimates are then multiplied to obtain the

regression-adjusted estimate for overall visits per person.

In general, Table 6 shows that differences between the regression-adjusted estimates and

the actual estimates were greatest for ED visits for adults, and particularly ED visits for

Medicaid/SCHIP. After accounting for differences in individual, community, and health system

characteristics, ED visits per person for Medicaid/SCHIP adults decreased by about one-third,

from 0.94 visits per person to 0.66 visits per person. The regression adjustments had less of an

effect on ED use for uninsured and privately insured adults. The result is that differences in ED

use between Medicaid/SCHIP and the other groups are considerably smaller after the regression

adjustments, although ED use for Medicaid/SCHIP enrollees continues to be higher. For

example, the difference in regression-adjusted ED visits between Medicaid/SCHIP and

uninsured (0.25 visits per person) is less than half that of the actual difference (0.56 visits per

person).

Differences between actual and regression-adjusted estimates of ED use are smaller for

children, as are differences in estimates for physician utilization. The result is that individual,

community, and health system factors account for less of the differences in physician utilization

by insurance status than they do for ED use, and that physician utilization is considerably higher

for Medicaid/SCHIP enrollees than for uninsured and even privately insured.
14

4.4 Effects of key health system factors on ED use, by insurance coverage

Table 7 shows selected results of separate regression models for Medicaid/SCHIP,

uninsured, and privately insured. The results show that the effects of key health system factors

on ED use differ depending on type of insurance coverage. Enrollment in HMOs decreases the

probability of ED use by about 5 percentage points for Medicaid/SCHIP adults, although this

effect is only significant at the .10 level. HMO enrollment had no statistically significant effect

on ED use by privately insured adults.

As one would expect, Medicaid acceptance rates had the largest effect on

Medicaid/SCHIP enrollees, in that higher acceptance rates significantly decreased the probability

of ED use among Medicaid/SCHIP adults. Higher CHC capacity decreased the probability of

ED use for adults across all three insurance groups, although the magnitude of the effect appears

to be larger for Medicaid/SCHIP enrollees. Longer distances to the ED decreased the

probability of ED use primarily for privately insured adults, although longer ED distances

decreased the number of ED visits for uninsured adults. Longer ED distances are associated

with a higher number of ED visits for Medicaid/SCHIP adults, although this anomalous is

statistically significant only at the .10 level.

The findings are more mixed for children, perhaps due in part to smaller sample sizes and

less statistical precision. Only a few results achieved a level of statistical significance at the .05

level. Greater CHC capacity decreased the probability of ED use for Medicaid/SCHIP children,

but slightly increased the probability of ED use for uninsured children. While Medicaid

acceptance rates had no affect on ED use for Medicaid/SCHIP children, higher acceptance rates

increased the number of ED visits for uninsured children.
15

4.5 Testing for the endogeneity of insurance coverage

As discussed earlier, an instrumental variable (IV) analysis that accounts for the possible

endogenous effects of insurance coverage on ED use was tested. This involved estimating

models of ED utilization using predicted measures of insurance coverage, derived from linear

models that included binomial measures of private insurance and Medicaid/SCHIP coverage as

separate dependent variables.3 Independent variables in these models included a set of

identifying variables that are important for predicting insurance coverage but not ED utilization.

The identifying variables included measures of Medicaid/SCHIP eligibility, county-level

measures of private insurance premiums, individual and family-level measures of employment

status and job characteristics that are related to the availability of ESI coverage (e.g. firm size,

self-employed status, industry, wage-rate). 4 Other variables include age, gender, family

income, education, and health status. The specification and results of these models are included

in Table 8 (for adults) and Table 9 (for children).

As the results in the table show, most of the identifying variables exhibited statistically

significant effects on coverage in the expected direction. In addition, tests to determine whether

the increment in the proportion of variance accounted for by the instrumental variables as a

group (i.e. change in R-square tests) were statistically significant at the .05 level in all four

models (Pedhazur 1982). A final set of specification tests determines the independence of the

identifying variables with the residuals of ED and physician use models that include the

predicted measures of insurance coverage (Hausman 1983). In all models, the test statistic was

small and not statistically different from zero at the .05 level, indicating little or no correlation of

the identifying variables with the residuals in the utilization models.
16

Table 10 compares the effect of the predicted insurance coverage measures on ED use

(i.e. from the IV analysis) with the simple OLS estimates. Although there are some large

differences in magnitude between some of the OLS and IV coefficients, the IV estimates have a

much lower degree of statistical precision (i.e. high standard errors), making it difficult to

conclude that the IV estimates are truly different from the OLS estimates. In fact, tests of

differences between the OLS and IV estimates show that for only one finding – the effect of

privately insured on ED use for adults – was the difference between OLS and IV estimates

statistically significant. While the OLS estimates show no difference between privately insured

and uninsured in probability of ED use, IV estimates show that privately insured are more likely

to have an ED visit than the uninsured. The standard errors for the IV estimates for children are

especially high. Overall, the results indicate that IV estimates are not an improvement on the

OLS estimates in this analysis, and that the low level of statistical precision of the IV estimates

makes them unsuitable for simulating the effects of coverage changes on aggregate ED

utilization.

While it is still possible that the OLS estimates in this analysis are biased due to

insurance-related selection or endogeneity, a recent review of research showed that the effects of

insurance coverage on utilization were quite consistent across three different methods, including

the same types of standard regression-based methods used in this analysis, instrumental variable

analysis, and longitudinal studies (Buchmueller et al., 2005). The broad and extensive set of

controls for individual, health system, and community characteristics used in this analysis also

reduces the potential for bias due to unobserved factors that are correlated with both insurance

coverage and utilization. Nevertheless, causality can only be inferred conclusively with

longitudinal data on individuals’ insurance coverage and use, which are not available nationally.
17

5. Conclusion

This research brief provides both a conceptual and empirical basis for estimating the

effects of insurance coverage on ED use for low income persons, and for estimating the effects

of coverage changes and reduced primary care capacity on aggregate ED use. The analysis is

unique in that it is derived from a large nationally representative survey of the U.S. population

and includes a broad and extensive array of individual, health system, and community

characteristics that are known and potential correlates of ED use. The analysis tested for the

effects of insurance coverage selection through an instrumental variable analysis, but did not find

that the results were an improvement on the OLS estimates. Although the possibility of

selection bias can’t be discounted, a recent review of research indicates that it is less of a

problem than commonly assumed (Buchmueller et al. 2005).

The results show that ED use by Medicaid/SCHIP adults is higher than for privately

insured and uninsured adults, even after controlling for the considerable differences in health

status and other characteristics. ED use by Medicaid/SCHIP children is slightly lower than

uninsured children, but still higher compared to privately insured children. Both

Medicaid/SCHIP adults and children have consistently higher levels of physician use compared

to privately insured and uninsured, while uninsured adults and children consistently have the

lowest levels of physician compared to persons with insurance coverage.

The results also show that individual characteristics are important correlates of both ED

and physician use, especially health status. In addition, several health system factors have

significant effects on ED use for Medicaid/SCHIP persons, including Medicaid acceptance rates

and CHC capacity (lower values of both are associated with higher probability of use). Lower

CHC capacity also increases the probability of ED use for uninsured adults. Longer distances to
18

hospital EDs has some effect on decreasing ED use, although these effects are somewhat

inconsistent across insurance groups.

ENDNOTES

1
The most recent physician survey was conducted in 2004, although the data were not available
at the time of this analysis.
2
The regression-adjusted estimates of ED and physician utilization in Table 6 were used to
compute the regression-adjusted differences in utilization between the insurance groups shown in
Exhibit 3 of the paper, “Do Reductions in Medicaid/SCHIP Enrollment Increase Hospital ED
Use Among Low Income Persons?” (Cunningham 2005)
3
Another possible approach was to estimate insurance coverage using a multinomial logistic
regression model. When such a model was used to generate predicted measures of public and
private insurance coverage, the effects of these predicted measures on ED use were roughly
similar to the predicted measures based on linear probability models.
4
Because a few of the identifying variables were available for use only with the 2003 survey, the
test is limited to the 2003 sample. OLS estimates for the 2003 sample are very similar to OLS
estimates for the combined 2000-01 and 2003 sample.
19

REFERENCES

American Hospital Association. 2003. Annual Survey of Hospitals, 2002. Chicago: American
Hospital Association.

Andersen, Ronald M. 1968. Behavioral Model of Families’ Use of Health Services. Research
Series No. 25. Chicago, IL: Center for Health Administration Studies, University of
Chicago.

Andersen, Ronald M. 1995. Revisiting the Behavioral Model and Access to Medical Care:
Does it Matter? Journal of Health and Social Behavior 36(March): 1-10.

Buchmueller, T., K. Grumbach, R. Kronick, and J.G. Kahn. 2005. The effect of health
insurance on medical care utilization and implications for insurance expansion: A review
of the literature. Medical Care Research and Review 62(1): 3-30.

Bureau of Primary Health Care. BPHC Uniform Data System Manual: 2004 Revision.
(http://bphc.hrsa.gov/uds).

Cunningham, P.J. 2005. Do reductions in Medicaid/SCHIP enrollment increase hospital ED
use among low income persons? Health Affairs 25(1):237-247.

Cunningham, P.J., and L.M. Nichols. 2005. The effects of Medicaid reimbursement on the
access to care of Medicaid enrollees: A community perspective. Medical Care Research
and Review 62(6): 676-696.

Cunningham, P.J., C.M. Clancy, J.W. Cohen, M. Wilets. 1995. The use of hospital emergency
departments for nonurgent health problems: A national perspective. Medical Care
Research and Review 52(4): 453-474.

Cunningham, Peter J. and Jack Hadley. 2004. Expanding care versus expanding coverage:
How to improve access to care. Health Affairs 23(4): 234-244.

Diaz-Tena, N., Potter, F., Strouse, R., Williams, S., and M. Ellrich. 2003. Community Tracking
Study, Physician Survey Methodology Report 2000-01 (Round 3). Technical Publication
No. 38. Washington, D.C.: Center for Studying Health System Change.

Duan, H., W.G. Manning Jr., C.N. Morris, and J.P. Newhouse. 1982. A Comparison of
Alternative Models for the Demand for Medical Care. Rand Health Insurance Experiment
Series (#R-2754-HHS) Santa Monica, CA: The Rand Corporation.

Dubay, L. and G. Kenney. Gains in children’s health insurance coverage but additional progress
needed, Pediatrics 114 (Nov): 1338-1340.
20

Grossman, L.K., L.N. Rich, and C. Johnson. 1998. Decreasing nonurgent emergency
department utilization by Medicaid children. Pediatrics 102: 20-24.

Guttman, N., Zimmerman, D.R., and M.S. Nelson. 2003. The many faces of access: reasons for
medically nonurgent emergency department visits. Journal of Health Politics, Policy, and
Law 28(6): 1089-1120.

Hadley, Jack, and Peter Cunningham. 2004. Availability of safety net providers and access to
care of uninsured persons. Health Services Research 39(5): 1527-1546.

Hausman, J. 1983. Specification and Estimation of Simultaneous Equation Models, in Z.
Griliches and M. Intrilligator (eds.), Handbook if Econometrics, Amsterdam: North
Holland.

Hurley, R.E., D.A. Freund, and Captain D.E. Taylor. 1989. Gatekeeping the emergency
department: Impact of a Medicaid primary care case management program. Health Care
Management Review 14(2): 63-71.

McLaughlin, C.G., and K. Mortensen. 2003. Who walks through the door? The effect of the
uninsured on hospital use. Health Affairs 22(6): 143-155.

Oster, A., and A.B. Bindman. 2003. Emergency department visits for ambulatory care sensitive
conditions. Medical Care 41(2): 198-207.

Pedhazur, E. 1983. Multiple Regression in Behavioral Research. New York: Holt, Rinehard,
and Winston.

Piehl, M.D., Clemens, C.J., and J.D. Jones. 2000. Decreasing emergency department use by
children enrolled in the Medicaid program by improving access to primary care. Archives
of Pediatric and Adolescent Medicine 154 (Aug): 791-795.

Powers, R.D. 2000. Emergency department use by adult Medicaid patients after implementation
of managed care. Academic Emergency Medicine. 7(December): 1416-1420.

Shah, B.V., Barnwell, B.G., and G.S. Bieler. 1996. SUDAAN User’s Manual, Release 7.0.
Research Triangle Park, N.C.: Research Triangle Institute.

Strouse, R., Carlson, B., and J. Hall. 2003. Community Tracking Study: Household Sruvey
Methodology Report 2000-01 (Round Three). Technical Publication No. 46. Washington,
D.C.: Center for Studying Health System Change.

Zuckerman, S., N. Brennan, and A. Yemane. 2002. Has Medicaid managed care affected
beneficiary access and use? Inquiry 39(3): 221-242.
21

Table 1. Means of dependent and indepenendent variables.

Low income adults Low income children
(18-64) (age 0-17)
Dependent variables
Probability of ED use 0.23 0.21
Number of ED visits per user 1.98 1.91
Probability of physician visit 0.67 0.79
Number of physician visits per user 5.69 3.78

Independent variables
Medicaid/SCHIP coverage 0.12 0.34
Private insurance 0.52 0.51
Other coverage 0.08 0.04
Uninsured 0.28 0.11

Enrolled in HMO 0.32 0.39

Medicaid acceptance rates1 62.8 62.6

CHC revenue within 5 miles2 86.29 75.64

Distance to nearest ED (miles)3 5.44 5.47

Age 0-6 -- 0.32
Age 7-12 -- 0.42
Age 13-17 -- 0.26
Age 18-34 0.44 --
Age 35-44 0.26 --
Age 45-54 0.17 --
Age 55-64 0.13 --

Female 0.54 0.49

Income < 100% of poverty 0.25 0.28
Income 100-199% of poverty 0.37 0.39
Income 200-299% of poverty 0.38 0.33

White 0.58 0.56
Black 0.16 0.17
Hispanic 0.21 0.22
Other 0.05 0.05

Less than H.S. 0.24 0.20
H.S. grad 0.40 0.40
Some college 0.24 0.26
College grad 0.12 0.13
22

Low income adults Low income children
(18-64) (age 0-17)
Single person 0.31 --
Married couple, no kids 0.15 --
Married couple with kids 0.40 0.63
Single parent with kids 0.15 0.37

Excellent, very good health 0.49 0.78
Good health 0.29 0.17
Fair or poor health 0.22 0.06

0 chronic conditions 0.69 --
1 chronic condition 0.18 --
2 or more chronic conditions 0.12 --

Community-level variables
Percent HMO in community 0.38 0.38
Number of PCPs per 1000 persons4 0.53 0.52
Percent uninsured in community 14.03 13.9
Average family income / 100 436 437
Percent black 0.12 0.12
Percent Hispanic 0.12 0.12
Percent in fair or poor health 0.13 0.13
Large MSA (> 200,000 persons) 0.71 0.72
Small MSA 0.06 0.05
Nonmetro 0.23 0.23

Indicator for 2003 survey 0.53 0.51

1
Predicted percent of physicians in the site accepting all or most new Medicaid patients
(Cunningham and Nichols 2005).
2
Data on CHC revenue from HRSA’s Uniform Data System linked to the CTS household survey
by zip code.
3
Data from the 2000 and 2002 American Hospital Association Annual Survey linked to the CTS
household survey by zip code.
4
2000-01 Community Tracking Study physician survey.

All other variables are based on the 2000-01 and 2003 Community Tracking Study household
survey.
23

Table 2. OLS regressions for ED use, low income adults (age 18-64, less than 300% of
poverty).

Probability of ED visit Number of ED visits
for persons with at
least one visit
Intercept 0.032 (0.094) 0.325 (1.06)

Medicaid/SCHIP coverage 0.097 (0.015)*** 0.219 (0.149)*
Private insurance 0.011 (0.014) -0.061 (0.118)
Other coverage 0.069 (0.019)** 0.167 (0.166)
Uninsured (reference group) 0 0

Enrolled in HMO 0.005 (0.009) 0.141 (0.137)
Medicaid acceptance rates by physicians1 0.0005 (0.0007) 0.004 (0.007)
CHC revenue within 5 miles2 -0.000032 (0.00001) *** 0.00016 (0.0001)
Distance to nearest ED (miles)3 -0.0013 (0.006) -0.013 (0.006) **

Age 18-34 0.042 (0.01)*** 0.068 (0.098)
Age 35-44 (reference group) 0 0
Age 45-54 -0.065 (0.014)*** -0.325 (0.138)**
Age 55-64 -0.111 (0.018)*** -0.61 (0.132) ***

Female 0.034 (0.008)*** 0.182 (0.08) **

Income < 100% of poverty (reference 0 0
group)
Income 100-199% of poverty -0.004 (0.012) -0.278 (0.142) *
Income 200-299% of poverty -0.016 (0.014) -0.429 (0.162) ***

White (reference group) 0 0
Black 0.019 (0.013) 0.042 (0.081)
Hispanic -0.024 (0.016) -0.174 (0.124)
Other 0.021 (0.02) 0.494 (0.28) *

Less than H.S. 0.046 (0.016)*** 0.375 (0.09) ***
H.S. grad 0.03 (0.014) ** 0.356 (0.086) ***
Some college 0.043 (0.014) *** 0.449 (0.13) ***
College grad (reference group) 0 0

Single person 0.024 (0.008) *** 0.077 (0.10)
Married couple, no kids -0.004 (0.013) 0.041 (0.137)
Married couple with kids (reference group) 0 0
Single parent with kids 0.067 (0.011) *** 0.021 (0.126)

Excellent, very good health (reference 0 0
group)
Good health 0.04 (0.006)*** 0.156 (0.127)
Fair or poor health 0.096 (0.009) *** 0.521 (0.177) ***
24

Probability of ED visit Number of ED visits
for persons with at
least one visit

0 chronic conditions (reference group) 0 0
1 chronic condition 0.084 (0.011) *** 0.323 (0.112) ***
2 or more chronic conditions 0.138 (0.011) *** 0.560 (0.107) ***

Other community-level variables
HMO penetration in site 0.005 (0.063) 1.147 (0.815) *
Number of PCPs per 1000 persons4 -0.175 (0.086) ** 0.652 (0.675)
Percent uninsured in community -0.0016 (0.0014) 0.01 (0.01)
Average family income 0.00009 (0.00008) 0.0001 (0.0008)
Percent black 0.239 (0.064) *** 2.317 (0.67) ***
Percent Hispanic -0.167 (0.065) ** -0.637 (0.578)
Percent in fair or poor health 0.168 (0.175) 0.322 (1.809)
Large MSA (> 200,000 persons) 0.02 (0.015) -0.446 (0.19) **
Small MSA 0.053 (0.018) *** -0.124 (0.14)

Binary indicator for 2003 survey -0.0003 (0.008) 0.11 (0.085)

Number of observations 26,200 5,720
R2 0.08 0.10
***p < .01, **p < .05, * p < .10
Standard errors of estimates are in parentheses.

Note: Binary variables for each of the states in the CTS were included in the regressions but are
not shown in the table.
1
Predicted percent of physicians in the site accepting all or most new Medicaid patients
(Cunningham and Nichols 2005).
2
Data on CHC revenue from HRSA’s Uniform Data System linked to the CTS household survey
by zip code.
3
Data from the 2000 and 2002 American Hospital Association Annual Survey linked to the CTS
household survey by zip code.
4
2000-01 Community Tracking Study physician survey.

All other variables are based on the 2000-01 and 2003 Community Tracking Study household
survey.
25

Table 3. OLS regressions for ED use, low income children (age 0-17) less than 300% of
poverty).

Probability of ED visit Number of ED visits
for persons with at
least one visit
Intercept -0.131 (0.184) -0.774 (1.34)

Medicaid/SCHIP coverage 0.019 (0.026) -0.476 (0.263)*
Private insurance -0.017 (0.023) -0.708 (0.263) **
Other coverage 0.073 (0.04) * -0.524 (0.469)
Uninsured (reference group) 0 0

Enrolled in HMO -0.001 (0.017) 0.232 (0.124) *
Medicaid acceptance rates by physicians1 0.002 (0.0016) 0.021 (0.011) *
CHC revenue within 5 miles2 -0.000036 (0.00003) 0.00017 (0.0002)
Distance to nearest ED (miles)3 -0.003 (0.001) ** -0.002 (0.009)

Age 0-5 (reference group) 0 0
Age 6-12 -0.075 (0.16) *** -0.131 (0.118)
Age 13-17 -0.053 (0.019) *** -0.483 (0.159) ***

Female -0.001 (0.016) 0.025 (0.14)

Income < 100% of poverty (reference 0 0
group)
Income 100-199% of poverty -0.002 (0.019) -0.341 (0.139) **
Income 200-299% of poverty -0.033 (0.022) -0.076 (0.198)

White (reference group) 0 0
Black -0.048 (0.022) ** 0.248 (0.163)
Hispanic -0.002 (0.021) -0.225 (0.205)
Other 0.077 (0.042) * 0.778 (0.40) *

Less than H.S. -0.045 (0.02)** 0.375 (0.09) ***
H.S. grad -0.021 (0.017) 0.356 (0.086) ***
Some college -0.009 (0.018) 0.449 (0.13) ***
College grad (reference group) 0 0

Single parent with kids 0.067 (0.011) *** 0.273 (0.108) **

Excellent, very good health (reference 0 0
group)
Good health 0.056 (0.018)*** 0.087 (0.125)
Fair or poor health 0.162 (0.029) *** 1.849 (0.477) ***

Other community-level variables
HMO penetration in site 0.157 (0.126) 0.034 (0.83)
Number of PCPs per 1000 persons4 0.068 (0.154) -2.29 (1.15) **
26

Probability of ED visit Number of ED visits
for persons with at
least one visit
Percent uninsured in community 0.002 (0.003) 0.052 (0.021) **
Average family income 0.00017 (0.00015) 0.0015 (0.001)
Percent black 0.116 (0.147) 1.041 (0.96)
Percent Hispanic -0.033 (0.089) -2.241 (0.983) **
Percent in fair or poor health 0.332 (0.304) 2.264 (2.011)
Large MSA (> 200,000 persons) 0.002 (0.034) 0.492 (0.19) ***
Small MSA 0.014 (0.032) 0.715 (0.316) **

Indicator for 2003 survey -0.02 (0.019) -0.013 (0.104)

Number of observations 8,008 1,764
R2 0.05 0.24
***p < .01, **p < .05, * p < .10
Standard errors of estimates are in parantheses.

All other variables are based on the 2000-01 and 2003 Community Tracking Study household
survey.
27

Table 4. OLS regressions for physician use, low income adults (age 18-64, less than 300% of
poverty).

Probability of physician Number of physician
visit visits for persons with
at least one visit
Intercept 0.47 (0.118) *** -4.45 (2.41) *

Medicaid/SCHIP coverage 0.295 (0.019)*** 2.80 (0.39) ***
Private insurance 0.237 (0.018) *** 0.32 (0.33)
Other coverage 0.271 (0.022) *** 2.20 (0.39) ***
Uninsured (reference group) 0 0

Enrolled in HMO 0.011 (0.01) 0.234 (0.205)
Medicaid acceptance rates by physicians1 -0.00002 (0.0009) 0.036 (0.016) **
CHC revenue within 5 miles2 -0.000003 (0.00002) 0.00004 (0.0003)
Distance to nearest ED (miles)3 0.012 (0.0068) * -0.033 (0.013) **

Age 18-34 0.024 (0.008)*** 0.335 (0.234)
Age 35-44 (reference group) 0 0
Age 45-54 0.011 (0.012) -0.729 (0.269) ***
Age 55-64 0.018 (0.013) -1.583 (0.304) ***

Female 0.161 (0.006)*** 1.319 (0.146) ***

Income < 100% of poverty (reference 0 0
group)
Income 100-199% of poverty 0.03 (0.011) *** 0.126 (0.25)
Income 200-299% of poverty 0.056 (0.012) *** -0.058 (0.27)

White (reference group) 0 0
Black -0.017 (0.01) -1.305 (0.289) ***
Hispanic -0.075 (0.016) *** -1.331 (0.278) ***
Other -0.036 (0.019) * 0.089 (0.553)

Less than H.S. -0.061 (0.018)*** -0.39 (0.273)
H.S. grad -0.053 (0.014) *** -0.446 (0.218)**
Some college -0.023 (0.016) 0.004 (0.24)
College grad (reference group) 0 0

Single person -0.018 (0.009) ** -0.368 (0.242)
Married couple, no kids 0.013 (0.013) -0.154 (0.23)
Married couple with kids (reference group) 0 0
Single parent with kids 0.013 (0.013) 0.055 (0.283)

Excellent, very good health (reference 0 0
group)
Good health 0.023 (0.009)*** 0.614 (0.143) ***
Fair or poor health 0.049 (0.01) *** 2.966 (0.247) ***
28

Probability of physician Number of physician
visit visits for persons with
at least one visit

0 chronic conditions (reference group) 0 0
1 chronic condition 0.191 (0.01) *** 2.185 (0.25) ***
2 or more chronic conditions 0.244 (0.01) *** 4.492 (0.242) ***

Other community-level variables
HMO penetration in site -0.107 (0.099) 5.222 (1.789) ***
Number of PCPs per 1000 persons4 -0.037 (0.116) -0.924 (2.654)
Percent uninsured in community -0.003 (0.002) -0.023 (0.03)
Average family income -0.0001 (0.0001) 0.004 (0.002)*
Percent black 0.128 (0.078)* 5.65 (1.29) ***
Percent Hispanic 0.055 (0.071) -4.09 (1.60)**
Percent in fair or poor health 0.140 (0.23) 7.75 (6.37)
Large MSA (> 200,000 persons) 0.003 (0.02) -0.545 (0.392)
Small MSA -0.008 (0.021) 1.254 (0.377)***

Binary indicator for 2003 survey -0.003 (0.01) 0.199 (0.164)

Number of observations 26,200 18,393
R2 0.22 0.13
***p < .01, **p < .05, * p < .10
Standard errors of estimates are in parentheses.

All other variables are based on the 2000-01 and 2003 Community Tracking Study household
survey.
29

Table 5. OLS regressions for physician use, low income children (age 0-17) less than 300% of
poverty).

Probability of physician Number of physician
visit visits for persons with
at least one visit
Intercept 0.731 (0.173) *** 3.248 (1.60) **

Medicaid/SCHIP coverage 0.277 (0.031) *** 1.177 (0.22) ***
Private insurance 0.235 (0.029) *** 0.711 (0.278) **
Other coverage 0.217 (0.043) *** 1.953 (0.455) ***
Uninsured (reference group) 0 0

Enrolled in HMO 0.003 (0.014) -0.184 (0.179)
Medicaid acceptance rates by physicians1 -0.001 (0.0015) 0.006 (0.012)
CHC revenue within 5 miles2 0.000034 (0.000015) ** -0.00005 (0.0002)
Distance to nearest ED (miles)3 0.002 (0.001) * 0.0006 (0.011)

Age 0-5 (reference group) 0 0
Age 6-12 -0.102 (0.014) *** -1.05 (0.162) ***
Age 13-17 -0.16 (0.014) *** -1.16 (0.219) ***

Female 0.007 (0.013) -0.193 (0.14)

Income < 100% of poverty (reference 0 0
group)
Income 100-199% of poverty 0.035 (0.02) * -0.025 (0.184)
Income 200-299% of poverty 0.063 (0.018) *** -0.141 (0.232)

White (reference group) 0 0
Black -0.046 (0.02) ** -0.875 (0.203) ***
Hispanic -0.07 (0.026) *** -0.643 (0.237) ***
Other -0.055 (0.022) ** -0.458 (0.485)

Less than H.S. -0.063 (0.029) ** -0.306 (0.304)
H.S. grad -0.036 (0.023) -0.242 (0.264)
Some college -0.007 (0.017) 0.172 (0.25)
College grad (reference group) 0 0

Single parent with kids 0.042 (0.014) *** 0.571 (0.20) ***

Excellent, very good health (reference 0 0
group)
Good health 0.064 (0.016)*** 1.48 (0.285) ***
Fair or poor health 0.095 (0.03) *** 2.92 (0.609) ***

Other community-level variables
HMO penetration in site 0.16 (0.11) -1.304 (1.19)
Number of PCPs per 1000 persons4 -0.024 (0.173) -1.697 (1.52)
30

Probability of physician Number of physician
visit visits for persons with
at least one visit
Percent uninsured in community 0.0001 (0.003) 0.004 (0.028)
Average family income -0.00001 (0.00014) 0.004 (0.028)
Percent black -0.230 (0.164) -0.241 (1.23)
Percent Hispanic 0.113 (0.116) -1.23 (0.88)
Percent in fair or poor health -0.305 (0.327) -1.205 (3.16)
Large MSA (> 200,000 persons) 0.056 (0.041) 0.403 (0.29)
Small MSA 0.049 (0.041) 1.02 (0.326) ***

Indicator for 2003 survey 0.017 (0.012) -0.141 (0.133)

Number of observations 8,008 6,400
R2 0.11 0.09
***p < .01, **p < .05, * p < .10
Standard errors of estimates are in parentheses.

All other variables are based on the 2000-01 and 2003 Community Tracking Study household
survey.
31

Table 6. Predicted ED and physician visits, adults and kids

Medicaid/SCHIP Uninsured Privately insured
Actual Regression- Actual Regression- Actual Regression-
adjusted1 adjusted1 adjusted1
ED visits – adults
Percent with visit 37.62 30.53 19.82 20.79 20.2 21.91
Number of visits for 2.51 2.16 1.92 1.95 1.73 1.88
users
ED visits per person 0.94 0.66 0.38 0.41 0.35 0.41

ED visits – children
Percent with visit 24.12 22.34 20.97 20.42 27.75 18.72
Number of visits for 2.16 1.95 2.38 2.43 1.94 1.72
users
ED visits per person 0.52 0.44 0.50 0.50 0.54 0.32

Physician visits – adults
Percent with visit 84.02 78.92 42.99 49.45 74.28 73.19
Number of visits for 8.82 7.68 4.53 4.88 4.84 5.20
users
ED visits per person 7.41 6.06 1.95 2.41 3.59 3.81

Physician visits –
children
Percent with visit 83.99 84.55 52.95 56.83 79.57 80.31
Number of visits for 4.23 4.08 2.89 2.91 5.15 3.62
users
ED visits per person 3.55 3.45 1.53 1.65 4.10 2.91
1
Predicted visits are based on the regression results in Table’s 2-5. Predicted use for
Medicaid/SCHIP and privately insured is computed by setting the indicator for the specific group
to “1”, setting the other coverage groups to 0, and using the actual values for all variables.
Predicted use for uninsured is computed by setting all insurance coverage indicators to 0.
32

Table 7. Effects of key health system factors on ED use based on separate models for
Medicaid/SCHIP, uninsured, and privately insured.

Medicaid/SCHIP Uninsured Privately
insured
Probability of ED use (adults)
Enrolled in HMO -0.049* -- 0.013
Medicaid acceptance rates by physicians1 -0.006*** 0.002 -0.00001
CHC revenue within 5 miles2 -0.0001*** 0.00004** -0.00002*
Distance to nearest ED (miles)3 -0.003 0.001 -0.002**

Number of ED visits (adults)
Enrolled in HMO 0.358 -- 0.09
Medicaid acceptance rates by physicians1 0.016 -0.001 0.013*
CHC revenue within 5 miles2 -0.0004 0.0002 0.0004*
Distance to nearest ED (miles)3 0.027* -0.037*** -0.013

Probability of ED use (children)
Enrolled in HMO -0.038 -- 0.034
Medicaid acceptance rates by physicians1 0.003 0.006 0.004*
CHC revenue within 5 miles2 -0.0002*** 0.0001* -0.00002
Distance to nearest ED (miles)3 -0.003 -0.009* -0.002

Number of visits (children)
Enrolled in HMO 0.333* -- 0.154
Medicaid acceptance rates by physicians1 0.017 0.145** 0.006
CHC revenue within 5 miles2 -0.0004 0.005 0.0001
Distance to nearest ED (miles)3 0.019 -0.087* -0.01
***p < .01, **p < .05, * p < .10
Standard errors of estimates are in parentheses.

Binary variables for each of the states in the CTS were included in the regression but not shown
in the table.
1
Predicted percent of physicians in the site accepting all or most new Medicaid patients
(Cunningham and Nichols 2005).
2
Data on CHC revenue from HRSA’s Uniform Data System linked to the CTS household survey
by zip code.
3
Data from the 2000 and 2002 American Hospital Association Annual Survey linked to the CTS
household survey by zip code.

All other measures based on the 2000-01 and 2003 Community Tracking Study household
survey.
33

Table 8. Linear probability models for predicting health insurance coverage for adults (ages 18-
64)

Probability of Probability of
private Medicaid/SCHIP
insurance coverage
Intercept 0.526* 0.335*
Age
18-34 -0.060* 0.021
35-44 0.000 0.000
45-54 0.006 -0.012
55-64 0.116* -0.081*
Female 0.016 0.027*
Family income relative to poverty
<100% 0.000 0.000
100-199% 0.122* -0.111*
200-299% 0.291* -0.168*
Race/ethnicity
White 0.000 0.000
Black -0.026 0.016
Hispanic -0.089* 0.034
Other -0.013 0.061
Education
Less than high school -0.191* 0.046*
High school grad -0.111* 0.020
Some college -0.056* 0.004
College grad 0.000 0.000
Family type
Single -0.148* 0.018
Married couple, no kids -0.086* -0.013
Married couple, with kids 0.000 0.000
Single person with kids -0.113* 0.026
Health status
Excellent, very good health 0.000 0.000
Good health -0.037* 0.018*
Fair/Poor health -0.110* 0.066*
Number of chronic conditions
0 chronic conditions 0.000 0.000
1 chronic conditions 0.024 0.069*
2+ chronic conditions 0.017 0.183*
Noncitizen -0.094* -0.066*
Likely to see doctor when sick 0.034* 0.040*
34

Probability of Probability of
private Medicaid/SCHIP
insurance coverage
Identifying variables
State Medicaid program eligibility1 0.022 0.170*
Average premium for family coverage2 -0.00001 -0.00003*
One worker in family 0.117* -0.107*
Two workers in family 0.104* -0.095*
At least one full-time worker 0.182* -0.075*
All workers in family in small firms ( < 25 workers) -0.265* 0.038*
* p < .05

Source: 2003 Community Tracking Study household survey.
1
Estimated proportion of adults in the state eligible for Medicaid, based on a standardized
population.
2
Estimated for counties based on a methodology developed by Dubay and Kenney (2005) using
state-level estimates of premiums from the Medical Expenditure Panel Survey – Insurance
Component.
35

Table 9. Linear probability models for predicting insurance coverage for children (less than age
18).

Probability of
Probability of Medicaid/SCHIP
private insurance coverage
Intercept 0.550* 0.401*
Age
0-5 0.000 0.000
6-12 0.042 -0.026
13-17 0.085* -0.084*
Female -0.024 0.009
Family income relative to poverty
<100% 0.000 0.000
100-199% 0.122* -0.128*
200-299% 0.343* -0.369*
Race,ethnicity
White 0.000 0.000
Black -0.063 0.038
Hispanic -0.159* 0.119*
Other -0.096* 0.083*
Education
Less than high school -0.241* 0.133*
High school grad -0.177* 0.115*
Some college -0.086* 0.065*
College grad 0.000 0.000
Single person with kids -0.101* 0.119*
Health status
Excellent, very good health 0.000 0.000
Good health -0.058* 0.017
Fair/Poor health -0.058 0.068
Noncitizen -0.172* -0.282*
Likely to see doctor when sick 0.033 -0.018

Identifying variables
State income eligibility levels for Medicaid/SCHIP1 -0.00003 0.00008
Average premium for family coverage2 -0.00003 0.00003
One worker in family 0.021 0.003
Two workers in family 0.026 0.004
At least one fulltime worker in family 0.196* -0.185*
All workers in small firms (less than 25 workers) -0.210* 0.149*
* p < .05

Source: 2003 Community Tracking Study household survey.
36

1
State-level income eligibility thresholds for Medicaid/SCHIP 2003, by age.
2
Estimated for counties based on a methodology developed by Dubay and Kenney (2005) using
state-level estimates of premiums from the Medical Expenditure Panel Survey – Insurance
Component.
37

Table 10. Comparison of the effects of insurance coverage on use based on OLS and
instrumental variable analysis.

OLS estimates IV estimates Difference between
OLS and IV estimate
is statistically
significant
(p < .05)

Probability of ED use -- Adults
Medicaid/SCHIP 0.124 (0.02) *** 0.294 (0.126) ** No
Privately insured 0.028 (0.02) 0.171 (0.067) ** Yes

Probability of ED use --
Children
Medicaid/SCHIP 0.001 (0.04) 0.682 (1.21) No
Privately insured -0.02 (0.038) 0.481 (0.925) No

Probability of physician use -- No
Adults
Medicaid/SCHIP 0.29 (0.02) *** 0.30 (0.157) No
Privately insured 0.234 (0.026)*** 0.092 (0.087) No

Probability of physician use --
Children
Medicaid/SCHIP 0.258 (0.052) *** 0.171 (1.23) No
Privately insured 0.237 (0.053) *** 0.131 (0.97) No
***Difference with uninsured is statistically significant at .01 level
**Difference with uninsured is statistically significant at .05 level.
*Difference with uninsured is statistically significant at .10 level.