Immigrants and Medicaid Enrollment: The Effect of Language Networks

Transcription

1 Immigrants and Medicaid Enrollment: The Effect of Language Networks Emily R. Gee Boston University G. Osea Giuntella Boston University February 9, 2011 Abstract Immigrants account for 22% of the uninsured population in the United States. While Medicaid and other public health insurance programs are available to lowincome individuals, language barriers often limit immigrants participation in or knowledge of public assistance. Using the 2009 American Community Survey, we show that linguistic networks facilitate Medicaid enrollment among non-englishspeaking adults. Our identification strategy employs local variation in the concentration of immigrant enclaves and in Medicaid enrollment rates. Using differencein-difference estimation as well as an instrumental variables approach, we find network effects in Medicaid take-up. In particular, among potentially eligible immigrants without private insurance, the probability of having Medicaid coverage is greater for individuals who live in an enclave and whose language group is characterized by a high Medicaid enrollment rate. Given a hypothetical policy to increase Medicaid enrollment, for every 1 percentage point of direct effect, networks would boost take-up by an additional 0.2 percentage points. Our results suggest that information attained via language and ethnic networks affects health insurance status. JEL Codes: F22, J15, R23. Department of Economics, Boston University, 270 Bay State Road, Boston, MA 02215, USA. egee@bu.edu. Department of Economics, Boston University, 270 Bay State Road, Boston, MA 02215, USA. osea@bu.edu. 1

2 1 Introduction Uninsured immigrants present a challenge in U.S. efforts to expand health insurance coverage. Compared with natives, immigrants are more likely to be uninsured and less likely to have regular sources of health care. Take-up rates among Medicaideligible immigrants are particularly low (KFF, 2008; Long et al., 2010). Previous studies have shown that language difficulties and distrust of authorities present barriers to immigrant enrollment in welfare programs (Ponce et al., 2006; Watson, 2010; Lavarreda et al., 2006), but that ethnic networks can help immigrants overcome these challenges by acting as conduits for information, encouraging participation in public programs (Bertrand et al., 2000) and increasing job opportunities (Munshi, 2004). In this paper, we analyze the influence of immigrant networks on participation in the Medicaid health insurance program. In the years since the welfare reform act of 1996, which tightened eligibility restrictions for various public assistance programs, the number of legal immigrants on Medicaid has declined and the gap in health insurance coverage between immigrants and natives has widened (Fremstad and Cox, 2004). The reform also led to greater variation in state-level eligibility rules, thereby increasing disparities among immigrants between states. Surveys of immigrants show that fear of authorities, communication problems, and confusion about health insurance eligibility all hinder access to the health care system. Immigrants are uninsured at a higher rate compared with native citizens. Non-citizens account for about a tenth of the U.S. population but 22% of the uninsured (KFF, 2008). They are more likely to lack employer-sponsored health insurance while also less likely to have public coverage. English proficiency is one of the main barriers to Medicaid enrollment. Brach and Chevarley (2008) use data from the Medical Expenditure Panel Survey to show that among non-elderly adults, Hispanics with low-english proficiency, or LEP, are more likely to be uninsured (60%) than Hispanics with a better grasp of the language (29%). The authors also find that Hispanics of LEP are less likely to use ambulatory services, prescription drugs and dental care. Parental language ability affects children s participation in Medicaid and the Children s Health Insurance Program, known as CHIP (Feinberg et al., 2002). Insurance issues aside, non-native speakers may be confused about how to find a suitable health care provider or be hesitant to approach one. Ponce et al. (2006), using the California Health Interview Survey, show that English proficiency explains disparities in health and health care. Many studies have shown how enclaves are associated with health status itself. Gresenz et al. (2009) show that Mexican-Americans who live in areas with higher concentrations of Spanishspeakers have better access to health care. The authors conclude this is likely due to communication problems, since the effect of living near other Spanish-speakers is of greater magnitude for recently arrived immigrants than more established ones. 2

3 Past Medicaid expansions have had only limited effects on covering the uninsured. Cutler and Gruber (2004) found a 24% take-up rate among children and 7% among pregnant women during the expansion. Lack of awareness about Medicaid is a chief reason why eligible people are not enrolled. Card and Shore-Sheppard (2004) also find low rates for take-up among uninsured low-income children, estimating 8% take-up for the 1990 expansion and none to 5% for Lack of awareness can inhibit the newly eligible from participating in Medicaid. A survey of parents of low-income uninsured children found that knowledge gaps were common reasons for not enrolling (Haley and Kenney, 2001). Among parents who did not inquire about Medicaid or CHIP, about one-third answered they did not know enough about the program believed their child was not eligible, and 40% said they did not want to enroll. Beyond legal and information barriers to Medicare enrollment, immigrants may also hesitate to sign up because they distrust government authorities or fear deportation. Previous literature found evidence that participation in public programs might be influenced by policy environment at the state (Borjas, 2003) as well as at the national level (Kaushal and Kaestner, 2005). Watson (2010) describes chilling effects amplify the effects of immigration policy on Medicaid enrollment across states. She examines the period around the 1996 welfare reform act. Where these welfare policy changes, which directly placed greater restrictions on eligibility for Medicaid, occurred contemporaneously with harsher enforcement of immigration laws, they also had an indirect effect, causing non-citizen participation in Medicaid decreased even further. Such cultural and language barriers are less problematic for immigrants living in ethnic enclaves or among those who can provide them with information about the Medicaid program, its application process, and the relationship between the immigration process and public assistance programs. There is a rich literature on the influence of ethnic groups and cultural networks on individual preferences and economic behavior (Borjas, 1995; Bertrand et al., 2000; Munshi, 2004; Andersson et al., 2010). In the context of social science, information and norms are believed to be the two main channels through which networks influence individual behavior. The network effect for health insurance take-up and similarly for welfare participation, as described by Bertrand et al. (2000) would operate through linguistic links as well as cultural attitudes toward public assistance: if non-native speakers belong to a dense network, they may be more likely to know that Medicaid exists, to figure out the eligibility criteria, and to enroll in it. We posit that even when immigrants are eligible for the Medicaid program, linguistic and cultural barriers may prevent them from actually enrolling, but membership in strong social networks may help immigrants surmount these barriers. Previous network effects literature has emphasized that both social pressure and information are the channels through which groups influence individual behavior. In practice, we are not able to pin down the mechanism for broader health insurance take-up, our empirical results suggest that information is a principal channel 3

4 through which networks influence Medicaid enrollment. Our method is to look for the differential effect across networks. We follow the approach of Bertrand et al. (2000), who studied network effects in participation in the welfare system. The authors used Census microdata to calculate the proportion of people receiving income from public assistance in various immigrant communities. They found that receipt of welfare income was more common among immigrants who lived in enclaves and who were of an ethnicity characterized by a relatively high welfare participation rate. In this paper, we exploit variation in the concentration of immigrant communities and in the average Medicaid enrollment rate of different ethnic groups. We found evidence that among potentially eligible immigrants who lack private insurance, the probability of Medicaid coverage is greater for individuals living in a linguistic enclave and whose language group is characterized by a high Medicaid enrollment rate. The structure of the paper is as follows. In Section 2, we describe the data. We explain our model in Section 3 and present the main results in Section 4. In Section 5, we conduct several robustness checks and some sensitivity analysis. Section 6 concludes. 2 Data We use data from the 2009 American Community Survey (ACS), obtained the data via the Integrated Public Use Microdata Series (Ruggles et al., 2010). The ACS is a random sampling of 3 million households conducted annually by the U.S. Bureau of the Census. With individual-level information on all members of each surveyed household, it is appropriate for our purposes because it contains information on location, language, and, starting in 2008, health insurance coverage. The ACS asks respondents what language they speak at home, allowing us to detect the presence of linguistic enclaves and test for the effect of these language groups people with a common mother tongue on Medicaid enrollment. Using answers from the questions Does this person speak a language other than English at home? What is this language?, we build two alternate measures of network availability. One is based on Public Use Microdata Areas (PUMA), the smallest geographic unit available to us in the ACS, and the other on the Metropolitan Statistical Areas (MSA). Strictly speaking, the networks we identify are not ethnically defined, but in most cases languageand culture-based groupings do coincide. For example, the Spanish group contains Spaniards as well as Mexicans and Salvadorians. We capture network strength as contact availability (CA), adopting the definition from Bertrand et al. (2000). CA is the logarithm of the local proportion of people speaking a given language, relative to the national share of that language s speakers. 4

5 Formally, the network density is defined as following: 1 ln C jk/a j L k /T (1) where C jk is the number of people in area j speaking language k, A j is the total population of area j and L k /T is the share of population living in the United States and speaking language k. The term in the denominator is a normalization to prevent less common languages from being underweighted, since even at their most dense, small language-groups will never account for more than a tiny proportion of their PUMA or MSA (Bertrand et al., 2000). Our calculation for CA is based on the entire ACS population, since the relevant network for information dissemination is all speakers of a language, not just those who are Medicaid-eligible. 2 For all other parts of our analysis, we pare down the ACS to potentially Medicaideligible immigrants, broadly defined. We focus on first-generation immigrants by excluding anyone born in United States. 3 Since we are interested in linguistic networks, we drop individuals whose primary language spoken at home is English. We then keep only individuals who are non-institutionalized, non-elderly adults (ages 19 to 64), and we must drop those living outside identifiable MSAs due to our identification procedure. Describing the remaining criteria for our sample necessitates a brief discussion of Medicaid eligibility rules. Medicaid was created to assist the low-income, low-asset population. While each state designs and administers its own program, all must comply with the federal minimum levels of provision, based on the federal poverty level. The poverty line is not a single number but a sliding scale based on income and household size, and the eligibility threshold also varies by demographic group. For example, the current federal minimum is 235% for children, 185% for pregnant women, 64% for working parents, 38% for non-working adults, and 75% for the elderly and disabled. The bulk of Medicaid enrollees are women and children, a legacy of the historical link between Medicaid eligibility rules and those of Aid to Families with Dependent Children. The elderly and disabled account for 70% of Medicare expenditures (KFF, 2010). The welfare reform act of 1996 bars most immigrants from receiving federally supported Medicaid until they have resided in the United States for 5 years. Any state that 1 Our results are not sensitive to the exact specification of this measure. In Table 13 we provide robustness checks using different measures of concentration, including a non-log measure and a nonnormalized measure. 2 Summary statistics on contact availability measures are presented in the lower panel of Table 1. These measures are consistent with other findings in the literature (Bertrand et al., 2000). 3 We tested our model on second-generation immigrants and found no significant network effect. This result, discussed later in the paper, is consistent with our prior that linguistic and ethnic networks should have less influence on the behavior of U.S.-born individuals. 5

6 wishes to cover immigrants during that period must do so with its own dollars. States are allowed use federal money to provide Medicaid to legal immigrants who have been in the U.S. for at least 5 years. Eight states do not provide coverage to immigrants even after the 5-year bar: Alabama, Idaho, Indiana, Mississippi, Minnesota, Ohio, Texas, Virginia and Wyoming (Chin et al., 2002). 4 We retain only individuals who have lived in the United States for 5 years or more. Unfortunately, we do not have the information needed to determine eligibility for Medicaid program with precision. In order to focus on individuals who potentially could choose Medicaid as their primary form of insurance, we further restrict the sample to those who report having either Medicaid or no insurance. 5 Finally, we restrict the analysis to language groups with more than 50 observations after imposing all other restrictions and exclude individuals with missing wages. 6 Our final sample comprises 36 language groups, 200 MSAs, 1360 PUMAs and 45,437 individuals. Within the sample, 15% are covered by Medicaid. We report descriptive statistics for our main variables of interest and our controls in Table 1 in the appendix. There is substantial variation in insurance status across language groups. In Table 2 we show at the population not covered by private insurance. The average Medicaid enrollment ranges from less than 10% for German, Japanese, Korean, Romanian and Thai to over 30% for Arabic, Bengali, Cushite, Cantonese, Cushite and Miao. Spanish speakers account for more than three-quarters of our final sample (34,310 people), and are far more numerous than next-largest language groups in our sample: Vietnamese (1,271), Chinese (1,096) 7 and Korean (997). Table 3 shows for each language the proportion of people covered by private insurance, enrolled in Medicaid or uninsured. Though our sample excludes individuals with private insurance, for Table 3 we considered all immigrants regardless of insurance status who live in identifiable MSAs and speak one of the 36 languages included in our sample. Spanish-speaking immigrants are characterized by a high percentage of uninsured individuals (40%). 3 Empirical Specification The network effect in our model is interaction between the density of people that share a language and the culture of these people with respect to Medicaid. Bertrand et al. (2000) suggest a network measure that can be thought of as capturing both the 4 Excluding from the sample residents of these states does not change our results much because these policy differences are picked up by the PUMA- or MSA-fixed effects already in the model. 5 In our robustness checks, we further restrict the sample below various multiples of the FPL. 6 Our results are not sensitive to raising or lowering this minimum size cutoff. 7 We follow the ACS categories and leave Chinese as a category distinct from Cantonese, Formosan and Mandarin. Grouping all these into a single Chinese-language group did not affect our results. 6

7 quantity and the quality of contacts in a given language group. express this as: Formally, we Netw jk = (CA ijk Medicaid k ) (2) CA jk is a measure of network availability (our network quantity measure). Network quality is represented by Medicaid k, the Medicaid enrollment rate of the language group expressed as a deviation from the average Medicaid use in the overall sample. Although the network effect occurs at the local level, we define Medicaid k as average Medicaid enrollment rate nationwide for each language because it captures something about the culture s attitude toward welfare. Taking the rate of area j alone would create potential for omitted variable bias (Bertrand et al., 2000). 8 Our main model specification is a linear probability model to capture the role of language networks on Medicaid enrollment: Medicaid ijk = (CA ijk Medicaid k( i) )α + X i jkβ + CA ijk θ + γ j + δ k + ɛ ijk (3) Again, subscript i is for individuals, j for geographic areas and k for languages. The binary dependent variable, Medicaid ijk, equals 1 if a person is covered by Medicaid. 9 As in Bertrand et al. (2000), we use fixed effects γ j for languages and δ k for geographic units, which allows us to control for potential unobserved differences between localities and between language groups. Controlling for network availability CA ijk we take into account potential selection bias from due to individual-level characteristics correlated with CA jk. This model would be still vulnerable to omitted variable bias if CA ijk *Medicaid k( i) were correlated with unobserved individual characteristics (Bertrand et al., 2000). Living in a low (or high) CA area could be associated with different unobserved individual characteristics depending on the language group (high- versus low-medicaid use). Another concern of peer-effects models is the reflection problem (Manski, 1993). There may be heterogeneous, unobserved shocks to language groups at the local level. While our model does not allow us to distinguish between endogenous effects (influenced by group behavior) and exogenous effects (influenced by group characteristics), the IV specification allow us to detect whether even in the presence of selection, there are network effects among language groups. We partially address the reflection problem by excluding individual i from the 8 The average enrollment rate of language group k in area j could be correlated with unobservable characteristics shared by the individual and the language group living in that area. 9 The ACS asks, Is this person currently covered by any of the following types of health insurance or health coverage plans? We consider the person to be on Medicaid if they selected the multiple-choice answer Medicaid, Medical Assistance, or any kind of government-assistance plan for those with low incomes or a disability. 7

8 computation of Medicaid k, 10 although as Angrist and Pischke (2008) points out, the presence of random group effects in the error term could affect statistical inference. 11 In our case, a shock common among members in language group k living in area j in the model above could create spurious correlation and bias our estimates. Therefore, to identify the model in the presence of such shocks and potential differential selection, we use instrumental variables. We instrument each PUMA-level network measure CA ijk with the CA of the sum of all other PUMAs in the MSA. In this way we take into account of commons shocks affecting only the PUMA j analyzed in a given MSA. 12 Our IV identification rests on the assumption that selection at the PUMA level is greater than selection at the MSA level. We do not believe that immigrants choose to settle based on the availability of Medicaid for a few reasons. First, due to federal policy immigrants are generally ineligible for the first 5 years in the United States. For newly arrived immigrants, the utility of possibility of qualifying for health insurance 5 years down the road is unlikely to weigh heavily in their choice of location. 4 Results 4.1 Difference-in-differences We begin our analysis with difference-in-difference estimation. We split the sample along two dimensions. First, we distinguish individuals belonging to a language groups with high Medicaid enrollment (i.e., those above the median group) from those belonging to a language group with low enrollment on a nationwide basis. Second, we divide individuals into high- and low-ca categories at the PUMA level, 13 resulting in four categories. Speakers of Vietnamese, for example, fall into the low-medicaid category. Those who live in Atlanta are in a low-ca area, while the concentration in Phoenix is high. Filipinos in San Diego are low-medicaid, high-ca; German-speakers in Atlanta are low-medicaid and low-ca. Spanish-speaking enclaves dominate the low-ca, high-medicaid category. We can control for variation between language groups by comparing enrollment across low- and high-ca areas. Taking differences between individuals of the same linguistic group but different geographic areas approximates the technique of using language fixed effects in ordinary least-squares Bertrand et al. (2000), and the difference- 10 In an abuse of notation, from here on we will call the mean Medicaid rate simply Medicaid k rather than Medicaid k( i). 11 Angrist suggests using a measure of ex ante peer characteristics to address common shocks. Later in our robustness section, we follow this approach and use past Medicaid enrollment as a proxy for current rates. 12 Bertrand et al. (2000) use as an IV an MSA-level CA that includes all PUMAs within the MSA. However, when using this instrument results were similar. 13 We obtained similar results when calculating CA at the MSA level. 8

9 in-difference on Medicaid enrollment is analogous to our main model s interaction term, CA jk Medicaid k. The difference-in-difference calculation supports the idea that CA has a positive and significant impact. These results, shown in Table 4, show that linguistic enclaves raise Medicaid enrollment more for groups characterized by a relatively high rate of enrollment in the program. The mean for low-ca, low-medicaid language groups is 12% at the PUMA level. For high-medicaid groups, the difference between high-ca areas and low-ca areas in the probability of having Medicaid 7.5%. When we control for standard socio-demographic variables, plus PUMA and language fixed effects, as in our main specification, the difference-in-difference estimate is slightly smaller (0.058, with standard error 0.02) but still positive and significant. 5 Main Results The main parameter of interest in our linear probability model is α, the coefficient on the interaction of CA jk with Medicaid k. We think of this term as capturing a multiplier effect that happens when an enclave is more dense and its members come from a culture of relatively widespread Mediciad use. We show estimates from the baseline model and its variants in Table 5. Standard errors are clustered at the PUMA or MSA level appropriate for the specification. 14 Our first step was to look for network effects at the PUMA level in an ordinary least squares regression, the results of which are shown in Column 1. The network effect has a coefficient of and is significant at the 1 percent level. We included in the model standard socioeconomic and demographic controls, language group fixed effects and geographic fixed effects. Specifically, our controls include age and its quadratic, the logarithm of the sum of weekly income of all family members in the household, number of years since arrival in United States, and dummies for education, marital status, single-motherhood, the presence of a child in the household, a the number of children ever born and race. As we expected, variables associated with lower socioeconomic status increased the likelihood of Medicaid coverage. High-school drop-outs are more likely to be on Medicaid, as are single mothers. Being white or of higher income is associated with lower probability of being on Medicaid, while having more kids and being divorced or never married are associated with a higher probability of enrollment. Following the methods from Bertrand et al. (2000) and Evans et al. (1992), we 14 All results were generated with Stata Statistical Software: Release 11 and with the ACS person weights using the aweight option. We also tried dual-level clustering by PUMA and by language, as well as by PUMA and by household. Our results were similar to the baseline and are available upon request. 9

10 employed for an IV approach both local levels of geographic designation available to us: the smaller PUMA and larger MSA. We instrumented the PUMA-level contact availability and the interaction term at the PUMA level with corresponding variables based on CA of all other PUMAs within the MSA. 15 Under the alternative hypothesis of no network effects, the OLS estimate is biased from selection at both the MSA and the PUMA level. Any bias in the coefficient from the IV results, presented in Column 2 of Table 5, would come from MSA-level selection. Instrumenting with MSA-level contact availability generates a network coefficient of and does not change the direction of the coefficients on the controls. The increase in the estimated size of the network effect anything it suggests OLS underestimates the network effect on Medicaid enrollment. The IV results support the hypothesis that living in an area with concentrations of high-enrollment groups raises the probability Medicaid coverage for immigrants without private insurance. Consistent with the literature, we adopt as our baseline specification the OLS model; this is the most conservative estimate of network effect and can be considered lower bound on the true effect (Bertrand et al., 2000). Although we can easily compare models based on the relative magnitude of the coefficient, α, on the network effect, its meaning of α itself is not straightforward to interpret. Therefore, we repeat the thought experiment presented in Bertrand et al. (2000). Imagine a policy shock that directly increases Medicaid participation by some amount λ. Mathematically, this is expressed in the model as (Bertrand et al., 2000): Medicaid ijk = λ + CA jk Medicaid k α + X i jkβ + CA jk θ + γ j + δ k + ɛ ijk. (4) In this formulation, a 1 percentage point increase in λ would increase Medicaid enrollment by 1 percentage point in the absence of network effects. However, an increase in λ also generates an indirect effect through the increase in Medicaid k. After taking the average with respect to k on both sides we can solve for the overall increase in Medicaid participation and Medicaid k and identify the network effect by differentiating and evaluating the expression using the weighted mean of CA over all language groups. Formally: dmedicaid k = 1 + αca jk dmedicaid k (5) dλ dλ Policy change boosting Medicaid enrollment by 1 percentage point in the absence of networks would actually increase enrollment for people belonging to language group k percentage points. by 1 1 αca k 15 With the IV used by Bertrand et al. (2000), who use the CA at the MSA level without excluding the relevant PUMA, the estimate is slightly smaller (0.22, with standard error 0.05). We also tried instrumenting the 2009 CA jk for each PUMA with past CA measures based on from the 1990 and 2000 Censuses and obtained similar results (Card, 2001, see). Results are available upon request. 10

11 Plugging in the results of our baseline OLS specification, we find that linguistic networks boost the impact of such a policy by 25 percent. By contrast, this multiplier effect is 31 percent when we calculate the contact availability at the MSA level, as shown in Column 3 of the table. To gain some perspective for the size of the coefficient, we estimate two unsophisticated network effects specifications that naively attribute all group-level differences in the dependent variable to network effects (Bertrand et al., 2000). In Column 4 we look at the effect of mean Medicaid enrollment use of the language group. In Column 5 we look at the Medicaid enrollment rates at the PUMA level. In these specifications we do not control for language and area fixed effects. The coefficients are much larger, at and 0.691, since these specifications by design suffer from omitted variables problems. 6 Robustness Checks In this section, we discuss the sensitivity of our results to other specifications. We tested other batteries of control variables, strengthened our criteria for our Medicaid-eligibility dummy, and ran the model on subsets of our data based on sex and language. Our results seem robust to these tweaks of our model. Our baseline specification uses a linear probability model, and we obtained similar results using probit estimation (see Table 13, Column 3). Our results are also robust to different definitions of contact availability. For example, instead of using CA as our measure for people speaking language k in area j, we replaced it with simple logarithm (ln(c jk )) and with the logarithm of the local share that speaking language (ln( C jk A j )) without normalizing by the national level (see Table 13, columns 1 and 2). In Table 6 we test the sensitivity of our estimates to different sets of control variables. Adding and removing these exogenous control variables (e.g., age, gender, race) as well as variables that are more likely endogenous (marital status or family size) did not substantially affect the coefficient. 6.1 Restricting the Sample by Age, Sex, and Language We then investigated how the size of the network effect varied across different samples. Since more than three-quarters of our sample were Spanish speakers, we check whether our results were driven by this group. Only 48% of Spanish speakers in the ACS have private insurance, a proportion much lower than that of the other language groups, compared with 79% among the English speakers and 74% in the overall sample. To check if our results held in the absence of the Spanish-speaking population, we ran the same set of regressions only non-spanish language groups. While the coefficient is significantly smaller for the non-spanish groups, the network effect was still positive 11

12 and significant at the 10% level. 16 We were also concerned that our measure of high-ca areas was due to not by a relatively high density of households but by the size of those households. We addressed this concern with specifications using households, rather than individuals, as the unit of analysis. To ensure that each household was represented only once in the data, we eliminated all men and retained only one woman per household. Since Medicaid rules are more generous toward women of childbearing age, so we also tried isolating various age groups. When we brought the upper bound on age, originally set at 64, down to 55 or 45, the effect was still strong and significant. However, when we limited the sample to the group, the sample size shrank and the network effect became weaker and lost statistical significance. We also found the network effect was smaller among women than among men, although though not statistically different (Table 7, panel A). One criticism of our experiment might be that the sample is overly broad. Our potentially Medicaid-eligible population includes all immigrants who have been in the United States more than 5 years and are not covered by private insurance. Ideally, we use in our sample to only those actually eligible for the program, but we cannot observe all the conditions for Medicaid eligibility in the ACS data. While we do not have the information needed to exactly determine eligibility for Medicaid, we can further restrict our sample to individuals with low family income (see Table 7, Panel B). 17 Network effects appeared stronger and still significant when we focused on individuals below the 300%, 200% and 150% of the FPL. Restricting the sample to people with a family income below the 100% of FPL, which reduced the sample size, made the coefficient lose statistical significance by increasing the variance of the point estimate. While the main criterion for Medicaid eligibility is income, eligibility also depends on sex, age, disability status, pregnancy, and assets. Further complicating matters, most states provide Medicaid or Medicaid/SCHIP coverage beyond the federal minimum. We attempted an exact imputation of eligibility, but the results were inconsistent with the survey data. We used the 2009 state-level Medicaid criteria summarized in Donna and Horn (2008) and compared those with ACS respondents information on state of residence, age, family size, income, and sex. The imputed eligibility differed greatly from reported Medicaid participation; almost half the individuals who reported Medicaid coverage were deemed not eligible by our imputation. Despite this inconsistency, the ACS data do, however, accurately describe other aspects of the Medicaid population. For example, we can use the ACS to confirm that enrollment rates in Medicaid are lower among immigrants. In Massachussetts, which has a relatively liberal eligibility regime, we found that 38% of potentially eligible immigrants in the We lost power in the estimation since eliminating such a large group drastically reduced the sample size. 17 In our sample, 50% of individuals covered by Medicaid were above the 134% of FPL, while 25% were above the 216% of FPL. 12

13 survey reported being uninsured, compared with 28% of natives. 6.2 Citizenship, Language Fluency and Length of Residence Since our main argument is that network effects might reduce information costs related to language barriers, we expected networks to be more influential for people of limited English proficiency or who arrived in the United States more recently. When we restricted our sample to the English-proficient, the network effect was lower than when we included those who are were less proficient. There is no network effect when we focus on second-generation immigrants (see Table 7, Panel C). This is consistent with our prior that networks affect insurance behavior by reducing language barriers and facilitating access to information, so those with the weakest English-language skills would benefit most from linguistic networks. At the same time, linguistic network should matter less for people who assimilated in the new culture and language through years. We also tested a specification that took into account the overall English proficiency of the whole language group. We changed the interaction term to include the length of residency in United States, a dummy for English fluency (equal to 1 for those who report speaking English well or very well ; 0 otherwise) and average skills of the language group (Table 8). The network effects appear weaker for both more recent cohorts of immigrants and for fluent speakers. The average English fluency of the network also reduces the network s effect. These results are consistent with the idea that the information advantage implied by network availability becomes less relevant for people who have been immigraned earlier or have less difficulty understanding English. 6.3 Medicaid Information in Other Languages An alternative explanation for the phenomenon that we ascribe to language networks would be that government authorities respond to large populations of language groups with outreach campaigns. A simple and relatively inexpensive policy solution to narrowing the take-up gap between immigrants and natives is to create information campaigns in languages besides English. States could increase awareness of Medicaid by providing radio advertisements, billboards, posters, or pamphlets in the languages of their local immigrant communities. Imagine a state that decides it will provide Medicaid information in a foreign language only if some critical mass of immigrants either resides there or is potentially eligible for Medicaid. As a result, it would appear in the data that immigrants in high-concentration states would be more likely to have Medicaid, even though the mechanism was not information-sharing but a bureaucratic channel (Bertrand et al., 2000). We do not, however, think this story about bureaucracy describes our findings. Our model uses PUMA fixed-effects, which will pick up any state-level differences in 13

14 policy. Moreover, not all immigrant waves have historically been drawn to the same cities or states. We tested for the effect of bureaucratic response as distinct from the language network effect. While we had no way of measuring the volume of state-provided information, we did want to capture the availability of Medicaid information to non-english speakers. In some state Medicaid agencies provide on their websites enrollment information and application forms in languages other than English. We realize that websites with English-language portals are likely not the primary source of information for LEP immigrants, our rationale is that the availability of Internet-based materials is correlated with the overall availability of foreign-language materials. Given that a state has already created a form or pamphlet in, for example, Vietnamese, the marginal cost of placing that information online is negligible. We visited each state s Medicaid website to check whether it included either an application form or an informational page in a language other than English (see Table 11). The vast majority of states did not provide any information in a language other than English or Spanish. A total of 30 state provide some sort of Spanish-language information, while 9 states had material in Vietnamese, the next most common language. Multilingual outliers were Washington state (12 languages besides English), California (10) and Kansas (9). In Table 10, we modify our baseline specification to include a control for the effect of having information in one s own language. We found a positive and significant coefficient for the effect of Internet-based information; the network effect coefficient was unchanged. We then interacted the our network measure with the own-language information dummy and found a positive and weakly significant effect. In general, immigrants whose state provides Internet resources in their native tongue are more likely to enroll in Medicaid. When controlling for web information, the impact of the network was stronger for individuals who are not fluent in English, though not significantly different for those who had no information in their own language. While the results of this website survey may not be sufficient to disentangle the mechanism underlying the network effect, they do suggest that publishing information in the languages of the immigrant community might have positive effects on take-up both directly and through the network s multiplier effect. 6.4 Contact Availability Based on Country of Origin While a common language makes communication easier, immigrants may associate with people who speak their language because they also share cultural norms. For example, perhaps what matters for Senegalese is not the local presence of other francophones but of others from Senegal. In the case of Medicaid, the relevant cultural norms would be attitudes toward the welfare state and stigma about government handouts. Immigrants 14

15 from some cultures may also be inclined to turn to friends or family networks rather than the state in times of need. We cannot definitively determine the relative importance of language-network effects vis-a-vis cultural networks, but we can test a model with another type of network. We used data from the same 2009 ACS to create ethnic networks that are instead defined by country of origin. We exclude from the ACS all immigrants with non-missing data on country of origin whose native countries are represented by fewer than 50 observations in our final sample and those from countries where English is spoken by a majority of the population. 18 Our econometric specification remains the same, except we define contact availability using country of origin instead of language spoken at home, and we replace language fixed effects with country-of-origin fixed effects. 19 With these changes, we still find evidence of significant ethnic network effects. As shown in Table 10, the network effects are smaller, but still positive and significant. A policy change boosting Medicaid enrollment by 1 percentage point in the absence of networks would actually increase enrollment for people belonging to the ethnic group by around 1.17 percentage points. As we noted earlier, an alternative explanation for the network effect might be bureaucratic response for example, an enclave of a low-income language group might prompt state administrators to hire Medicaid staff who speak that language. However, if this was the case, we should not see any effect when we consider network effects based on country of origin within a given language group. In Column 2, we show that network effects remain significant when we restrict the analysis to Spanish speakers. This result supports the idea that our main result is driven by some quality of the networks themselves, rather than by bureaucratic response. Using country-of-origin networks also allows us to use past data on average enrollment in Medicaid as an instrument for current Medicaid enrollment. The ACS only recently began collecting health insurance data, so we use the 1994 CPS to calculate a Medicaid k1994 for each country-based network k. While past Medicaid enrollment does capture information about cultural attitudes toward public assistance, it should be less correlated with any random shocks affecting group and individual behavior in Since our baseline model used language-based networks, we cannot directly compare them to the country-of-origin results, but we do find that the network effect with the past-medicaid proxy is positive and significant both among all immigrants and within 18 We take our list of majority-english countries from the United Kingdom Border Agency s regulations for non-european Union migrants exempt from language examinations, and add to it the U.K. and Ireland. The list is available at 19 Our results are essentially unchanged if we use language instead of country-of-origin fixed effects. 20 We choose 1994 because it is before the 1996 welfare reform, which changed the eligibility criteria for Medicaid. Immigrant enrollment rates substantially declined after the reform. 15

16 the Spanish-speaking population. 7 Conclusion We find strong evidence that immigrant networks boost take-up of Medicaid health insurance. Although we are mainly concerned with language-based networks, which overlap with ethnic networks, the fact that the network effect is decreasing in language fluency suggests that information dissemination is the mechanism for the network effect. We cannot pin down the exact contribution of shared language versus other sorts of social capital, but we do conduct several robustness checks that support that language networks are most influential for the limited-english-proficient and the more recently arrived. The network effect is significant even when controlling for state-provided foreign-language information. Our results support the theory that information attained through ethnic/language networks affect access to health care and the magnitude of the effect is consistent with previous literature on network effects and public program take-up. While the network effect works in favor of those living in enclaves, variation in network density also represents a further source of disparities in insurance coverage among immigrant groups. As more immigrants move into states that do not provide coverage, these disparities are expected to worsen. States efforts to provide multilingual Medicaid information is associated with higher take-up rates. We believe further efforts to reduce language barriers, clarify eligibility criteria, and simplify the application process are needed. While a policy focus on large enclaves seems intuitive, our results suggest that the multiplier effect of immigrant groups means that persons of small, relatively dispersed ethnicities may be at greatest risk for being Medicaid-eligible yet uninsured. References Andersson, Fredrik, Monica Garcia-Perez, John C. Haltiwanger, Kristin McCue, and Seth Sanders (2010) Workplace concentration of immigrants. NBER Working Paper Angrist, Joshua D., and Jorn-Steffen Pischke (2008) Mostly Harmless Econometrics: An empiricist s companion (Princeton, NJ: Princeton University Press) Bertrand, Marianne, Erzo F. P. Luttmer, and Sendhil Mullainathan (2000) Network effects and welfare cultures. The Quarterly Journal of Economics 115(3), Borjas, George (1995) Ethnicity, neighborhoods and human capital externalities. American Economic Review 85(3),

17 (2003) Welfare reform, labor supply, and health insurance in the immigrant population. Journal of Health Economics 22(3), Brach, Cindy, and Frances Chevarley (2008) Demographics and health care access and utilization of limited-english-proficient and english-proficient Hispanics. Research Findings, Agency for Healthcare Research and Quality Card, David (2001) Immigrant inflows, native outflows, and the local labor market impacts of higher immigration. Journal of Labor Economics 19(1), Card, David, and Lara Shore-Sheppard (2004) Using discontinuous eligibility rules to identify the effects of the federal Medicaid expansions on low-income children. Review of Economics and Statistics 86(3), Chin, Kimberley, Stacy Dean, and Kathy Patchan (2002) How have states responded to the eligibility restrictions on legal immigrants in Medicaid and SCHIP? Center on Budget and Policy Priorities for the Kaiser Commission on Medicaid and the Uninsured Cutler, David M., and Jonathan Gruber (2004) Does public insurance crowd out private insurance? The Quarterly Journal of Economics 111(2), Donna, Cohen Ross, and Aleya Horn (2008) Health coverage for children and families in Medicaid and SCHIP: State efforts face new hurdles. Kaiser Commission on Medicaid and the Uninsured Evans, William N., Wallace E. Oates, and Robert M. Schwab (1992) Measuring peer group effects: A study of teenage behavior. Journal of Political Economy 119(C), Feinberg, Emily, Katherine Swartz, Alan M. Zaslavsky, Jane Gardner, and Deborah Klein Walker (2002) Language proficiency and the enrollment of Medicaideligible children in publicly funded health insurance programs. Maternal and Child Health Journal 6(1), 5 18 Fremstad, Shawn, and Laura Cox (2004) Covering new Americans: A review of federal and state policies related to immigrants eligibility and access to publicly funded health insurance. Kaiser Commission on Medicaid and the Uninsured Gresenz, Carole Roan, Jeanette Rogowski, and Jose J. Escarce (2009) Community demographics and access to health care among U.S. Hispanics. HSR: Health Services Research 44(5),

18 Haley, Jennifer M., and Genevieve M. Kenney (2001) Why aren t more uninsured children enrolled in Medicaid or SCHIP? New Federalism: National Survey of America s Families Kaushal, Neeraj, and Robert Kaestner (2005) Immigrant and native responses to welfare reform. Journal of Population Economics 18(1), KFF, Kaiser Family Foundation (2008), Summary: Five basic facts on immigrants and their health care. Kaiser Commission on Medicaid and the Uninsured KFF, Kaiser Family Foundation (2010), Medicaid: A Primer. Kaiser Commission on Medicaid and the Uninsured Lavarreda, Shana Alex, E. Richard Brown, Jean Yoon, and Sungching Glenn (2006) More than half of California s uninsured children eligible for public programs but not enrolled. Mimeo, University of California Long, Sharon K., Lokendra Phadera, and Victoria Lynch (2010) Massachusetts Health Reform in 2008: Who are the Remaining Uninsured Adults? State Health Access Reform Evaluation, Robert Wood Johnson Foundation Manski, Charles F. (1993) Identification of endogenous social effects. Review of Economic Studies 60, Munshi, Kaivan (2004) Networks in the modern economy: Mexican migrants in the U.S. labor market. Quarterly Journal of Economics 118(2), Ponce, Ninez A., Leighton Ku, William E. Cunningham, and E. Richard Brown (2006) Language barriers to health care access among Medicare beneficiaries. Inquiry 43(1), Ruggles, Steven, J. Trent Alexander, Katie Genadek, Ronald Goeken, Matthew B. Schroeder, and Matthew Sobek (2010) Integrated Public Use Microdata Series: Version 5.0 [machine-readable database]. Minneapolis: University of Minnesota Watson, Tara (2010) Inside the refrigerator: Immigration enforcement and chilling effects in Medicaid participation National Bureau of Economic Research working paper 18

19 Table 1: Summary Statistics Variable Mean Std.Dev. Min. Max. enrolled in Medicaid male age age 2 / high school dropout high school graduate some college college and more married (present) married (absent) widowed divorced separated never married number of children in HH child present single mother log(family wage) white black years in U.S English fluency MSA CA PUMA CA log MSA CA log PUMA CA Notes: Data Source: 2009 American Community Survey. For the top portion of the table, the sample is composed of non-english speaking immigrants ages 19 to 64 who have been in U.S. for at least 5 years, have valid wage information, are not covered by private insurance and who belong to a language group with more than 50 observations remaining after imposing all other criteria. Individuals living in non-identified MSAs and the institutional population are excluded from the sample. Family wage is the sum of all wage income for family members in the household. In the lower portion of the table, the contact availability measures are calculated as described in Section 2. Childpresent is a dummy that equals 1 if an individual has any children. F luent equals 1 for individuals who speak English well or very well and 0 for those who speak it not well or not at all. 19