America’s Uninsured Children A Failure of State Policy Part 2

4. DATA and EMPIRICAL STRATEGY 4.1: Data The primary dataset used for this paper was the Annual Social and Economic (ASEC) Supplement to the US Census Bureau’s Current Population Survey (CPS) The CPS provides information on population and employment in the United States, and the ASEC Supplement (formerly known as the March Supplement) contains added information on variables such as the individual’s source of health insurance coverage, age, state of residence, family income, and participation in other government programs. The survey samples individuals of all different ages, and includes 67,240 children aged 18 or below. All of the individual level regressions in this paper were run using this sample of children. Point-in-time observations (such as location and age) are taken from March 2005. Longterm data observations (for variables such as annual family income, or total time unemployed) are taken over the course of the previous year from March 2004 to March 2005. Included in the data set was a variable that attached a sampling weight to each individual. The individual level regressions were all weighted using this variable in order to generate a nationally representative sample. The US Census Bureau also publishes aggregated state-level data, which is estimated using the CPS survey. This state level data was used for the macro-level regression in Table I. The individual level data set did not provide any unique variable identifying individuals as being either insured or uninsured. However, it contained several dummy variables indicating coverage from various types of insurance (such as Medicaid, employer-based private coverage, and independently purchased private coverage). I labeled individuals who were not covered by any of the insurance types listed as uninsured. The data set also did not have any variable indicating eligibility for public insurance. In order to determine eligibility, I used data from the Kaiser Family Foundation website, which lists the income cutoffs for both Medicaid and SCHIP eligibility in each state. If the family income listed for a given child fell below the cutoff level for either Medicaid or SCHIP in his or her state of residence, I labeled the child eligible for public 15 insurance. On the whole, 28,726 children (approximately 43% of the child population) were determined to be eligible for public insurance. The following table presents summary values of the data. Data Summary Table I II Public Insurance eligible population (n = 28,726) III Eligible Insured population (n = 23,799) IV Eligible Uninsured population (n = 4,927) 66,679 (70,231) 23,270 (15,989) 23,916 (16,061) 20,148 (15,249) 17 39.8 38.7 45.4 38.1 86.6 86 89.3 9.26 (5.37) 8.75 (5.33) 8.54 (5.27) 9.74 (5.51) 2.17 (1.23) 2.37 (1.41) 2.43 (1.38) 2.03 (1.48) 62.6 18.7 12.7 45.4 29.7 19.1 47.6 26.9 19.8 34.9 43 15.9 25.6 48.8 58.9 0 10.9 17.2 0 100 42.6 100 100 100 General Population (n = 67,420) Family Income ($) % Below Poverty Line % Below 200% of Poverty Line Age Number of Children under 18 in Family % White % Hispanic % Black % Covered by Public Insurance % Uninsured % Eligible for Public Insurance Standard error in parentheses for those values not expressed as a percentage. Sample restricted to children under age 18. The above table provides details on the sample surveyed by the data set. Accordingly, sampling weights were not used. Column I of the above table analyzes the entire population of children in the sample. Column II examines the population of children eligible for public insurance. Columns III and IV divide the population of eligible children into those who are covered by some sort of insurance, and those who are not, respectively. Certain expected trends emerge when comparing the general population with the eligible population. For one, the average income of individuals in the eligible population is markedly lower than that of those in the general population, with a far greater proportion of children in the eligible population finding themselves below the poverty line. Minority individuals (of Hispanic or Black background)10 are predictably over-represented in the 10 Children identified as Hispanic and another ethnicity were classified only as Hispanic for the purposes of this paper. Mixed race Black children were classified as Black in every case, except for when Black and Hispanic identity were mixed. In this case, children were classified as Hispanic. 16 eligible population. The rate of uninsurance is also markedly higher in the eligible population, indicating that as noted in the Theory section, access to health insurance is often a problem for poor individuals. A comparison of the eligible insured and eligible uninsured populations reveals additional observations of interest. First, the average age of children in the eligible uninsured population is 1.3 years higher than that of children in the eligible insured population. While these two averages are not statistically distinct from one another, their difference nonetheless suggests that older children are disproportionately represented in the eligible uninsured population. Additionally, the percentage of Hispanics in the eligible uninsured population is markedly higher than in the eligible insured population. Interestingly, the percentage of Black individuals in the eligible uninsured population is smaller than the percentage in the eligible insured one. This indicates that Hispanic individuals in particular have a high probability of uninsurance. Much of the analysis in the following Results section examines this effect, and factors that may underlie it. 4.2: Empirical Strategy For each of the individual level regressions in the following Results section, the dependent variable is a dummy indicating whether or not a given individual is uninsured. The six enrollment procedure variables mentioned previously are the main explanatory variables of interest. I used the previously mentioned Kaiser Family Foundation website to obtain data on which enrollment procedures were used in the various states. I created each enrollment procedure variable as a dummy taking on a value of 1 for those children living in states implementing the given policy. For example, the “continuous eligibility” dummy would take on a value of 1 for a child living in California (which uses continuous eligibility for both Medicaid and SCHIP) and a value of 0 for a child living in Texas (which does not use this policy). In certain states, a given policy was used for Medicaid enrollment but not for SCHIP (or vice versa). In these cases, the dummy variable took on a value of 1 only for states using the policies for both programs, unless otherwise noted. 17 Different enrollment procedure variables are used in different regressions in the following Results section. Given that the dependent variable for all the individual level regressions takes on values of only 0 and 1, a dprobit regression model is used to estimate coefficients on the right hand side variables. For each regression, controls are included for the race or ethnicity, age, and family income of each subject. In addition, a series of state dummies are used to control for state fixed effects. 5. RESULTS and EMPIRICAL ANALYSIS I begin the empirical portion of this project by testing the underlying assumption of this paper, which is that state level variations in policy implementation can influence the access that individuals have to public insurance. Prior to examining the main CPS dataset, I experiment with aggregated state-level data, and estimate the effect of the six previously mentioned enrollment procedures on state wide rates of uninsurance among poor individuals. I then turn to the CPS data and estimate the effect that the enrollment procedures have on the likelihood of a particular individual being uninsured. The latter data allows for more precise estimates of the effects of the six enrollment procedures, as it allows me to control for several individual characteristics (such age, family income, and ethnic background) that could potentially be correlated with the probability of uninsurance. In both cases, dummy variables indicating the use of each of the six previously mentioned enrollment procedures in each state serve as the key explanatory variables. Next, I examine the effects of Black and Hispanic background, and their interactions with age. In particular, I analyze whether the increased uninsurance rates among Hispanic individuals observed in Table I are observed for Hispanic children of all ages. I then proceed to focus on two specific enrollment procedures: continuous eligibility, and use of a face-to-face interview requirement. I study whether these two policies may have different effects on children of different ages and different racial or ethnic background. Finally, I test whether two enrollment procedures – 12-month continuous eligibility and self declaration of income – have different effects on children of different income levels. 18 5.1 State Level Regression: Enrollment Procedures There are six enrollment procedure variables that I test in the state level regression below, each of which can possibly influence access to public insurance. The first indicates whether or not a certain state conducts an asset test on prospective Medicaid enrollees.11 By adding an extra level of complexity to the enrollment procedure for public insurance, states requiring asset tests will likely curtail access to Medicaid among eligible individuals. I thus hypothesize that the “asset test” variable will be associated with increased uninsurance among poor individuals. The “initial face-to-face interview” variable indicates whether a given state requires that prospective applicants complete a face-to-face interview in order to first enroll in Medicaid or SCHIP. The “re-enrollment interview” variable identifies states requiring that enrollees seeking to renew their coverage complete a face-to-face interview with a program representative. By increasing the cost (as far as time and convenience is concerned) of obtaining public insurance, these interviews can make enrollment and re-enrollment more difficult for eligible individuals. On the other hand though, these interviews can act as a highly effective means of disseminating valuable information about enrolling in a given state’s public insurance programs. Accordingly, requiring prospective enrollees or returning enrollees to undergo face-to-face interviews can serve to increase access to public insurance. The overall effect of interview requirements on enrollment thus depends on which of these two effects predominate. The “presumptive eligibility” variable can also conceivably reinforce or hinder public insurance take-up. By making it possible for healthcare providers to determine eligibility, and help with the initial enrollment procedure, this policy ostensibly increases access to public insurance for eligible individuals. However, as I mentioned in the Theory section, it is also possible that the use of this policy will lead individuals to believe that they are conditionally insured, and able to obtain treatment covered by public insurance programs even if they are not 11 Only one state (Texas) employs the use of an asset test for both Medicaid and SCHIP. Thus, for the sake of generating more variation in the “asset test” dummy, I allowed it to take on a value of 1 for those states using an asset test for Medicaid and not SCHIP (Colorado, Montana, Nevada, and Utah) as well. 19 enrolled in advance. This could potentially outweigh the effect of improved public insurance access, and thus increase the probability of eligible uninsurance. I hypothesize that the remaining two variables, which indicate the use of self-declaration of income and continuous eligibility, will both be associated with increased access to public insurance. Both policies are aimed at facilitating public insurance enrollment, and lowering the barriers to obtaining it. Hence, I predict that both will have negative coefficients. The dependent variable in the state level regression measures the percentage of poor or near-poor children (defined by the Census Bureau as those with family incomes below 200% of the FPL) who were uninsured in the year 2004. This variable is not a perfect indicator of the number of eligible uninsured, given that in certain states some individuals within this population are not eligible for public insurance (and in others, the eligible population extends beyond those with family incomes under 200% of the FPL). Nonetheless though, I believe that it is a good proxy, given that the majority of states use the 200% mark as the cutoff for determining public insurance eligibility. The main purpose of the state level regression was to gauge the potential effects of the enrollment procedure variables. Because the dependent variable is a continuous one, a linear regression model estimated using OLS was employed. The results of the regression are shown in Table I. State Level Regression Model % uninsured below 200% FPL = β1*(Self-declaration of income) + β2*(Asset test) + β3*(Presumptive eligibility) + β4*(Continuous eligibility) + β5*(Initial enrollment interview) + β6* (Re-enrollment interview) + ε In the above regression, the coefficient on the “continuous eligibility” variable is the only one that is not of the expected sign. This coefficient, however, is not close to being statistically significant at any level, indicating that this variable does not appear to have any significant correlation with the dependent variable. 20 Table I State Level Enrollment Procedure Regression Self Declaration of Income Asset Test Presumptive Eligibility Continuous Eligibility Face to Face Enrollment Interview Face to Face ReEnrollment Interview -0.03426 (1.71)* 0.08702 (3.05)** 0.03450 (1.20) 0.00323 (0.17) -0.05468 (1.87)* 0.06060 (1.63) N = 51 Adjusted R^2: 0.1874 Cross sectional regression, in which the dependent variable was the percentage of those below 200% of the FPL in a particular state without Health Insurance in 2004. Independent variables are dummies indicating whether the above procedures were used in the given state. Coefficients on the above explanatory variables estimated using linear regression. T-statistics in parentheses. The “presumptive eligibility” and “re-enrollment interview” variables both have positive coefficients, with the coefficient on the latter being nearly significant at the 90% level. It appears, therefore, that having already gone through the enrollment process once, individuals gain little added information from an additional round of interviews, but experience a significant amount of inconvenience from them. The positive coefficient on the presumptive eligibility variable suggests that the phenomenon of conditional insurance caused by this policy outweighs the increased access to public insurance that it creates. This coefficient is not statistically significant, however, making it difficult to conclude this with any certainty. The “asset test” variable also has 21 a coefficient that is positive and statistically significant, indicating that as I had hypothesized, the use of this policy serves to decrease public insurance access. The “initial face-to-face interview” variable has a negative coefficient that is significant at the 90% level. This result indicates that the informational gain from the interview appears to outweigh the inconvenience that the requirement poses. As hypothesized, the “self-declaration of income” variable has a negative coefficient that is significant at the 90% level, indicating that this policy increases public insurance access through simplifying enrollment procedure that a prospective enrollee must complete. The most important observation to take from the above macro-level regressions, however, is the fact that the variations in how the different states implement policies do indeed exert an influence on public insurance take-up among eligible individuals. The statistically significant coefficients on several of those variables, combined with the significant explanatory power of the regression (explaining 18.7% of the variation in uninsurance rates among the poor) strongly suggest that the implementation policies enacted by various states can play a very significant role in analyzing the differing rates of eligible uninsurance across the states. Having established the linkage between the procedures and their impact on enrollment, I now proceed to use individual level data to take a closer look at these enrollment procedures, and specifically analyze their impact on different sectors of the population. 5.2: Individual Level Regression: Enrollment Procedures I next examine whether the findings from the state level regression are applicable to individual level data as well. As was mentioned in the Data section, the dependent variable for each of the following individual level regressions is a dummy variable indicating whether or not a particular child was uninsured. I began by restricting the sample to those children eligible for public insurance in their state of residence, and then individually regress each enrollment procedure dummy on the dependent variable. Next, I regress all six enrollment procedure variables jointly on the uninsurance dummy. In both regressions, I control for the age, racial or 22 ethnic background, family income, and state fixed effects of every individual in the sample. For all of the individual level regressions (which use CPS data from March 2005), data on the enrollment procedures used in the various states is taken from the fiscal year beginning on July 1, 2004, and ending on June 30, 2005. Column I-VI Uninsured = β1i*(State Policy Dummy)i + λ*Demographic controls + γ*State fixed effects + ε Column VII Uninsured = β1*(SD of income) + β2*(No Asset test) + β3*(Presumptive eligibility) + β4*(Continuous eligibility) + β5*(FF Interview) + β6* (FF re-enrollment interview) + λ*Demographic controls + γ*State fixed effects + ε As was the case in the previous regression, I hypothesize that three of the enrollment procedures will have a positive impact on public insurance take-up: the use of self declaration of income, absence of an asset test, and use of continuous eligibility. In the above regressions, in which a binary indicator of uninsurance serves as the dependent variable, I predict that the variables corresponding to these three enrollment procedures will have negative coefficients. As was discussed in the previous section, the three remaining enrollment procedure variables – “face-to-face enrollment interview,” “face-to-face re-enrollment interview,” and “presumptive eligibility” – can potentially work in two different directions. It is thus difficult to definitively predict whether their coefficients will be positive or negative. In the table below, the “self-declaration of income” and “no asset test” variables have negative coefficients in both models. These coefficients are statistically significant at the 90% or 95% level, indicating a strong negative correlation with the probability of uninsurance. These two results mirror those from the state level regression, and are in line with the theory predicting that both variables would be associated with a decreased probability of uninsurance. The “reenrollment interview” variable, meanwhile, has a positive coefficient in both models. Both coefficients are again statistically significant at the 90% or 95% level, indicating that the use of 23 this policy is associated with a significant increase in the probability of uninsurance. This result was observed in the macro-level regression as well, and indicates that the inconvenience and hassle caused by the re-enrollment interview requirement outweighs any informational gain from the actual interview. Table II Individual Level Enrollment Procedure Regressions I II III IV V VI Self Declaration of Income -0.1246 (4.57)** -- -- -- -- -- No Asset Test -- -0.1285 (3.05)** -- -- -- -- -- -- -0.0088 (0.26) -- -- -- -- -- -- -0.0817 (2.54)** -- -- Face to Face Enrollment Interview -- -- -- -- 0.0396 (1.23) -- -0.0663 (2.38)** Face to Face ReEnrollment Interview -- -- -- -- -- 0.0658 (1.84)* 0.1924 (3.51)** Presumptive Eligibility Continuous Eligibility VII -0.0378 (1.78)* -0.2066 (4.91)** 0.0505 (2.16)** 0.0369 (1.62) For each regression, n = 28,726 (*) denotes significance at 90% level, (**) denotes significance at 95% level. Z-statistics in parentheses. Sample restricted to children eligible for public insurance. Each column above denotes a different regression. In each regression, the dependent variable is a dummy taking on a value of 1 if the individual is uninsured, 0 otherwise. Dprobit regression was used to estimate coefficients on above explanatory variables. The above explanatory variables were first regressed individually on this dependent variable (Columns I-VI), then jointly (Column VII).Controls were included for subject's age, race, family income relative to FPL, and state fixed effects in each regression. The other three variables yield somewhat ambiguous results, with coefficients of different sign in the individual and joint regressions. Given that certain enrollment procedure variables are correlated with one another, I believe that the most accurate coefficients come from the model in which each enrollment procedure variable is regressed individually (with controls) against the probability of uninsurance. Indeed, collinearity between the six enrollment procedure variables makes it difficult to isolate the specific effect of one variable when all six are run jointly. Given this reasoning, the negative and statistically significant coefficient on the “continuous eligibility” variable from the individual model is likely to be the most accurate. This is again in line with the 24 theory that the use of this policy will lower the administrative hassle associated with maintaining public insurance coverage, and will thus lower the probability of uninsurance. The coefficient on presumptive eligibility in the individual model is negative, but not close to being statistically significant. This finding is in line with the theory stating that this policy can conceivably serve to either reinforce or deter enrollment based on the circumstances of the individual in question. The fact that the coefficient is ambiguous likely indicates that neither effect predominates. The “face- to-face interview” variable also demonstrates ambiguous results in Table III that make it difficult to pinpoint its effect on uninsurance. I believe that a major factor contributing to the ambiguity of these coefficients is the fact that the face-to-face interview requirement may exert opposite effects on different groups of individuals within the sample. Later in this Results section, I specifically examine the different impact that the interview requirement may have on Black and Hispanic individuals. 5.3: Age, Race, and Ethnicity Before analyzing whether certain enrollment procedures have different effects on different subsets of the eligible uninsured population, it is important to present a clear picture of this population. The Data Summary Table indicates that older children and Hispanics are disproportionately represented in this group. In this section, I run three regressions that quantify these two observations. The dependent variable is the same binary indicator of uninsurance. In Column I, having controlled for family income and state fixed effects, I use as explanatory variables two dummies indicating Black or Hispanic identity, and a series of four dummies indicating the age bracket under which a particular individual falls. In Column II, I interact the “Hispanic” dummy with the four age bracket dummies. The regression in column III is structured in the same way, but includes interaction terms involving the “Black” dummy and four age bracket dummies. Finally, in column IV, I include both sets of dummy variables. These 25 interaction terms are useful in estimating the effect of age on uninsurance within the Black and Hispanic populations. Column I: Uninsured = β1i*(Age bracket)i + β2*(Hispanic) + β3*(Black) + λ*Demographic controls + γ*State fixed effects + ε Column II: Uninsured = β1i*(Age bracket)i*(Hispanic) + β2i*(Age bracket)i + β3*(Hispanic) + β4*(Black) + λ*Demographic controls + γ*State fixed effects + ε Column IV: Uninsured = β1i*(Age bracket)i*(Hispanic) + β2i*(Age bracket)i*(Black) + β3i*(Age bracket)i + β4*(Hispanic) + β5*(Black) + λ*Demographic controls + γ*State fixed effects + ε Column I in the table below demonstrates that as expected, older children are far more likely to be uninsured, as the coefficients on the age dummies grow increasingly positive and significant for children of older age groups. The coefficients on the two oldest age brackets are both positive and statistically significant at the 95% level, indicating that children belonging to these two groups are significantly more likely to be uninsured when compared to 0-1 year old children (the omitted age category in the above regression). A 16-18 year old child has the highest probability of uninsurance, and is nearly 12.5 percentage points more likely to be uninsured when compared to an infant in the sample. The phenomenon of attrition mentioned in the Theory section is a likely mechanism causing these higher rates of uninsurance among older children is attrition. It may be the case that at each age level, a certain number of children are simply not re-enrolled in public insurance programs that they are initially covered by. In older age groups, a larger group of drop-outs is likely to have accumulated, perhaps explaining the increased probability of uninsurance. 26 Table III Individual Level Age, Race, and Ethnicity Regression Age1to5 I II III -0.0159 (1.16) -0.0432 (2.66)** -0.0030 (0.19) -0.0345 (1.75)* IV Age6to9 0.0135 (0.95) -0.0258 (1.56) 0.0245 (1.51) -0.0256 (1.28) Age10to15 0.0298 (2.13)** -0.0101 (0.62) 0.0468 (2.96)** -0.0003 (0.02) Age16to18 0.1249 (7.59)** 0.0710 (3.77)** 0.1467 (7.80)** 0.0817 (3.57)** Hispanic 0.0849 (11.77)** -0.0177 (0.66) 0.0852 (11.79)** -0.0115 (0.40) Black -0.0057 (0.75) -0.0060 (0.79) 0.0540 (1.65)* 0.0148 (0.44) Hispanic* Age1to5 0.0907 (2.79)** 0.0805 (2.33)** Hispanic* Age6to9 0.1293 (3.75)** 0.1288 (3.48)** Hispanic* Age10to15 0.1274 (3.84)** 0.1158 (3.28)** Hispanic* Age16to18 0.1529 (4.20)** 0.1418 (3.67)** Black* Age1to5 -0.0507 (1.66)* -0.0258 (0.74) Black* Age6to9 -0.0412 (1.30) -0.0006 (0.02) Black* Age10to15 -0.0614 (2.09)** -0.0274 (0.81) Black* Age16to18 -0.0628 (2.06)** -0.0254 (0.71) For each regression, n = 28,726 (*) denotes significance at 90% level, (**) denotes significance at 95% level. Z-statistics in parentheses. Sample restricted to children eligible for public insurance. Each column above denotes a different regression. In each regression, the dependent variable is a dummy taking on a value of 1 if the individual is uninsured, 0 otherwise. Dprobit model used to estimate the coefficients on the above explanatory variables. Column I includes only non-interacted age, Hispanic, and Black dummies. Column II includes interaction terms involving Hispanic dummy and age dummies. Column III includes interaction terms involving Black dummy and age dummies. Controls were included for subject's family income relative to FPL, and state fixed effects in each regression. Controls for Asian or Native American background were included in each regression, but are not reported above. The results from Column II indicate that this effect is particularly pronounced within the Hispanic population. All of the interaction terms involving the “Hispanic” variable and Age dummies have positive and statistically significant coefficients at the 95% confidence level, indicating that Hispanic children of each age bracket are significantly more likely to be uninsured when compared to Hispanic newborns. These coefficients display a general trend of growing more positive with increased age (though the interaction term involving the “Age 6-9” dummy does have a more positive coefficient than the interaction term corresponding to the “Age 10-15” dummy). Controlling for age effects among Hispanic children markedly lowers the attrition effect 27 observed in the general population; children aged 16-18 in the general population are now only 8.2 percentage points more likely to be uninsured compared to infants. Controlling for attrition among Hispanics in column II also renders the Hispanic dummy statistically insignificant. This indicates that after controlling for Hispanic age effects, Hispanic individuals are no more likely to be uninsured than other individuals in the general population. Hence, it appears that Hispanic newborns enter the world with an equal probability of insurance compared to non-Hispanics in the eligible population. The high rate of uninsurance in this ethnic group thus appears to be driven almost entirely by attrition. These results contrast with those observed for Black individuals. In column III, none of the interaction terms involving the age dummies and “Black” dummy are positive (indeed, two of the negative coefficients are statistically significant at the 95% level). This indicates that while the general trend of increasing uninsurance with age may still be present in the Black population, Black identity does not appear to exacerbate this attrition, as Hispanic identity appears to do. Column IV, which features both sets of interaction terms, allows for an easy comparison of the different attrition rates of Black and Hispanic children. In order to compare the probability of uninsurance of a 16-18 year old Hispanic child with that of a newborn Hispanic child, it is necessary to add the coefficients from the non-interacted “Age 16-18” dummy and the interaction term involving the “Age 16-18” dummy and the “Hispanic” dummy. Adding the two coefficients, I find that a 16-18 year old Hispanic child is 22.4 percentage points more likely to be uninsured when compared to a Hispanic newborn. Conducting the same comparison between a Black child aged 16-18 and a Black newborn, I find that the Black 16-18 year old is only 5.6 percentage points more likely to be uninsured when compared to a Black newborn. Carrying out this calculation across the four income brackets, I observe that Hispanic children of all four age brackets have a higher probability of uninsurance when compared to Hispanic newborns. This probability of uninsurance relative to that of a Hispanic newborn increases sharply with age. For Black children, only those in the “Age 16-18” bracket have a higher probability of uninsurance 28 when compared to Black newborns. It is thus clear that attrition has a prominent effect on Hispanic children, and that this effect is not nearly as strong in the Black population. It is possible that those states with large Hispanic populations feature policies or other state specific factors that promote attrition. In this case, the attrition effect could primarily be driven by individuals in these states, and may not be observed throughout the general population. To test this possibility, I removed children from California and Texas – the two states with the largest Hispanic populations – from the sample, and re-ran the regressions above. When these children are removed, the coefficients on the age dummies grow slightly less positive, providing less dramatic, but still solid evidence for attrition (for example, children aged 16-18 are now 11.1 percentage points more likely to be uninsured compared to 0-1 year olds, rather than 12.5 percentage points). This slight decrease in attrition is likely because the large population of Hispanic individuals in California and Texas, which is now removed from the sample, helps to drive the effect in the overall population. Re-running Regressions II, III, and IV with this smaller sample, meanwhile, does not cause the interaction terms to change in any consistent manner. As a result, the attrition hypothesis appears robust to removing these two states from the sample. The bulk of the remaining analysis in this paper, therefore, focuses on analyzing in greater depth factors that may cause children – particularly those of Hispanic background – to be uninsured in larger and larger numbers as they grow older. In particular, I focus on the enrollment procedures analyzed in sections 5.1 and 5.2, and examine their effects on attrition among children of Hispanic origin. As a comparison, I also examine their effects on children in the Black population. 5.4 Face-to-face Interview and Uninsurance The results from Tables I and II indicate that the re-enrollment interview requirement increases the probability of uninsurance, ostensibly by adding significantly to the cost (in terms of time and effort) of re-enrollment. Given that many Hispanics do not speak English as a first language, I hypothesized that individuals within this group would find complicated enrollment 29 procedures to be particularly burdensome. I thus predicted that Hispanic individuals in particular would be disproportionately affected by the use of a re-enrollment interview requirement. However, regression analysis examining the impact of the re-enrollment interview on Hispanic individuals revealed no significant relationship. Most likely, this is due to the fact that only two states – Tennessee and Mississippi – require re-enrollment interviews for both Medicaid and SCHIP. This does not provide much variation in state policy with which to test these hypotheses regarding the re-enrollment interview requirement. Perhaps a larger sample size of children living in states requiring re-enrollment interviews would provide support for this hypothesis. There are six states that require interviews for initial enrollment in Medicaid, SCHIP, or both.12 Four of these states require interviews for both policies. Two other states (New York and Alabama) require face-to-face interviews for Medicaid, but not for SCHIP. I allowed the “face-toface interview” dummy to take on a value of 1 for children living in all 6 states. While this variable is no longer fully accurate (given that not all eligible children in New York and Alabama were subject to the policy), I believe that it is still a useful one, given that Medicaid is by far the larger of the two programs in both New York and Alabama.13 While these six states again do not provide a tremendous amount of variation in state policy, they do provide a better starting point for analyzing the impact of the face-to-face interview requirement. I hypothesized that the initial enrollment interview would have a far smaller impact promoting drop-out when compared to the re-enrollment interview. This is because this initial interview represents for many the first direct contact with the relevant program office. As mentioned previously, this direct contact can serve as a means of disseminating useful information about the procedures for public insurance 12 In three of these states – Alabama, Tennessee, and Utah – families are permitted to complete the interview by phone rather than face to face. In this model though, I did not differentiate between those permitting a phone interview and those requiring face to face contact, and categorized all six states as having an interview requirement. 13 According to estimates from the US Census Bureau, there were approximately 1,538,000 child enrollees in Medicaid in New York and 367,000 in Alabama as of March 2005. According to a 2004 report from the Kaiser Family Foundation, SCHIP enrollment in December 2004 was 452,932 in New York, and 62,817 in Alabama. Assuming that there was no significant change in SCHIP enrollment between December 2004 and March 2005, the above numbers indicate that as of March 2005, approximately 77% of public insurance enrollees in New York and 85% in Alabama participated in Medicaid. 30 enrollment, and can thus increase program take-up. Additionally, the initial enrollment interview is a procedure needing to be completed only once, while the re-enrollment interview must be conducted year after year. As a result, the re-enrollment interview requirement would ostensibly place a larger burden on applicant families. Table IV tests these hypotheses. Column I examines the impact of the face-to-face interview requirement on attrition for the general population, using as right-hand side variables the face to face interview dummy, the four age dummies, and four interaction terms involving these two variables. Column II examines whether the interview requirement increases attrition among Hispanic individuals in particular. This regression includes all the variables from the previous one, and additionally contains a set of four triple interaction terms between the age dummies, the face-to-face interview dummy, and the Hispanic dummy. Column III is structured in the same way, but includes triple interaction terms that measure the marginal effect of the faceto-face interview requirement on Black individuals across the age spectrum. The regression in Column IV includes both sets of interaction terms. Column I Uninsured = β1*(FF Interview) + β2i*(Age bracket)i + β3i*(FF Interview)(Age bracket)i + λ*Demographic controls + γ*State fixed effects + ε Column II Uninsured = β1*(FF Interview) + β2i*(Age bracket)i + β3*(Hispanic) + β4i*(FF interview)(Age bracket)i + β5i*(Age bracket)i(Hispanic) + β6*(FF interview)(Hispanic) + β7i*(FF interview)(Age bracket)i(Hispanic) + λ*Demographic controls + γ*State fixed effects + ε Column IV Uninsured = β1*(FF Interview) + β2i*(Age bracket)i + β3* (Hispanic) + β4*(Black) + β5i*(FF interview)*(Age bracket)i + β6i*(Age bracket)i (Hispanic) + β7i*(Age bracket)i*(Black) + β8*(FF interview)(Hispanic) + β9*(FF interview)*(Black) + β10i*(FF interview)(Age bracket)i(Hispanic) + β11i*(FF interview)*(Age bracket)i*(Black) + λ*Demographic controls + γ*State fixed effects + ε 31