Technical Appendix B: Immigration Status Imputation

To impute documentation status, we draw on the methods underlying the 2013 analysis by the State Health Access Data Assistance Center (SHADAC) and the recommendations made by Van Hook et. al..^1,2 This approach uses the Survey of Income and Program Participation (SIPP) to develop a model that predicts immigration status for each person in the sample; it then applies the model to a second data source, controlling to state-level estimates of total undocumented population as well as the undocumented population in the labor force from the Pew Research Center.³ Below we describe how we developed the regression model and applied it to the American Community Survey (ACS). We also describe how the model may be applied to other data sets. The programming code, written using the statistical computing package R v.4.2.2, is available upon request for people interested in replicating this approach for their own analysis.

Data Sources

We used the second wave of the 2008 Survey of Income and Program Participation (SIPP) panel data to build the regression model. The SIPP Wave Two dataset contains questions on migration history at the person level.

The regression model is designed to be applied to other datasets in order to impute legal immigration status in surveys that do not ask about migration status. The code mentioned above includes programming to apply the model to either the SIPP Core file, ACS, or the Current Population Survey (CPS). Because the SIPP Core file contains different survey questions and variable specifications from the ACS and CPS, we create unique regression models to apply the model to each dataset. For the analysis underlying this brief and other KFF estimates of eligibility for ACA coverage, we apply the regression model to the 2013 ACS and then each subsequent year of the ACS.

Due to underreporting of legal immigration status in the SIPP, in imputing immigration status we control to state and national-level estimates of the total undocumented population and also the undocumented population in the labor force from the Pew Research Center. Pew reports these estimates for all states and the District of Columbia.⁴

Construction of Regression Model

We use the SIPP Wave Two to create a binomial, dependent variable that identifies a respondent as a potential unauthorized immigrant. The dependent variable is constructed based on the following factors:

1. Respondent was not a United States (US) citizen,
2. Respondent did not have permanent resident status upon US entry,
3. Respondent’s immigration status did not change to permanent resident since US entry, and
4. Respondent does not have other indicators that imply legal status.⁵

We use the following independent variables to predict unauthorized immigrant status:

1. Year of US entry,
2. Job industry classification,
3. State of residence,
4. Family Poverty Level,
5. Ownership or rental of residence,
6. Presence of at least one citizen in household,
7. Number of occupants in the household (< or >= six occupants),
8. Whether all household occupants are related,
9. Number of workers in household,
10. Health insurance coverage status,
11. Sex, and
12. Ethnicity.

The regression model was sub-populated to remove respondents who could not be considered unauthorized. People who could not be considered unauthorized include people who 1) were born in the US, 2) are US citizens, or 3) have other indicators that imply legal status.

Imputing Unauthorized Immigrants in Other Datasets

We use the Pew estimates as targets for the total number of unauthorized immigrants that the imputation generates. We first apply this strategy to the 2013 ACS, which contains health insurance information prior to the ACA’s coverage expansions. We stratify the targets by state and the District of Columbia and by participation in the labor force. We impute immigration status within each of these 102 strata.⁶

To generate the imputed immigration status variable, we first calculated the probability that each person in the dataset was unauthorized based on the SIPP regression model. Next, we isolated the dataset to each individual stratum described above. Within each stratum, we sampled the data using the probability of being unauthorized for each person. After sampling, we summed the person weights until reaching the Pew population estimate for each stratum. The records that fell within the Pew population estimate were considered to be unauthorized immigrants. We repeated the process of sampling using the probability of being unauthorized and subsequently summing the person weights to reach Pew targets five times, creating five different unauthorized variables per record. These five imputed authorization status variables were then incorporated into a standard multiple imputation algorithm, closely matching the imputed variable analysis techniques used by the Centers for Disease Control and Prevention for the National Health Interview Survey.⁷

We used this first pass on the ACS 2013 to inform our sampling targets for the latest available microdata (ACS 2021). Looking at the results of our undocumented imputation on the ACS 2013, we calculated the share of undocumented immigrants lacking health insurance within each of those 102 strata prior to the ACA’s coverage expansions and transferred that information into a new dimension of sampling strata for the ACS 2021. We split each of the 102 sampling strata used on the pre-ACA ACS 2013 into uninsured versus insured categories, resulting in 204 sampling strata for subsequent years. We then repeated our imputation on the ACS 2021 with the newly-divided strata, allowing for a small decline in the undocumented uninsured rate based off of the percent drop in the uninsured rate we see in the KFF’s Survey of the Low-Income Population and the ACA.⁸

To easily apply the regression model to other data sets, we created a function that applies this approach to a chosen data set. The function first loads the dataset of choice, then standardizes the data to match the independent variables from the SIPP regression model, and finally applies the multiple imputation to generate a variable for legal immigration status.