Development and internal validation of three multivariable prediction models of Hispanic immigrant documentation status utilizing two samples from the southwestern United States (Preprint)

Abstract

Research on undocumented populations is sparse, both because documentation status is rarely ascertained in studies, and because undocumented individuals are less likely overall to participate in research studies. A better understanding of the health of these individuals is important, though, and one method that has been proposed to examine outcomes for this group without putting them at additional risk is through predictive modelling.

Develop and internally validate a predictive model of documentation status.

Utilizing combined data from two population-based samples in Arizona in which documentation status were collected, three prediction models of documentation status were created and internally validated. Models created used multiple imputation by chained equations (MICE), and two machine learning algorithms: random forests machine learning and support vector machine. Predicted documentation status was modelled from demographic, health behavior, and disease status indicators commonly collected in public health research studies. The performance of each model was assessed in terms of accuracy, precision, recall, and Matthews correlation coefficient (MCC).

The combined sample consisted of 473 individuals, of whom 85 were undocumented. The sample included 209 men and 263 women, aged 20 to 83 years of age at the time of their participation in the respective studies. The MICE model had the lowest rates of accuracy (64.3%), precision (0.81), recall (0.74), and MCC (-0.07). The random forests model was slightly better with an accuracy of 73.1%, precision of 0.82, recall of 0.86, and MCC of 0.002. The support vector machine was the strongest in all categories, with an accuracy of 90.7%, precision of 0.90, recall of 0.99, and MCC of 0.66.

Utilizing predictive models could protect the safety of vulnerable individuals as well as make possible investigation into potentially undocumented individuals in settings where it would be impractical or impossible to gather information on documentation status first-hand. The results of the current analysis demonstrate some promise and further investigation, particularly the support vector machine model.