Imputation strategies for child poverty level

Background: Most researchers using survey data encounter incomplete observations. The nature of missing values may generate unwanted bias and diminish power, affecting the standard error and model estimates.

Objective: All survey data have missing information on socio-demographic variables— especially poverty level. Because most of the national children's health surveys are available for public use and poverty level is a key demographic variable, it is important to develop imputation strategies for child poverty level.

Methods: Poverty status was grouped as more or less than 200% of the federal poverty level and single imputation (multivariable logistic regression) was performed for complete cases. Variables associated with poverty level, but not related to other outcomes of interest in our hypotheses (to avoid potential co-linearity) were chosen for the model. Since the actual values for complete cases are known, this provided a basis for assessing the imputation method.

Results: The imputation produced a correct classification rate of 80% for all complete cases. We then predict values for the 8.4% of cases with incomplete poverty level data. Among other variables, parents' health insurance status, parents' health condition, and family structure were used. After the imputation procedure, 25% and 75% of the incomplete cases assigned to less than and more than 200% of poverty level, respectively.

Conclusion: Choosing the most suitable imputation method depends on the purpose and structure of each specific dataset. Since it is almost impossible to avoid missing values with survey data, correctly using statistical imputation can be useful to overcome this problem.