Identifying Unvaccinated Individuals in Canada: A Predictive Model

TL;DR

This study develops a predictive model using Canadian health survey data to identify unvaccinated individuals and improve vaccination outreach efforts, achieving high accuracy with minimal variables.

Contribution

The paper introduces a novel combined approach using probit and Random Forest models to predict vaccination status with high accuracy using only 3.6% of available variables.

Findings

Achieved 87.8% overall accuracy in predicting vaccination status.

Identified key variables associated with vaccination behavior.

Model requires only 3.6% of survey variables for effective prediction.

Abstract

Recently, the media and public health officials have become increasingly aware of the rise in anti-vaccine sentiment. Vaccinations have numerous health benefits for immunized individuals as well as for the general public through herd immunity. Given the rise in immunization-preventable diseases, a consequence of people opting out of their routine vaccinations, we determined that Canadian health data can identify individuals over the age of 60 who chose not to get vaccinated (80.1% negative predictive value) and individuals under the age of 60 who have recently been vaccinated (96.4% positive predictive value). Using the 2009-2014 Canadian Community Health Surveys (CCHS), a probit model identified the variables that were most commonly associated with flu vaccination outcomes. Of 1,381 variables, 47 with the most significant marginal effects were selected, including the presence of diseases (e.g. diabetes and cancer), behavioral characteristics (e.g. smoking and exercise), exposure to the medical system (e.g. whether the individual gets a regular check-up), and a person's living situation (e.g. having young children in the household). These variables were then used to generate a Random Forest classification model, trained on the 2009-2013 dataset, and tested on the 2014 dataset. We achieved an overall accuracy of 87.8% between the two final models, each using 25 classification trees with bounded depth of 20 nodes, randomly selecting from all 47 variables. With the two proposed policies, this model can be leveraged to efficiently allocate vaccination promotion efforts. Additionally, it can be applied to future surveys, only requiring 3.6% of the variables in the CCHS for successful prediction.