Connecting Statistical and Machine Learning Models in Social Work Research: An Introduction

Social work researchers have primarily relied on statistical models to engage in quantitative inquiries to advance social justice. As the broader field of artificial intelligence (AI) has gained significant popularity in scientific investigations, it becomes critical for the profession to understand and evaluate the conceptual and methodological compatibility between AI-driven methods and existing frameworks of quantitative approaches. Machine learning (also known as statistical learning) models have become increasingly relevant in social and health science research. Machine learning models are a set of statistical methods and algorithms used to predict outcomes and understand the underlying structure of the data. The primary goal of this workshop is to bridge the gap between statistical models and machine learning models, encouraging social work researchers to integrate machine learning techniques into their quantitative inquiries.

The workshop will start with a high-level review of explanatory (statistical) models versus predictive (machine learning) models and their conceptual relevance to social work research. Key concepts such as supervised, unsupervised, or semi-supervised learning will also be covered to establish the foundation of the workshop. We will then extend the linear regression models to predictive machine learning methods/algorithms, specifically including the Multivariate Adaptive Regression Splines (MARS) and Shrinkage/Regularization methods (including LASSO regression). MARS is a non-parametric regression algorithm for modeling complex and non-linear relationships, especially using adaptive relationships between variables to improve model predictions. Shrinkage/Regularization methods are a set of techniques that are natural extensions of the classic regression models with added penalty terms to the loss function to prevent overfitting and model generalizability.

In addition, tree-based methods, specifically the single decision trees and the Bootstrap Aggregating (Bagging) methods, will be covered. A decision tree is a non-parametric approach to supervised learning for both classification and regression tasks, whereas the Bagging method is an ensemble technique that can be used to improve the stability and accuracy of decision trees. Our review of the tree-based methods will focus on both technical details as well as conceptual significance/relevance to social work research. We will conclude the workshop by demonstrating our team’s process of constructing a peer-reviewed journal article that is relevant to health social work research, including a step-by-step tutorial of codes running the machine learning models. If time permits, we will include a brief introduction to Random Forests, another ensemble learning algorithm that builds on both the decision tree and the bagging method.

The workshop will include a mixture of lectures, interactive discussions, code demonstrations, and questions. Slides and computer analytical codes will be provided to attendees. No background in machine learning is assumed, though a solid foundation in linear regression and generalized models is expected to fully benefit from the workshop.