Paper Summary
Share...

Direct link:

A Monte Carlo Simulation on Nonlinear Effects of Likert-Scaled Items With Machine Learning

Tue, April 26, 4:15 to 5:45pm PDT (4:15 to 5:45pm PDT), Division Virtual Rooms, Division D - Section 2: Quantitative Methods and Statistical Theory Virtual Roundtable Session Room

Abstract

Although linear effects of predictors are predominant in social science research, nonlinear effects are often reported. Focusing on the nonlinear effects of Likert-scaled items, we investigated how different coding schemes (dummy coding or not) and machine learning methods (group Mnet or random forest) affect models’ prediction measures and variable selection in a Monte Carlo simulation. Other conditions to emulate large-scale survey data included linearity (linear, quadratic, cubic, and piecewise constant), density of true predictors in item parcels (concentrated or scattered) and symmetry in responses (symmetric or asymmetric). Beyond providing interpretable prediction models as a linear method, group Mnet either outperformed or was on par with random forest in terms of prediction and variable selection regardless of condition combinations.

Authors