Comparison of the linear regression, multinomial logit, and ordered probability models for predicting the distribution of thermal sensation

Abstract

This study compares the linear regression model, ordered probability model, and multinomial logit model for prediction of the individual thermal sensation votes (TSVs) and TSV distributions under given conditions. Two thermal comfort datasets were used to develop and evaluate the models. One dataset was taken from an indoor thermal comfort survey conducted in Pakistan, and the other was taken from an outdoor thermal comfort survey conducted in Tianjin, China. The data were divided into training and validation datasets. The training datasets were used for model development. The developed models were then used to predict new cases in the validation dataset. The predictive capability of the three models were systematically evaluated and compared to examine how well the developed models predicted individual TSVs and TSV distributions for the validation dataset. The results showed that the ordered probability model and the multinomial logit model correctly predicted around 50% of the individual TSVs, whereas the accuracy of the linear regression model was only around 20 to 40%. In addition, the chi-square statistics show that the ordered probability model and the multinomial logit model better predicted the TSV distributions than the linear regression model.