How much do attitudinal variables improve travel demand models? Evaluation using an overlap sample from an attitude-rich survey and the 2017 National Household Travel Survey
利用两份调查的重叠样本，用弹性网回归从标记变量插补态度因子得分，评估加入态度变量对出行行为模型（如车辆拥有量、行驶里程、混动/电动车采用）的预测改进效果。
• Advocates adding selected attitudinal marker statements to travel surveys. • Uses the rare opportunity of having an overlapping sample from two surveys. • Imputes attitudes from one dataset to another using machine learning. • Confirms excellent imputation performance of elastic net regression. • Explores how including imputed attitudes benefits travel behavior models. This study evaluates the effectiveness of adding a few attitudinal marker statements to transportation surveys (instead of designing, deploying, and factor-analyzing a full set of attitudinal variables). We exploit the rare opportunity offered by a 2017 Georgia Department of Transportation (GDOT)-funded survey and the 2017 Georgia add-on to the National Household Travel Survey (NHTS) having 1245 respondents in common. The non-overlap GDOT survey dataset (N = 2043) is the donor sample, based on which elastic net regression models are trained for imputation of attitudinal factor scores using marker variables. The overlap NHTS dataset – the recipient sample (N = 1245) – is treated as if it has only marker variables, with attitude scores needing to be imputed using those variables. We find that the elastic net regression models predict the factor scores very well in both the donor and recipient datasets, while using marker variables alone also offers excellent prediction performance. Three travel behavior variables in the recipient dataset – household vehicle count, (personal yearly) vehicle miles driven, and hybrid/electric vehicle adoption – are then modeled with alternative specifications involving no attitudes, predicted attitude scores, and marker variables. For each dependent variable, several attitudes show statistical significance, although their contributions to model fit vary. Overall, including attitudes leads to (a) better prediction of less-common alternatives (zero vehicles and hybrid/electric vehicle adoption), primarily by improving the prediction of the groups most likely to select such alternatives, and (b) discovery of additional non-attitude variables that would have been considered insignificant otherwise.