Modelling non-response in establishment-based freight surveys: A sampling tool for statewide freight data collection in middle-income countries

Abstract

Non-response is unavoidable in any disaggregate-level survey efforts where information is sought from individuals. There is a strong need to understand the contributing factors of non-response behavior in freight surveys as there is very little information on what sort of establishments respond positively to survey inquiries. To meet this research gap, this paper analyzes the non-response behavior in an establishment-based freight survey (EBFS) conducted in Kerala, India. The analysis involves usage of five prominent classifiers: classification trees, random forests, Naïve Bayes classifier, logistic regression and K nearest neighbors. A closer look at the results revealed that the industrial classes handling commodities with high value density (machinery, electronics and electrical equipment, other manufacturing products, textiles) exhibit high non-response probability for freight surveys. This may be explained in terms of the limited logistic information shared by establishments handling high valued products due to less attention given to logistics strategies. Another reason could be that the amount of capital and opportunity costs tied up with these commodities are high, and therefore, their information tend to be more proprietary in nature. The analysis also reveals that the location of establishment plays a modest, yet, noticeable effect on non-response behavior. The establishments from cities located closer to Ports with high per-capita income are less likely to respond to survey requests, possibly due to the higher opportunity cost for time and increased concerns of intrusion into privacy. Finally, the comparison of ROC curves suggests that KNN algorithm is most suitable for modelling non-response behavior. These models are expected to be useful for developing sample weighting schemes and targeted incentive strategies that can improve the state of practice of freight data collection in middle-income countries like India. The study findings are expected to be useful for developing more elaborate and dynamically responsive survey designs for EBFS, in which sample recruitment strategies can be adapted real time during data collection. This study also informs policymakers that the apparent trends in non-response, if not arrested, are likely to weaken the validity of inferences drawn from estimates based on EBFS and undermine, perhaps fatally, the potential of EBFS to guide facility planning and policy interventions.