MULTINOMIAL LOGISTIC REGRESSIONS: FACTORS AND PREDICTION ON MALAYSIAN FILM VIEWERS

Abstract

Many researchers are currently interested in conducting a technique in data analysis with qualitative dependent variables involving more than two categories, known as Multinomial Logistic Regressions (MLR). This study too had conducted using MLR to examine the significant factors based on the frequency of watching categories, such as 1) once or nil in a month, 2) twice a month, 3) three to four times in a month and 4) more than five times per month. From the frequency of watching categories (1, 2, 3 & 4), the frequency of watching more than five times per month (Cat 4) was referred to as the reference group, while the other categories had exhibited 255 MLR models each. Statistical tests, modelling procedures and models’ goodness-of-fit tests were carried out on a total of 765 models from the 3 categories on film watching frequencies. In order to obtain a set of selected models (with significant variables), a progressive elimination (one by one, least significant first) of the insignificant variables was employed at Phase 2 of the model building procedures involving three types of tests namely NPC/NPM, multicollinearity and coefficient tests. Criteria based on pseudo R-squared were proposed consisting of Cox & Snell, Nagelkerke and McFadden, to finally single out the best model. The important findings highlighted in this study were the best model validation using the Mean Absolute Percentage Error (MAPE). Via the best models from each category, the model-building approach in Multinomial Logistic Regression analysis was established, and prediction using MAPE was done. Findings showed that the best models from all the respective categories (1, 2 & 3) had two common significant factors on the dependent variable. The results also showed that the best model from Cat 1 had the least MAPE (6.57%, thus indicated it was excellent to be used for prediction. Based on this, it is suggested that to attract more viewers, less films should be produced in a year, however, the allocated budget for film making should be focused on producing films which conformed to the identified significant factors that would attract more viewers. By using the best model, film viewing frequencies and number of film viewers can thus be predicted, and the expected revenue for the film industry can thus be estimated