Research On The Predictive Models Of Middle School Students' Active Smoking Behaviors

Objective: To explore the methods of assessing and building the models on ordinal response data and provide a reference for building the models on ordinal response data based on building the predictive models of mmiddle school students'active smoking behaviors with the epidemical data. Methods and data: The cumulative logit,probit and cloglog(or gompit) models are used to build the predictive models of middle school students'active smoking behaviors based on the selected primary related factors on the middle school students'active smoking,and explore the methodological problems on building the cumulative models.The data were collected in Jiulongpo county,Chongqing according to the questionair ,《survey on the effect of tobacco advertisement on adolesent's acquiring active smoking behaviors》,designed by former Chinese Academy of Preventive Medicine .Results: The rate of active smoking in adolesent is 9.5%,and 18.28% for male,0.85% for female,the rate of male is obviously higher than that of female;The rate of active smoking among the four types schools is statistical significance,the same for the different grades;The everyday rate of smoking is 3.85%,the weekly is 3.25%,the occasional is 2.40% and the none is 90.50%;The average age of onset cigarette is 13.0 years old.The factors related to the active smoking include age,sex,culture degree of mother, study score,pocket money,influence from movies ,perception tosmoking behaviors,whether mind other smoking around,whether smoking when friends pass cigarette and smoking proportion of friends.The cumulative logit,probit and cloglog models are built respectively with the 10 related factors ,and the model fitting information criteria ,Goodness-of-fit,generalized R2 and association predicted probabilities and observed response display that the fitting results of three cumulative models are good and they can interpret the variation well.The generalized adjusted R2 of 200 simulated data sets models all are less than that of the actual data models.Conclusions: The model fitting information criteria ,goodness-of-fit ,generalized R2 and association predicted probabilities and observed response indicate that the three cumulative models fit well.The data must be ordinal response , enough sample size and not too low event rate for the cumulative models. The monte carlo simulation is a good method for assessing the utility of the cumulative models based on the ordinal response data when the sample size is not enough and the event rate is rather low. For interpretion of parameters and pepole's understanding according to odds ratio ,the best model is the cumulative logit model when the values of assessing index have little difference.