How to improve my classification tree ? R

Question

I have a database with 1200 observations and 14 variables and I'am trying to do a classification tree for my dependent nominal variable who hase 4 modality

    > table(testarbre2$Q99)

  Autres       Nahdha Ne pas voter Nidaa Tounes 
     248          351          303          298 

at firt i tried to do a multinom logistic regression but i got the mojority of my predictor variables non significant. it seems that Even with 1200 people I was trying to fit a model for which I don't have sufficient data. so i tried to do a classification tree using the package rpart from R but the problem is that the error is so high about 65% and more, and the missclassification is about 70% this is the code R that i used

   #preparation of the data
   set.seed(26)
   train=sample(1:nrow(testarbre2),nrow(testarbre2)*7/10)
   test=-train
   training_data=testarbre2[train,]
   testing_data=testarbre2[test,]
   testing_vote=vote[test]

   #fitting the model
   library(rpart)
   library(rpart.plot)
   Tree <- rpart(Q99~.,data=training_data)
   rpart.plot(Tree)
   printcp(Tree)
   plotcp(Tree)

    #Construction of the complete tree
  Tree <-rpart(Q99~.,data=training_data,control=rpart.control(minsplit=50,cp=0))

     #Prune the tree
    treeOptimal <- prune(Tree,cp=Tree$cptable[which.min(Tree$cptable[,4]),1])
    rpart.plot(treeOptimal)

   #Prediction
   a=predict(ptitanicOptimal,testing_data2,type = "class")
   mc=table(a,testing_vote2)

I don't know if i missed a step or i used a wrong approach in the construction of my classification tree or the database is causing the problem

Please someone help me to understand what's wrong with my model

Answer 1

Classification trees require sometimes ten times the sample size of logistic regression, and you will be quite disappointed in the stability of the tree. Bootstrap the process for a few resamples and you will see the tree topology change quite a bit. Simplicity in single trees is more of an illusion than a reality. Trees seem simple when you select one tree from many competitors that are very difficult to choose from. In addition you have chosen a discontinuous improper accuracy scoring rule which is optimized by bogus predictions, i.e., optimized by using the wrong model for the data.

Lack of significance is not a reason to change methods. Instead consider data reduction masked to $Y$, or use penalized maximum likelihood estimation to deal with your relatively small sample size.