0
$\begingroup$

I have a data set to predict how long in years does a student stay in college. The dependent has value like 0.5, 1, 1.5.etc,. I used classification methods like SVM, decision tree but the accuracy is like 30%, which is so low. I tried to round up 0.5 to decrease the levels but still got low accuracy.

So I doubt if it is a classification.

The frequency of number of years is like this

   0.5    1  1.5    2 2.25  2.5    3  3.5    4 4.25  4.5    5 5.25  5.5 
   758  223  357  118    2  182  120  840  287    1  576  158    1  261

enter image description here

I also tried linear regression and svm. RMSE is like 1.4 which translate to 1.4 years? I am not sure if the accuracy here is very bad or bad?

And also, I tried Cox PH regression, the question is that some students graduated before say 4 years and I do not know what is the right censoring time here.

Improve this question
$\endgroup$
3
  • 1
    $\begingroup$ It doesn't seem to be a problem of the estimator/classifier you're using. It just seems that you need more predictor variables and maybe more observations. $\endgroup$
    – Yannis Vassiliadis
    Commented Jun 13, 2017 at 5:32
  • 1
    $\begingroup$ Also there seems to be something weird going on with your data. More than half your observations stay only 0.5 a year? 3080 leave after a year, but only 2340 leave (presumably graduate) after 4? Given that, it's not surprising that accuracy is like 30%. $\endgroup$
    – Yannis Vassiliadis
    Commented Jun 13, 2017 at 5:38
  • $\begingroup$ @YannisVassiliadis I just updated the frequency table. yes, there are lots of 1st year dropout or transfer. $\endgroup$
    – janicebaratheon
    Commented Jun 13, 2017 at 15:26

1 Answer 1

Reset to default
0
$\begingroup$

I think you should use linear regression, at least to start, and examining the assumptions using the usual graphical methods.

I don't think you have censoring, unless I have misunderstood. Do you have people who are still in college when you collect data? Then they would be censored because you would not know how long they will stay.

To evaluate the model, I'd look at actual values vs. predicted values. See how big the differences are. But, if your results are poor, it may just be that your variables don't do a good job of predicting years in college.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.