I am running a CHAID classification tree on SPSS to classify my data set. I have a couple independent variables including categorical and continuous ones. For continuous variables, I've noticed that SPSS (and other software as well) technically binned the continuous variables into groups having similar observations. My question is why it is necessary for the groups to have similar observations?

To make it more sense, I am thinking of using Jenks algorithm to categorize my continuous variables into categories as inputs to the CHAID tree. Do you think such approach is good?

Also, could you tell me how to determine the number of parent and child nodes as well as the tree depth?

Thanks all.

  • $\begingroup$ technically binned the continuous variables into groups having similar observations CHAID, actually, does not operate on continuous predictors, it treats such predictors as categorical ordinal. $\endgroup$ – ttnphns Oct 2 '14 at 18:03
  • $\begingroup$ Thanks. I understand that CHAID actually prefers categorical variables. But what I don't know is why the binning approach is based on the number of observations. $\endgroup$ – Tran Oct 2 '14 at 18:05
  • $\begingroup$ What do you mean? $\endgroup$ – ttnphns Oct 2 '14 at 18:15
  • $\begingroup$ I mean I don't know why the number of observations should be similar for all groups. In other words, why quantile classification is used to categorize continuous variables in SPSS. Is the number of observation important in CHAID? $\endgroup$ – Tran Oct 2 '14 at 18:48
  • $\begingroup$ Hmm, I would say it's not so. I often got final grouping of a predictor (whatever scale type) into unequal-sized groups. However, getting about equal-sized groups is quite likely, because chi-sq test is more powerful (hence more significant) when groups are about equal-sized. $\endgroup$ – ttnphns Oct 2 '14 at 19:30

You can find the exact binning algorithm via Help > Algorithms, but note that you can control the number of bins for (each) continuous variable. It is often reasonable to expect that the effect of a continuous variable varies slowly and may be assumed constant within each interval. Intervals will be grouped as appropriate. By specifying a larger number of bins, you can account for more sharply varying effects.

You could also use the OPTIMAL BINNING procedure to prebin a continuous variable.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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