determine significant cut-off points

[ShawnZ]

Hi all,

I'm trying to determine the relationship between student absent days and their proficiency level. I want to know at what point/percentile of absence days would there be a significant proficiency change.

For example, the proficiency can be binomial (pass, fail), the absent days level can be a continuous variable at range of 0%-100%. What would be the best steps to determine the cut-off point for absent days level where you can see the most significant change in proficiency?

I think 'proc logistic' would fit the problem but I'm not sure how to get the right cut-off point.

Any ideas or comments will be much appreciated!

Thank you!

[Reeza]

I think it's more like a CART with only one predictor. 

I don't know how to implement that in BASE SAS though.

[PGStats · Accepted solution]

Proficiency should decrease smoothly with increasing absence, you can request to estimate the absence level for different proficiency tresholds with PROC PROBIT. Something like:

proc probit data=myData plot=predplot;

class passFail;

model passFail = absence / lackfit inversecl;

run;

PG

[1zmm]

The MODEL statement of PROC LOGISTIC has a CTABLE option that estimates and prints the sensitivity, the specificity, the percent false positives, and the percent false negatives at each cutpoint of the probability predicted from the model.  The OUTPUT statement can write these predicted probabilities corresponding to each level of absence day.  Therefore, you can appreciate the trade-offs at different cutpoints of the predicted probabilities.

[ShawnZ]

Thanks all, that was very helpful!

I tried both PG and 1zmm's methods. They both worked great. IMO I think PROC PROBIT provide a clearer result to me as well as nice probability plot. Although I think PROC LOGISTIC can do probability plot as well, using ESTPROB option maybe?

I also find that specify certain cut-off point in CONTRAST in PROC LOGISTIC can also be helpful for such case. Like:

model passFail  = absent / link= probit;

contrast '1%' intercept 1 absent 0.01 / estimate=prob;

contrast '10%' intercept 1 absent 0.10 / estimate=prob;

...

Anyway, thanks for the help again!