02-10-2015 08:57 AM
I am facing a difficulty in limiting the distributions obtained by PROC UNIVARIATE i.e. I want to have an option to set lower and upper limits within which the losses for each percentile (for each distribution) fall.
Can anyone please suggest something on the same.
02-10-2015 10:09 AM
Thanks Reeza for your reply , but my problem is not related to limit the data. I will give you example . Now suppose
In case of the sample data I am using, the lowest claim is $968, but in case of fitted Normal Distribution (using PROC UNIVARIATE), the 1st, 5th and 10th percentile are coming out to be negative. Let say, had we limited the lower tail of distribution to $100, all these percentiles would been more than $100. There should be option for similar upper limit too. Please do let me know if this helps.
02-10-2015 11:33 AM
If a value in the 10th percentile is not "valid" then your data very likely isn't normally distributed. Monetary values often have a skewed distribution.
You might need to investigate to get a better idea which distribution does fit.
02-10-2015 11:38 AM
I think that PROC UNIVARIATE is letting you know that a normal distribution does not fit your data. Are you asking for options to fit bounded distributions?
If you know that your data are always positive, you can choose to fit a bounded distribution to your data. PROC UNIVARIATE supports the lognormal and gamma distributions, and some others. Since you mention "claims data," you might want to look into PROC GENMOD which enables you to fit various nonnegative distributions, including count data and the Tweedie distribution.
02-10-2015 12:55 PM
Also, distribution free percentiles are available in PROC UNIVARIATE. Try:
proc univariate data=yourdsn CIQUANTDF(type=asymmetric);
02-11-2015 02:33 AM
Thank you all for your responses. This is very helpful. Let me propose all the solutions to my team and see what happens. I will keep uou all posted.