PDA

View Full Version : Testing Goodness of Fit and Misc


Weikko Jaross
June 8th, 2001, 05:33 PM
I am very new to this parameter estimation stuff, so any and all advice I will appreciate.

I am working with data that of all things represents elements of cycle time observations for logging machinery (time to success, ha! ha!).

For example, I am taking the time observations (sec) for setting a choker and attempting to fit distributions to these observations using Weibill 6. Say for 30 to 1044 independent observations, depending on how I stratify my data.

I have been working under the impression that the goodness of fit is very important, esp. if I want to demonstrate that I know what I am doing to my MS committee. Also, I am not psyched about making the wrong choice of population distribution. This would not help be successful.

When I first got started, I was estimating parameters using the distribution wizard with the MLE technique. I then went through each individual distribution and viewed the goodness of fit results for both the KS and X^2. I got horrible results, in sum cases 99% for both and for all distributions.

SO then I got the bright idea to begin grouping the data by progressively larger buckets in an attempt to smooth the data until I got a KS < 5%.

I would like some advice on how I really should be proceeding.

Dr. Dave Olwell
June 21st, 2001, 08:35 AM
I am not sure why you are binning your data. Weibull++ does not need the data binned; you enter the actual failure times.

Did you not believe the distribution wizard? Were the probability plots suspect? I don't understand why you started binning the data instead of analyzing the complete set. Binning throws away information.

Dr. Dave Olwell