PDA

View Full Version : Calculation Choices & Errors


Aurora Tiffany-Davis
August 15th, 2007, 05:00 PM
I have 2 sets of data each comprising 90 datapoints.

I am using the Weibull-2 distribution (as a default). Given the large number of samples, I chose the MLE analysis method and the Fisher Matrix confidence bound method. Because there are no intervals, I chose the SRM regression method.

Attempting this calculation, I get 2 error messages. One is that Beta > 50 and thus W2 is not the most appropriate distribution, the other reads, "MLE Method cannot converge to reasonable values. This might be due to unreasonable or insufficient data."

If I run the Distribution Wizard, W2 comes up as #5 preference for distributions. However, if I simply change from MLE to RRX, the calculation is successful. Alternatively, if I change from W2 to either Normal or Logistic, but keep MLE, the calculation is successful. Note: the W2-calculated PDF for either dataset does not look exactly "Normal". For this reason I have chosen to keep W2 and change to RRX.

Ultimately though, I don't know *why*. I just fiddle with the settings until it works. Since the results of my analysis will impact timeline and budget, I would like to be more confident in my own analysis, especially when deviating from the recommended settings for the type of dataset I am working with. Any suggestions?

So, now with W2, RRX, FM, SRM settings in place, I compare the results of the 2 datasets. All looks okay except that the 2nd dataset is unable to produce a contour plot, at any confidence level. The only available explanation is "Error In Calculations". The engineers I will report my findings to are somewhat familiar with contour plots and although the Test of Comparison Tool says that "... will last longer with a probability of 94%..." it would be great to also show dataset divergence with a Contour Plot. Not to mention, the simple fact that I cannot produce one with my dataset and settings further erodes my confidence.

So, I can keep fiddling around and changing this and that until I find settings that, for both datasets, successfully calculate and can produce all plots. But somehow that seems like cheating, and a little less than scientific. Any help?

Raw data is attached.

Adamantios
August 16th, 2007, 04:42 PM
I assume that you are using W++ version 6. In version 6 since when it was developed there were still 16bit operating systems, we had a lower threshold on some parameters to prevent big numbers that could not be handled. However, since many systems were already 32bit, we had the oportunity to increase this threshold. So we added an option in the User Setup (File->User Setup...). If you go to the Calculations Page in the User Setup, select the option "Use Higher Numerical Precision on Calculations." If you do that and recalculate your dataset, MLE will converge with no problems.

Regarding the Contour Plots, in the dataset that you attached, i was able to obtain Contour Plots with no problem. So I guess the problem is in the other dataset, but I don't have it to check what is going on there. If you want you can attach it here, or email it to us at: support@reliasoft.com.

Adamantios
August 16th, 2007, 04:51 PM
PS on the contour plots: Please remember that Contour plots are actually valid when using MLE, and if you use regression the contours may not always be able to generate. This is inherent when using Likelihood theory with regression, but if they can be generated they can still provide means of comparison and not be restricted to always use MLE if you want to plot the contours.

I hope this helps, and please let us know if you have nay further questions. Our support line is available to answer any of your questions. You can call us toll free at 520-886-0366, or email us at support@reliasoft.com.

Thanks.

Aurora Tiffany-Davis
August 18th, 2007, 11:45 AM
Thank you for the suggestion but I am using Weibull ++7 and that option is not available.

David
August 21st, 2007, 07:48 AM
Hi Aurora,

Do you have any additional questions regarding your analysis? I just want to make sure that any confusion has been cleared up.