JMP Time Series Outlier Analysis

[#1]

Hi,

I was wondering if JMP has time series outlier analysis. I have a set of data for failures vs. time and like to identify  the outlier.

Thanks

Adam

[#2 · Accepted solution]

Hi Adam,
I'd recommend the following and let me know if it works for you.
Use the Principal Components platform and input your failure and time variables as the Y variables (both need to be numeric continuous).
From the PCA report menu, select Outlier Analysis. The T² plot shown will give you the distance to the multivariate mean accounting for the structure of the failures vs time data. Points with high T² values should likely be the outliers (above the UCL line).
Hope this helps!
Mark

[#3]

Hi Mark,

Thanks so much for your quick response and solution. I followed your suggestion and I can get the outliers.

What is the advantage of PCA outlier detection vs. doing IQR analysis? IQR also finds the same outlier. 

If I do PCA outside of JMP, how do I find the outliers? 

Thanks again.

Regards,

Adam

[#4]

Hi Adam,
The PCA will account for the covariance along the principal component axes where the IQR does not. Basically it's the difference between identifying outliers in a multivariate space (where one dimension is time) vs a univariate space.

Just to note: the PCA method may not be effective if there is a lot of non-linear behavior in your time series. Another way which will be more flexible is fit a split model using Fit Y by X (or Graph Builder). From the Fit Y by X Bivariate menu, select Flexible > Kernel Smoother. Choose a smoothness you prefer, then from the red triangle next to the Local Smoother in the legend, select save residuals. You can then apply a data filter on those residuals to filter outliers with a more flexible model than using principal components.

Regards and stay safe,

Mark

[#5]

Hi Mark,

Thanks for providing the explanation for using PCA vs IQR and also pointing the importance of nonlinearity. I tried your suggestion for kernel smoother and got the residual and then it used outlier analysis on residuals and got the outlier point.

Thanks again for your help and suggestions.

Regards,

Adam