hello
I'm working on a new technology in urology, which can significantly shorten operation time.
I have a scatterplot and I have created LOESS ( locally estimated scatterplot smoothing ) , with an obvious break-point. There is a sharp change in slopes before and after this point.
What I'm looking for is a automatic, objective way to identify this point, rather pointing it visually only. I'm working in SPSS 23.0
1. Is there any ready syntax for SPSS to do it?
2. Is it makes any different if the slopes are linear\non-linear?
3. Is there any way to compare the slopes before and after the break-point, with a p-value for significance?
thank you all
This is a pretty comprehensive tutorial on piecewise regression (using SAS)
https://www.fs.fed.us/rm/pubs/rmrs_gtr189.pdf
And break points are a classic problem that is now often analysed by Bayesian procedures
Article An objective Bayesian analysis of the change point problem
including discrete outcomes and multiple change points:
https://www.nyu.edu/projects/spirling/documents/PAadvance.pdf
It is possible to do a two line regression. The ability to fit will depend upon the scatter in the data. Not familiar with functions in SPSS that might perform such a fit.
I don't think there is a native SPSS command to do this. One way to do it is described in this Tech Note:
- http://www-01.ibm.com/support/docview.wss?uid=swg21476694
It may be that there are other methods involving Python or R programmability. I suggest that you post your question to the SPSS mailing list (see link below). Some members of that forum may be able to help you. Click on "more options" at the top of the page for instructions on how to join the list.
- http://spssx-discussion.1045642.n5.nabble.com/
HTH.
Are you wanting to do this with loess or with a hockey/broken stick model of the form: y_i = B0 + B1 x_i + B2 (x_i > B3) + e_i? (assuming x_i scaled just the one intercept)
Framed like this you are trying to estimate the optimal breakpoint (B3) and then you are wanting to test if B2 = 0 for this value. Is this what you are after? As Bruce Weaver shows in those links, there are ways to get at these constrained non-linear regressions. With loess much depends on how you are using that.
This is a pretty comprehensive tutorial on piecewise regression (using SAS)
https://www.fs.fed.us/rm/pubs/rmrs_gtr189.pdf
And break points are a classic problem that is now often analysed by Bayesian procedures
Article An objective Bayesian analysis of the change point problem
including discrete outcomes and multiple change points:
https://www.nyu.edu/projects/spirling/documents/PAadvance.pdf
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS