Different Slopes and Intercepts for RobReg (Expressions) and Interpretation regressions

[AllenKennedy]

Below is ExpoTh peek results, and a regression of the same two parameters under Interpretations. There is a difference in the calculated Slope and Intercept of these two regressions. I tested this when both parameters are highly correlated and the slope is close to 1.0, such as Pb/Th vs Pb/ThO and there still is a difference in outcomes. I assume the regression equations are different and we need an explanation for the users.

[bowring]

This is another instance of the random number generator problem that we are seeking to resolve. In addition the 2-D regression shown in the "plot any two" is generated using Noah McLean's (@noahmclean) algorithms and not using RobReg. You have clearly identified a problem we need to resolve.

[bowring]

I experimented by setting the uncertainties to zero prior to using McLean Regression in plot-any-two and get very similar results to RobReg. So it seems that in addition to solving the RobReg issue, we need to decide whether plot-any-two regressions should use McLean or RobReg or both (with a choice). Thanks to @AllenKennedy for your continued testing!

[NicoleRayner]

So "plot any two" is a McLean regression? I think that we should definitely include a RobReg as this is what folks are going to assume is the same as the Expo expression.

Relatedly - and this was the case in squid2.5, you can't exclude analyses from RobReg (manually or auto reject) used to calculate Expo - you can exclude things from the weighted mean of the calib constant, but not the regression.

This can be problematic when you have a RM analysis with a major analytical problem (like a duo flame out) but now in squid3 at least it is easy to remove this analysis from processing in the data audit (or manage samples) windows. I don't think we need to provide this functionality here but a warning to people who see the weighted mean rejects excluded in their plot LnUO/U vs LnPb/U and expect that this will align with their RobReg in expressions.

[noahmclean]

Can I be of any help? What's RobReg? Is McLean Regression (in 2D, same as York and many others) producing different outputs?

[bowring]

@noahmclean - RobReg is the Isoplot version of robust regression that does not include uncertainties and does include the use of random numbers to perturb the slope as discussed in issue #615. There are two basic issues - 1) should RobReg even use random numbers and if so, should it use the same series each time to guarantee reproducibility? and 2) provide both RobReg and McLean regression for users of plot-any2.

[noahmclean]

Got it, thanks, I'll check out RobReg. There are certainly lots of robust regression algorithms out there.

…

________________________________ From: Jim Bowring ***@***.***> Sent: Saturday, May 8, 2021 8:37 AM To: CIRDLES/Squid ***@***.***> Cc: Noah McLean ***@***.***>; Mention ***@***.***> Subject: Re: [CIRDLES/Squid] Different Slopes and Intercepts for RobReg (Expressions) and Interpretation regressions (#617) @noahmclean<https://github.com/noahmclean> - RobReg is the Isoplot version of robust regression that does not include uncertainties and does include the use of random numbers to perturb the slope as discussed in issue #615<#615>. There are two basic issues - 1) should RobReg even use random numbers and if so, should it use the same series each time to guarantee reproducibility? and 2) provide both RobReg and McLean regression for users of plot-any2. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub<#617 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AB2PQSEKB5GYPDXHDCBSF3LTMU5AVANCNFSM44CHSAFA>.