compute_roots for only some of the roots?

[BarkleyBG]

I think there may be a situation where someone has, say,

• GLM model (p=2 params)
• another parameter (i.e. delta) to estimate that may be most easily handled with an equation solver.

For the model the easiest thing is to probably fit the model & pass it into geex. But for delta., you may want to use the root solver.
It may be nice in this case to allow for compute_roots = c(FALSE, FALSE, TRUE) so that the (p=2) model parameters are ignored in the solver, but the solver does solve for delta_hat.

This is a future-feature; nothing urgent. Your thoughts?

[bsaul]

This feature would really depend more on the root finding function than geex itself I think. It's this functionality built into rootSolve? To be honest, I've looked at the documentation that closely, but I have a vague recollection of this feature.

[BarkleyBG]

I am nearly certain that rootSolve has this feature. I will reply with more info in a few days. - Brian Sent from my phone; please excuse any typos On Sep 15, 2017, at 11:34 AM, Bradley Saul <notifications@github.com<mailto:notifications@github.com>> wrote: This feature would really depend more on the root finding function than geex itself I think. It's this functionality built into rootSolve? To be honest, I've looked at the documentation that closely, but I have a vague recollection of this feature. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#39 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AMTSsraRNAIs7jXlXBGbuaMz4WqUZSc4ks5siqcngaJpZM4PSCU1>.

[BarkleyBG]

Briefly looking into the docs, maybe `rootSolve` doesn’t have this feature added in. It does have the argument `positive`, which I think I was thinking of. In the docs for `rootSolve::stode()` it says:

positive: either a logical or a vector with indices of the state variables that have to be nonnegative;

if TRUE, all state variables y are forced to be non-negative numbers. However, I think that in this issue ticket we would need to devise a function that takes `compute_roots = c(FALSE, FALSE, TRUE)` and splits the parameters manually. I think that it can be done with `multiroot()` but each other solver would need its own helper function, I think. Does that seem right? - - - - - - - - - - - - - - - - - - Brian Barkley, MS PhD Candidate, Department of Biostatistics UNC Gillings School of Global Public Health Email: Brian.Barkley@UNC.edu From: Bradley Saul [mailto:notifications@github.com] Sent: Friday, September 15, 2017 12:35 PM To: bsaul/geex Cc: Barkley, Brian G; Author Subject: Re: [bsaul/geex] compute_roots for only some of the roots? (#39) This feature would really depend more on the root finding function than geex itself I think. It's this functionality built into rootSolve? To be honest, I've looked at the documentation that closely, but I have a vague recollection of this feature. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#39 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AMTSsraRNAIs7jXlXBGbuaMz4WqUZSc4ks5siqcngaJpZM4PSCU1>.

[bsaul]

On further reflection, I don't know if this is possible with multiroot or even in general. The issue: in order to find roots using Newton-Raphson, multiroot is computing the Jacobian of the G_m = sum(psi(theta)) function. If some element of theta is "fixed", then the root solver can't compute the derivative with respect to that element of theta. How might this affect finding root with respect to the non-fixed elements of theta?

[BarkleyBG]

Perhaps I’m reading this wrong. Let’s say theta is of length 2 and you want to solve for the 2nd element while keeping the 1st element fixed. Could you supply the 1st element as a fixed parameter/argument to the solver, and then ask to solve only for the 2nd element? In crappy pseudo-code, ##theta == c(fixed, to_solve_for) solved <- solver( starting=to_solve_for, function=myFun, other_args = list(fixed=fixed)) - - - - - - - - - - - - - - - - - - Brian Barkley, MS PhD Candidate, Department of Biostatistics UNC Gillings School of Global Public Health Email: Brian.Barkley@UNC.edu From: Bradley Saul [mailto:notifications@github.com] Sent: Monday, September 18, 2017 11:00 AM To: bsaul/geex Cc: Brian G. Barkley; Author Subject: Re: [bsaul/geex] compute_roots for only some of the roots? (#39) On further reflection, I don't know if this is possible with multiroot or even in general. The issue: in order to find roots using Newton-Raphson, multiroot is computing the Jacobian of the G_m = sum(psi(theta)) function. If some element of theta is "fixed", then the root solver can't compute the derivative with respect to that element of theta. How might this affect finding root with respect to the non-fixed elements of theta? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#39 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AMTSsqHMJj_Qb2vhYPymZVGF_TEC_h80ks5sjoVogaJpZM4PSCU1>.

[bsaul]

Let's discuss in person tomorrow. I need to discuss using math.

…

On Mon, Sep 18, 2017 at 11:08 AM, Brian G. Barkley ***@***.*** > wrote: Perhaps I’m reading this wrong. Let’s say theta is of length 2 and you want to solve for the 2nd element while keeping the 1st element fixed. Could you supply the 1st element as a fixed parameter/argument to the solver, and then ask to solve only for the 2nd element? In crappy pseudo-code, ##theta == c(fixed, to_solve_for) solved <- solver( starting=to_solve_for, function=myFun, other_args = list(fixed=fixed)) - - - - - - - - - - - - - - - - - - Brian Barkley, MS PhD Candidate, Department of Biostatistics UNC Gillings School of Global Public Health Email: ***@***.*** From: Bradley Saul ***@***.*** Sent: Monday, September 18, 2017 11:00 AM To: bsaul/geex Cc: Brian G. Barkley; Author Subject: Re: [bsaul/geex] compute_roots for only some of the roots? (#39) On further reflection, I don't know if this is possible with multiroot or even in general. The issue: in order to find roots using Newton-Raphson, multiroot is computing the Jacobian of the G_m = sum(psi(theta)) function. If some element of theta is "fixed", then the root solver can't compute the derivative with respect to that element of theta. How might this affect finding root with respect to the non-fixed elements of theta? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<https://github.com/ bsaul/geex#39#issuecomment-330251094>, or mute the thread< https://github.com/notifications/unsubscribe- auth/AMTSsqHMJj_Qb2vhYPymZVGF_TEC_h80ks5sjoVogaJpZM4PSCU1>. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#39 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AEqCLh0L9M1A6v6Nz9V6z4iUI3eRZImzks5sjodWgaJpZM4PSCU1> .