[Rd] Potential bug in fitted.nls

[J C Nash]

nls() actually uses different modeling formulas depending on the 'algorithm', and
there is, in my view as a long time nonlinear modeling person, an unfortunate
structural issue that likely cannot be resolved simply. This is because for nonlinear
modeling programs we really should be using explicit model statements
e.g., a linear model should be y ~ a * x + b where x is the (independent)
variable and a and b are parameters. But we put in y ~ x as per lm().
Partially linear approaches and indexed parameter models add complexity and
inconsistency. We're pushing structures beyond their design.

This is one of the topics in a paper currently undergoing
final edit on "Improving nls()" that came out of a Google Summer of Code project.
The desired improvements in nls() were mostly frustrated by entanglements in the code,
but they have led to a lot of tweaks to package nlsr. Perhaps someone more facile
with the intricacies of R internals can succeed. For the paper and nlsr,
all the bits should get sent to CRAN and elsewhere in the next month or so,
(co-author is newly started on a Ph D) but if anyone is anxious to try, they can email
me. The nlsr code has been stable for several months, but some documentation still
being considered.

Sorting out how to deal with the model expression for nls() and related tools
is a worthwhile goal, but not one that can be settled here. It could make a good
review project for a senior undergrad or master's level, and I'd be happy to
join the discussion.

Cheers, JN


On 2023-01-26 12:55, Bill Dunlap wrote:
> Doesn't nls() expect that the lengths of vectors on both sides of the
> formula match (if both are supplied)?  Perhaps it should check for that.
> 
> -Bill
> 
> On Thu, Jan 26, 2023 at 12:17 AM Dave Armstrong <darmst46 using uwo.ca> wrote:
> 
>> Dear Colleagues,
>>
>> I recently answered [this question]() on StackOverflow that identified
>> what seems to be unusual behaviour with `stats:::nls.fitted()`. In
>> particular, a null model returns a single fitted value rather than a
>> vector of the same fitted value of `length(y)`.  The documentation
>> doesn’t make it seem like this is the intended behaviour, so I’m not
>> sure if it’s a bug, a “Wishlist” item or something that is working
>> as intended even though it seems unusual to me.  I looked through the
>> bug reporting page on the R project website and it suggested contacting
>> the R-devel list in cases where the behaviour is not obviously a bug to
>> see whether others find the behaviour equally unusual and I should
>> submit a Wishlist item through Bugzilla.
>>
>> Below is a reprex that shows how the fitted values of a model with just
>> a single parameter is length 1, but if I multiply that constant by a
>> vector of ones, then the fitted values are of `length(y)`.  Is this
>> something that should be reported?
>>
>> ``` r
>> dat <-
>> data.frame(y=c(80,251,304,482,401,141,242,221,304,243,544,669,638),
>>                     ones = rep(1, 13))
>> mNull1 <- nls(y ~ a, data=dat, start=c(a=mean(dat$y)))
>> fitted(mNull1)
>> #> [1] 347.6923
>> #> attr(,"label")
>> #> [1] "Fitted values"
>>
>> mNull2 <- nls(y ~ a*ones, data=dat, start=c(a=mean(dat$y)))
>> fitted(mNull2)
>> #>  [1] 347.6923 347.6923 347.6923 347.6923 347.6923 347.6923 347.6923
>> 347.6923
>> #>  [9] 347.6923 347.6923 347.6923 347.6923 347.6923
>> #> attr(,"label")
>> #> [1] "Fitted values"
>> ```
>>
>> Created on 2023-01-25 by the [reprex
>> package](https://reprex.tidyverse.org) (v2.0.1)
>>          [[alternative HTML version deleted]]
>>
>> ______________________________________________
>> R-devel using r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-devel
>>
> 
> 	[[alternative HTML version deleted]]
> 
> ______________________________________________
> R-devel using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel