Skip to content

gam.check() NA results (k-index, p-value) of a gam logistic regression model

3 messages · Simon Wood, Fotis Fotiadis

#
Hello all

I am using bam for a mixec-effects logistic regression model:

b0<-bam(acc~ 1 + igc + s(ctrial, by=igc) + s(sbj, bs="re") + s(ctrial, sbj,
bs="re") , data=data, family=binomial)
Family: binomial
Link function: logit

Formula:
acc ~ 1 + igc + s(ctrial, by = igc) + s(sbj, bs = "re") + s(ctrial,
    sbj, bs = "re")

Parametric coefficients:
              Estimate Std. Error z value Pr(>|z|)
(Intercept)     2.8334     0.2030  13.955  < 2e-16 ***
igcPA.pseudo    0.4692     0.1285   3.650 0.000262 ***
igcCAT.ideo     0.3276     0.2906   1.127 0.259734
igcCAT.pseudo   0.6701     0.2945   2.275 0.022888 *
---
Signif. codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1

Approximate significance of smooth terms:
                           edf Ref.df Chi.sq  p-value
s(ctrial):igcPA.ideo     3.827  4.733  295.0  < 2e-16 ***
s(ctrial):igcPA.pseudo   3.317  4.110  356.1  < 2e-16 ***
s(ctrial):igcCAT.ideo    3.979  4.911  308.6  < 2e-16 ***
s(ctrial):igcCAT.pseudo  4.937  5.974  383.8  < 2e-16 ***
s(sbj)                  54.326 62.000 3032.8  < 2e-16 ***
s(ctrial,sbj)           43.045 62.000 2706.6 1.31e-08 ***
---
Signif. codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1

R-sq.(adj) =  0.362   Deviance explained = 38.9%
fREML =  25436  Scale est. = 1         n = 18417


I want to know if the wigglyness of the smooths [s(ctrial, by=igc)] is
appropriate, so I used the gam.check() function. The values though for
k-index and p-value are NAs:
Method: fREML   Optimizer: perf newton
full convergence after 5 iterations.
Gradient range [-7.60152e-08,8.12795e-06]
(score 25436.12 & scale 1).
Hessian positive definite, eigenvalue range [0.6271375,24.46625].
Model rank =  168 / 168

Basis dimension (k) checking results. Low p-value (k-index<1) may
indicate that k is too low, especially if edf is close to k'.

                           k'   edf k-index p-value
s(ctrial):igcPA.ideo     9.00  3.83      NA      NA
s(ctrial):igcPA.pseudo   9.00  3.32      NA      NA
s(ctrial):igcCAT.ideo    9.00  3.98      NA      NA
s(ctrial):igcCAT.pseudo  9.00  4.94      NA      NA
s(sbj)                  64.00 54.33      NA      NA
s(ctrial,sbj)           64.00 43.04      NA      NA

Does anyone know why is this?

Thank you in advance for your time,
Fotis

P.S. I am using RStudio Version 0.99.896, R 3.3.0, and mgcv package version
1.8.12.
--
PhD Candidate
Department of Philosophy and History of Science
University of Athens, Greece.
http://users.uoa.gr/~aprotopapas/LLL/en/members.html#fotisfotiadis

Notice: Please do not use this account for social networks invitations, for
sending chain-mails to me, or as it were a facebook account. Thank you for
respecting my privacy.

<https://www.avast.com/sig-email?utm_medium=email&utm_source=link&utm_campaign=sig-email&utm_content=webmail>
???????????
??? ????. www.avast.com
<https://www.avast.com/sig-email?utm_medium=email&utm_source=link&utm_campaign=sig-email&utm_content=webmail>
<#DDB4FAA8-2DD7-40BB-A1B8-4E2AA1F9FDF2>
#
Dear Fotis,

The test is a randomization test, based on comparing differences of 
residuals, ordered with respect to the covariate of the smooth, to 
differences of residuals in randomized order. Random effect terms are 
excluded because there is not basis size to choose. Currently smooths 
with factor by variables are also excluded for reasons of maintainer 
laziness, as this would require special case code to exclude the 
covariate values that are irrelevant given the factor level. Sorry about 
that.

My guess is that you don't have a problem here anyway, given the fairly 
low edfs relative to the basis dimension. In general as a double check I 
would plot the residuals against ctrial, colour coded by level of igc, 
just to check that there doesn't seem to be missed pattern in them. 
However with binary residuals you are unlikely to see much.

best,
Simon
On 17/05/16 20:39, Fotis Fotiadis wrote:

  
    
#
Dear Prof. Wood,

Thank you for your reply!

Best,
Fotis
On Wed, May 18, 2016 at 12:05 PM, Simon Wood <simon.wood at bath.edu> wrote: