Help for AIC and REML score

[iandanilevicz]

Hi all,

I have a problem with the AIC when I make a pffr model. I fitted a model like this:

pffr_adj_fit1 = pffr(MIMS ~ age + gender + BMI, data = nhanes_ave[1:500,])

and I found the AIC value like this:

pffr_adj_fit1$aic

However, how is this value calculated? Because I found an astronomical value that doesn't coincide with the formula presented by Krivobokova (2007) or Crainiceanu (2024).

n=500

n*log(pffr_adj_fit1$deviance) + 2*sum(pffr_adj_fit1$edf) # Krivobokova equation 4

log(pffr_adj_fit1$deviance) + 2*sum(pffr_adj_fit1$edf)/n # Crainiceanu page 42

I have a second issue regarding how the REML score is calculated. I can find it, but it doesn't match any proxy that I have.

pffr_adj_fit1$gcv.ubre

References: Krivobokova, T., & Kauermann, G. (2007). “A Note on Penalized Spline Smoothing with Correlated Errors.” Journal of the American Statistical Association, 102(480), 1328–1337, doi: 10.1198/016214507000000978

Crainiceanu, C.M. et al. (2024). "Functional Data Analysis with R". CRC Press, doi: 10.1201/9781003278726

[jaimelopez98]

Hi, I tried to achieve the values of AIC and GCV with a simple pffr model using the notation introduced in the help (gam {mgcv}) without success. My code is:

n = 2684 mod1 = pffr(y ~x, method = "GCV.Cp" mod1$aic #143591443 mod1$deviance / n+ 2*mod1$scale.estimated*sum(mod1$edf)/ n- mod1$scale.estimated # 252615346 mod1$gcv.ubre #65786.27 (n * mod1$deviance) /((n-sum(mod1$edf))^2) #268332499

To check my reference, it is available in the R: ?mgcv::gam

PS. Unfortunately, I didn't solve the problem.