Extend FACE method to irregularly observed data

[mynanshan]

refund/R/fpca.face.R

Line 159 in eb6af5c

stopifnot(!is.null(Y))

All FPCA functions in the package refund provide two arguments for data input, Y for a observation matrix observed on a common grid and ydata for irregularly observed data. However, the ydata option is not always available, for example, when using fpca.face. I guess the reason is that the method proposed by Xiao 2016 is only developed for regular data.

According to Xiao 2013 and a more recent paper, Xiao 2020, the bivariate smoothing can also be applied to irregularly observed data. Therefore, I wonder if the fpca.face method can be extended accordingly.

[fabian-s]

Hi @mynanshan -- they put this stuff in their own separate package, you can find a FACE implementation for sparse/irregular data here: https://cran.r-project.org/web/packages/face/index.html

[fabian-s]

@jeff-goldsmith @julia-wrobel: are we ever going to re-implement this or import and re-export the face-package?
If not, feel free to close this issue.

(Also: I think it might be worthwhile to do a review/benchmarking paper on methods for functional covariance operator estimation sometime, there are SO MANY and there are lots of interesting complications/nuances......)

[mynanshan]

Hi @mynanshan -- they put this stuff in their own separate package, you can find a FACE implementation for sparse/irregular data here: https://cran.r-project.org/web/packages/face/index.html

Thanks for the reference!

[jeff-goldsmith]

I feel a bit strange having a somewhat-out-of-date and unsupported function in refund. I suppose we could remove fpca.face from refund (in a staggered way, with a message and pointer to the face package).

I'd be okay with importing and exporting the face package as part of refund -- already have oodles of dependencies and this seems like a function folks use -- but don't know if that's necessary.