============
Description:
============

  This is the main function to perform Functional Principal Component Analysis
  for generalized logitudinal observations. The non-Gaussian observations Y, 
  which are repeatedly collected for a sample of individuals over time are
  modeled with latent Gaussian processes X.

Includes the following steps:

1) Estimation of the mean function for Y and X

2) Estimation of the covariance functions for Y and X

3) Calculation of the number of selected components and the overdispersion parameter

4) Computing the smoothed individual trajectories for X and Y

========
Usage:
========

    [yy] = FPCfglm(y,t,p)

======
Input: 
======
     y:    1*n cell array, y{i} is the vector of non-Gaussian observations for the ith subject, i=1,...,n.
     t:    1*n cell array, t{i} is the vector of time points for the ith subject on which
           corresponding observations y{i} are taken, i=1,...,n.
     p:    a struct obtained from setOptions.m sets the rest of arguments for PCAfglm.m
           ex:
           >> p = setOptions();
           type 'names(p)' to see the option names included in p.
           Most of the options are similar to those used in FPCA.m; please refer to the help file of PAC.m.
           Special options for FPCfglm.m:
               family: a character string naming the distribution of the repeated observations.
                       Accepted families include: 'binomial', 'poisson'.
               param:  parameter in the designated family. default is [] (no estra parameter needed).

=======
Output:  
=======  
    yy:    a cell array that contains the following values:
               no_opt:   integer, automatically or subjectively selected
                         value of K, the number of selected components.

               sigma:    scalar, estimate of overdispersion parameter.

               xlambda:  1*K vector, estimated eigenvalues of latent Gaussian process X
                         (variances of functional principal components scores).

               xphi:     N*K matrix, estimated principal component
                         functions for latent Gaussian process X,
                         valued at distinct input time points with ascending
                         order of all subjects, corresponding to out1.

               xeigen:   ngrid*K matrix, estimated principal component functions 
                         for latent Gaussian process X,
                         valued at out21, ngrid of the pooled distinct time points
                         with ascending order of all subjects,
                         xphi is an interpolated version of xeigen at out1.

               xi_est:   n*K matrix, predictions for random coeffecients (PC
                         scores) for n subjects.

               ymu:       1*N vector, estimated mean functions for observed non-Gaussian process Y,
                          valued at distinct input time points (newdata = []), in ascending order from
                          all subjects, corresponding to out1; when newdata is defined,
                          corresponds to the time points from newdata, same as out1.

               xmu:       1*N vector, estimated mean functions for latent Gaussian process X,
                          valued at distinct input time points (newdata = []), in ascending order from
                          all subjects, corresponding to out1; when newdata is defined,
                          corresponds to the time points from newdata, same as out1.

               bw_mu:     scalar(>0), automatically or subjectively selected
                          bandwidth for smoothing mean curve.

               ycov:      ngrid*ngrid matrix, smoothed covariance surface for observed non-Gaussian 
                          process Y,corresponding to out21.

               xcov:      ngrid*ngrid matrix, smoothed covariance surface for latent Gaussian process X,
                          corresponding to out21.

               bw_cov:    1*2 vector(>0), automatically or subjectively selected
                          bandwidths for smoothing covariance surface.

               xcovfit:   ngrid * ngrid matrix, fitted covariance surface for latent Gaussian process X, 
                          based on truncated estimate of eigenvalues ("xlambda") and
                          principal component functions ("xeigen"), corresponding to out21.
                          
               AIC:      1*maxk vector, AIC values obtained when choosing K from
                         K=1 to K=maxk, where AIC(K) is the minimum. If AIC
                         method is not applied, it is [].

               BIC:      1*maxk vector, BIC values obtained when choosing K from
                         K=1 to K=maxk, where BIC(K) is the minimum. If BIC
                         method is not applied, it is [].

               FVE:      1*ngrid vector of fraction of variance explained.

               x_pred:   1*n cell array, x_pred{i} is the vector of predictions for
                         the latent Gaussian process X for the ith subject,
                         evaluated at time points from the output grid vector "out1".

               y_pred:   1*n cell array, y_pred{i} is the vector of predictions for
                         the observed non-Gaussian process Y for the ith subject,
                         evaluated at time points from the output grid vector "out1".

            x_predOrig:  1*n cell array, x_predOrig{i} is the vector of predictions
                         for the latent Gaussian process X for the ith subject,
                         evaluated at the same time points as the input.

            y_predOrig:  1*n cell array, y_predOrig{i} is the vector of predictions
                         for the observed non-Gaussian process Y for the ith subject,
                         evaluated at the same time points as the input.

               out1:     1*N vector, distinct input time points with ascending
                         order from all subjects if p.newdata = []; otherwise, it
                         is the same as p.newdata.

               out21:    1*ngrid vector, a grid of time points for which the
                         smoothed covariance surface assumes values, i.e.,
                         ngrids from out1.

  To see an example, check with exampleFglm.m