============ Description: ============ This is the main function to implement the Functional Additive Model (FAM), where the predictor is a function X(t_x) and the response can be either a function Y(t_y) or a scalar. Reference: M\"uller, H.G., Yao, F. (2008). Functional additive models. J. American Statistical Association 103, 1534–-1544. IMPORTANT: This model is intended primarily for densely sampled trajectories Includes the following steps: 1) FPCA using the PACE method for X(t_x) and Y(t_y) (in the case that the response is functional) 2) Computation of the additive model components 3) Prediction of response functions for given predictor values 4) Computation of (generalized) R-squared. ======== Usage: ======== [res, xx, yy] = FPCfam(x, t_x, y, t_y, param_X, param_Y, K_x, K_y, GCV, kern, isNewSub) ======= Input: ======= x : 1*n cell array for predictor function x, where x{i} is the row vector of measurements for the ith subject, i=1,...,n. It may contain data for subjects that are used for prediction; this is controlled by "isNewSub" which is either a vector consisting of 0's and 1's according to whether subject is used for prediction (0) or estimation (1), or is controlled by a positive integer nn. In this case, nn is the number of subjects to be used for estimation and n-nn is the number of remaining subjects to be used for prediction, corresponding to the last n-nn data rows. When "isNewSub" is set to [], all n subjects are used for estimation and no prediction will be calculated; see "isNewSub" for more details. t_x : 1*n cell array, t_x{i} is the row vector of time points for the ith subject at which corresponding measurements x{i} are taken, i=1,...,n. It contains subjects that are used for prediction. See above for 2 different cases of "isNewSub" and the definition of "isNewSub" for more details. y : i) When no prediction is requested, that is, isNewSub = [] or isNewSub = 0: 1*n cell array for response function y, y{i} is the response row vector for the ith subject, i = 1,...,n, or a 1*n vector for scalar response y, then y(i) is the response value for the ith subject, i = 1,...,n. ii) When prediction is requested, that is isNewSub is either a vector of 0's (subjects used for estimation) and 1's (subjects used for prediction but not estimation) or a positive integer: Cell array of dim 1*nn (functional response case) or a vector of dim 1*nn (scalar response case). In either case, nn is the number of subjects used for estimation. iii) When prediction is requested and y is 1*n, it will be truncated to 1*nn according to the isNewSub definition below. (see "isNewSub" for more details). t_y : 1*n cell array, t_y{i} is the row vector of time points for the ith subject at which corresponding measurements, y{i} are taken (for each response trajectory), i = 1,...,n, or [] if y is a scalar response. See above for two different cases of "isNewSub" and the definition of "isNewSub" for more details. param_X: an object that is returned by setOptions() that sets the input arguments for FPCA() of the X (predictor) functions (for default, set param_X = []). The default method for choosing the number of principal components is 'AIC_R', which uses the regression AIC criterion, selecting the number of principal components based on the linear relationship between predictor and response (Note that if FIT = -1, the number of principal components of X and Y are jointly selected by 'AIC_R'). If the main goal of the analysis is the estimation of the regression parameter function or surface, suggested settings are FIT = 0 and 'BIC1' for 'selection_k' in param_X. For optimal prediction, it is better to use the default options. For other default values, see setOptions(). param_Y: an object that is returned by setOptions(), that sets the input arguments for FPCA() of the Y (response) functions (for default, set param_Y = []). The default method for choosing the number of principal components for the response trajectories is 'BIC1'. See setOptions() for more details. When y is a scalar, this object will be ignored. K_x: positive integer, not required. Number of principal components of predictor x used in regression, must be smaller than or equal to the maximum number of principal components available from FPCA. Default K_x = [], in this case K_x is the number of components selected by FPCA.m for predictor functions. K_y: positive integer, not required. No effect for the case of a scalar response. The number of principal components for the response y, used for functional regression, must be smaller than or equal to the maximum number of principal components available from FPCA. Default K_y = [], Here K_y is the number of components selected by FPCA.m for response functions. GCV: 1 use generalized cross-validation (GCV) to choose the bandwidth when estimating the additive model components (default); 0 use cross-validation (CV) to choose the bandwidth when estimating the additive model components. kern: a character string for the kernel to be used when estimating the additive model components 'epan': Epanechikov kernel 'gauss': Gaussian kernel (default) 'gausvar': variant of Gaussian kernel 'rect': rectangular kernel 'quar': quartic kernel isNewSub: i) 1*n vector of 0s or 1s, where 1 : the data for the corresponding subject, i.e., x(isNewSub == 1), t_x(isNewSub==1), t_y(isNewSub == 1) are used for prediction only; The count is n-nn for subjects with isNewSub = 1. 0 : the data for the corresponding subject, i.e., x(isNewSub == 0), t_x(isNewSub == 0), y and t_y(isNewSub == 0) are used for estimation only. The count is nn for subjects with isNewSub = 0 This option is convenient for computing leave-one-out prediction if desired. ii) If it is a positive integer, say 5, then the last 5 subjects (in the order from top to bottom within the array x) and their values for t_x, t_y are used for prediction. In other words, when choosing this option, one would append the ``new'' subjects for which one desires prediction of the response to occur at the end of x, t_x and t_y. Then the first nn rows of the arrays x, t_x,t_y and the entire array y (of length nn) are used for estimation and the remainder (last n-nn rows) of x,t_x and t_y (of length n-nn) will be used for prediction. Example: isNewSub = 5; n = length(x) x(1:(n-5)), t_x(1:(n-5)), y, t_y(1:(n-5)) are used for estimation. x((n-4):n), t_x((n-4):n) and t_y((n-4):n) are used for prediction. This option is convenient for obtaining predictions for a set of new subjects. iii) set isNewSub = [] for the default value, which is no prediction. Note: o nn > n - nn, the number of subjects used for estimation should generally be larger than number of subjects used for prediction. o When no prediction is requested, x,t_x, y, t_y will be of length n = nn. o When prediction is requested, x, t_x, t_y will be of length n, which includes nn subjects for estimation and n-nn subjects for prediction. Here y is always of length nn (< n). o When prediction is requested and y is of length n < nn, a warning message will be given and y will be truncated to length nn. This assumes that only nn among the n available data for y are used for estimation and the remaining n-nn observations will be ignored. Details: i) There are no default values for the first 3 arguments, that is, they are always part of the input arguments. Note, t_y can be set to [] when y is a scalar response; ii) Any unspecified or optional arguments can be set to "[]" for default values; iii) FPCA() calls PCA(), so setOptions() sets the input arguments for PCA() and the returned object contains all values returned by PCA(); iv) Names of objects can be retrieved through names() function i.e., names(xx), names(yy), names(res) and the actual values can be retrieved through the getVal() function, example: getVal(res, 'BETA'), getVal(res, 'newy') etc. v) When isNewSub is set to be [], no prediction will be performed, and the number of subjects in x is the same as the number of subjects in y; when isNewSub is either a vector of 0's and 1's or a positive integer, the number of subjects in x is larger than the number of subjects of y, since the former contains new subjects for prediction only, whose response will not be used in the model estimation process. Vi)Default: no prediction will be performed. ======= Output: ======= res: an aggregated object that contains fxi, spcx, newx, new_tx, newy, new_ty, r2, fitted_y, K_x, K_y, isYFun. 1) fxi: a nn*K_x*K_y array containing the estimated additive model components. fxi(:,i,j) is the smoothed function between the j-th component of Y and the i-th component of X. 2) spcx: a nn*K_y matrix containing the sorted component scores of X. 3) newx: 1*numNewSub cell array that contains measurements for new x (predictor) functions. 4) new_tx: 1*numNewSub cell array that contains time points corresponding to newx. 5) newy: 1*numNewSub cell array that contains predicted measurements for corresponding newx. 6) new_ty: 1*numNewSub cell array that contains time points corresponding to newy. 7) r2: the (generalized) R-square If Y is a scalar, r2 = 1-sum((Y_i-Yhat_i)^2)/sum((Y_i-mean(Y))^2) If Y is a function, r2 = 1-\int((Y-Yhat)^2)/\int((Y-mean(Y))^2) NOTE: This is a measure of the fraction of the variation of the responses that is explained by the regression. Similar to the Quasi R-square in FPCreg. 9) fitted_y: 1*nn (same length of y) cell array containing fitted measurements for corresponding y values that were used for estimation; when y is scalar, it is a 1*nn vector. 10) K_x: number of principal components used in regression for predictor X. 11) K_y: number of principal components used in regression for response Y. xx: an aggregated object that contains the returned values from FPCA(x,t_x,param_X). See PCA() or pcaHELP.txt for more details. Use getVal() to retrieve the individual values. yy: an aggregated object that contains the returned values from FPCA(y,t_y,param_Y) or returns "[]" if y is a scalar response (see PCA() or pcaHELP.txt for more details). Use getVal() to retrieve the individual values. o See exampleFam.m for an example of the functional response case, where both predictors and responses are dense, but irregular. See also PCA, FPCA, FPCreg