R: Impact point selection with PVS

PVS.fit {fsemipar}

R Documentation

Impact point selection with PVS

Description

This function implements the Partitioning Variable Selection (PVS) algorithm. This algorithm is specifically designed for estimating multivarite linear models, where the scalar covariates are derived from the discretisation of a curve.

PVS is a two-stage procedure that selects the impact points of the discretised curve and estimates the model. The algorithm employs a penalised least-squares regularisation procedure. Additionally, it utilises an objective criterion (criterion) to determine the initial number of covariates in the reduced model (w.opt) of the first stage, and the penalisation parameter (lambda.opt).

Usage

PVS.fit(z, y, train.1 = NULL, train.2 = NULL, lambda.min = NULL, 
lambda.min.h = NULL, lambda.min.l = NULL, factor.pn = 1, nlambda = 100, 
vn = ncol(z), nfolds = 10, seed = 123, wn = c(10, 15, 20), range.grid = NULL, 
criterion = "GCV", penalty = "grSCAD", max.iter = 1000)

Arguments

`z`	Matrix containing the observations of the functional covariate collected by row (linear component).
`y`	Vector containing the scalar response.
`train.1`	Positions of the data that are used as the training sample in the 1st step. The default setting is `train.1<-1:ceiling(n/2)`.
`train.2`	Positions of the data that are used as the training sample in the 2nd step. The default setting is `train.2<-(ceiling(n/2)+1):n`.
`lambda.min`	The smallest value for lambda (i. e., the lower endpoint of the sequence in which `lambda.opt` is selected), as fraction of `lambda.max`. The defaults is `lambda.min.l` if the sample size is larger than `factor.pn` times the number of linear covariates and `lambda.min.h` otherwise.
`lambda.min.h`	The lower endpoint of the sequence in which `lambda.opt` is selected if the sample size is smaller than `factor.pn` times the number of linear covariates. The default is 0.05.
`lambda.min.l`	The lower endpoint of the sequence in which `lambda.opt` is selected if the sample size is larger than `factor.pn` times the number of linear covariates. The default is 0.0001.
`factor.pn`	Positive integer used to set `lambda.min`. The default value is 1.
`nlambda`	Number of values in the sequence from which `lambda.opt` is selected. The default is 100.
`vn`	Positive integer or vector of positive integers indicating the number of groups of consecutive variables to be penalised together. The default value is `vn=ncol(z)`, resulting in the individual penalization of each scalar covariate.
`nfolds`	Number of cross-validation folds (used when `criterion="k-fold-CV"`). Default is 10.
`seed`	You may set the seed for the random number generator to ensure reproducible results (applicable when `criterion="k-fold-CV"` is used). The default seed value is 123.
`wn`	A vector of positive integers indicating the eligible number of covariates in the reduced model. For more information, refer to the section `Details`. The default is `c(10,15,20)`.
`range.grid`	Vector of length 2 containing the endpoints of the grid at which the observations of the functional covariate `x` are evaluated (i.e. the range of the discretisation). If `range.grid=NULL`, then `range.grid=c(1,p)` is considered, where `p` is the discretisation size of `x` (i.e. `ncol(x))`.
`criterion`	The criterion used to select the tuning and regularisation parameters: `wn.opt` and `lambda.opt` (also `vn.opt` if needed). Options include `"GCV"`, `"BIC"`, `"AIC"`, or `"k-fold-CV"`. The default setting is `"GCV"`.
`penalty`	The penalty function applied in the penalised least-squares procedure. Currently, only "grLasso" and "grSCAD" are implemented. The default is "grSCAD".
`max.iter`	Maximum number of iterations allowed across the entire path. The default value is 1000.

Details

The sparse linear model with covariates coming from the discretization of a curve is given by the expression

Y_i=\sum_{j=1}^{p_n}\beta_{0j}\zeta_i(t_j)+\varepsilon_i,\ \ \ (i=1,\dots,n)

where

Y_i is a real random response and \zeta_i is assumed to be a random curve defined on some interval [a,b], which is observed at the points a\leq t_1<\dots<t_{p_n}\leq b.
\mathbf{\beta}_0=(\beta_{01},\dots,\beta_{0p_n})^{\top} is a vector of unknown real coefficients.
\varepsilon_i denotes the random error.

In this model, it is assumed that only a few scalar variables from the set \{\zeta(t_1),\dots,\zeta(t_{p_n})\} are part of the model. Therefore, the relevant variables (the impact points of the curve \zeta on the response) must be selected, and the model estimated.

In this function, this model is fitted using the PVS. The PVS is a two-steps procedure. So we divide the sample into two independent subsamples, each asymptotically half the size of the original sample (n_1\sim n_2\sim n/2). One subsample is used in the first stage of the method, and the other in the second stage.The subsamples are defined as follows:

\mathcal{E}^{\mathbf{1}}=\{(\zeta_i,\mathcal{X}_i,Y_i),\quad i=1,\dots,n_1\},

\mathcal{E}^{\mathbf{2}}=\{(\zeta_i,\mathcal{X}_i,Y_i),\quad i=n_1+1,\dots,n_1+n_2=n\}.

Note that these two subsamples are specified to the program through the arguments train.1 and train.2. The superscript \mathbf{s}, where \mathbf{s}=\mathbf{1},\mathbf{2}, indicates the stage of the method in which the sample, function, variable, or parameter is involved.

To explain the algorithm, we assume that the number p_n of linear covariates can be expressed as follows: p_n=q_nw_n, with q_n and w_n being integers.

First step. A reduced model is considered, discarding many linear covariates. The penalised least-squares procedure is applied to the reduced model using only the subsample \mathcal{E}^{\mathbf{1}}. Specifically:
- Consider a subset of the initial p_n linear covariates, containing only w_n equally spaced discretized observations of \zeta covering the interval [a,b]. This subset is the following:
  
  \mathcal{R}_n^{\mathbf{1}}=\left\{\zeta\left(t_k^{\mathbf{1}}\right),\ \ k=1,\dots,w_n\right\},
  
  where t_k^{\mathbf{1}}=t_{\left[(2k-1)q_n/2\right]} and \left[z\right] denotes the smallest integer not less than the real number z. The size (cardinality) of this subset is provided to the program in the argument wn (which contains a sequence of eligible sizes).
- Consider the following reduced model involving only the w_n linear covariates from \mathcal{R}_n^{\mathbf{1}}: \mathcal{R}_n^{\mathbf{1}}:
  
  Y_i=\sum_{k=1}^{w_n}\beta_{0k}^{\mathbf{1}}\zeta_i(t_k^{\mathbf{1}})+\varepsilon_i^{\mathbf{1}}.
  
  The penalised least-squares variable selection procedure is applied to the reduced model using the function lm.pels.fit, which requires the remaining arguments (for details, see the documentation of the function lm.pels.fit). The estimates obtained are the outputs of the first step of the algorithm.
Second step. The variables selected in the first step, along with the variables in their neighborhood, are included. Then the penalised least-squares procedure is carried out again considering only the subsample \mathcal{E}^{\mathbf{2}}. Specifically:
- Consider a new set of variables :
  
  \mathcal{R}_n^{\mathbf{2}}=\bigcup_{\left\{k,\widehat{\beta}_{0k}^{\mathbf{1}}\not=0\right\}}\left\{\zeta(t_{(k-1)q_n+1}),\dots,\zeta(t_{kq_n})\right\}.
  
  Denoting by r_n=\sharp(\mathcal{R}_n^{\mathbf{2}}), we can rename the variables in \mathcal{R}_n^{\mathbf{2}} as follows:
  
  \mathcal{R}_n^{\mathbf{2}}=\left\{\zeta(t_1^{\mathbf{2}}),\dots,\zeta(t_{r_n}^{\mathbf{2}})\right\},
- Consider the following model, which involves only the linear covariates belonging to \mathcal{R}_n^{\mathbf{2}}
  
  Y_i=\sum_{k=1}^{r_n}\beta_{0k}^{\mathbf{2}}\zeta_i(t_k^{\mathbf{2}})+\varepsilon_i^{\mathbf{2}}.
  
  The penalised least-squares variable selection procedure is applied to this model using lm.pels.fit.

The outputs of the second step are the estimates of the model. For further details on this algorithm, see Aneiros and Vieu (2014).

Remark: If the condition p_n=w_n q_n is not met (then p_n/w_n is not an integer), the function considers variable q_n=q_{n,k} values k=1,\dots,w_n. Specifically:

q_{n,k}= \left\{\begin{array}{ll} [p_n/w_n]+1 & k\in\{1,\dots,p_n-w_n[p_n/w_n]\},\\ {[p_n/w_n]} & k\in\{p_n-w_n[p_n/w_n]+1,\dots,w_n\}, \end{array} \right.

where [z] denotes the integer part of the real number z.

Value

`call`	The matched call.
`fitted.values`	Estimated scalar response.
`residuals`	Differences between `y` and the `fitted.values`.
`beta.est`	`\hat{\mathbf{\beta}}` (i. e. estimate of `\mathbf{\beta}_0` when the optimal tuning parameters `w.opt` and `lambda.opt` are used).
`indexes.beta.nonnull`	Indexes of the non-zero `\hat{\beta_{j}}`.
`w.opt`	Selected size for `\mathcal{R}_n^{\mathbf{1}}`.
`lambda.opt`	Selected value of the penalisation parameter `\lambda` (when `w.opt` is considered).
`IC`	Value of the criterion function considered to select `w.opt` and `lambda.opt`.
`beta2`	Estimate of `\mathbf{\beta}_0^{\mathbf{2}}` for each value of the sequence `wn`.
`indexes.beta.nonnull2`	Indexes of the non-zero linear coefficients after the step 2 of the method for each value of the sequence `wn`.
`IC2`	Optimal value of the criterion function in the second step for each value of the sequence `wn`.
`lambda2`	Selected value of penalisation parameter in the second step for each value of the sequence `wn`.
`index02`	Indexes of the covariates (in the entire set of `p_n`) used to build `\mathcal{R}_n^{\mathbf{2}}` for each value of the sequence `wn`.
`beta1`	Estimate of `\mathbf{\beta}_0^{\mathbf{1}}` for each value of the sequence `wn`.
`IC1`	Optimal value of the criterion function in the first step for each value of the sequence `wn`.
`lambda1`	Selected value of penalisation parameter in the first step for each value of the sequence `wn`.
`index01`	Indexes of the covariates (in the entire set of `p_n`) used to build `\mathcal{R}_n^{\mathbf{1}}` for each value of the sequence `wn`.
`index1`	Indexes of the non-zero linear coefficients after the step 1 of the method for each value of the sequence `wn`.
`...`

Author(s)

German Aneiros Perez german.aneiros@udc.es

Silvia Novo Diaz snovo@est-econ.uc3m.es

References

Aneiros, G. and Vieu, P. (2014) Variable selection in infinite-dimensional problems. Statistics & Probability Letters, 94, 12–20, doi:10.1016/j.spl.2014.06.025.

Examples

data(Sugar)

y<-Sugar$ash
z<-Sugar$wave.240

#Outliers
index.y.25 <- y > 25
index.atip <- index.y.25
(1:268)[index.atip]


#Dataset to model
z.sug<- z[!index.atip,]
y.sug <- y[!index.atip]

train<-1:216

ptm=proc.time()
fit<- PVS.fit(z=z.sug[train,], y=y.sug[train],train.1=1:108,train.2=109:216,
        lambda.min.h=0.2,criterion="BIC", max.iter=5000)
proc.time()-ptm

fit 
names(fit)

[Package fsemipar version 1.1.1 Index]