HFR is a regularized regression estimator that decomposes a least squares regression along a supervised hierarchical graph, and shrinks the edges of the estimated graph to regularize parameters. The algorithm leads to group shrinkage in the regression parameters and a reduction in the effective model degrees of freedom.
Usage
hfr(
x,
y,
weights = NULL,
kappa = 1,
q = NULL,
intercept = TRUE,
standardize = TRUE,
partial_method = c("pairwise", "shrinkage"),
l2_penalty = 0,
...
)Arguments
- x
Input matrix or data.frame, of dimension \((N\times p)\); each row is an observation vector.
- y
Response variable.
- weights
an optional vector of weights to be used in the fitting process. Should be NULL or a numeric vector. If non-NULL, weighted least squares is used for the level-specific regressions.
- kappa
The target effective degrees of freedom of the regression as a percentage of \(p\).
- q
Thinning parameter representing the quantile cut-off (in terms of contributed variance) above which to consider levels in the hierarchy. This can used to reduce the number of levels in high-dimensional problems. Default is no thinning.
- intercept
Should intercept be fitted. Default is
intercept=TRUE.- standardize
Logical flag for x variable standardization prior to fitting the model. The coefficients are always returned on the original scale. Default is
standardize=TRUE.- partial_method
Indicate whether to use pairwise partial correlations, or shrinkage partial correlations.
- l2_penalty
Optional penalty for level-specific regressions (useful in high-dimensional case)
- ...
Additional arguments passed to
hclust.
Details
Shrinkage can be imposed by targeting an explicit effective degrees of freedom.
Setting the argument kappa to a value between 0 and 1 controls
the effective degrees of freedom of the fitted object as a percentage of \(p\).
When kappa is 1 the result is equivalent to the result from an ordinary
least squares regression (no shrinkage). Conversely, kappa set to 0
represents maximum shrinkage.
When \(p > N\) kappa is a percentage of \((N - 2)\).
If no kappa is set, a linear regression with kappa = 1 is
estimated.
Hierarchical clustering is performed using hclust. The default is set to
ward.D2 clustering but can be overridden by passing a method argument to ....
For high-dimensional problems, the hierarchy becomes very large. Setting q to a value below 1
reduces the number of levels used in the hierarchy. q represents a quantile-cutoff of the amount of
variation contributed by the levels. The default (q = NULL) considers all levels.
When data exhibits multicollinearity it can be useful to include a penalty on the l2 norm in the level-specific regressions.
This can be achieved by setting the l2_penalty parameter.
References
Pfitzinger, Johann (2024). Cluster Regularization via a Hierarchical Feature Regression. _Econometrics and Statistics_ (in press). URL https://doi.org/10.1016/j.ecosta.2024.01.003.
Examples
x = matrix(rnorm(100 * 20), 100, 20)
y = rnorm(100)
fit = hfr(x, y, kappa = 0.5)
coef(fit)
#> (Intercept) V1 V2 V3 V4 V5
#> -0.045528849 -0.044402118 -0.047856870 -0.092421394 -0.126072592 -0.009288087
#> V6 V7 V8 V9 V10 V11
#> 0.011728797 -0.144757423 0.167845451 0.055608271 -0.102802516 -0.124190165
#> V12 V13 V14 V15 V16 V17
#> 0.012207753 0.009962975 -0.124733367 0.055228331 0.144381079 -0.135998132
#> V18 V19 V20
#> 0.039950932 0.063773122 0.009338699