Package {KRLS}

Gaussian kernel bandwidth by maximum off-diagonal var(K)

Description

Pick the Gaussian kernel bandwidth \sigma that maximizes the variance of the off-diagonal entries of the kernel matrix K. The idea is that this choice makes the columns of K most informative: at very small or very large \sigma, the off-diagonal entries collapse toward 0 or 1 respectively and carry little discriminating signal.

This is the bandwidth-selection convention used in kbal::b_maxvarK and the GPSS package; in KRLS 1.7+ it is also the default when sigma = NULL in krls.

b_maxvarK_nystrom is the Nystrom-aware variant: it evaluates the variance of the entries of the n by m cross-kernel C = K(X, Z) between observations and landmarks instead of forming the full n by n kernel.

Usage

b_maxvarK(X_proc, search_lower = 1e-6, search_upper = NULL,
          tol = .Machine$double.eps^0.25)

b_maxvarK_nystrom(X_proc, Z_proc, search_lower = 1e-6,
                  search_upper = NULL,
                  tol = .Machine$double.eps^0.25)

Arguments

Value

A list with components

References

Hazlett, C. (2020). Kernel Balancing: A Flexible Non-Parametric Weighting Procedure for Estimating Causal Effects. Statistica Sinica.

See Also

krls, gausskernel.

Compute first differences with KRLS

Description

Internal function that is called by krls to computes first differences for binary predictors in the X matrix. It would normally not be called by the user directly.

Usage

fdskrls(object,...)

Arguments

Value

A object of class krls where the derivatives, average derivatives, and the varinaces of the average derivatives are replaced with the first differences for binary predictors. The binaryindicator is also updated and set to TRUE for binary predictors.

Author(s)

Jens Hainmueller (Stanford) and Chad Hazlett (UCLA)

See Also

krls

Gaussian Kernel Distance Computation

Description

Given a N by D numeric data matrix, this function computes the N by N distance matrix with the pairwise distances between the rows of the data matrix as measured by a Gaussian Kernel.

Usage

gausskernel(X = NULL, sigma = NULL)

Arguments

Details

Given two D dimensional vectors x_i and x_j. The Gaussian kernel is defined as

k(x_i,x_j)=exp(\frac{-|| x_i - x_j ||^2}{\sigma^2})

where ||x_i - x_j|| is the Euclidean distance given by

||x_i - x_j||=((x_i1-x_j1)^2 + (x_i2-x_j2)^2 + ... + (x_iD-x_jD)^2)^.5

and \sigma^2 is the bandwidth of the kernel.

Note that the Gaussian kernel is a measure of similarity between x_i and x_j. It evalues to 1 if the x_i and x_j are identical, and approaches 0 as x_i and x_j move further apart.

The function relies on the dist function in the stats package for an initial estimate of the euclidean distance.

Value

An N by N numeric distance matrix that contains the pairwise distances between the rows in X.

Author(s)

Jens Hainmueller (Stanford) and Chad Hazlett (UCLA)

See Also

dist function in the stats package.

Examples

X <- matrix(rnorm(6),ncol=2)
gausskernel(X=X,sigma=1)

Extract the landmark coordinates from a Nystrom KRLS fit

Description

Returns the landmark coordinate matrix used by an approx = "nystrom" fit. Landmarks are stored internally in the standardized X-space the model operates in; by default this accessor un-standardizes them so the returned matrix is in the original X units and can be passed back through krls(..., landmarks = ...) on a comparable dataset without a standardize-twice bug.

Usage

get_landmarks(fit, scale = c("original", "standardized"))

Arguments

Value

An m by D numeric matrix of landmark coordinates with column names inherited from fit$X.

Author(s)

Jens Hainmueller (Stanford) and Chad Hazlett (UCLA)

See Also

krls

Examples

## Not run: 
fit <- krls(X, y, approx = "nystrom", nystrom_m = 30)
Z   <- get_landmarks(fit)        # original-scale m x D matrix
# Reuse the same landmarks on a sensitivity-check refit with a new y:
fit2 <- krls(X, y_alt, approx = "nystrom", landmarks = Z)

## End(Not run)

Kernel-based Regularized Least Squares (KRLS)

Description

Function implements Kernel-Based Regularized Least Squares (KRLS), a machine learning method described in Hainmueller and Hazlett (2014) that allows users to solve regression and classification problems without manual specification search and strong functional form assumptions. KRLS finds the best fitting function by minimizing a Tikhonov regularization problem with a squared loss, using Gaussian Kernels as radial basis functions. KRLS reduces misspecification bias since it learns the functional form from the data. Yet, it nevertheless allows for interpretability and inference in ways similar to ordinary regression models. In particular, KRLS provides closed-form estimates for the predicted values, variances, and the pointwise partial derivatives that characterize the marginal effects of each independent variable at each data point in the covariate space. The distribution of pointwise marginal effects can be used to examine effect heterogeneity and or interactions.

Usage

krls(X = NULL, y = NULL, whichkernel = "gaussian", lambda = NULL,
sigma = NULL, derivative = TRUE, binary= TRUE, vcov=TRUE,
print.level = 1,L=NULL,U=NULL,tol=NULL,eigtrunc=NULL,data=NULL,
approx = c("auto", "none", "nystrom"), nystrom_m = NULL,
landmarks = NULL, landmark_method = c("random", "kmeans"),
nystrom_eps = sqrt(.Machine$double.eps),
landmark_seed = NULL,
lambda_method = c("loo", "gcv"),
cat_columns = NULL)

Arguments

Details

krls implements the Kernel-based Regularized Least Squares (KRLS) estimator as described in Hainmueller and Hazlett (2014). Please consult this reference for any details.

Kernel-based Regularized Least Squares (KRLS) arises as a Tikhonov minimization problem with a squared loss. Assume we have data of the from y_i,\,x_i where i indexes observations, y_i \in R is the outcome and x_i \in R^D is a D-dimensional vector of predictor values. Then KRLS searches over a space of functions H and chooses the best fitting function f according to the rule:

argmin_{f \in H} \sum_i^N (y_i - f(x_i))^2 + \lambda ||f||_{H^2}

where (y_i - f(x_i))^2 is a loss function that computes how ‘wrong’ the function is at each observation i and || f ||_{H^2} is the regularizer that measures the complexity of the function according to the L_2 norm ||f||^2 = \int f(x)^2 dx. \lambda is the scalar regularization parameter that governs the tradeoff between model fit and complexity. By default, \lambda is chosen by minimizing the sum of the squared leave-one-out errors, but it can also be specified by the user in the lambda argument to implement other approaches.

Under fairly general conditions, the function that minimizes the regularized loss within the hypothesis space established by the choice of a (positive semidefinite) kernel function k(x_i,x_j) is of the form

f(x_j)= \sum_i^N c_i k(x_i,x_j)

where the kernel function k(x_i,x_j) measures the distance between two observations x_i and x_j and c_i is the choice coefficient for each observation i. Let K be the N by N kernel matrix with all pairwise distances K_ij=k(x_i,x_j) and c be the N by 1 vector of choice coefficients for all observations then in matrix notation the space is y=Kc.

Accordingly, the krls function solves the following minimization problem

argmin_{f \in H} \sum_i^n (y - Kc)'(y-Kc)+ \lambda c'Kc

which is convex in c and solved by c=(K +\lambda I)^-1 y where I is the identity matrix. Note that this linear solution provides a flexible fitted response surface that typically reduces misspecification bias because it can learn a wide range of nonlinear and or nonadditive functions of the predictors.

If vcov=TRUE is specified, exact KRLS also computes the variance-covariance matrix for the choice coefficients c and fitted values y=Kc based on a variance estimator developed in Hainmueller and Hazlett (2014). Note that both exact matrices are N by N and therefore this results in increased memory and computing time. Under approx = "nystrom", vcov=TRUE computes conditional approximate variances in the m-landmark feature space and does not store the full fitted-value covariance matrix.

By default, krls uses the Gaussian Kernel (whichkernel = "gaussian") given by

k(x_i,x_j)=exp(\frac{-|| x_i - x_j ||^2}{\sigma^2})

where ||x_i - x_j|| is the Euclidean distance. The kernel bandwidth \sigma^2 is set to D, the number of dimensions, by default, but the user can also specify other values using the sigma argument to implement other approaches.

If derivative=TRUE is specified, krls also computes the pointwise partial derivatives of the fitted function wrt to each predictor using the estimators developed in Hainmueller and Hazlett (2014). These can be used to examine the marginal effects of each predictor and how the marginal effects vary across the covariate space. Average derivatives are also computed with variances when vcov=TRUE. Note that the derivative=TRUE option results in increased computing time and is only supported for the Gaussian kernel, i.e. when whichkernel = "gaussian". For the exact path, derivative=TRUE requires vcov=TRUE; under approx = "nystrom", derivative=TRUE, vcov=FALSE returns derivative point estimates without standard errors.

If binary=TRUE is also specified, the function will identify binary predictors and return first differences for these predictors instead of partial derivatives. First differences are computed going from the minimum to the maximum value of each binary predictor. Note that first differences are more appropriate to summarize the effects for binary predictors (see Hainmueller and Hazlett (2014) for details).

A few other kernels are also implemented, but derivatives are currently not supported for these: "linear": k(x_i,x_j)=x_i'x_j, "poly1", "poly2", "poly3", "poly4" are polynomial kernels based on k(x_i,x_j)=(x_i'x_j +1)^p where p is the order.

Value

A list object of class krls with the following elements:

Note

The function requires the storage of a N by N kernel matrix and can therefore exceed the memory limits for very large datasets.

Setting derivative=FALSE and vcov=FALSE is useful to reduce computing time if pointwise partial derivatives and or variance covariance matrices are not needed.

Author(s)

Jens Hainmueller (Stanford) and Chad Hazlett (UCLA)

References

Jeremy Ferwerda, Jens Hainmueller, Chad J. Hazlett (2017). Kernel-Based Regularized Least Squares in R (KRLS) and Stata (krls). Journal of Statistical Software, 79(3), 1-26. doi:10.18637/jss.v079.i03

Hainmueller, J. and Hazlett, C. (2014). Kernel Regularized Least Squares: Reducing Misspecification Bias with a Flexible and Interpretable Machine Learning Approach. Political Analysis, 22(2)

Rifkin, R. 2002. Everything Old is New Again: A fresh look at historical approaches in machine learning. Thesis, MIT. September, 2002.

Evgeniou, T., Pontil, M., and Poggio, T. (2000). Regularization networks and support vector machines. Advances In Computational Mathematics, 13(1):1-50.

Schoelkopf, B., Herbrich, R. and Smola, A.J. (2001) A generalized representer theorem. In 14th Annual Conference on Computational Learning Theory, pages 416-426.

Kimeldorf, G.S. Wahba, G. 1971. Some results on Tchebycheffian spline functions. Journal of Mathematical Analysis and Applications, 33:82-95.

See Also

predict.krls for fitted values and predictions. summary.krls for summary of the fit. plot.krls for plots of the fit.

Examples

# Linear example
# set up data
N <- 200
x1 <- rnorm(N)
x2 <- rbinom(N,size=1,prob=.2)
y <- x1 + .5*x2 + rnorm(N,0,.15)
X <- cbind(x1,x2)
# fit model
krlsout <- krls(X=X,y=y)
# summarize marginal effects and contribution of each variable
summary(krlsout)
# plot marginal effects and conditional expectation plots
plot(krlsout)


# non-linear example
# set up data
N <- 200
x1 <- rnorm(N)
x2 <- rbinom(N,size=1,prob=.2)
y <- x1^3 + .5*x2 + rnorm(N,0,.15)
X <- cbind(x1,x2)

# fit model
krlsout <- krls(X=X,y=y)
# summarize marginal effects and contribution of each variable
summary(krlsout)
# plot marginal effects and conditional expectation plots
plot(krlsout)

## 2D example:
# predictor data
X <- matrix(seq(-3,3,.1))
# true function
Ytrue <- sin(X)
# add noise 
Y     <- sin(X) + rnorm(length(X),sd=.3)
# approximate function using KRLS
out <- krls(y=Y,X=X)
# get fitted values and ses
fit <- predict(out,newdata=X,se.fit=TRUE)
# results
par(mfrow=c(2,1))
plot(y=Ytrue,x=X,type="l",col="red",ylim=c(-1.2,1.2),lwd=2,main="f(x)")
points(y=fit$fit,X,col="blue",pch=19)
arrows(y1=fit$fit+1.96*fit$se.fit,
       y0=fit$fit-1.96*fit$se.fit,
       x1=X,x0=X,col="blue",length=0)
legend("bottomright",legend=c("true f(x)=sin(x)","KRLS fitted f(x)"),
       lty=c(1,NA),pch=c(NA,19),lwd=c(2,NA),col=c("red","blue"),cex=.8)

plot(y=cos(X),x=X,type="l",col="red",ylim=c(-1.2,1.2),lwd=2,main="df(x)/dx")
points(y=out$derivatives,X,col="blue",pch=19)

legend("bottomright",legend=c("true df(x)/dx=cos(x)","KRLS fitted df(x)/dx"),
       lty=c(1,NA),pch=c(NA,19),lwd=c(2,NA),col=c("red","blue"),,cex=.8)

## 3D example
# plot true function
par(mfrow=c(1,2))
f<-function(x1,x2){ sin(x1)*cos(x2)}
x1 <- x2 <-seq(0,2*pi,.2)
z   <-outer(x1,x2,f)
persp(x1, x2, z,theta=30,main="true f(x1,x2)=sin(x1)cos(x2)")
# approximate function with KRLS
# data and outcomes
X <- cbind(sample(x1,200,replace=TRUE),sample(x2,200,replace=TRUE))
y   <- f(X[,1],X[,2])+ runif(nrow(X))
# fit surface
krlsout <- krls(X=X,y=y)
# plot fitted surface
ff  <- function(x1i,x2i,krlsout){predict(object=krlsout,newdata=cbind(x1i,x2i))$fit}
z   <- outer(x1,x2,ff,krlsout=krlsout)
persp(x1, x2, z,theta=30,main="KRLS fitted f(x1,x2)")

Leave-one-out optimization to find \lambda

Description

Function conducts leave-one-out optimization to find \lambda using a golden search search with caching. This function is called internally by krls. It would normally not be called by the user directly.

Usage

lambdasearch(L=NULL,
           U=NULL,
           y=NULL,
           Eigenobject=NULL,
           tol=NULL,
           noisy=FALSE,
           eigtrunc=NULL,
           lambda_method = c("loo", "gcv"))

Arguments

Details

By default, upper bound is found as follows: Set j to n, decrease by one until the following is longer true: sum(EigenValues / (EigenValues + j)) < 1.

By default, upper bound is found as follows: Get the position, q, of the eigenvalue that is closest to max(Eigenvalue)/1000. Set j to 0, increase in steps of 0.05 until the below is longer true: sum(EigenValues / (EigenValues + j)) > q.

Value

A scalar that contains the \lambda that minimizes the sum of squared leave-one-out errors.

Author(s)

Jens Hainmueller (Stanford) and Chad Hazlett (UCLA)

See Also

krls

Loss Function for Leave One Out Error

Description

Internal function that computes Leave-On-Out (LOO) Error for KRLS given a fixed value for lambda (the parameter that governs the tradeoff between model fit and complexity in KRLS). This function is called internally by krls to find value of lambda that minimizes the LOO error. It would normally not be called by the user directly.

Usage

looloss(y = NULL, Eigenobject = NULL,
lambda = NULL,eigtrunc=NULL)

Arguments

Value

Scalar value for LOO error.

Author(s)

Jens Hainmueller (Stanford) and Chad Hazlett (UCLA)

See Also

krls

Plot method for Kernel-based Regularized Least Squares (KRLS) Model Fits

Description

Produces two types of plots. The first type of plot shows histograms for the pointwise partial derivatives to examine the heterogeneity in the marginal effects of each predictor (which==1). The second type of plot shows estimates of the conditional expectation functions of E[Y|X] for each predictor (which==2). For each plot, the predictor of interest varies from its 1st to its 3rd quartile values, while the other predictors are kept at the means (or other values specified in setx). For binary varibales the E[Y|X] are predicted at the max and the min value of the predictor (instead of the range from the 1st to the 3rd quantile).

Usage

 ## S3 method for class 'krls'
plot(x,which=c(1:2),
 main="distributions of pointwise marginal effects",
 setx="mean",ask = prod(par("mfcol")) < nplots,nvalues=50,probs=c(.25,.75),...)

Arguments

Details

Notice that the historgrams for the partial derivatives can only be plotted if the KRLS object was computed with krls(,derivatives=TRUE).

Author(s)

Jens Hainmueller (Stanford) and Chad Hazlett (UCLA)

See Also

krls

Examples

# non-linear example
# set up data
N <- 200
x1 <- rnorm(N)
x2 <- rbinom(N,size=1,prob=.2)
y <- x1^3 + .5*x2 + rnorm(N,0,.15)
X <- cbind(x1,x2)

# fit model
krlsout <- krls(X=X,y=y)
# summarize marginal effects and contribution of each variable
summary(krlsout)
# plot marginal effects and conditional expectation plots
plot(krlsout)

Predict method for Kernel-based Regularized Least Squares (KRLS) Model Fits

Description

Predicted values and standard errors based on krls model object.

Usage

 ## S3 method for class 'krls'
predict(object, newdata, se.fit = FALSE , ...)

Arguments

Details

Function produces predicted values, obtained by evaluating the fitted krls function with the newdata (ie. the test points). The prediction at a new test point x_i is based on

f(x_i)= \sum_j=1^n c_j K_{x_j}(x_i)

where K is the kernel matrix and thus K_{x_j} is a vector whose j-th entry is K(x_j,x_i) (e.g. the distance between the test point x_i and the training point x_j). The training points are passed to the function with the krls fit in object.

When data are missing in newdata during prediction, the value of each k(x_i,x_j) is computed by using an adjusted Euclidean distance in the kernel definition. Assume x is D-dimensional but a given pair of observations x_i and x_j have only D' < D non-missing dimensions in common. The adjusted Euclidean distance computes the sum of squared differences over the D' non-missing dimensions, rescales this sum by D/D', and takes the square root. The result corresponds to an assumption that conditional on the observed data, the missing values would not have contributed new information predictive of the outcome.

Value

Author(s)

Jens Hainmueller (Stanford) and Chad Hazlett (UCLA)

See Also

krls

Examples

# make up data
X <- seq(-3,3,.1)
Y <- sin(X) + rnorm(length(X),.1)

# fit krls
krlsout <- krls(y=Y,X=X)

# get in-sample prediction 
predin <- predict(krlsout,newdata=X,se.fit=TRUE)

# get out-of-sample prediction
X2 <- runif(5)
predout <- predict(krlsout,newdata=X2,se.fit=TRUE)

# plot true function and predictions
plot(y=sin(X),x=X,type="l",col="red",ylim=c(-1.8,1.8),lwd=2,ylab="f(X)")
points(y=predin$fit,x=X,col="blue",pch=19)
arrows(y1=predin$fit+2*predin$se.fit,
       y0=predin$fit-2*predin$se.fit,
       x1=X,x0=X,col="blue",length=0)
       
points(y=predout$fit,x=X2,col="green",pch=17)
arrows(y1=predout$fit+2*predout $se.fit,
       y0=predout$fit-2*predout $se.fit,
       x1=X2,x0=X2,col="green",length=0)

legend("bottomright",
       legend=c("true f(x)=sin(X)",
                "KRLS fitted in-sample",
                "KRLS fitted out-of-sample"),
       lty=c(1,NA,NA),pch=c(NA,19,17),
       lwd=c(2,NA,NA),
       col=c("red","blue","green"),
       cex=.8)

Solve for Choice Coefficients in KRLS

Description

Internal function that computes choice coefficients for KRLS given a fixed value for lambda (the parameter that governs the tradeoff between model fit and complexity in KRLS). This function is called internally by krls. It would normally not be called by the user directly.

Usage

solveforc(y = NULL, Eigenobject = NULL,
lambda = NULL,eigtrunc=NULL)

Arguments

Details

Function relies on fast eigenvalue decomposition method described in method Rifkin and Lippert (2007).

Value

Author(s)

Jens Hainmueller (Stanford) and Chad Hazlett (UCLA)

References

Rifkin, Ryan M. and Lippert, Ross A. (2007). Notes on Regularized Least Squares. MIT-CSAIL-TR-2007-025. CBCL-268

See Also

krls

Summary method for Kernel-based Regularized Least Squares (KRLS) Model Fits

Description

Summarizes average partial derivatives (i.e. marginal effects) and the distribution of the partial derivatives for each predictor. For binary predictors, the marginal effects are the first differences if krls(,derivatives=TRUE,binary=TRUE) was specified.

Usage

 ## S3 method for class 'krls'
summary(object, probs=c(.25,.5,.75),...)

Arguments

Details

Notice that the partial derivatives can only be summarized if the krls object was computed with krls(,derivatives=TRUE).

Value

Author(s)

Jens Hainmueller (Stanford) and Chad Hazlett (UCLA)

See Also

krls

Examples

# non-linear example
# set up data
N <- 200
x1 <- rnorm(N)
x2 <- rbinom(N,size=1,prob=.2)
y <- x1^3 + .5*x2 + rnorm(N,0,.15)
X <- cbind(x1,x2)

# fit model
krlsout <- krls(X=X,y=y)
# summarize marginal effects and contribution of each variable
summary(krlsout)
# plot marginal effects and conditional expectation plots
plot(krlsout)