evmix (version 1.0)

fkden: Cross-validation MLE Fitting of Kernel Density Estimator Using normal Kernel

Description

Maximum likelihood estimation for fitting kernel density estimator using a normal kernels, by treating it as a mixture model.

Usage

fkden(x, linit = NULL, extracentres = NULL,
    add.jitter = FALSE, factor = 0.1, amount = NULL,
    std.err = TRUE, method = "BFGS",
    control = list(maxit = 10000), finitelik = TRUE, ...)

Arguments

linit

initial value for bandwidth parameter or NULL

extracentres

extra kernel centres used in KDE, but likelihood contribution not evaluated, or NULL

add.jitter

logical, whether jitter is needed for rounded data

factor

see jitter

amount

see jitter

x

vector of sample data

std.err

logical, should standard errors be calculated

method

optimisation method (see optim)

control

optimisation control list (see optim)

finitelik

logical, should log-likelihood return finite value for invalid parameters

...

optional inputs passed to optim

Value

• Returns a simple list with the following elements ll{ call: optim call x: (jittered) data vector x kerncentres: actual kernel centres used x init: linit optim: complete optim output mle: vector of MLE of bandwidth cov: variance of MLE of bandwidth se: standard error of MLE of bandwidth nllh: minimum negative cross-validation log-likelihood n: total sample size lambda: MLE of bandwidth } The output list has some duplicate entries and repeats some of the inputs to both provide similar items to those from fpot and to make it as useable as possible.

Warning

Two important practical issues arise with MLE for the kernel bandwidth: 1) Cross-validation likelihood is needed for the KDE bandwidth parameter as the usual likelihood degenerates, so that the MLE $\hat{\lambda} \rightarrow 0$ as $n \rightarrow \infty$, thus giving a negative bias towards a small bandwidth. Leave one out cross-validation essentially ensures that some smoothing between the kernel centres is required (i.e. a non-zero bandwidth), otherwise the resultant density estimates would always be zero if the bandwidth was zero. This problem occassionally rears its ugly head for data which has been heavily rounded, as even when using cross-validation the density can be non-zero even if the bandwidth is zero. To overcome this issue an option to add a small jitter should be added to the data (x only) has been included in the fitting inputs, using the jitter function, to remove the ties. The default options red in the jitter are specified above, but the user can override these. Notice the default scaling factor=0.1, which is a tenth of the default value in the jitter function itself. A warning message is given if the data appear to be rounded (i.e. more than 5 estimated bandwidth is too small, then data rounding is the likely culprit. Only use the jittering when the MLE of the bandwidth is far too small. 2) For heavy tailed populations the bandwidth is positively biased, giving oversmoothing (see example). The bias is due to the distance between the upper (or lower) order statistics not necessarily decaying to zero as the sample size tends to infinity. Essentially, as the distance between the two largest (or smallest) sample datapoints does not decay to zero, some smoothing between them is required (i.e. bandwidth cannot be zero). One solution to this problem is to splice the GPD at a suitable threshold to remove the problematic tail from the inference for the bandwidth, using either the kdengpd function for a single heavy tail or the kdengng function if both tails are heavy. See MacDonald et al (2013).

Details

The kernel density estimator with a normal kernel is fitted to the entire dataset using maximum cross-validation likelihood estimation. The estimated bandwidth, variance and standard error are automatically output. Missing values (NA and NaN) are assumed to be invalid data so are ignored. Normally for likelihood estimation of the bandwidth the kernel centres and the data where the likelihood is evaluated are the same. However, when using KDE for extreme value mixture modelling the likelihood only those data in the bulk of the distribution should contribute to the likelihood, but all the data (including those beyond the threshold) should contribute to the density estimate. The extracentres option allows the use to specify extra kernel centres used in estimating the density, but not evaluated in the likelihood. The default is to just use the existing data, so extracentres=NULL. The default optimisation algorithm is "BFGS", which requires a finite negative log-likelihood function evaluation finitelik=TRUE. For invalid parameters, a zero likelihood is replaced with exp(-1e6). The "BFGS" optimisation algorithms require finite values for likelihood, so any user input for finitelik will be overridden and set to finitelik=TRUE if either of these optimisation methods is chosen. It will display a warning for non-zero convergence result comes from optim function call. If the hessian is of reduced rank then the variance (from inverse hessian) and standard error of bandwidth parameter cannot be calculated, then by default std.err=TRUE and the function will stop. If you want the bandwidth estimate even if the hessian is of reduced rank (e.g. in a simulation study) then set std.err=FALSE.

References

http://en.wikipedia.org/wiki/Kernel_density_estimation http://en.wikipedia.org/wiki/Cross-validation_(statistics) Scarrott, C.J. and MacDonald, A. (2012). A review of extreme value threshold estimation and uncertainty quantification. REVSTAT - Statistical Journal 10(1), 33-59. Available from http://www.ine.pt/revstat/pdf/rs120102.pdf Bowman, A.W. (1984). An alternative method of cross-validation for the smoothing of density estimates. Biometrika 71(2), 353-360. Duin, R.P.W. (1976). On the choice of smoothing parameters for Parzen estimators of probability density functions. IEEE Transactions on Computers C25(11), 1175-1179. MacDonald, A., Scarrott, C.J., Lee, D., Darlow, B., Reale, M. and Russell, G. (2011). A flexible extreme value mixture model. Computational Statistics and Data Analysis 55(6), 2137-2157. MacDonald, A., C. J. Scarrott, and D. S. Lee (2011). Boundary correction, consistency and robustness of kernel densities using extreme value theory. Submitted. Available from: http://www.math.canterbury.ac.nz/~c.scarrott.

See Also

jitter, density and bw.nrd0 Other kden: dkden, kden, lkden, nlkden, pkden, qkden, rkden

Examples

nk=50
x = rnorm(nk)
xx = seq(-5, 5, 0.01)
fit = fkden(x)
hist(x, nk/5, freq = FALSE, xlim = c(-5, 7))
rug(x)
for (i in 1:nk) lines(xx, dnorm(xx, x[i], sd = fit$lambda)*0.05)
lines(xx,dnorm(xx), col = "black")
lines(xx, dkden(xx, x, lambda = fit$lambda), lwd = 2, col = "red")
lines(density(x), lty = 2, lwd = 2, col = "green")
lines(density(x, bw = fit$lambda), lwd = 2, lty = 2,  col = "blue")
legend("topright", c("True Density", "KDE fitted evmix",
"KDE Using density, default bandwidth", "KDE Using density, c-v likelihood bandwidth"),
lty = c(1, 1, 2, 2), lwd = c(1, 2, 2, 2), col = c("black", "red", "green", "blue"))

# bandwidth is biased towards oversmoothing for heavy tails
nk=100
x = rt(nk, df = 2)
xx = seq(-8, 8, 0.01)
fit = fkden(x)
hist(x, seq(floor(min(x)), ceiling(max(x)), 0.5), freq = FALSE, xlim = c(-8, 10))
rug(x)
for (i in 1:nk) lines(xx, dnorm(xx, x[i], sd = fit$lambda)*0.05)
lines(xx,dt(xx , df = 2), col = "black")
lines(xx, dkden(xx, x, lambda = fit$lambda), lwd = 2, col = "red")
legend("topright", c("True Density", "KDE fitted evmix, c-v likelihood bandwidth"),
lty = c(1, 1), lwd = c(1, 2), col = c("black", "red"))

# remove heavy tails from likelihood evaluation, but still include used in KDE within likelihood
# often gives better bandwidth
nk=100
x = rt(nk, df = 2)
xx = seq(-8, 8, 0.01)
fit2 = fkden(x[(x > -4) & (x < 4)], extracentres = x[(x <= -4) | (x >= 4)])
hist(x, seq(floor(min(x)), ceiling(max(x)), 0.5), freq = FALSE, xlim = c(-8, 10))
rug(x)
for (i in 1:nk) lines(xx, dnorm(xx, x[i], sd = fit2$lambda)*0.05)
lines(xx,dt(xx , df = 2), col = "black")
lines(xx, dkden(xx, x, lambda = fit2$lambda), lwd = 2, col = "red")
lines(xx, dkden(xx, x, lambda = fit$lambda), lwd = 2, col = "blue")
legend("topright", c("True Density", "KDE fitted evmix, tails removed",
"KDE fitted evmix, tails included"),
lty = c(1, 1, 1), lwd = c(1, 2, 2), col = c("black", "red", "blue"))