This is the second of two vignettes on the crplot
function. It details advanced features which are helpful when
troubleshooting plot issues, focusing on its jump-center algorithm
parameters. For a basic comprehensive summary of crplot
usage, please reference the first vignette titled crplot (users
should familiarize with it prior to viewing this vignette). Both
vignettes are available via links on the conf package
webpage.
The crplot
function generates a two-dimensional
confidence region plot for the specified two-parameter parametric
distribution, fitted to a dataset. Details of the crplot
algorithm are available in its corresponding publication1.
A jump-center is necessary to plot confidence region boundary points
that are inaccessible via the radial profile technique (plotting
algorithm of crplot
) because a line joining them to the MLE
crosses the confidence region boundary multiple times. The jump-center
algorithm “repairs” the inaccessible region by providing an alternate
point-of-reference (or alternate “center-point”, hence its name) from
which to execute the radial profile plotting technique. See
crplot
vignette 1 of 2 for additional details.
Many advanced crplot
options involve customizing its
jump-center parameters. Before exploring those methods, it is first
important to familiarize with jump-center components. The plot
annotations below will aid in those definitions.
(note: select word coloring below is chosen to correspond with the appropriate reference points above)
Jump-Center & its Left and Right Boundaries. Customization is informed by the default jump-center (JC) and its JC left and JC right boundary reference points (pictured above). The “left” and “right” directional labels represent their location relative to the jump-center (when looking “out” from the jump-center into the inaccessible region). Maximum angle tolerance constraints are not satisfied at the left and/or right inaccessible region border points, and the jump-center repair aims to plot points within this uncharted region.
Quadrant. The jump-center position with respect to the MLE (\(+\)) is significant. As a convenience, quadrant notation relative to the MLE is used. For example, the jump-center pictured above lies in the fourth quadrant relative to the MLE. All future references to quadrants within this vignette adopt this perspective (quadrants defined as if the MLE were the origin).
Gap. Two inaccessible region orientations are possible: a vertical (shown above) or horizontal gap. The vertical gap shown above has an inaccessible region orientated counterclockwise from the MLE (\(+\)). For a jump-center in the fourth quadrant, a clockwise inaccessible region orientation corresponds to a horizontally oriented gap.
Jump-Shift. The direction of the jump-center, relative to the MLE (\(+\)), intercepts the horizontal or vertical gap according to a jump-shift parameter. The jump-shift parameter represents the fraction of its gap where an azimuth to the jump-center crosses. The default, pictured above, is a jump-shift = 0.5 (or bisecting its gap). For this example, decreasing this value moves the azimuth from the MLE to the jump-center nearer its right boundary reference point. Increasing its value would move it nearer the left boundary.
Jump-Uphill. The jump-center must locate inside of the existing confidence region boundary. Considering the log-likelihood function governs the landscape from which the confidence region contour is found, points inside its region are those “uphill” from its border. The jump-uphill parameter quantifies how far uphill (given as a measure of level of significance) from its confidence region boundary the jump-center locates, and thus determines its distance from the MLE (or confidence region boundary). For example, consider an \(89\%\) versus a \(90\%\) confidence region (or levels of significance of \(0.11\) and \(0.1\), respectively). The \(89\%\) confidence region is smaller; its boundary is inside that of the \(90\%\) region and represents a “jump-uphill” measure of \(0.01\) (in level of significance). The default parameterization for jump-uphill is \(\min(\alpha, 0.01)\).
Plot issues using crplot
are relatively uncommon, but
possible. Certain datasets and/or distributions pose greater plot
challenges than others. In particular, small datasets and/or heavily
censored datasets can result in unusual (strong deviations from
elliptical, even to the point of non-convex) shapes, making their
boundary difficult to map using the radial profile log likelihood
technique2 employed by crplot
.
When the crplot
algorithm is stretched beyond its
default limitations, adjusting a series of optional arguments can
further extend its practical reach. Optional arguments in the
crplot
function enable algorithm customization, which is
often enough to overcome plot issues when they arise.
This heavily right-censored example produces a particularly
challenging confidence region shape, enabling us to illustrate a suite
of optional arguments available to its user to customize the
crplot
algorithm. Understanding the impact of optional
argument adjustments will facilitate their use elsewhere when
appropriate.
Bearing cage fracture data is taken from Abernethy et. al. (1983)3. Among
1703 samples there are six observed failures. The remaining 1697
right-censored samples occur over a variety of right-censored values
between 50–2050 hours. Both data subsets are given below with the
variables bc_obs
and bc_cen
, respectively.
<- c(230, 334, 423, 990, 1009, 1510)
bc_obs <- c(rep(50, 288), rep(150, 148), rep(250, 124), rep(350, 111), rep(450, 106),
bc_cen rep(550, 99), rep(650, 110), rep(750, 114), rep(850, 119), rep(950, 127),
rep(1050, 123), rep(1150, 93), rep(1250, 47), rep(1350, 41), rep(1450, 27),
rep(1550, 11), rep(1650, 6), rep(1850, 1), rep(2050, 2))
<- c(bc_obs, bc_cen)
bc <- c(rep(1, length(bc_obs)), rep(0, length(bc_cen)))
cen print(length(bc))
#> [1] 1703
Confidence regions corresponding to a variety of parametric
distributions for this heavily right-censored dataset are given below.
Each plot uses a level of significance of alpha = 0.1
(a
90% confidence region).
<- c("cauchy", "gamma", "llogis", "logis", "norm", "weibull")
distns par(mfrow = c(2, 3)) # plot in a 2-row, 3-column grid
for (i in 1:length(distns)) {
try(crplot(dataset = bc, alpha = 0.1, distn = distns[i], cen = cen,
sf = c(0, 2), ylas = 1, main = distns[i]))
}
The gamma distribution confidence region plot for the bearing cage fracture dataset takes a suspect shape. Nearer the bottom of its plot there are long segments that are seemingly missing confidence region boundary points. Relatively sharp vertex angles flank each of these line segments, a characteristic indicative of a plot issue.
To assess what is happening, jump-center reference points are plot
using showjump = TRUE
. Additionally, its plot information
is returned using the optional argument jumpinfo = TRUE
.
The jumpinfo
argument is used in lieu of its alternative
info = TRUE
because we want jump-center repair information
returned in addition to the final confidence region plot information. It
is then available for subsequent use, as shown by plotting those points
on the “without repairs” plot below.
<- crplot(bc, alpha = 0.1, distn = "gamma", cen = cen, jumpinfo = TRUE, showjump = TRUE,
x main = "with repairs")
#> [1] "Confidence region plot complete; made using 276 boundary points."
crplot(bc, alpha = 0.1, distn = "gamma", cen = cen, repair = FALSE,
main = "without repairs")
#> [1] "Confidence region plot complete; made using 134 boundary points."
<- t(matrix(c(x$repair$q2jumpxy, x$repair$q2jumpL, x$repair$q2jumpR,
jumppoints $repair$q4jumpxy, x$repair$q4jumpL, x$repair$q4jumpR), nrow = 2))
xpoints(jumppoints, pch = 24, bg = rep(c("red", "green", "yellow")))
print(labels(x$repair))
#> [1] "q2jumpuphill" "q4jumpuphill" "q2jumpshift" "q4jumpshift"
#> [5] "q2jumpxy" "q4jumpxy" "q2jumpL" "q4jumpL"
#> [9] "q2jumpR" "q4jumpR" "q2gaptype" "q4gaptype"
Setting jumpinfo = TRUE
returns a repair
attribute, which is a list relevant to all jump-center repairs. The
repair
labels are output to the console above, and given in
greater detail next.
Prefixes associated with repair
are:
q1
, q2
, q3
, q4
.
Each label begins with an abbreviation specifying what quadrant (with
respect to the MLE) the jump-center repair is made. For example, a
q1
prefix indicates a jump-center repair located in the 1st
quadrant (the jump-center horizontal and vertical values are both
greater than the MLE’s respective values). In our example,
q2
and q4
prefixes indicate jump-center
repairs in the 2nd and 4th quadrant relative to the MLE (to its
upper-left and lower-right, respectively).Suffixes associated with repair
are:
jumpuphill
quantifies the amount uphill (with regards
to level of significance) from the confidence region boundary that the
jump-center will locate,jumpshift
quantifies the point along the vertical or
horizontal “gap” (as a fractional value between 0 and 1) that the
azimuth from the MLE to the jump-center will cross,jumpxy
stores the coordinates for the jump-center,jumpL
stores the left (with perspective taken from the
jump-center location towards its inaccessible region) confidence region
boundary point bordering its inaccessible region, andjumpR
stores the left (with perspective taken from the
jump-center location towards its inaccessible region) confidence region
boundary point bordering its inaccessible region.Seeing its jump-center location enables a targeted strategy—using the optional arguments given next—to improve the plot when necessary.
Optional arguments whose adjustment can refine crplot
results include:
jumpuphill
jumpshift
ellipse_n
maxcount
Optional argument adjustments can be applied separately or in
combination, each serving a unique purpose. Combinations of
jumpshift
and/or jumpuphill
are the
recommended first option to address plot regions that remain
inaccessible to its jump-center.
The fourth quadrant repairs (with respect to the MLE) of the aforementioned example will illustrate optional argument uses below since they are more easily visible than its second quadrant repairs.
jumpuphill
Increasing this optional argument from its default value,
jumpuphill = min(alpha, 0.01)
, moves the jump-center nearer
the MLE and further from its confidence region boundary. Doing so in our
example will enable the fourth quadrant jump-center to achieve a better
line-of-sight to the bottom portion of its confidence region (where a
perceived lack of plot points appears problematic).
crplot(bc, alpha = 0.1, distn = "gamma", cen = cen, jumpuphill = 0.1, showjump = TRUE)
#> [1] "Confidence region plot complete; made using 277 boundary points."
The confidence region’s bottom-most curve benefits substantially from
this jumpuphill
adjustment, but a problematic area (lacking
plot points) remains in the upper curve at higher \(\theta\) values.
jumpshift
Adjusting this optional argument from its default value,
jumpshift = 0.5
, shifts the radial azimuth (with respect to
the MLE) to its jump-center. Decreasing its value generally “pushes” the
jump-center location further along its confidence region boundary,
nearer the inaccessible region. Doing so in our example enables the
fourth quadrant jump-center to achieve a better line-of-sight to the
upper portion of its confidence region curve at higher \(\theta\) values (where a perceived lack of
plot points appears problematic).
crplot(bc, alpha = 0.1, distn = "gamma", cen = cen, jumpshift = 0.1, showjump = TRUE)
#> [1] "Confidence region plot complete; made using 292 boundary points."
The confidence region’s upper curve (at higher \(\theta\) values) benefits substantially
from this jumpshift
adjustment, but does so at the expense
of its bottom curve. The bottom curve (at minimal \(\kappa\) values) plot difficulties are
compounded by the perspective taken by this new jump-center
location.
Caution: although decreasing jumpshift
generally improves access to inaccessible regions, inevitably a
jumpshift
value too near zero will result in multiple
solutions for its jump-center at the specified jumpuphill
value. Under those circumstance, the point nearer its MLE is typically
returned, resulting in a jump-center whose location is not between the
inaccessible boundary points, and is therefore unable to plot its
inaccessible region. Using jumpshift = 0.01
rather than
jumpshift = 0.1
with the syntax above demonstrates this
undesirable paradigm, and is left as an exercise to the reader. Another
example of a jumpshift
value that is prohibitively low for
its circumstance is given later in this vignette, in the Optional
Argument Combinations section.
ellipse_n
Adjusting this optional argument from its default value,
ellipse_n = 4
, combines the elliptically-oriented plot
technique with the default smoothing-search algorithm. Details of the
elliptically-oriented plot algorithm are given in the first
crplot
vignette. It increases the final number of plot
points, albeit at a computational expense. Used here, it improves but
does not entirely correct our example’s plot concerns.
crplot(bc, alpha = 0.1, distn = "gamma", cen = cen, showjump = TRUE, ellipse_n = 80)
#> [1] "Confidence region plot complete; made using 387 boundary points."
maxcount
The maxcount
parameter specifies the number of
smoothing-search iterations crplot
performs before
terminating its algorithm. The default value (30) is typically
sufficient to plot all accessible regions to its maximum degree
tolerance constraint (maxdeg
), so reaching this limit is
indicative of an inaccessible region plot issue. On rare occasion
however, increasing maxcount
may improvement select plot
regions, albeit at a noteworthy computational cost. This is true for our
example plot, whose bottom region continues to (slowly) populate with
relevant points as maxcount
increases.
crplot(bc, alpha = 0.1, distn = "gamma", cen = cen, maxcount = 50)
#> [1] "Confidence region plot complete; made using 476 boundary points."
Increasing maxcount
is effective at identifying
additional relevant points here due to poor circumstance surrounding its
fourth quadrant jump-center position. The smoothing search algorithm
iteratively locates additional plot points based on azimuths (or angles)
associated with the mid-point of its existing points (wherever
maxdeg
angle constraints are not yet met). The mid-point
between adjacent points along its bottom curve are so offset and spread
from the jump-center that, despite aiming towards the mid-point of its
long vacant region, additional points are successively added much nearer
its very far-right existing point.
Combinations of the aforementioned optional arguments are typically
capable of correcting the most challenging plot circumstance. For the
bearing cage fracture data example, jumpuphill
and
jumpshift
parameter adjustments—influencing the jump-center
distance and direction from the MLE, respectively—prove sufficient to
adequately repair plot issues in its fourth quadrant. One solution,
using jumpshift = 0.07
and jumpuphill = 0.05
,
is given below.
<- crplot(bc, alpha = 0.1, distn = "gamma", cen = cen, showjump = TRUE,
x jumpshift = 0.07, jumpuphill = 0.05, jumpinfo = TRUE)
#> [1] "Confidence region plot complete; made using 240 boundary points."
plot(c(x$theta, x$theta[1]), c(x$kappa, x$kappa[1]), type = "l", axes = FALSE,
xlab = expression(theta), ylab = expression(kappa))
axis(side = 1, at = range(x$theta))
axis(side = 2, at = range(x$kappa))
A closer review of its lower-right and upper-left plot regions is
appropriate to ensure those extremes are adequately captured. Those
plots reveal that the quadrant four repairs were successful, but
applying its customizations uniformly for all jump-center locations
results in an inaccurate plot region for quadrant two. It encounters an
issue discussed as “Caution: …” in the jumpshift
section above. Specifically, its decreased jumpshift
value
causes the MLE-to-jump-center azimuth to pass too near its inaccessible
region right boundary. It subsequently fails to locate its jump-center
by the “far” confidence region boundary, and instead crosses another
solution for the jumpuphill
parameter before its
inaccessible region.
par(mar = c(4, 4, 1.5, 1))
# top-left confidence region extreme (zoomed-in)
plot(x$theta, x$kappa, xlim = c(min(x$theta), x$thetahat), ylim = c(x$kappahat, max(x$kappa)),
type = 'l', xlab = expression(theta), ylab = expression(kappa),
main = "top-left (zoom-in)")
# augment the top-left region graphic with plot points and jump-center information
points(x$theta, x$kappa)
<- t(matrix(c(x$repair$q2jumpxy, x$repair$q2jumpL, x$repair$q2jumpR), nrow = 2))
jumppoints points(jumppoints, pch = 24, bg = rep(c("red", "green", "yellow")))
# bottom-right confidence region extreme (zoomed-in)
plot(x$theta, x$kappa, ylim = c(min(x$kappa), 0.95), type = 'l',
xlab = expression(theta), ylab = expression(kappa), main = "bottom-right (zoom-in)")
points(x$theta, x$kappa)
Different jumpshift
and jumpuphill
values
for quadrant two and four repairs are necessary to adequately plot each
of its extreme regions to a high level of detail. Itemizing
customizations in this manner is possible, and detailed in the next
section.
jumpshift
and jumpuphill
values by
quadrant.A plot with multiple jump-center repairs may warrant a different
jumpshift
and jumpuphill
value for each
respective jump-center. Given scalar jumpshift
and/or
jumpuphill
arguments, the crplot
algorithm
assumes its value applies regardless of the jump-center repair location
(or quadrant). To customize these parameters to jump-center
location-specific values, assign jumpshift
and/or
jumpuphill
a vector of length four; its values then
correspond uniquely to the four respective quadrants surrounding the
MLE.
When providing a quadrant-specific vector of jumpshift
or jumpuphill
values, any placeholder value is acceptable
for quadrants with no jump-center. For our example, quadrants one and
three contain no jump-center repairs and we can therefore arbitrarily
assign the default values of 0.5 and 0.01 for jumpshift
and
jumpuphill
in those respective locations.
Through separate analysis (analogous to that for quadrant four), we
determine that the quadrant two jump-center adequately plots its
associated region using jumpuphill = 0.25
(keeping the
default value for jumpshift = 0.5
). Notice how the maximum
\(\kappa\) value increases as a result.
This solution is combined with the quadrant four solution
(jumpshift = 0.07
and jumpuphill = 0.05
)
below.
# top-left confidence region extreme (zoomed-in)
<- crplot(bc, alpha = 0.1, distn = "gamma", cen = cen, showjump = TRUE, jumpinfo = TRUE,
x jumpshift = c(0.5, 0.5, 0.5, 0.07), jumpuphill = c(0.01, 0.25, 0.01, 0.05),
xlim = c(min(x$theta), x$thetahat), ylim = c(x$kappahat, 4.15),
main = "top-left (zoom-in)")
#> [1] "Confidence region plot complete; made using 299 boundary points."
# bottom-right confidence region extreme (zoomed-in)
plot(x$theta, x$kappa, ylim = c(min(x$kappa), 0.95), type = 'l',
xlab = expression(theta), ylab = expression(kappa), main = "bottom-right (zoom-in)")
points(x$theta, x$kappa)
Satisfied with plot accuracy in all regions, the final \(90\%%\) confidence region for the heavily right-censored bearing fracture data is given:
# final plot, with all necessary repairs complete
plot(c(x$theta, x$theta[1]), c(x$kappa, x$kappa[1]), type = "l", axes = FALSE,
xlab = expression(theta), ylab = expression(kappa), main = "gamma 90% CR for BF data (final)")
axis(side = 1, at = range(x$theta))
axis(side = 2, at = range(x$kappa))
points(x$thetahat, x$kappahat, pch = 3)
Albeit uncommon, if crplot
is unable to return a
confidence region plot the following message prints to the console:
#> [1] "--------------------------------------------------------------------------------------"
#> [1] "R uniroot failure searching for confidence region boundary---challenging parameters or shape."
#> [1] "Unable to produce a confidence region for the given sample and/or parameterization. "
#> [1] "--------------------------------------------------------------------------------------"
The R uniroot
function numerically solves for confidence
region boundary points within crplot
. No plot returns when
this numeric method fails. Failures can occur due-to, or despite
uniroot
parameters upper
and
lower
properly bracketing the confidence region (uniroot
solves using a bisection method). This typically occurs under the most
challenging plot circumstance such as heavily censored and/or small
sample sizes, or scales that approach the limits of R’s finite precision
arithmetic.
Good options to troubleshoot this error include: increasing
alpha
, setting repair = FALSE
, and adjusting
the jumpshift
and/or jumpuphill
parameters.
Increasing alpha
is effective because it reduces the
confidence region size, and this is can remedy uniroot
upper-bound issues. The latter two are effective options because both
influence the jump-center repairs algorithm. The jump-center repairs
algorithm is usually the source of uniroot errors since it addresses the
most “extreme” corners of the confidence region plot, where
uniroot
numeric issues tend to occur. Setting
repair = FALSE
may enable return of a working-copy of the
confidence region (complete with exception of its radially inaccessible
regions). To attain a complete confidence region, adjusting
jumpshift
and/or jumpuphill
inherently will
assign its jump-center a different location, from where its
uniroot
calculations may not experience similar numeric
difficulties.
The example below demonstrates circumstance resulting in a plot error, and attainable solutions using the strategy described above.
# crplot is unable to plot this 98% confidence region
crplot(dataset = c(1.5, 2), alpha = 0.01, distn = "invgauss")
#> [1] "--------------------------------------------------------------------------------------"
#> [1] "R uniroot failure searching for confidence region boundary---challenging parameters and/or shape."
#> [1] "Unable to produce a confidence region for the given sample and/or parameterization. "
#> [1] "--------------------------------------------------------------------------------------"
The three troubleshooting options given above are illustrated next. For this example, each successfully produces a confidence region plot.
# a plot without jump-center repairs is attainable, but its 3rd and 4th quadrants,
# relative to the MLE, are in need of jump-center repairs
crplot(dataset = c(1.5, 2), alpha = 0.01, distn = "invgauss", repair = FALSE,
sf = c(3, 3), main = "without repairs")
#> [1] "99% confidence region plot complete; made using 131 boundary points."
# a complete plot is returned by increasing alpha to values >= 0.03
crplot(dataset = c(1.5, 2), alpha = 0.05, distn = "invgauss", main = "95% CR",
sf = c(3, 3))
#> [1] "95% confidence region plot complete; made using 276 boundary points."
# adjusting jumpshift and jumpuphill parameters
<- crplot(dataset = c(1.5, 2), alpha = 0.02, "invgauss", jumpinfo = TRUE,
x sf = c(3, 3), jumpshift = 0.1, jumpuphill = 0.001, main = "98% CR")
#> [1] "98% confidence region plot complete; made using 234 boundary points."
plot(c(x$mu, x$mu[1]), c(x$lambda, x$lambda[1]), type = "l", axes = FALSE, xlab = expression(mu),
ylab = expression(lambda), main = "98% CR")
axis(side = 1, at = format(range(x$mu), digits = 3))
axis(side = 2, at = format(range(x$lambda), digits = 3))
Weld, C., Loh, A., and Leemis, L. (in press), “Plotting Likelihood-Ratio Based Confidence Regions for Two-Parameter Univariate Probability Models”, The American Statistician↩︎
Jaeger, A. (2016), “Computation of Two- and Three-Dimensional Confidence Regions With the Likelihood Ratio”, The American Statistician, 70, 395–398↩︎
Abernethy, R. B., Breneman, J. E., Medlin, C. H., and Reinman, G. L. (1983), “Weibull Analysis Handbook”, Aero Propulsion Laboratory, 43–47↩︎