For those plots where a confidence cone is plotted, the statistic type used in computing that mean direction and confidence cone are boldfaced.
Because of some problems in implementing the correct approximations for
computing for collections of directions with small
(
< 5) and a small number of directions (less than 30), the
program presently averages the approximation in Fisher et al., 1987 for
>5 with that for n>30. You will hear a beep whenever
this is done, for a maximum of four (Least Squares) or six (Locality) beeps. The confidence cones in question
can be identified by the low
and n.
A slight variation of Fisher statistics was suggested by McFadden and McElhinny (1988). This allows the combination of line and plane or circle fits in obtaining a Fisher mean; of greater impact (and the reason that this was in fact added) is the ability to restrict the plane or circle fits to those arcs that are reasonable given the paleomagnetic observations. In this code, this arc is taken to extend from the "farthest" point along the plane or circle (generally the last point used) to the antipode of the "nearest" point on the plane or circle (generally the first point used; see the figure in the description of the .lsq file format, below). Any such combination of line and plane fits must be approached with caution, as noted by Bailey (1990); it seems clear that some work remains to separate the best estimate of the mean and its uncertainty (McFadden and McElhinny's formulation) from the best estimate of the uncertainty of the mean that includes non-Fisherian biases inherent in the data or its reduction (Bailey's objection and the Bingham and non-parametric statistics below). The algorithm of McFadden and McElhinny has been coded from their paper and is not their original code or a modification; it was tested against the example of McFadden and McElhinny and found to match their estimated means and uncertainties within about 0.1-0.2°ree;. Note that McFadden and McElhinny noted that the uncertainties, while probably good approximations, are not rigorously derived.
Bingham statistics (Onstott, 1980) are available whereever Fisher statistics
are. Although similar to Fisher stats in use, certain points should be noted.
The code does not perform the complete integration to get exact concentration
parameters; instead it uses a lookup table from Mardia and Zemroch (1977) with
extensions by Gillett (1986). This should be adequate for most purposes.
However, we have found that the approximations that are commonly used in
computing Bingham statistics (and are used at present in the code) are poor for
calculating 's for poles to planes or circles (the
|
condition is violated). This can be seen by
using the "Combine lines and planes" option in the Least-Squares plot for a
collection of best-fit lines and comparing the result with that obtained using
the "Treat as lines" option.
Because of some of the problems with the Bingham distribution, non-parametric
bipolar/girdle statistics were added to the Least Squares Equal Area Plot's
elliptical statistics option and, for convenience (and because Watson has
discussed these uncertainty calculations), have been dubbed "Watson" statistics
despite the fact these are NOT the symmetrical Watson bipolar or girdle
distributions (Fisher et al., 1987, p. 155ff.). This statistic appears to
yield a more stable error ellipse regardless of the type of data (line data
will have the same ellipse when either "Treat as lines" or "Combine lines and
planes" are chosen). However, this statistic, being non-parametric, lacks
estimated concentration parameters as described in Fisher et al. [1987]. This
statistic is approximate for less than 25 directions. For poorly grouped data
the 95% confidence cone might exceed 90°ree; in general the statistics on such
data are meaningless. But to provide some comparative measure, the n
cone is displayed, where n is the greatest confidence level producing a cone
with semi-major axis of 90°ree;. The text in the upper right corner will
indicate n.
Statistics presented in the window include the two 's of the plotted
points (for the Bingham case), the
's of the major and minor axis of
the error ellipse, the mean direction, and the VGP latitude and longitude. The
"Oval Azimuth" is the angle measured clockwise from a line through the mean
direction and the stereonet's center to the major axis of the error ellipse.
The statistic uses the eigenvalues (e1>e2>e3) and associated eigenvectors
(
,
,
) of a
orientation or principal component matrix constructed from the selected
paleomagnetic directions. "Plot Girdle" to Bingham or Watson error ellipses are
simply 2 error ellipses with the same mean and oval azimuth but with axes equal
to 90°ree;+
and 90°ree;-
, where
is the either the
semi-major or semi-minor axis.
The performance and use of the Bingham option is usually straightforward, but
some circumstances require special attention to the values actually calculated
(see below). For the elliptical option in the "Locality Equal Area..." plot
(which still uses Bingham statistics), the program determines if the selected
points seem to describe a single direction (e1 > 2e2) or a girdle (e1 < 2e2).
If a single direction, the error ellipse about the mean direction
() is plotted. If a girdle, the error
ellipse about the pole (
) of the best
fitting plane through the points is plotted, as are the bounding ellipses
(girdles) of the distribution. One could interpret these girdles as the
confidence bounds on the possible positions of two different components, where
each of the measured directions reflect a different linear combination of the
two components.
For the "Least Squares Equal Area...," the Bingham or Watson statistics are manually controlled in a dialog box that pops up after selecting the directions to plot. Generally, if all the directions selected were derived using line fits, use the "Treat as Lines" option; if all the directions were plane or circle fits, use the "Treat as Planes" option. If you wish to combine planes and lines, use the "Combine lines and planes" option.
There are some instances where you might wish to override the choices outlined above; for instance, if you have a set of plane-fits from data that clearly represent the overlap of two distinct (and unvarying) directions, all of the plane fits will have similar poles. Using "Treat as Planes" will produce an elongated ellipse centered on the data's mean; this might have little meaning in terms of confidence regions of either underlying direction. In this case, using the "Treat as Lines" option with the "Plot Girdle" option will give the direction and uncertainty of the pole to the plane containing both directions; the girdle should enclose the region containing both directions.
For reference, when data are "treated as lines," each direction is added to the
orientation (or principal component) matrix as usual, the mean direction and
associated uncertainties are for . When
"treated as planes," directions are again added as usual, but the mean
direction and its uncertainties now are of the
vector of the moment-of-inertia matrix. When
"combining lines and planes," plane fits (identified by a "P" or "C" in the id
in the least-squares file .LSQ) are added as usual, but line fits are instead
added as two mutually orthogonal directions, each orthogonal to the line-fit
direction. The plotted mean and uncertainties again are of
. This is "Method 2" of Kirschvink (1980). NOTE:
There are several errors in Kirschvink (1980) regarding this method of
combining lines and planes; that paper implies that this technique is only
approximate, changes the mean direction of line data, and is probably inferior
to "Method 1." In fact it appears that there were bugs in the codes used by
Kirschvink and Method 2 is in fact exact both in producing a mean and in
producing error ellipses with the non-parametric statistic (with the one change
that each of the two orthogonal directions are added at full weight, not half
weight as suggested by Kirschvink (1980)). Using either "Treat as lines" or
"Combine lines and planes" will produce identical means if all the selected
directions are line-fits. As noted above, the "Watson" statistics were added
because they are more consistent than the Bingham statistics when comparing
line data analyzed as lines with line data transformed to pairs of planes; the
"Watson" and Bingham mean directions are the same. A discussion of these
statistics has been prepared and is available from