Variations in shear-wave splitting parameters.
Backazimuthal variability of shear-wave splitting parameters occurs
at all sites in the composite network that have a sufficient number of
observations.
Plotting fast directions for all stations as a function of
backazimuth (Figure 8) reveals an overall similarity of the pattern.
The data is matched well by the pattern predicted for the model of
anisotropy distribution under HRV derived by Levin et al., [1999].
Figure 9 and Table 3, taken from Levin et al. [1999], illustrate
this two-layer model.
In backazimuthal range 260 
- 290 two branches of the data
pattern overlap.
This overlap results from modest regional variations in the splitting
pattern, and are clear when only highest quality data is retained
(Figure 10).
The break in the pattern occurs at 260 for the stations close to
the Atlantic coast (PAL, LSCT, MM01, HRV), is at 290 for the
westernmost stations (BINY, MM03, MM04), and is not seen at all (or is
very much shifted) for the stations over the Grenville province (PACK,
BLUE, KEEN, GAC ) and station LBNH. Such "drift" of the break in the pattern can be explained by 
km thickness fluctuations in the layers of the HRV model (Figure 11).
The strong variation of splitting parameters with backazimuth can easily be mistaken for lateral heterogeneity, especially when different sets of events are used to estimate splitting at different stations. This effect may account for the early interpretation that the anisotropy in the northeastern North America has strong lateral variation. The difference in anisotropy between neighboring stations can be overestimated even when a common set of events is analyzed, if the event-set spans too small a range of backazimuths. A small rotation between azimuthally varying patterns could be mistaken for a large difference between the overall patterns. In light of the evidence gathered for this study, Levin et al. [1996] overestimated the heterogeneity between the Adirondack and New Hampshire regions.
Variation in Seismic Velocity.
Traveltime residuals indicate that shear velocity fluctuates locally
over the study region.
Stations separated by 100 km or less (e.g., MM01 and PAL, ADVT and
LBNH) display strong differences in their relative delay patterns
(Figure 7).
The distribution of sources in our data set is fairly uneven, and the
resulting sampling under the composite network by ray paths is not
optimal for tomographic imaging.
Nevertheless, the projection of delays onto the ray set using a simple
model provides useful constraints on the scale length and depth
distribution of the heterogeneity responsible for the delays.
We approximated the volume under the region with a rectangular box
divided into horizontal layers 100 km thick.
Each layer was divided into a number of boxes.
Using IASPEI91 velocity model with slight modifications, we traced S
rays through this model.
Hit counts in individual boxes and resolution tests with simple shapes
indicate that the layer between 100 and 200 km under the central part
of the network is sampled sufficiently well to resolve velocity
features of scale length 100-200 km.
Figure 12 shows S velocity distribution in the central part of this
layer.
The resolution of features outside the region
immediately under the network was poor.
The exact shapes of the
anomalies are strongly dependent on the number of boxes in the layer, and therefore are not well constrained.
Nevertheless the general locations of anomalies and an overall 
small-scale variation in velocity values persist with different
choices of parameterization.
The slow feature under the Adirondack Mountains dominates the image,
and is the most robust element of the inversion.
Overall, there appears to be little correlation between the splitting parameters estimated from the core-refracted shear phases and the distribution of the shear velocity in the volume under the region. While anisotropy indicators depend on back azimuth at all sites, the relative traveltime delays exhibit significant azimuthal variability at only 3 sites (ADVT, HRV and MM01). Only a slight variation of the two-layer anisotropic model is needed to accommodate regional variations in the azimuthal pattern of fast directions, but such modest variability will not predict the velocity perturbations (Figure 12) required to generate the observed delays.