During the spring and summer of 1995 three independent broadband networks operated in the Northeastern US, complementing the existing permanent observatories (Figure 2). Earthquakes recorded by this composite network are shown on Figure 3 and are also listed in Table 1. Stations are listed in Table 2. Data distribution is heavily biased towards the Pacific region, with only a handful of sources in other quadrants. The composite network contained a variety of seismic sensors. In order to homogenize the dataset we corrected observed waveforms for the respective instrument responses, filtered resulting time series between 0.01 and 0.2 Hz and resampled them at 10 samples per second. Seismic phases of interest were first identified by visual inspection, with rough (within a few seconds) time alignment done by picking a characteristic part of the broad-band waveform. The time window selected for the subsequent analysis varied between 25 and 75 s, depending on the length and clarity of the phase.
Shear wave splitting.
Measurements of shear-wave splitting were performed using
cross-correlation of horizontal components of motion
[Levin et al, 1999].
Briefly, we assume that the effect of propagation through the
anisotropic medium leads to the splitting of an initially rectilinear
S wave into two orthogonally polarized time-shifted versions of it.
We therefore seek a rotation 
and a delay 
that would
result in an optimally similar particle motion on two components, as
measured by cross-correlation.
We use core-refracted phases (SKS, SKKS and PKS) which contain
information about anisotropy only along the path from the core-mantle
boundary to the surface [Vinnik et al, 1984].
Figure 4 shows waveforms for two such phases that arrive from different
backazimuths, and maps with estimated splitting parameters.
Fast-axis estimates from the
same phase are similar at the majority of sites.
Fast-axis estimates from two different phases at a single site are
often significantly different.
In contrast, estimates of the splitting delay tend to vary more across
the composite network for single phases and between the two events.
Figure 5 shows all shear-wave splitting observations.
Splitting parameters at all stations exhibit a strong dependence on the
backazimuth.
Relative Traveltime Delays.
Relative traveltime delays were determined for all phases used in
shear-wave splitting analyses, and also for all clear S phases. The pulse shape of a typical shear phase we
observed does not vary significantly over the region, allowing us to
estimate relative travel-time delays using cross-correlation.
For core-refracted phases like SKS we used the radial components of
motion.
Experiments with other rotations (e.g., fast directions for the
particular phase) showed no significant difference in estimated
delays.
For S and S 
phases we chose either radial or transverse
component, depending on the clarity of the signal.
In cases where signal was equally clear on both components we used
both, and found close similarity in estimated delays.
Most earthquakes in our dataset were recorded by a subset of the full
composite network.
We took a number of steps, described below, to compensate as much as
possible for the changes that may be attributable to the variations in
network configuration.
Determination of traveltime delays was done using windowed time series that had already been aligned roughly. From the group of observations we constructed an average waveform. Using cross-correlation of time series, we aligned the records of a phase relative to one of the observations in the group, stacked and normalized them to produce an average pulse. We repeated this operation, using a different observation in the group as a base. All average pulses produced were aligned and stacked once more. The average waveform was related to the hypothetical location within the region determined as a geometric center of mass for the subset of stations observing the phase, and an average starting time was assigned to it. Constructed in this manner, the average waveform should be representative of the true pulse shape, as distortions due to the noise and small-scale structures at individual stations should be suppressed. We computed time shifts of individual phases relative to this average waveform, and corrected them for the time differences predicted by the IASPEI91 model given variations in the source-receiver distance. At this stage most groups of delays contained a linear trend against source-receiver distance, implying a deviation of the velocity structure from IASPEI91 along the source-receiver path. We estimated this trend by fitting a straight line to the relative delays, and removed it. Residual delays were generally less than 1.5 sec in magnitude, and were normally distributed (Figure 6). Most stations in the Adirondack mountains (BLUE, PACK, NCB, KEEN) have slow delays associated with them, stations GAC, PAL and LSCT are almost always fast (Figure 7). Stations ADVT, HRV, MM01 show strong azimuthal dependence of delays.