Coefficients not reported since purpose is not to develop models for seismic hazard assessments but to
derive confidence limits on median PGA and thereafter to examine possible regional dependence of ground
motions.
Uses same site classifications as original studies. Si= 1 for site class i and 0 otherwise.
Adopts a simple linear functional form and standard one-stage regression method so that the covariance
matrices can be easily computed.
Assumes a fixed coefficient of 5km (a rough average value for this coefficient for most models using adopted
functional form) inside square root to make function linear.
Examines 95% confidence limits on PGA since it is standard to use 5% significance levels when testing
null hypotheses. Plots predicted median PGAs and their confidence limits for Mw5, 6.5 and 8.0 up to
200km to show effects of extrapolation outside range of applicability of models. Finds that confidence
limits for models derived using limited data (Ulusay et al., 2004; Kalkan and Gülkan, 2004b; Sabetta
and Pugliese, 1987) are wider than models derived using large well-distributed datasets (Joyner and
Boore, 1981; Boore et al., 1993, 1997; Ambraseys et al., 1996, 2005a). Notes that for 5.5 < Mw< 7
and 10 ≤ df≤ 60km the 95%-confidence limits of the median are narrow and within bands 10–30% from
the median but for other magnitudes and distances (away from the centroid of data) they are much wider
(bands of 100% from the median). Notes that inclusion of data from large magnitude events decreases the
width of the confidence limits of the model derived using the data of Boore et al. (1993, 1997) compared
with that derived using the data of Joyner and Boore (1981) and similarly that derived with the data of
Ambraseys et al. (2005a) compared with that derived using the data of Ambraseys et al. (1996).