- Ground-motion model is:
where Y is in cm∕s2, a = 1.92909, b = 0.21829, c = 0.00328, d = -1.06750, e = 0.01016, h = 0.01005, f

_{1}= 0.09664, f_{2}= 0.08438, e_{1}= 0.12297, e_{2}= 0.09175 and σ = 0.35530. - Use 3 site classes using the NEHRP classification:
- Rock
- Classes A and B (V
_{s}> 760m∕s). ST_{s}= SF_{s}= 0. - Stiff soil
- Class C (360 < V
_{s}≤ 760m∕s). ST_{s}= 1, SF_{s}= 0. - Soft soil
- Classes D and E (V
_{s,30}≤ 360m∕s. SF_{s}= 1, ST_{s}= 0.

Combine classes A and B and D and E together because of lack of records in classes A and E.

- Use 3 faulting mechanisms:
- Normal
- SS = RS = 0.
- Strike-slip
- SS = 1, RS = 0.
- Reverse
- RS = 1, SS = 0.

- Include focal depth H (1–30km) because it reduces σ.
- Model developed to demonstrate advantage of regression technique based on genetic algorithm with initial population development, using Latin Hypercube sampling over standard nonlinear least-square regression techniques. Compare results for PGA and simplified model using various regression techniques and adopted method.
- Unprocessed records taken from the Internet Site for European Strong-motion Data (Ambraseys et al., 2004), Geodynamic Institute of the National Observatory of Athens and ITSAK. Data from Greece, Italy, Turkey and Iran. Use metadata from ISC to reduce epistemic uncertainty.
- Use only data from free-field and basement-level (of buildings up to 2-storeys) stations.
- Records resampled to 200Hz and rejected poor-quality time-histories. Instrument-correct records from analogue instruments. Baseline-correct records (using pre-event mean for digital records and entire record for analogue records). Bandpass filter using zero-phase fourth-order Butterworth with optimal cut-off frequencies (depending of individual components).
- Use r
_{epi}because information on location of fault rupture not available for majority of events. - Note that alternative site classification should be investigated because current scheme does not account for depth of sediment.
- Find that using standard one-step random-effects technique that difficult to find physically-reasonable coefficients and the site and mechanism coefficients are poorly determined because of poor data distribution. Do not use a two-step approach because of the large number of singly-recorded events.
- Most data from r
_{epi}> 10km. - Plot residuals w.r.t. magnitude and fit trends. Find overestimation up to M
_{w}5, which relate to poor metadata. - Compute σs for sites with measured and estimated V
_{s,30}and for the three site classes. Find σ 10% lower for sites with measurements compared to those with only estimates. Find σ for rock sites 20% lower than for other classes. - Plot residuals w.r.t. distance and find no correlation.