- Ground-motion model is:
where Y is in g, a = -3.2191 ± 0.16, b = 0.7194 ± 0.025, c = -1.7521 ± 0.075, d = 0.1780 and σ = 0.282.

- Originally use three site classes based on Eurocode 8:
- A
- Rock, V
_{s,30}> 800m∕s. Marine clay or other rocks (Lower Pleistocene and Pliocene), volcanic rock and deposits. 11 stations. 833 records. - B
- Stiff soil, 360 < V
_{s,30}< 800m∕s. Colluvial, alluvial, lacustrine, beach, fluvial terraces, glacial deposits and clay (Middle-Upper Pleistocene). Sand and loose conglomerate (Pleistocene and Pliocene). Travertine (Pleistocene and Holocene). 6 stations. 163 records. - C
- Soft soil, V
_{s,30}< 360m∕s. Colluvial, alluvial, lacustrine, beach and fluvial terrace deposits (Holocene). 3 stations. 67 records.

Classify stations using geological maps. Find that results obtained using this classification are not realistic because of some stations on very thick (> 1000m) sedimentary deposits whose amplification factors are small. Therefore, use two site classes using H/V ratios both using noise and earthquake records. Confirm H/V results by computing magnitude residuals at each station.

Final site classes are:

- Rock
- Site amplification factors < 2 at all considered frequencies from H/V analysis. 422 records. S
_{soil}= 0. - Soil
- Site amplification factors > 2. 641 records. S
_{soil}= 1.

- Use data from velocimeters (31 stations) and accelerometers (2 stations) from 33 sites with sampling rates of 62.5samples∕s.
- Relocate events and calculate M
_{L}. - Exclude data from M
_{L}< 2.5 and r_{hypo}> 300km. - Few near-source records (r
_{hypo}< 150km) from M_{L}> 4 but for M_{L}< 4 distances from 0 to 300km well represented. - Exclude records with signal-to-noise ratios < 10dB.
- Correct for instrument response and bandpass filter between 0.5 and 25Hz and then the velocimetric records have been differentiated to obtain acceleration.
- Visually inspect records to check for saturated signals and noisy records.
- Compare records from co-located velocimetric and accelerometric instruments and find that they are very similar.
- Compare PGAs using larger horizontal component, geometric mean of the two horizontal components and the resolved component. Find that results are similar and that the records are not affected by bias due to orientation of sensors installed in field.
- Try including a quadratic magnitude term but find that it does not reduce uncertainties and therefore remove it.
- Try including an anelastic attenuation term but find that the coefficient is not statistically significant and that the coefficient is positive and close to zero and therefore remove this term.
- Try using a term clog
_{10}rather than clog_{10}(R) but find that h is not well constrained and hence PGAs for distances < 50km underpredicted. - Find that using a maximum-likelihood regression technique leads to very similar results to the one-stage least-squares technique adopted, which relate to lack of correlation between magnitudes and distances in dataset.
- Find site coefficients via regression following the derivation of a, b and c using the 422 rock records.
- Compare observed and predicted ground motions for events in narrow (usually 0.3 units) magnitude bands. Find good match.
- Examine residuals w.r.t. magnitude and distance and find no significant trends except for slight underestimation for short distances and large magnitudes. Also check residuals for different magnitude ranges. Check for bias due to non-triggering stations.
- Compare predicted PGAs to observations for 69 records from central northern Italy from magnitudes
5.0–6.3 and find good match except for r
_{hypo}< 10km where ground motions overpredicted, which relate to lack of near-source data.