- Ground-motion model is:
where y is in g, a = 0.46, b = 0.35, c = 0.07, d = -4.79, e = 0.60, h = 8.9km and σ = 0.33.

- Investigates effect of nontriggering stations on derivation of empirical Ground-motion model based on the assumption that the triggering level is known (or can be estimated from data) but do not know which stations triggered (called left truncated data).
- Develops mathematical theory and computational method (after trying various alternative methods) for truncated regression analysis (TRA) and randomly truncated regression analysis (RTRA) (where triggering level changes with time).
- Tests developed methods on 1000 lognormally-distributed synthetic data points simulated using the
equation of Ambraseys et al. (1996) for 4 ≤ M
_{s}≤ 7 and 1 ≤ d_{f}≤ 100km. A fixed triggering threshold of 0.02g is imposed. Regresses remaining 908 samples using TRA and RTRA. Finds a very similar equation using TRA but large differences for d_{f}> 20km by using standard regression analysis (SRA) due to slower attenuation. Also apply TRA to randomly truncated synthetic data and find a close match to original curve, which is not found using SRA. - Applies method to 189 records from rock sites downloaded from ISESD with M > 4.5 (scale not specified)
and d < 80km (scale not specified) using functional form: log
_{10}(y) = a + bm + clog_{10}(). Uses these selection criteria to allow use of simple functional form and to avoid complications due to crustal reflections that reduce attenuation. Discards the five points with PGA < 0.01g (assumed threshold of SMA-1s). Applies TRA and SRA. Finds both M-scaling and distance attenuation are larger with TRA than with SRA because TRA accounts for larger spread in original (not truncated) data. Differences are relevant for M < 6 and d > 20km. - Applies method to dataset including, in addition, non-rock records (456 in total). Finds no differences between TRA and SRA results. Believes that this is due to lack of data in range possibly affected by truncation (small M and large d). Finds similar results to Ambraseys et al. (1996).
- Applies method to NE Italian data from seven seismometric and ten accelerometric digital stations
assuming: log
_{10}(y) = a + bm + clog_{10}(). Accelerometric stations used usually trigger at 0.001g. Seismometric stations used trigger based on ratio of short-term and long-term averages (STA/LTA), which varies from station to station and acts like a random threshold. Firstly neglects randomness and assumes trigger level of each station equals lowest recorded PGA and applies TRA and SRA. Finds small differences for d < 8km and d > 30km. - Applies method using functional form above, which believes is more physically justified. SRA does not converge. Studies reason for this by regressing on data from M intervals of 0.3 units wide. Finds behaviour of PGAs inverts for M < 3. Finds increasing σ with decreasing M for M > 3. TRA does converge and shows stronger magnitude saturation than SRA.
- Notes that application of RTRA to model effect of STA/LTA for used data is not realistic since probably not enough data to constrain all 23 parameters and to computational expensive using adopted maximization technique for RTRA.
- Estimates the random truncation parameters for one station (Zoufplan) and finds that the fixed threshold assumption made is acceptable since estimated random truncation parameters predict that only 14% of observations are lost at the earlier assumed fixed threshold level (the lowest PGA recorded).