## Abstract

Slip rate, locking width and ground-motion prediction equation (GMPE) selection are important in seismic hazard analysis because they are used to estimate earthquake recurrence, to limit the maximum magnitude in an earthquake source and to estimate earthquake ground shaking. In this study, we examine the sensitivity of probabilistic seismic hazard analysis (PSHA) to fault slip rates, fault locking width and the selection of GMPEs for the Aceh Fault Segment, Indonesia. The hazard level differences vary considerably owing to changes in these three parameters. Therefore, careful consideration is needed in applying PSHA in areas of high fault parameter uncertainty.

The Sumatran Fault System (SFS) in Indonesia is a major right-lateral system consisting of 19 segments. It runs parallel to the Sumatra Subduction Zone, where the Australian Plate subducts beneath the Burma Microplate and Sunda Block (Fig. 1). Sumatra is a classical example of a partitioned subduction zone, where much of the trench-normal plate motion is accommodated by the megathrust plate boundary, but much of the trench-parallel motion is accommodated by the SFS. According to historical and instrumental data, most of the 19 segments have ruptured within the last century. The largest recorded magnitudes for events on the segments vary between M_{w} 6.5 and 7.7 (Natawidjaja & Triyoso 2007).

Knowledge of the slip rate and locking depth along the SFS is sparse, and has changed little since the geological study of Sieh & Natawidjaja (2000) and the geodetic study of Genrich *et al.* (2000), and has been summarized more recently by Natawidjaja & Triyoso (2007). Geological studies generally suggest an increase in slip rate from 11 mm a^{−1} below the equator to 27 mm a^{−1} above it, while geodetic estimates of slip vary between 19 and 26 mm a^{−1}, with no apparent trend along strike. Locking depths determined from GPS measurements (Genrich *et al.* 2000) vary between 9 and 56 km. Knowledge of slip rate and locking depth is particularly sparse for the northernmost tip of the SFS, where it splits into two fault segments: the Aceh and the Seulimeum faults (Fig. 1). Slip rates for this segment vary from as low as 5 mm a^{−1} (Genrich *et al.* 2000) to 38 mm a^{−1} (Bennett *et al.* 1981). A more recent study by Ito *et al.* (2012) suggests a locking width of between 1 and 30 km, with a slip rate of 20±6 mm a^{−1}.

This uncertainty in fault parameters is potentially important in seismic hazard analyses, both probabilistic and deterministic, which requires an assessment of the future earthquake potential in a region. In this paper, we explore the sensitivity of probabilistic seismic hazard analysis (PSHA) on the Aceh segment to three key parameters: slip rate; maximum magnitude as determined by locking width; and the selection of ground-motion prediction equations (GMPEs: also known as attenuation functions). Slip rate is important in determining seismic hazard because it is used to estimate earthquake recurrence (how often a fault is expected to host an earthquake). Earthquake hazard estimates are sensitive to slip rate because slip rate controls the overall rate of earthquake activity on a fault (e.g. Youngs & Coppersmith 1985; Kramer 1996).

Estimating the maximum expected earthquake on a specific fault or earthquake source is also important in PSHA (e.g. Beauval & Scotti 2004). It is rare, however, that the largest possible earthquakes along individual faults have occurred during the historical recording period. Therefore, maximum magnitude is often evaluated from estimates of fault dimensions (e.g. Wells & Coppersmith 1994; Leonard *et al.* 2007), and these dimensions are usually determined from a combination of geological and geodetic observations.

The latest Indonesian seismic hazard map (Irsyam *et al.* 2010) assumed a constant locking width of 17 km (depth of 3–20 km) for all SFS segments. However, Ito *et al.* (2012) used GPS to study the locking width of the Aceh Fault Segment and found uncertainty in its size, which ranged from a few kilometres up to 35 km, twice that used by Irsyam *et al.* (2010). Since earthquake magnitude depends on the area of rupture, which is the product of fault length and locking width, the maximum magnitude on a fault is expected to vary with locking width. However, the uncertainty in fault locking width is rarely considered in PSHA.

Another important aspect to consider in PSHA is the selection of the GMPEs. There are no GMPEs derived for the Indonesian region. Consequently, Indonesia's seismic hazard analyses adopt GMPEs from other regions. For example, Irsyam *et al.* (2010) used Boore & Atkinson (2008), Campbell & Bozorgnia (2003) and Chiou & Youngs (2008) in equal weights to estimate ground motion from crustal fault sources. Sabetta *et al.* (2005) demonstrated that the selection of GMPEs has a greater impact than the expert judgement applied in assigning relative weights to the GMPEs.

## Sources of uncertainty in seismic hazard analysis

Seismic hazard may be analysed deterministically (e.g. when a particular earthquake scenario is assumed) or probabilistically, in which all possible earthquakes of various magnitudes, locations and probabilities of occurrence are explicitly considered (Kramer 1996). We focus here on PSHA, which can be summarized in four steps (Reiter 1990):

Identification and characterization of earthquake sources. These include:

area source zones: in which an earthquake can occur anywhere in a certain area with a probability that is typically determined from an analysis of earthquake catalogue data (i.e. a recurrence relationship);

fault sources: where earthquakes can occur on particular faults with a specified probability typically determined from either catalogue data, or from slip rate and locking width estimated using either geological or geodetic observations.

Application of GMPEs to forecast the level of shaking or propagation of ground motion. GMPEs are determined from analyses of strong-motion data.

Integration (or summation) over the uncertainties in earthquake location, earthquake size and ground-motion parameters to obtain the probability that the ground-motion parameter will be exceeded during a particular time period.

Steps (1a), (b) and (2) all introduce uncertainty into the PSHA analysis. The GMPEs of step (2) are each associated with an aleatory uncertainty that reflects the observed variability in the ground-motion measurements and is formally included in the PSHA. In addition, the epistemic uncertainty associated with the lack of knowledge about which GMPE best describes the actual ground motion is usually accounted for via a logic tree.

However, none of the parameters describing the recurrence relationships of step (1a) and the fault behaviour in step (1b) are known perfectly, but their uncertainly is rarely considered (except for maximum magnitude, for which alternative values are sometimes considered in a logic tree). In this study, we explore the sensitivity of the PSHA results to the uncertainty in fault parameters, namely fault slip rate and maximum magnitude (as determined by locking width). Although PSHA results are also affected by uncertainty in recurrence parameters for aerial source zones (step 1a), an assessment of this sensitivity is beyond the scope of this paper.

In order to account for uncertainty in fault parameters in PSHA, we need to consider whether this uncertainty is best described as aleatory or epistemic. Aleatory is often defined as the natural variability of a quantity that cannot be reduced by making additional measurements (McGuire 2004). The uncertainty in GMPEs is normally regarded as an aleatory uncertainty. Epistemic uncertainty, however, is usually regarded as uncertainty due to a lack of knowledge and can, in principle, be reduced if additional data become available. Uncertainty in maximum magnitude, and that represented by choice of GMPE, are often regarded as epistemic uncertainties that are accounted for via the use of logic trees. We prefer the more practical definitions of Bommer (2003), in which aleatory uncertainty is defined as uncertainty that can be measured, while epistemic uncertainty is that which must be judged.

The estimations of the fault parameters slip rate and locking width used in this paper are based on measurements of offsets in dated geological formations and geodetic strain rates, respectively. Both types of observations are associated with measurable uncertainties. In the case of fault slip, the uncertainties are estimated from the measurement error in offset length and in the dating, and the uncertainty in locking width is estimated from the analysis of GPS velocities. We therefore consider the uncertainty in fault parameters as aleatory, and attempt to formally include this as part of the PSHA in a way analogous to that used for GMPEs.

## Methodology for sensitivity analysis

The modelling in this study utilizes the Earthquake Risk Model (EQRM), which is a computer package for estimating earthquake hazard and earthquake risk (Robinson *et al.* 2005, 2006). The EQRM is developed by Geoscience Australia and can be freely downloaded from http:/code.google.com/p/eqrm/source/checkout

The EQRM uses an event-based approach. This approach differs from the traditional ‘site-based’ approach to PSHA integration over the magnitude and distance distributions on a site-by-site basis. The traditional approach was introduced by Cornell (1968) and was summarized by McGuire & Arabasz (1990). In contrast, an event-based analysis with the EQRM begins with the generation of a single simulated event catalogue, which is, in turn, used to generate ground-motion fields at all sites. Both the ‘site-based’ and ‘event-based’ approaches solve the same equations – they differ only in numerical mechanics. The generation of the synthetic event catalogue relies on a model for the seismicity in the region. Typically, the model of seismicity comes from an interpretation of historical earthquakes, geology and neotectonics (Robinson *et al.* 2005). In this paper, the seismicity model consists of the geometry and slip rate of the Aceh Fault Segment.

We focus on the Aceh Fault Segment, located in northern Sumatra, which typically hosts events with a strike-slip mechanism. Its length is approximately 200 km (Fig. 2). In the Irsyam *et al.* (2010) seismic hazard map, the slip of the Aceh segment is 2 mm a^{−1}, the locking width is 17 km and the hazard is derived using three GMPEs, all with equal weights. The GMPEs used are Boore & Atkinson (2008), Campbell & Bozorgnia (2003) and Chiou & Youngs (2008) (Irsyam *et al.* 2010).

We study the sensitivity of fault slip rates, uncertainty in locking width and different GMPEs for PSHA by comparing and analysing hazard curves at two different sites. Both sites are located near the middle of the fault trace, with Site 1 located on top of the fault line (0 km) and Site 2 located 20 km perpendicularly from the segment mid-point (see Fig. 2). We also compare the hazard maps for Aceh using different input parameters.

## Slip-rate sensitivity analysis

Reported slip rates for the Sumatran Fault System (SFS) are based on geological and geodetic studies. The latest seismic hazard map of Indonesia (Irsyam *et al.* 2010) assigned slip of 2 mm a^{−1} to the Aceh Fault. However, recent work by Ito *et al.* (2012) estimated the slip rate in the middle of the Aceh segment (Profile A in Fig. 2) to be about 20±6 mm a^{−1}. Since slip-rate estimates are used to determine the recurrence interval, it is important to study how hazard varies with slip rate. Our study focuses on the sensitivity of PSHA to fault slip rate. We do not attempt to determine which slip-rate estimates are most suitable for the Aceh segment.

Figure 3a, b illustrates a comparison between the hazard curves for two different slip rates at sites 1 and 2. The green and red dashed lines are the hazard curves for 2 and 20 mm a^{−1} slip, respectively. Figure 3a shows the hazard curves comparison at a site located in the middle of the fault (Site 1). The hazard level is significantly higher for larger slip rate at all return periods. At a 500 year return period, which is commonly used in PSHA, the hazard discrepancy is significant: 0.83*g* for 20 mm a^{−1} slip and 0.18*g* for 2 mm a^{−1} slip. This represents a variation in hazard of approximately 400%. The comparison at the second site, located 20 km from the fault, also indicates a similar trend (Fig. 3b). At a 500 year return period, the hazard level for a higher slip rate reaches 0.31*g*, while the hazard level for a lower slip reaches 0.09*g*. This represents a 300% variation. As expected, the hazard level at the second site illustrates that hazard decreases as we move away from the fault. These hazard curve comparisons show that seismic hazard analysis is sensitive to fault slip rate (in fact, it is linear in slip rate) and demonstrate that the distance from an earthquake source plays an important role in the hazard level.

We also compare the hazard maps derived using these two slip rates for the Aceh Fault Segment. Figure 4 shows a comparison for the Aceh hazard maps at a 500 year return period. Figure 4a is the hazard map using 2 mm a^{−1}, while Figure 4b is the hazard map using 20 mm a^{−1}. Both maps demonstrate that the highest hazard level is located along the fault trace, and that hazard decreases as we move further from the fault. The main differences between these two maps are the maximum hazard level and the area impacted. A higher slip rate clearly leads to a greater maximum of ground shaking and a larger area impacted by future earthquakes. Consequently, the recent work by Ito *et al.* (2012) suggests that the hazard level in the Aceh area (Irsyam *et al.* 2010) is underestimated.

## Locking width (maximum magnitude) sensitivity analysis

Another key parameter in PSHA is the maximum magnitude, which can be related to the locking width. Ito *et al.* (2012) studied GPS constraints on locking width for two cross-sections on the Aceh Fault Segment (Fig. 2). Their results are presented as locking width probability density functions (PDFs) for two profiles in Figure 5. We use the locking width data from Ito *et al.* (2012) to estimate a PDF for the maximum magnitude using the formula from Wells & Coppersmith (1994) (Table 1) that relate rupture width to magnitude. For the purpose of this exercise, we assume a mapped length of 200 km for the Aceh Fault Segment.

To study how broadly the hazard level might vary due to aleatory uncertainty in the locking width, we undertook a detailed sensitivity study by sampling the maximum magnitude PDF tightly (1000 samples) and then plotted all the hazard curves (the grey shaded hazard curves in Fig. 3c, d). The hazard curve from the largest maximum magnitude in the PDF is presented as the dashed red line, while the smallest maximum magnitude is presented as the magenta dashed line. These hazard curves indicate a large variation in hazard at all return periods. However, the largest discrepancies occur at longer return periods.

The maximum magnitudes based on the Ito *et al.* (2012) locking widths are also plotted as a maximum magnitude PDF (Fig. 6). The mean value for the maximum magnitude from this PDF is M_{w} 6.13, which is considerably smaller than the M_{w} 7.7 used by Team-9 (Irsyam *et al.* 2010).

Standard PSHA tools, such as the EQRM we use here, do not provide for the incorporation of aleatory uncertainty in fault parameters, but allow only for epistemic uncertainty to be included using logic trees. Therefore, in order to incorporate the maximum magnitude aleatory uncertainty into our PSHA, we had to borrow the tools of epistemic uncertainty and then apply them to an aleatory uncertainty: that is, we used a logic tree approach with parameters chosen in such a way as to account for the aleatory uncertainty in the maximum magnitude. To achieve this, we sampled the maximum magnitude PDF by dividing it into five different quantiles with an equal one-fifth area below the PDF (Fig. 6). The median values for each of the five quantiles can be seen in Table 2.

As described above, the five median (i.e. from each quantile) maximum magnitudes were used separately in the hazard calculations and used to derive the weighted hazard curve. The weighted hazard curve is the weighted sum of the five hazard curves using equal weights of one-fifth (see Table 2). We also compared the difference between the weighted hazard curves using three, five and seven quantiles. As the latter two results were similar, we reasoned that five quantiles are sufficient to represent the aleatory uncertainty.

To assess the differences between hazard curves utilizing different maximum magnitude estimates and the hazard curve that accounts for aleatory uncertainty by combining the five quantiles, we compared the hazard curves for two different locations in the Aceh Fault Segment (Fig. 1). The comparison between the hazard curves for the five median maximum magnitudes and also the corresponding weighted hazard curve derived from these five median hazard curves at sites 1 and 2 can be seen in Figure 3c, d.

Figure 3c shows a comparison for Site 1, located in the middle of the Aceh Fault. The hazard curves for different maximum magnitudes differ widely at most return periods. At shorter return periods, the hazard level for lower maximum magnitude is higher than the hazard curve with the largest maximum magnitude. For instance, a comparison between the M_{w} 4.8 and M_{w} 7.1 hazard curves at a 100 year return period show that the peak ground acceleration (PGA) for the lower magnitude is 0.65*g*, and 0.56*g* for the higher magnitude. However, at a longer return period (100 kyr), the hazard level for the lower magnitude reaches 0.9*g*, while the hazard level for a higher magnitude is 1.84*g*. This result is consistent with that of Youngs & Coppersmith (1985), who studied a similar variation in hazard curve changes with maximum magnitude given a constant moment rate. They found that lowering the maximum magnitude with constant moment rate forces the activity rates of smaller earthquakes to increase, so that hazard levels increase for short return periods. The approach that we used does not assume a constant moment rate because, for each different maximum magnitude, we also change the moment rate based on the locking width for each magnitude. Nevertheless, we observe a similar effect.

The hazard curve at Site 2, located 20 km from the fault, is shown in Figure 3d. The hazard curves show a similar trend to those of the previous comparison. The main difference is that the hazard level decreases significantly at each maximum magnitude. This is simply due to the increased distance from the fault.

We are also interested in understanding the implication of uncertainty in the maximum magnitude on our understanding of hazard, as illustrated in the latest Indonesian seismic hazard map. To explore this, we considered the hazard contribution from the Aceh Fault Segment using the maximum magnitudes of the Aceh Fault corresponding to the M_{w} 7.7 used by Irsyam *et al.* (2010), the mean of the maximum magnitude PDF of M_{w} 6.13 of Ito *et al.* (2012), the weighted hazard curve using the five quantile maximum magnitudes (as given in Table 2) and results from 1000 maximum magnitudes sampled from the maximum magnitude PDF of Ito *et al.* (2012). These results for sites 1 and 2 (Fig. 2) can be seen in Figure 3c, d.

The first hazard curve comparison for Site 1 (Fig. 3c) shows that the hazard level for higher maximum magnitudes is smaller at shorter return periods. At a 500 year return period, the hazard level for M_{w} 7.7 is 0.85*g*, while the weighted magnitude and the M_{w} 6.13 are 0.99*g* and 1.08*g*. The difference is significant, reaching up to 0.23*g*. At a 1000 year return period, the difference in hazard level between these three different magnitudes becomes smaller, about 0.03*g*–0.17*g*. At a 2500 year return period, the hazard differences are much smaller, at between 0.01*g* and 0.07*g*. However, at very long return periods (100 kyr), the hazard level from the largest maximum magnitude (M_{w} 7.7) reaches 1.83*g*, which is larger than the weighted magnitude and M_{w} 6.13, which have hazard levels of 1.59*g* and 1.66*g*, respectively. The hazed grey lines, which represent the sampled maximum magnitudes from the maximum magnitude PDF derived from the results of Ito *et al.* (2012), show how wide the uncertainty in the hazard level might be due to the uncertainty in the locking width from the geodetic observations in northern Sumatra. The maximum hazard level from these data varies between 0.7*g* and 1.8*g*. These hazed grey plots also show a similar trend. The hazard level for a lower maximum magnitude at shorter return period would be higher than the hazard level for higher magnitude at a shorter return period.

The second comparison at Site 2 (Fig. 3d) shows a similar trend with the comparison at Site 1 (Fig. 3c). However, the probability for exceeding a certain ground-motion level (e.g. 0.5*g*) occurred at a longer return period. The difference varies between 0.04*g* and 0.06*g* at a return period of 500 years. At a return period of 1000 years, the difference is about 0.1*g*: while at a 2500 year return period, it reaches 0.16*g*. The overall hazard level has decreased, which is simply owing to the increased distance from the fault.

Since the hazard map consists of many sites at a single return period, it is important to note that the hazard level for lower maximum magnitude is higher than the hazard level for higher maximum magnitude at shorter period. Based on this result, a careful consideration is needed before we assign a certain maximum magnitude for PSHA work.

## Ground-motion prediction equation (GMPE) sensitivity analysis

The other aspect that we study is the impact from the selection of GMPEs. Many studies mention the importance of the GMPE selection for PSHA (e.g. Lombardi *et al.* 2005; Sabetta *et al.* 2005; Bommer & Abrahamson 2006). The problem of GMPE selection is further complicated in Indonesia as there are no GMPEs derived from Indonesian earthquakes. The GMPEs used in the most recent hazard map of Irsyam *et al.* (2010) are Boore & Atkinson (2008), Campbell & Bozorgnia (2003) and Chiou & Youngs (2008). The seismic hazard map of Irsyam *et al.* (2010) used an equal weight for each of these GMPEs. We examine the contribution from these three GMPEs at sites 1 and 2, and compare the results with the weighted results. We also compare the hazard curves and hazard maps using each of the GMPEs used in the Irsyam *et al.* (2010) seismic hazard map. As with Irsyam *et al.* (2010), we use equal weights when combining the GMPEs.

Figure 3e, f show the hazard curve comparisons at sites 1 and 2. The first comparison at Site 1 (0 km) shows that Boore & Atkinson (2008) (BA08) and Chiou & Youngs (2008) (CY08) lead to larger hazard estimates than Campbell & Bozorgnia (2003) (CB08). The result for the weighted combination of GMPEs sits in the middle. The hazard level for the weighted combination is about 23% larger than CB08. However, compared to BA08 and CY08, the weighted combination hazard level is about 13 and 11% smaller, respectively. The hazard level for BA08 is 5% smaller than CY08 for shorter return periods (<600 years), and larger by approximately 10% for longer return period (>800 years).

The second comparison at Site 2 (20 km) shows a different result (Fig. 3f). The hazard level for CY08 is larger than BA08 and CB08 for all return periods. This shows that the hazard level dependence on GMPE depends on distance, highlighting the important role of GMPE selection on seismic hazard analysis. The weighted combination GMPEs in this case also sits in the middle between CB08, BA08 and CY08. The main difference compared to the first site is that the overall hazard level decreases owing to the increased distance from the fault.

## Discussion

The sensitivity analysis for probabilistic seismic hazard analysis (PSHA) for two different sites near the Aceh Fault Segment, Indonesia, shows a significant sensitivity in hazard level to the parameters that we analysed: slip rates, locking widths (M_{max}) and GMPEs.

The hazard curves comparison for the Aceh segment using two different slip rates show significant differences in hazard level. The previous slip rate used for this fault segment (Irsyam *et al.* 2010) is only 10% of that suggested by more recent work (Ito *et al.* 2012). We clearly show that seismic hazard analysis is sensitive to such a large difference in fault slip rates, with a higher slip rate leading to a higher maximum hazard, and a larger area subject to high hazard levels. In general, geodetic studies in the northern part of Sumatra suffer from limited measurements in the far field because of the narrow land area available for measurements. Therefore, it seems important to account for uncertainty in geodetic measurements of slip rate there.

Fault geometry, particularly the locking width, influences the maximum potential magnitude of a fault and is, hence, important for hazard. It is common to use geological data to estimate maximum magnitude: however, recent developments in geodesy (Ito *et al.* 2012) can provide researchers with a PDF for locking width and, hence, a new option for determining the maximum magnitude on faults. The analysis for different maximum magnitudes show that the hazard level for lower maximum magnitude (a shorter return period) is higher than the hazard level for higher magnitude at shorter return periods. While this has been found to be the case when moment rate is held constant (Youngs & Coppersmith 1985), we find that it is still true for the case of the Aceh Fault, even when the moment rate increases as the locking width increases.

GMPE selection, however, shows a very significant difference in the hazard level, and we found that the difference may vary at difference distances. A GMPE that results in the highest hazard level at a certain distance might have smaller hazard level at a different distance. However, given the fact that there is no single GMPE derived for Indonesia, a further study is needed in order to select the most reliable GMPE(s) for use in seismic hazard analysis in Indonesia.

## Conclusions

In this study, we present a sensitivity analysis for three parameters used in probabilistic seismic hazard analysis. We analysed hazard curves and maps for two different sites in the Aceh Fault Segment, Indonesia. The parameters that we studied were slip rates, locking widths and GMPEs.

The comparison between different slip rates shows a significant hazard difference between smaller (2 mm a^{−1}) and larger (20 mm a^{−1}) slip rates. The study at Site 1 shows that at a 500 year return period, the hazard level from smaller slip is 0.06*g*, while the larger slip hazard level is 0.55*g* (Fig. 3a), resulting in a difference of 0.49*g*, which is significant. Another comparison at a 2500 year return period shows that the hazard level for smaller slip reaches 0.45*g*, while the larger slip produces 1.13*g*, which still gives a difference of approximately 0.68*g*. A similar result also appears at Site 2 (Fig. 3b): however, owing to the increased distance from the fault, the hazard level is smaller.

Conversely, the maximum magnitude (as determined from the fault locking width) also plays an important role in determining the hazard level. The comparison between M_{w} 4.6 (magenta dashed line) and M_{w} 7.7 (red line) at Site 1 (Fig. 3c) shows two important points. The first point is that the hazard difference is 0.08*g* at a return period of 500 years. The second point is that the hazard level from a smaller maximum magnitude results in a higher hazard level at a return period of <700 years. At longer return periods, the hazard level from the higher maximum magnitude shows a significant difference. For example, at a 2500 year return period, a M_{w} 7.7 maximum magnitude results in a 1.06*g* exceedance in ground shaking, at M_{w} 4.6, however, the maximum magnitude results in only 0.68*g*. The comparison at Site 2 (Fig. 3d) also shows a similar result, but the hazard level is smaller owing to the increased distance.

The study of different GMPEs also shows the importance of the correct selection of the GMPE. At smaller return periods, the difference is relatively small, while at longer return periods the difference is larger, as expected. At a 500 year return period (Fig. 3e), CY08 resulted in a similar hazard level to BA08: for CB08, however, the hazard level is smaller by 0.21*g*. At a 2500 year return period, the hazard level difference with CY08 and BA08 was still small, but the difference with CB08 increased to about 0.41*g*. However, at a return period of >2500 years, the hazard difference between CY08 and BA08 is larger. The comparison at Site 2 (Fig. 3f) for the GMPE selection shows a different trend. The hazard level between the three GMPEs can be seen clearly at both short and long periods. The main difference compared to Site 1 is that the hazard level is smaller owing to the increased distance from the fault.

The results show that each of the parameters studied are equally important and affect the hazard level significantly. Therefore, careful consideration is needed as we work on seismic hazard analysis in areas where there is significant uncertainty in fault parameters and GMPEs. However, it is important to note that PSHA also depends on other factors, such as earthquake catalogue data and source zonation, that have not been considered in this study.

## Acknowledgments

This work was supported by the Australian Department of Foreign Affairs Australian Aid Program and Australian Research Council Linkage Project LP11010525.

- © 2017 The Author(s). Published by The Geological Society of London. All rights reserved