Hazard magnitude scales are widely adopted to facilitate
communication regarding hazard events and the corresponding decision making
for emergency management. A hazard magnitude scale measures the strength of
a hazard event considering the natural forcing phenomena and the severity of the event with respect to average entities at risk. However, existing hazard
magnitude scales cannot be easily adapted for comparative analysis across
different hazard types. Here, we propose an equivalent hazard magnitude
scale to measure the hazard strength of an event across multiple types of
hazards. We name the scale the
Natural hazards pose significant challenges to human societies around the world. Between 2000 and 2020, natural hazard events caused over USD 130 billion in losses and 64 695 fatalities, and they affected more than 196 million people on average each year (Guha-Sapir et al., 2021). Hazardous events, such as earthquakes, floods, and forest fires, can inflict heavy losses on communities when people and property are exposed to the natural forces of these events. The impacts of events, whatever their type, can be quantified directly (e.g. by financial loss; Hillier et al., 2015) or estimated on a scale. To estimate the impacts of an event with the consideration of its hazard strength, various impact scales have been proposed, including the Bradford disaster scale (Keller et al., 1992, 1997), unified localizable crisis scale (Rohn and Blackmore, 2009, 2015), disaster impact index (Gardoni and Murphy, 2010), and cascading disaster magnitude (Alexander, 2018). However, a hazard strength scale is not the same as a hazard impact scale, as impacts are also driven by the exposure and vulnerability of entities, such as individuals, communities, and infrastructures, to an event. This makes it difficult to use impact scales to compare hazard strengths across natural hazard types. For example, the 2011 Christchurch earthquake was one of the most destructive earthquakes in New Zealand, albeit with a medium hazard strength of 6.2 in terms of its moment magnitude (Kaiser et al., 2012). Meanwhile, the 1964 Alaskan earthquake, with a larger moment magnitude of 9.2, resulted in fewer casualties and less economic damage than the Christchurch earthquake (USGS, 2021).
Hazard scientists have long called for separation of natural forcing phenomena (Bensi et al., 2020) from the study of disasters to better understand the causes of impacts rooted in the social and economic fabric of entities exposed to natural hazards (e.g. O'Keefe et al., 1976; Wisner et al., 2004). In this regard, quantifying hazard strength helps separate the natural force from other social, environmental, and engineering or built environmental factors that may drive impacts. Yet, despite the large volume of research that focuses on hazard strength for singular natural hazard types such as earthquake (e.g. Wood and Neumann, 1931; Richter, 1935; Kanamori, 1977; Katsumata, 1996; Grünthal, 1998; Wald et al., 2006; Rautian et al., 2007; Serva et al., 2016), tropical cyclone (e.g. Simpson and Saffir, 1974; Bell et al., 2000; Emanuel, 2005; Powell and Reinhold, 2007; Hebert et al., 2008), tornado (e.g. Fujita, 1971, 1981; Meaden et al., 2007; Potter, 2007; Dotzek, 2009), and drought (e.g. Palmer, 1965, 1968; Shafer and Dezman, 1982; McKee et al., 1993; Byun and Wilhite, 1999; Shukla and Wood, 2008; Hunt et al., 2009), few have quantified or modelled hazard strengths across multiple hazard types.
To quantify hazard strengths for cross-hazard comparison, impacts can be used to explore similarities between multiple hazards (e.g. Hillier et al., 2015; Hillier and Dixon, 2020). As an example, insurance professionals often leverage loss metrics to understand the relative significance of various hazards (see, e.g. Mitchell-Wallace et al., 2017). Such cross-hazard practices of risk aggregation and accumulation are intentionally focused on the exposed values and observed impacts, rather than hazard strengths. In contrast, risk quantification for nuclear facilities requires consideration of hazard strengths across multiple hazard types to facilitate probabilistic safety assessment within a multi-hazard context (see, e.g. Choi et al., 2021). Indices regarding hazard strengths have also been created and adopted for extreme meteorological events across multiple hazard types (see, e.g. Malherbe et al., 2020). When quantifying hazard strengths within a multi-hazard context, a calibration of hazard strength to the expectation of impact may be used to create impact-based proxies for hazard strengths, linking two extremes and allowing them to be studied in a way that is relevant to risk assessment and yet decoupled from the detail of exposed values and vulnerability (Hillier et al., 2020). Nevertheless, there is not yet a general metric that facilitates the comparison of events of different hazard types in terms of potential to cause damage in a way that is as decoupled as possible from exposed values and vulnerability.
To enable evaluation of event-wise hazard strengths across different hazard
types, in this article, we propose a multi-hazard
The subsequent sections are organized as follows. First, we provide a brief theoretical background for this study. We then introduce our methodology, including data processing, to derive the equivalent hazard magnitude on the Gardoni Scale. Next, we describe the results of applying our methodology and compare natural hazard types regarding the derived equivalent hazard magnitudes. Finally, we discuss the potential contributions and limitations of the proposed scale before concluding the article.
In natural hazards research, theoretical frameworks are often based on basic concepts, such as hazard, impact, exposure, vulnerability, recovery, and resilience, that have overlapping or discipline-specific definitions (see, e.g. Klijn et al., 2015). These inconsistencies across disciplines often result in confusion in quantitative modelling. Herein, the impacts of an event are the result of strength of the hazard agent, value of entities exposed to the event, and vulnerability of the exposed entities to hazard impacts (Nigg and Mileti, 1997; Coburn and Spence, 2002; Wisner et al., 2004; Dilley et al., 2005; McEntire, 2005; Adger, 2006; Peduzzi et al., 2009; Burton, 2010; Lindell, 2013; Birkmann et al., 2014; Highfield et al., 2014; van de Lindt et al., 2020; Wang et al., 2020; Wang and Sebastian, 2021). As shown in Fig. 1, hazard strength of an event is one of the main drivers, albeit not the sole driver, of impacts.
Hazard event impacts as a result of hazard strength, exposed value, and vulnerability of exposed entities.
Hazard strength is often referred to as the hazard magnitude or hazard
intensity (Blong, 2003; Alexander, 2018). However, these two concepts are
not equivalent. Hazard magnitude is a measure of the size of, or the total
energy involved in, the entirety of a hazard event (Blong, 2003; Alexander,
2018), whereas hazard intensity is often a measure of the strength of an
event with respect to a given location or area and/or a moment or period.
Recently, Wang and Sebastian (2022) identified two defining dimensions,
i.e. the spatial and temporal dimensions, to categorize existing hazard
strength scales. These scales can be classified as
To quantify hazard strength in terms of equivalent hazard magnitude, we considered 12 hazard types: cold wave, convective storm, drought, earthquake, extra-tropical storm, flash flood, forest fire, heat wave, riverine flood, tornado, tropical cyclone, and tsunami. A general standardized metric of impact was created by combining three loss measures from the EM-DAT database (Guha-Sapir et al., 2021): fatality, total affected population, and total damage. The impact metric was then related to an indicator of hazard strength, such as the Richter magnitude, for each hazard type via linear regression. The expectation of impact metric for each hazard type was linearly scaled and adopted as the equivalent hazard magnitude. Here, two assumptions were made. First, we assumed that the EM-DAT records were not significantly biased across similar hazard events. Second, we assumed that the derivation of expectation of impact metric cancelled out all local factors of exposed value and vulnerability. The following sections outline the method in detail.
To reduce the biases in model calibration due to different protocols for data collection across different types of natural hazards, we only used data gathered from the EM-DAT database (Guha-Sapir et al., 2021). To be included in the EM-DAT database, a hazard event must meet at least one of three criteria, i.e. 10 or more human fatalities, 100 or more people affected by the event, or a declaration of a state of emergency or an appeal for international assistance by a country (Guha-Sapir et al., 2021). For this study, we downloaded the entire EM-DAT datasets on all types of natural hazards. However, since some records of hazard magnitude indicators of events for some hazard types (e.g. the volcanic activities and landslides) were missing, we only included 12 hazard types. The final dataset for deriving the equivalent hazard magnitudes contained a total of 3844 data points, each representing one unique hazard event.
The 12 considered hazard types include convective storm, extra-tropical
storm, tornado, tropical cyclone (wind speed is used as hazard magnitude
indicator), cold wave, heat wave (temperature), drought, flash flood, forest
fire, riverine flood (affected area), earthquake, and tsunami (Richter
magnitude). For data quality control, we removed data points with
questionable values of hazard magnitude indicators. For cold wave events, we only included data points with a minimum temperature of
To facilitate regression modelling, we logarithmically transformed values of
hazard magnitude indicators to be close to a Gaussian distribution within
the theoretical range
We designed the impact metric as the principal component (Jolliffe, 2002;
Jolliffe and Cadima, 2016) of three logarithmically transformed and
standardized impact variables. The selected impact variables represented
three major impact dimensions as defined by the EM-DAT database (Guha-Sapir
et al., 2021). The first variable, fatality, indicated the number of people
who perished as the result of a hazard event. The second variable, total
affected population, referred to the total number of individuals injured,
made homeless, or affected by the event. The third variable, total
damage, indicated the total amount of damage to property, crops, and
livestock in 2019 USD caused by the event. The values of the impact
variables were logarithmically transformed to be within the range
Means and standard deviations of original and logarithmically
transformed impact variables
To reduce the bias associated with factors of exposed value and
vulnerability (Fig. 1), we included all available data points at the
country–year level for countries around the world and hazard events from
1900 to 2020. To compute the impact metric, we only kept data points
(
For each considered hazard type, we established the relationship between its
hazard magnitude indicator and hazard impact metric via linear regression
Statistics of parameters of 12 simple linear regression models for
deriving equivalent hazard magnitudes
Visualization of the distribution of data points with respect to the impact
variables and impact metric (Fig. 2a, d, h, and m) shows that the
empirical marginal distributions of the logarithmically transformed and
standardized impact variables and the impact metric appear to be
approximately Gaussian. The standardized natural logarithms of impact
variables are positively correlated with each other (Fig. 2c, f, and g;
also see Appendix A). Results of the linear regression modelling with two
independent variables (see Appendix A) indicate that each of the
standardized natural logarithms of impact variables is positively associated
with the other two logarithmically transformed and standardized impact
variables with a positive
Impact variables and impact metric.
Figure 3 demonstrates that the proposed methodology for deriving an equivalent
hazard magnitude of an event is effective in decoupling the natural force,
manifested in hazard strength, from other factors of impacts of natural
hazard events to support studies on exposed value and vulnerability. The
results of the calibration of linear regression models for 12 individual
hazards (Fig. 3 and Table 2) show that the direction of the coefficient of
hazard magnitude indicator in each model is consistent with expectation. In
particular, the estimates of coefficients of hazard magnitude indicators for
convective storm (Fig. 3b), drought (Fig. 3c), earthquake (Fig. 3d), flash
flood (Fig. 3f), forest fire (Fig. 3g), riverine flood (Fig. 3i), tropical
cyclone (Fig. 3k), and tsunami (Fig. 3l) are all statistically significant
at
Simple linear regressions on impact metric against magnitude
indicators. Impact metric is regressed on
Using the proposed methodology, we can plot all the data points onto one figure (Fig. 4), allowing us to compare equivalent hazard magnitudes of events across different hazard types on the Gardoni Scale. Each data point on Fig. 4 corresponds to a record of hazard event and all plotted data points are associated with impacts above the threshold defined by the EM-DAT database (Guha-Sapir et al., 2021).
Impact metric versus equivalent hazard magnitude on the Gardoni Scale. The expectation line shows values of the expected impact metric with respect to equivalent hazard magnitude.
Within the datasets for this study, all 37 events with the largest
equivalent hazard magnitudes are either a tsunami or a drought. Their
equivalent hazard magnitudes range
Figure 5 compares hazard magnitudes of events of four hazard types, i.e. earthquake, tornado, forest fire, and tropical cyclone, with ranges of hazard magnitudes adjusted according to the earthquake Richter magnitude scale. The figure shows that tornadoes tend to have a smaller hazard magnitude than large earthquakes and tropical cyclones. Most of the recorded tornadoes have a hazard magnitude equivalent to an earthquake Richter magnitude between 5 and 6. Compared with tropical cyclones in terms of peak sustained wind speed on the Saffir–Simpson hurricane wind scale, these tornadoes are similar in hazard magnitude to a tropical storm but not a hurricane. This indicates that hazard strength of an entire tornado event may be much smaller than the one for a large earthquake or tropical cyclone, even though tornadoes can still cause significant damage locally as in the case of the 2013 El Reno tornado. Meanwhile, the wide spread of data points of tornadoes with respect to hazard magnitude on Fig. 5a suggests that exposed value and vulnerability of exposed entities may be much stronger predictors of hazard impacts than hazard magnitude for tornado events.
Comparisons of hazard magnitudes of four hazard types.
Compared to earthquakes, tropical cyclones that reach a hurricane level on the Saffir–Simpson scale are equivalent in hazard magnitude to an earthquake with a Richter magnitude greater than 6.5. A magnitude 8 earthquake on the Richter scale has a similar size in hazard magnitude as a tropical cyclone labelled with a peak category 5 on the Saffir–Simpson scale. Within the datasets for this study, Typhoon Meranti is the tropical cyclone with the largest equivalent hazard magnitude at 5.66. Although the typhoon was strong and affected the Philippines, Taiwan, mainland China, and South Korea in September 2016, it only resulted in a total economic loss of around 2019 USD 70 million, according to the EM-DAT database (Guha-Sapir et al., 2021).
In addition to earthquake and tropical cyclone, forest fire is another hazard type with a statistically significant estimate of coefficient of hazard magnitude indicator (Table 2). However, forest fires tend to have smaller equivalent magnitudes than large earthquakes and tropical cyclones (Fig. 4b). The two largest forest fires within the datasets had an equivalent hazard magnitude of 4.33. They occurred in Russia and Mongolia in 1996, resulting in 19 and 25 fatalities, respectively (Guha-Sapir et al., 2021). Both forest fires were equivalent to a tropical cyclone with its peak sustained wind speed reaching category 1 on the Saffir–Simpson scale. They were also equivalent in hazard magnitude to an earthquake with a Richter magnitude between 6.5 and 7.
With Fig. 6, we can compare the hazard magnitudes of cold wave and heat wave
events. Both hazard types have a narrow range of equivalent hazard magnitude
of events, with
Cold wave minimum temperature versus heat wave maximum temperature.
Comparison of hazard magnitudes can also be conducted between riverine flood
and drought events (Fig. 7). Among hazard events included in the datasets
for this study, drought has a large range of equivalent hazard magnitude of
Riverine flood area versus drought area.
In this study, the impact metric was constructed as the principal component of three transformed impact variables. The sum of squares of weights of transformed impact variables within the impact metric equalled one. We conducted a visual sensitivity analysis to examine if altering the weights of transformed impact variables within the impact metric had any significant effect on the relative comparison of hazard magnitudes across hazard types. For this sensitivity analysis, we first kept the sum of squares of all weights of transformed impact variables equal to one. Second, we maintained an equal ratio of squares of weights between two transformed impact variables. Third, we changed the weight of the third transformed impact variable and adjusted the weights of the other two transformed impact variables according to the first two rules.
Figure 8 shows the result of a sensitivity analysis with data points of tsunami and flash flood as a demonstrative example. Data points are plotted based on their equivalent hazard magnitudes with a fixed scale of the hazard magnitude indicator of tsunami. When the weight of each of the transformed impact variables of fatality (Fig. 8a–d), total affected population (Fig. 8e–h), and total economic damage (Fig. 8i–l) is shifted from zero to one, there are identifiable increasing or decreasing trends of alterations of the distributions of data points as well as the deviations between clusters of data points of the two different hazard types. However, when weights of transformed impact variables are far away from the extreme value of zero or one, there is no significant change regarding the distribution of data points with respect to equivalent hazard magnitude (see Fig. 6b, c, f, g, j, and k). This result indicates desirable performance of the proposed methodology for deriving equivalent hazard magnitude of an event on the Gardoni Scale.
Results of visual sensitivity analysis regarding effects of
altering weight of one transformed impact variable within impact metric on
equivalent magnitudes of tsunami and flash flood events. Weight of fatality
equals zero,
To our knowledge, this study represents the first attempt to produce an
equivalent hazard magnitude scale, i.e. the Gardoni Scale, to quantify
agential–durational hazard strengths for hazard events across multiple
hazard types. The proposed scale has several merits. First, professionals in
natural hazard and emergency management could use equivalent hazard
magnitudes on the Gardoni Scale to facilitate hazard communication among
various stakeholders. Similarly, journalists and news media could adopt the
Gardoni Scale for news reporting on natural disasters to the public. When
events of different hazard types are described as equivalent to each other
in terms of their natural forces, we can use the proposed methodology to
compute the equivalent hazard magnitudes of these events on the Gardoni
Scale to confirm such equivalency. For example, if we adopt the minimum
temperature of
Beside its utility for emergency management, computation of equivalent hazard strengths of events can enhance hazard profiling and risk analysis within a multi-hazard context. When hazard strengths can be evaluated comparatively across hazard types, we can model hazard frequency and exposure regarding multiple types of hazards simultaneously and create multi-hazard hazard maps. With quantified hazard equivalency, it will also be possible to derive loss ratio curves with respect to a uniform equivalent hazard strength measure to indicate the differences in vulnerability and resilience of individuals, communities, and infrastructures facing hazards across different hazard types. Such multi-hazard quantification of hazard, exposure, vulnerability, and resilience can be integrated to facilitate risk analysis to predict future losses and loss ratios without additional efforts to develop sophisticated models for each individual hazard type. Thus, management of perceived and engineered risks due to natural hazard events would be facilitated by the proposed hazard equivalency methodology. To achieve such multi-hazard quantifications of risks of natural hazard events, more research is needed not only to improve the proposed Gardoni Scale for equivalent agential–durational hazard strengths, but also to explore the modelling of equivalency of other types of hazard strengths, particularly the locational hazard strengths, for hazard management at the local level.
As shown in the previous section, data points in this study can be visualized as centred along the expectation line, albeit with a large variation (Fig. 4). This implies that the derived equivalent hazard magnitudes may correspond well to the expectation of hazard impacts but without precision. Such a lack of precision is not a limitation. On the contrary, it suggests that impacts of hazard events are not only the result of hazard strength but also correlated with environmental, societal, and infrastructural factors that affect the exposed value and vulnerability of exposed entities within a natural hazard context (Fig. 1). Because of the effect of these factors other than hazard strength, however, the mere inclusion of, or the complete exclusion of, data points with a unique bias toward one direction of these factors will result in biased derivation of equivalent hazard strength metric. To reduce such a bias, in this study we included all available data points of hazard events worldwide and from a long period of 1900–2020. However, there may still be bias due to spatial or temporal concentrations of data points regarding certain hazard types, for example, events that have large hazard magnitudes but small impacts (due to, e.g. no exposed entities or low vulnerability, or under reporting, see, e.g. Paprotny et al., 2018). Future work should examine how to further reduce this potential bias caused by factors of exposed value and vulnerability of exposed entities.
To demonstrate the implementation of the proposed methodology for deriving equivalent hazard magnitudes of events, we only considered one hazard magnitude indicator for each hazard type. For many hazard types, one indicator cannot represent the true hazard magnitude of an event which may arise due to multiple forcings. For example, both wind and precipitation contribute significantly to damages associated with tropical cyclone events (Mudd et al., 2017). Selection of hazard magnitude indicators in this study was also limited by the adopted datasets. As an example, the earthquake Richter magnitude (Richter, 1935) was the only recorded hazard magnitude indicator in the datasets of this study. However, the EM-DAT database reported generically as “Richter magnitude” estimates for earthquake events, even though such estimates may include moment magnitude as well. In addition, regarding tsunami, the mere inclusion of earthquake magnitude of a tsunami-triggering earthquake as the magnitude indicator ignores the fact that tsunamis can also be caused by non-seismic events, such as volcanic island collapses and large coastal landslides. For flood hazards, as another example, there is a lack of established methods to quantify the agential–durational hazard strength metrics. In this study, we used the flooded area as the hazard magnitude indicator for the flood hazards in accordance with the procedure used to create the EM-DAT database (Guha-Sapir et al., 2021). However, the definition of such flooded area is still vague and deserves more research. An ideal agential–durational hazard strength metric for a flood event should integrate multiple flood intensity measures, such as water depth, flood volume, and flow velocity, over the entire flooded area and duration of the event to correspond to the total energy released by the natural force of the event. More effort, therefore, is needed to study, select, and quantify the appropriate hazard magnitude indicators for deriving equivalent hazard magnitudes of events on the Gardoni Scale.
In addition to hazard magnitude indicators, the construction of the impact metric is important for the calibration of regression models and for the derivation of equivalent hazard magnitudes as it is end-user specific. For example, insurance professionals may be interested in an equivalent hazard magnitude that is derived from data on financial and property loss, whereas environmental scientists may be more interested in an impact metric based on ecological damage. Herein, we derived a general metric of impact for equivalent hazard magnitude based on key indicators of societal impact: fatalities, damages, and number of affected individuals. However, hazard events can affect a variety of sectors resulting in impacts to physical, social, economic, and environmental well-being (Lindell and Prater, 2003; Gardoni and Murphy, 2010; Alexander, 2013; Wang et al., 2016, 2021). To advance methodological development for the proposed Gardoni Scale and quantification of other equivalent hazard strength metrics for various stakeholders, future work should scrutinize different indicators as impact variables of events and seek the optimal models to combine impact variables to inform the level of impacts of events for different hazard types.
To support modelling with consideration of hazard magnitude indicators and the impact metric, more statistical, machine learning, and other quantitative models should be pursued to establish the mapping between an equivalent hazard magnitude and the expectation of impacts of hazard events. When data on hazard events with little or zero impacts become available for modelling, we may also apply zero-inflated techniques or other methods to consider the effect of data points with zero impacts to improve the derivation of equivalent hazard magnitudes of events within a multi-hazard context.
Beside these above-mentioned issues, the inclusion and exclusion of certain data points based on values of variables may also affect the results of derivation of equivalency of hazard magnitude. First, in this study, a set of thresholds were adopted to filter out records of events with extremely small and large measures of magnitude indicators. However, some events with magnitude indicator measures barely inside the thresholds, such as the magnitude 3 earthquake in southern Russia in 1999, were still included in the data for modelling. On the other hand, because the EM-DAT database only included events with loss records beyond a set of criteria, numerous events with lesser impacts were not recorded for model calibration in the study. Such exclusion of events with lesser impacts caused the empirical marginal distributions of the logarithmically transformed and standardized impact variables and the impact metric to appear to be approximately Gaussian. Future work should explore to what extent the computation of equivalent hazard magnitude is sensitive to the inclusion and exclusion of data points of events of an either small or large size in terms of both the magnitude indicators and adverse impacts.
In this article, we proposed an equivalent hazard magnitude scale, called the Gardoni Scale, to measure the strength of natural force involved in the entirety of a natural hazard event for comparative analysis across different hazard types. A computational methodology based on PCA and regression modelling was introduced and implemented to demonstrate the methodological utility in derivation of the equivalent hazard magnitudes of events for 12 natural hazard types. The proposed equivalent hazard magnitudes of events on the Gardoni Scale are recommended to be adopted for hazard communication by various stakeholders including news media, decision makers, industry professionals, academic personnel, and the public. By applying the proposed Gardoni Scale, we can also help quantitatively decouple the natural forces of hazard events from the environmental, societal, and infrastructural factors of hazard impacts to support social scientific and engineering research in natural hazard phenomena with a multi-hazard approach. We anticipate that this study on equivalent hazard magnitude will be extended to comparative modelling of other types of hazard strengths of events in a multi-hazard manner to consolidate the foundations for quantifying and studying exposure, vulnerability, recovery, resilience, and other conditions for disaster risk reduction due to natural hazards at both local and global levels.
Six simple linear regression models and three multiple linear regression
models with two independent variables were calibrated with the same data
points for derivation of the impact metric. These regression models were
created to fill in missing values of impact variables for data points with
at most two empty entries among the three impact variables. Within each of
these nine linear regression models, the dependent variable was one of the
three impact variables. For each of the six simple linear regression models,
the independent variable was one of the two impact variables that were not
used as the dependent variable. The simple linear regression models had the
form
Statistics of parameters of six simple linear regression models
for filling in missing values of impact variables
Statistics of parameters of three multiple linear regression
models with two independent variables for filling in missing values of
impact variables
Python codes and data that support this study are available at
Video S1 shows the distribution of data points with respect to impact variables and the impact metric. The supplement related to this article is available online at:
YVW was responsible for design of the study, data collection, data processing, and coding. Data analysis, drafting, and critical review of the paper were undertaken by both authors.
The contact author has declared that neither of the authors has any competing interests.
Publisher’s note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Yi Victor Wang would like to thank Paolo Gardoni and Colleen Murphy for inspiring discussions and suggestions.
The article processing charges are covered by the Institute for Earth, Computing, Human and Observing (ECHO) at Chapman University.
This paper was edited by Mario Parise and reviewed by John K. Hillier and three anonymous referees.