We use a harmonised database of disasters for 89 countries, DesInventar (Disaster Inventory System), to investigate the relation between landslide building damage and deaths . DesInventar includes data from the mid-20th century to the present, with degree of completeness and metadata varying widely between countries. Nevertheless, it remains one of the most spatially and temporally comprehensive global-scale disaster databases . Commonly used indicators for damage are recorded, including the number of buildings destroyed and damaged, the number of lives lost, and the number declared missing for each disaster . There is substantial work on the differing degrees of damage that landslides can do to buildings , but these datasets do not allow for this level of detailed analysis, and we simplify here to a binary damaged or not damaged classification. For a nationwide case study for Nepal, we supplement the DesInventar event catalogue (which ends in 2013) with comparable data from the Nepal National Disaster Risk Reduction and Management Authority (NDRRMA, from 2014) available through the Building Information Platform Against Disaster (BIPAD) portal (NDRRMA, from 2020).

We begin with a systematic search of all keywords in the DesInventar database to identify words for “landslide” in different languages (e.g., “deslizamiento de tierra”), spelling errors (e.g., “landside”), or different terms for a similar physical process (e.g., “rock slide”). A full list of the keywords used is available in the Supplement. We find a total of 47 213 individual landslides, of which we exclude 25 030 which resulted in no deaths, missing people, or building damage, and we include 22 183 for further analysis which record losses in at least one of those categories (deaths, missing people, and/or building damage). We regress the total human loss from each landslide (sum of deaths and number of missing persons) against the total building damage (sum of buildings destroyed and damaged) for each country with sufficient data (> 10 landslides over the entire record), yielding estimates for a total of 44 countries. For Nepal, we run this workflow on both the DesInventar data alone and on a merged database comprising both DesInventar and BIPAD data.

We use different metrics to evaluate the power of total building damage (predictive variable) towards total human loss (dependent variable): the coefficient of determination (R2), the coefficient of estimation (CE), reduction of error (RE), and root mean squared error of prediction RMSEP;. The CE, RE, RMSEP, and relative RMSEP specifically test whether or not building damage can be used as a meaningful predictor for human damage. We adopt the formulae of to calculate the CE and RE, using bootstrapping to separate the data into validation and calibration datasets, randomly sampled (with repetition, both to the same size as the original dataset) 100 times to calculate the mean and standard deviation for the CE and RE metrics. We calculate the RMSEP and relative RMSEP from the same bootstrapped datasets. To contextualise the landslide results, we repeat the above analysis for several other disasters in the DesInventar database: floods, volcanic eruptions, earthquakes, and storms.

The correlation between human loss and number of affected buildings from landslides is good in national averages but negligible when disaggregated to individual events. Both human loss and total building damage follow long-tailed distributions, with most damage being accounted for by a small number of events. However, in the case of landslides, we find a poor correspondence between the events causing high levels of human loss and high levels of building damage. A comparison with other disasters shows higher correlation between these extremes for other disaster types, such as earthquakes and tsunamis. Averaging the impacts of multiple events, either as temporal averages (total damage per year) or geographic averages (total damage per country), can mask this lack of correlation for individual events by cumulating both high-death low-building damage events and low-death high-building damage events.

All databases provide an imperfect and biased record of the impacts of disasters. Disasters will only be recorded if they are large or damaging enough to be noteworthy, and the size of inventories varies drastically between the 89 countries included in the DesInventar database . Additionally, the number of deaths, missing people, and buildings damaged may be either overestimated or underestimated for events that are recorded, and the accuracy of these estimates will vary spatially and temporally . Our study design partly mitigates these limitations, as we investigate the damage per disaster instead of the total number of events (or total damage). Even though the datasets are incomplete, we can still make inferences about the relation between total human loss and total building damage from the events that were recorded. A comparison between landslides and other disasters provides a further test, as the relation between building damage and human loss varies for different disasters. This relation is expected to be strong for earthquakes, where a large proportion of human loss is directly caused by building collapse , and particularly weak for lightning, where buildings may shield inhabitants from damage. Both of these expectations are supported by the DesInventar database (Fig. S1 in the Supplement).

Our finding that damage to properties and human loss are poorly related is consistent with the complex and geographically dispersed nature of landslides and our current understanding of their links to human mortality. Different impacts may indeed be expected depending on landslide type; for instance, large landslides with clear warning signs may damage many buildings without loss of life, while small rockfalls may cause many fatalities without damaging any buildings. Similarly, landslide early warning systems may enable effective evacuations in some parts of the world, preventing fatalities but not building damage. The exact causes of this discrepancy, including different landslide processes, effective mitigation strategies, and spatially concentrated exposure and vulnerability, are likely to vary widely across the world and within this dataset. showed the importance of demographic, situational, and particularly behavioural factors in determining landslide morbidity, arguing that the relationship between building damage and morbidity is therefore complex. A low correlation has previously been noted between landslide-related deaths and the economic cost of landslides , although this study was limited to Switzerland alone. Creating an accurate risk map relies on a combination of two components: a map of the spatial distribution of the hazard and a measure of the exposure to this hazard. This exposure layer will depend on the objective of the risk map; for instance, insurance disaster risk maps will often be based on infrastructure value maps. In the case of landslides, our results show that building maps or databases – for example, MSBuildings or OpenStreetMap – are an inadequate proxy for the total human loss from landslides and should not be relied upon solely to estimate risk to human lives from landsliding.

Our results show that landslide disaster mitigation strategies using risk maps constructed from infrastructure assets or building datasets may implicitly prioritise monitoring or mitigation of high-building damage events instead of high total human loss events. This raises ethical and practical issues and is generally at odds with the primary objective of disaster risk reduction programmes. Two different and complementary approaches stand out to improve our understanding and representation of human loss from landsliding. The first involves building a detailed understanding of local conditions through consultations and interviews, and the second involves large-scale (and ideally dynamic) population or exposure modelling. Both methods fall outside the traditional remit of landslide science, and they highlight the need for transdisciplinary collaboration for effective landslide risk reduction.

Improving our understanding of disaster risk involves examining local risk perceptions, daily routine variations, and mitigation strategies. Local risk perceptions and actions may explain the low correlation between human casualties and building damage, either due to effective evacuation strategies reducing exposure or inadvertent actions increasing exposure . Additionally, considering local perspectives may reveal effective risk-reduction strategies and increase community involvement in mitigation efforts. The factors contributing to the low correlation between human casualties and the number of affected buildings are likely to vary across the 89 represented countries in the DesInventar database. Conducting informal and semi-structured interviews with local residents could help shed light on why building damage might exceed fatalities in certain landslides and vice versa within specific regions.

On a larger scale, population density and dynamic exposure maps offer an alternative perspective, albeit with some limitations. These population density maps often have low spatial resolution or rely on building data for interpolation . Additionally, static population density maps fail to capture substantial transient changes in population density occurring on a daily, seasonal, or interannual basis. In scenarios such as landslide risk assessment, where evacuation may be impractical and vulnerability is high , dynamic exposure becomes a critical element of risk-to-life modelling. Innovative modelling approaches, such as agent-based models e.g., have the potential to account for spatio-temporal population movements across different timescales. However, these methods are relatively untested in the context of disaster risk reduction, presenting an open science challenge in need of further development.

This study shows a complex relationship between building damage and human loss from landslides. We find that, despite moderate correlation for national averages, the two are uncorrelated in individual events at both a global scale and in Nepal specifically. Therefore, building damage from landslides is not an effective predictor of the number of fatalities from the same event. Comparative analysis with other disasters highlights the contrasts between them, with a stronger link between building damage and human loss present for earthquakes and tsunamis but not for other geohazards such as floods, avalanches, or lightning strikes. There is a need to develop exposure layers beyond simple building databases, encompassing localised insights into risk factors and dynamic population models, to improve mitigation of the deadliest landslides.