Our analyses show that crisis severity can be measured best through use of only 11 of the 35 GCSI indicators. Put another way, 24 of the 35 indicators used in the GCSI model do not contribute useful numerical information. In our final model, the strongest predictors of severity were a suite of indicators related to social structure/governance of a given nation state (rule of law, freedom, gender inequality, and empowerment), followed by indicators that were proxy measurements of humanitarian access/safety (number of people killed, restricted movement, and obstructed access to assistance). Weaker, although still relevant, predictors were related to the crisis impact on people and the environment. Overall, this analysis suggests that most of the key variables to estimate the severity of humanitarian need can be assessed globally, can be collected in a comparable way from one country to another, and do not depend on sudden changes to local conditions that would be unavailable to those making the calculations. In short, despite changing conditions and limited or imperfect information, we can use existing data sources to make reasonable estimates of severity around the world. Refinements to the existing GCSI model will make it easier and more reliable to make these estimates.

Holistically, the 11 selected indicators suggest that fragile states with limited accessibility for humanitarian actors have worse humanitarian conditions. This final model aligns with humanitarian actors’ experiences. Indeed, good governance is intrinsically related to avoiding or mitigating a humanitarian crisis [12]. We tested the role of governance in our models (see Additional file 1: Appendix 8) and found it to be a key latent constructure of severity, but only when crisis related variables were also included in the model. Broadly, economic and political stability are key components to this success, with inequality between social groups cited as a driver of crises and conflicts [25]. It is unsurprising that humanitarian practitioners call for more robust inclusion of conflict early warning into preparedness systems for humanitarian crises [26]. Indeed, considerable funding has been provided to post-conflict states for democracy development and peacebuilding, albeit with mixed success [27, 28]. Ample evidence supports these patterns, as data from the last 15 years show most humanitarian crises are re-occurring in the same countries, many of which are fragile states [13]. Chad, the Central African Republic, the Democratic Republic of Congo, Somalia, and Sudan have all had 15 crises between 2005 and 2015.

Beyond governance, access to reach those in need is also important to reducing crisis severity. Humanitarian access, the ability to reach the most vulnerable, can be limited through various mechanisms. Restricted movements, which are common in conflicts and complex humanitarian crises, inhibit connections between aid workers and communities [29]. Access can also be reduced through violence and obstruction. Within armed conflicts, bureaucratic and security constraints, and violence against aid workers and facilities distributing aid, have been cited as rationale for greatly reduced humanitarian access [29, 30]. For example, in the Syrian crisis, which is considered one of the worst in the world by humanitarian experts, UNOCHA reported that 1.1 million people were in need of humanitarian assistance in hard-to-reach-places in 2018; during this same year, access was inhibited by 142 attacks on health facilities, with 102 people dead and 189 injured [31]. Thus, it is not surprising that our model results give weight to indicators reflecting the quality of humanitarian access (e.g., restricted movements, obstructed access, and number of people killed) for a given crisis.

Importantly, our final model differs from the original GCSI in two fundamental ways. First, we presented a parsimonious model, which removed 24 GCSI indicators. Using the reduced set of 11 indicators, the model showed acceptable fit, but had slightly higher error than the standard cut points; however, some debate exists on the usefulness of applying a single heuristic to assess model fit within factor analysis [32]. The original GCSI was calculated using inconsistent approaches, and notably, does not account for basic statistical properties of correlated data. The high correlation in the dataset inhibits meaningful interpretation of combined values from the indicators. Second, we removed an entire GCSI pillar (‘conditions of the people’) as a result of insights from the EFA models, which has programmatic significance. Indeed, the data underlying the excluded indicators are routinely collected to estimate the number of people in need of humanitarian assistance. Given the strong value of these indicators to practitioners, we re-ran the final model and included these two indicators as standalone independent variables (Additional file 1: Appendix 5). Of note, we did not include the two indicators as latent constructs, as our EFA analyses showed that they were not correlated. This sensitivity analysis suggested that a model including the number of people affected indicator has comparable model fit and yields similar severity scores to the second-order CFA model.

Our analysis, however, is limited by the data available for inclusion. First, the index includes a combination of static and dynamic variables. It is possible that static variables, such as those used to estimate social structure/governance are distal determinants of a crisis, rather than proximal measures. Additionally, we used population average data, which masks any disparities experienced within a population. Several population groups, namely, children, women, and the disabled, have worse crisis-related health outcomes than the rest of the population. Moreover, data from humanitarian crises are difficult to obtain, highly inaccurate, and highly correlated. While our sensitivity analyses assessing data quality suggest that our final model contained data that was no more or less reliable than the indicators excluded (Additional file 1: Appendix 6), we cannot account for the lack of precision within the dataset. We included two indicators based on expert assessment of qualitative information (restricted movement, and obstructed access to assistance), which may be subject to imprecision or bias. Likewise, mortality estimates, which we also included in the final model, have been contested for accuracy in past crises [33, 34]. Additionally, the indicator for ‘relative people living in the affected area’ is highly correlated with many of the other variables in the final model. In an ideal scenario, this indicator would be removed from the model, however, when it was, the models did not converge. Thus, one limitation of retaining the variable is a slightly higher error than desired. Finally, we are limited in our ability to test the generalizability of the model given the small sample size and lack of additional data for testing. Nevertheless, our comparison of the model fit statistics and factor loadings to suggest that the model performance is consistent and unlikely overfit to the data (Additional file 1: Appendix 10).

Importantly, a gold standard for crisis severity is unavailable to validate our model results and out-of-sample data were not available to assess predictions. In lieu of traditional validation, we compared the latent severity scores to the original GCSI scores (Additional file 1: Appendix 7). This robustness check suggested that the latent severity score may be a closer measure to true crisis severity than the original GCSI. Despite the limitations with data availability and independent data source for validation, we emphasize that this work is a first step towards improving crisis severity measurement. Because the calculations are derived from a model that weights indicators based on their correlations, estimating severity for a new crisis would require re-running the final CFA. Further research is needed to assess the feasibility of linking this framework with a field friendly application for humanitarian actors after additional analyses have been conducted.