Barriers and Possibilities for Interdisciplinary Disaster Science Research: Critical Appraisal of the Literature

As the predictive power of modeling grows, the interest in developing comprehensive and representative models grows as well. The completeness of models now largely relies on the degree to which interdisciplinarity can be achieved from project conception through completion. The primary drivers of interdisciplinary research include the investigation of overlap and interaction between disciplines and the incorporation of emerging technology to fully understand the complexity of social and natural systems, especially as they impact overarching societal problems (National Academy of Sciences et al. 2005). Within the context of these drivers, fostering community resilience in disaster research necessitates an interdisciplinary approach due to several characteristic features: “many hazards and built and social systems are involved; many types of interacting impacts are involved; many interacting strategies exist to improve resilience; the problem is uncertain, spatial, and dynamic; many diverse stakeholders are involved; and effective communication about risk and resilience is difficult” (Davidson 2015). Furthermore, Peek et al. (2020) suggest that interdisciplinarity is a requisite element of convergence research; however, convergence research extends beyond interdisciplinary research in its explicit pursuit of solutions to identified problems. Indeed, disaster research can advance the implementation of convergence research, and by extension interdisciplinary practice, because it is so readily positioned at the intersection of multiple disciplines and has a self-evident problem to address (Peek et al. 2020). Thus, disaster research can serve as a platform for new integration strategies that can be applied to, or at least inform, many other research areas.

However, this path is not without its challenges. Previous efforts at interdisciplinary research from a variety of fields have noted several barriers to effective integration, most of which are universally felt across many disciplinary overlaps. In this paper these barriers will be discussed and the frequently encouraged solutions to alleviating these barriers will be highlighted. While many of these suggestions seem to adequately address the challenges at hand, data integration remains an area for which there are several avenues forward, but none have been sufficiently applied to the case of disaster research yet. Thus, next steps for disaster research data integration merit further discussion.

In working on interdisciplinary projects, past research has noted several barriers to complete integration. These barriers include differences in language. Bracken and Oughton (2006) categorized distinct facets of disciplinary language. The first challenge is that of dialect, or the variations in meaning of a word across disciplines as well as between specialized and everyday use. Secondly, disciplines often use contextually reduced or explanatory metaphors that are unique to that field and that perpetuate discipline-specific frames of thought. Solutions to these dialectical differences across disciplines are investigated by ontologists within the field of data science as well, though they typically use “semantics” in place of “dialects.” Within this subfield, ontologists are interested in linking data from various sources. To do this they must first determine the meaning of the words in a given context to determine where overlaps exist and to understand the network of ideas in order to avoid the creation of inconsistencies when the information is integrated. The implications and potential of semantic technologies will be discussed further in the discussion on data integration due to the inherent link between data and the meaning assigned to it (Shvaiko and Euzenat 2013). Also, differences in language pose a clerical issue. Research is difficult to find when different terminology is used across disciplines yet for the same topic (Sherman-Morris et al. 2018). These differences in terminology allude to the epistemological divides that have proven difficult to overcome.

Another key barrier is that of institutionally reinforced disciplinary boundaries. Because institutions are developed departmentally, there is often not the institutional social infrastructure, physical layout of facilities, or reward systems conducive to interdisciplinary research (Davidson 2015). If there are not established avenues for interaction and agreed upon methodologies, then the time-sensitive work of post-disaster research may not have the necessary structure for interdisciplinarity (Ganapti and Mostafavi 2018). Aside from the threats that not integrating disciplinary disaster research poses to the effectiveness of research in the field, it also draws into question their academic contribution. When this work is not institutionally supported and accepted, it can lead to duplicated and disjointed publications regarding the same topic but from other disciplinary lenses (Sherman-Morris et al. 2018). Current organizational structures preserve disciplinary boundaries even when to do so hampers the progress of representative, meaningful research. This leads to researchers from different fields struggling to agree on shared epistemologies for interdisciplinary research, resulting in a time-intensive integration process that is repeated every time a new interdisciplinary project occurs. This costly time commitment can dissuade some researchers from embarking on this type of research.

Another commonly reflected upon issue within interdisciplinary research is that of disciplinary superiority. According to Green and Anderson (2019), disciplinary imperialism describes the situation in which disciplines attempt to collaborate, but one discipline, knowingly or not, will work to impose its methodology on the work of the other disciplines. The threat of this phenomenon becomes more pronounced and is further exacerbated by one of the most fundamental issues that can arise in interdisciplinary research. That is the issue of distrust. When individuals do not have the social infrastructure mentioned previously to establish connections outside of the research, they do not afford themselves the opportunity to develop trust with the other researchers. When the researchers do not have complete knowledge of the other fields with which they are working, they will feel compelled to translate the information into their own field, especially if that researcher has not first established professional and interpersonal trust with their disciplinarily disparate colleagues.

One of the primary and consistent tools suggested in bridging disciplinary gaps is sustained and intensive communication with colleagues from various disciplines. Several methods for communication have been recommended. Bracken and Oughton (2006) suggest the use of articulation to bridge disciplinary language barriers. Articulation involves creating a coherent message across disciplines with different dialects or metaphors, thus creating commonality to bridge differences in disciplinary language. Similarly, metaphors can be repurposed as a helpful bridging tool between disciplines, instead of a barrier to integration as mentioned previously, when appropriate context is applied. Additionally, the practice of simply having agreed upon definitions in collaborative research is an imperative tool for convenient communication (Newgreen et al. 2019).

Boundary objects are another commonly recommended method for facilitated conversation. Boundary objects are any points at which disciplines overlap and cross-cutting conversations are easily fostered. One such boundary object is scenario development, which has been commonly implemented in a variety of fields including business, government, and sustainability planning. Scenario development, or scenario planning, centers on the idea of considering “a variety of possible futures that include many of the important uncertainties in the system rather than to focus on the accurate prediction of a single outcome” (Peterson et al. 2003). Hence this form of anticipatory planning is also referred to as futures thinking due to its usefulness in determining probable future scenarios, or simply futures. Because “scenarios are built from various assumptions, theories, and methods for studying the world” (Kröger and Schäfer 2016), individuals can be brought together to develop and defend scenarios with individuals of disparate perspectives in order to find commonalities as well as the source of differences. This work found increased cognitive and social integration as researchers from various disciplines began to create common perspectives and derive agreed upon results (Kröger and Schäfer 2016). Within another interesting scenario development study, five contrasting scenarios were used to predict the future of Brazil’s agriculture based on a series of interdisciplinary variables, a method which will be discussed further in the data integration discussion (Gomes et al. 2020). This method of creating a few characteristic scenario outcomes from multiple natural and societal factors may be applicable in the field of disaster research as well.

A very similar boundary condition is that of disaster simulation, which is commonly used in disaster research today due to its inherent ability to create increasingly realistic results without having to wait for another disaster to occur to collect data. These simulations have been used for evacuation planning (Wang et al. 2016), nurse development training (Zapko et al. 2015), and a host of other applications to anticipate shortcomings of present systems in disaster scenarios and suggest modifications. It is often used in tandem with modeling software to increase predictive strength. Simulation has proven to be a useful tool for determining best practices, necessary policy changes, and physical infrastructure changes. Other researched boundary conditions include the sharing of experiential stories (Moezzi and Peek 2019) and the alignment of desired outcomes (Lynam et al. 2019). These methods of social and academic integration can also provide the foundation for trust building, an imperative element of interdisciplinary work (Bracken and Oughton 2006).

When considering communication between peers from different disciplines, it is also important for researchers to have some familiarity with the other disciplines with which they hope to collaborate. An interdisciplinary researcher requires two types of expertise: contributory expertise in their field of study and interactional expertise in fields of study with which they frequently collaborate. The first references the researcher’s ability to create original work, and the second references their ability to understand work and communicate with experts from another field. As well as interactional expertise, it is advantageous to build a tacit knowledge in fields with which one plans to work consistently (Collins and Evans 2008; Gilligan 2019). Tacit knowledge is largely experiential, unwritten knowledge attained from working and interacting with experts from a given field. Thus, the conversations previously described provide an ideal space to develop and test this tacit knowledge and interactional expertise. Alternatively, some of this expertise can be built through interdisciplinary courses with instructors who team-teach the content for their individual areas of expertise (Gilligan 2019; Lattuca 2001). Additional long-term educational solutions for bridging disciplines are addressed in more detail in the following section.

Providing more institutional and structural support is another means of increasing and improving interdisciplinary work. This institutional support reduces the time required to begin research and data collection efforts, as is especially important in the field of disaster research. When colleagues from various fields already have established means of communication and relationships, they are much more likely to rapidly establish the shared perspective necessary for disaster response research, unlike newly formed interdisciplinary teams (Ge et al. 2019). Similarly, Faber et al. (2014) suggest that although communication is key in interdisciplinary work, it would be more effective to implement tracks of study that formally establish the frameworks necessary for successful transdisciplinary work. This strategy would eliminate the need to cultivate interdisciplinary connections anew for every project, in turn reducing the time investment required for a given project. Increasingly, collaborations that started as a series of individual projects are becoming established fields, thus remedying some of the issues previously mentioned as well as offering a level of longevity for the new field not afforded to all interdisciplinary efforts. Such is the case for the budding field of mathematical biology (Newgreen et al. 2019).

Further validating the barrier unnecessarily created by traditional institutional frameworks, a survey within the field of sustainability found that institutional responsibilities and traditional departmental frameworks sometimes prevent commitment to more interdisciplinary work. Meanwhile interdisciplinary research centers cultivated the collaborative projects that traditional institutions did not have the flexibility or interest in doing. When we provide institutional frameworks that foster interdisciplinary work, it follows that more such work will be pursued (Nastar et al. 2018). It is important to acknowledge several key continued initiatives for interdisciplinary research. Along with some other organizations, the US National Science Foundation sees the potential in these interdisciplinary efforts and has established dedicated funding and resources for researchers working in areas that exceed the bounds of a single discipline. They have included “Growing Convergent Research at NSF” as one of the 10 Big Ideas for Future NSF Investments, a portfolio of concepts “that will drive NSF’s long-term research agenda” (National Science Foundation 2016). Various universities, health organizations, and others also offer funding opportunities specifically for interdisciplinary projects.

Nevertheless, these initiatives for supporting interdisciplinary efforts do not serve to fully alleviate the increased difficulty in procuring funding for interdisciplinary projects. In fact, the rate of success in receiving funding and the degree of interdisciplinarity has been found to be inversely proportional (Bromham et al. 2016). This could be due in part to the difficulties funding organizations encounter when analyzing interdisciplinary projects. As noted in Facilitating Interdisciplinary Research, “Effective review of interdisciplinary research (IDR) proposals may not be possible with traditional peer review that relies primarily on experts in a single discipline” (National Academy of Sciences et al. 2005). For this same reason, interdisciplinary researchers may encounter barriers to publishing in journals with strong disciplinary scopes, which in turn could reduce their chances of receiving tenure (Bruzzese et al. 2020). Additionally, the traditional metrics of publication and funding success may be detrimental to the establishment of new interdisciplinary researchers if they have faced disciplinary gatekeeping when their interest areas exist in places of disciplinary overlap. Thus, it is important to simultaneously acknowledge the efforts that have been made thus far to bolster interdisciplinary efforts, while still understanding that interdisciplinary research is not yet enjoying the academic prestige of disciplinary work. With institutional support and restructuring, interdisciplinary research will gain a robustness, longevity, and academic trust that it cannot otherwise attain.

Research has documented several other best practices for alleviating interdisciplinary stresses. Metacognition is presented as a means of achieving more fruitful interdisciplinary work. By implementing practices that support “thinking about our thinking” throughout the planning and implementation of a project, many of the cognitive disciplinary barriers mentioned previously can be reduced if not eliminated. This use of metacognition also allows for more efficacy and creativity, improved team functioning, and increased adherence to project objectives (Ganapti and Mostafavi 2018).

Another frequent recommendation for dissolving disciplinary silos is epistemological pluralism. This concept is founded on the idea that for most research tasks there are several means of viewing the topic of interest. Different disciplinary backgrounds offer these distinct and complimentary viewpoints. Furthermore, this pluralism suggests a need to collectively determine the objectives of an interdisciplinary project as well as the means that will be used to achieve these objectives. In turn this allows for a restructuring of academic understanding not limited to disciplinary boundaries (Miller et al. 2008). This method is largely universalizable, though by the same token has very little specificity in terms of how to apply such a pluralism aside from its adherence to iterative discussion and evaluation until a collective understanding is reached. Not surprisingly, recognition of epistemological pluralities allows for the development of more robust theories in disaster research (Sherman-Morris et al. 2018).

A final method encountered for combating interdisciplinary barriers is the iterative solution design articulated by Subedi et al. (2018), which involves epistemological pluralism but also includes the stakeholders in the iterative process. In adding stakeholders, the solution design becomes community-specific, yet the design process remains universalizable. The looped phases of this design consist of disciplinary data collection, the formation of a shared interdisciplinary interpretation of the data, and finally community feedback.

In working across disciplines, different methodologies often present points of difficulty, many of which have already been mentioned. Common differences not discussed elsewhere in this review include the extent of community stakeholder involvement, reducibility of models, and the objectivity possible or desired in research (Eigenbrode et al. 2007). Additional differences include: “research design, sampling, data collection, analysis, and interpretation of results” (Jakobsen et al. 2004). A commonality in these differences is the difficulty in data collection and analysis. “Sampling to answer ‘what,’ ‘how,’ and ‘why’ questions using data from a single, integrated instrument remains an unresolved challenge” (Lynam et al. 2019). This issue of data integration is a common point of consideration for researchers in a wide range of interdisciplinary fields, not only disaster research. Within the field of biomathematics, for example, researchers struggle to agree upon the data requirements as well as the scale and level of this data (Newgreen et al. 2019). Although disaster research is not unique in its struggle with data integration, it does have the added pressure of a short time horizon for collecting data post-event. The perishability and dimensionality of this data requires that common understandings of data requirements need to be reached for long-term theory and not on a case-by-case basis (Ge et al. 2019). Hence more interdisciplinary work must be done specifically in the field of disaster research to develop a robust and generalizable theory for data integration. However, the lessons to be learned in data integration should not only come from the field of disaster research. Many studies that involve data integration may offer advice on how best to accomplish this within disaster research.

To more thoroughly understand the variety of methods used in data integration and their commonalities, a review of 29 papers that involve some level of data integration is conducted. The comprehensive list of these papers has been included in the Appendix. These papers include a wide range of topics from socio-hydrology to systems biology, almost all deriving from disciplinary researchers who have begun efforts to produce collaborative integration across disciplines or interest areas. The breadth of papers sampled was deemed important based on the goal of finding the common features of research as it transcends any disciplinary boundaries. Furthermore, the goals of the studies are drastically different. Some are aimed at producing a means of data storage or indexing, while others are interested in producing integrated, predictive models. By reviewing this variety of integration efforts, the trends presented in the subsequent graphs can be deemed more representative of data integration endeavors as a whole instead of being attributable to the unique features of a given focus area.

The selection process for these papers involved a general database query for data integration from which papers were analyzed for true data integration, and not merely compilation, between disciplines and/or stakeholders. From this query, papers that involved data integration, but which had minimal discussion of their method’s shortcomings and widely applicable implications for future work, were also eliminated because they did not provide consequential information for the scope of this review. Additionally, papers that documented original research were primarily selected, with the inclusion of only three literature reviews which were shown to present insightful observations and actionable next steps for data integration based on the extent of their scope and strength of their conclusions.

It is worth noting here as well that many papers involved discussion of data integration methods that were deeply entrenched in the field of data science. These were occasionally within the scope of this review; however, the main interest of this review is to document cases in which disciplinary researchers had reached across disciplines to interact with other fields, not in which data scientists have taken the noncollaborative end results and meshed them together. Consequently, many of these data science papers did not offer conclusions with noteworthy implications for interdisciplinary efforts, so in alignment with the previously stated criteria, they were eliminated. Nevertheless, when a data integration system met the requirements of collaboration and widely applicable suggestions, it was included. From these papers, data was compiled regarding the larger implications for data integration. This data has been collected based on a system involving the following steps. First, the areas of interest for this review were created as separate categories, each of which will be presented individually in the following subsections. Then generalizable features of data integration were tabulated and sorted into the appropriate category; these features were condensed into groupings of relative commonality in order to explore trends more easily. Next, a binary system was applied in which each paper was given a value of one if that feature was present or a value of zero if it was not. The resulting summations for each feature resulted in the graphs throughout this section. By including a wide variety of topics and types of data integration, the resultant data can represent more holistically the prevalent features of data integration generally and not only through a single lens of interest.

The reasons for and benefits of data integration are shown in Fig. 1. The most referenced benefit is the means by which data integration allows for more realistic portrayals of data and the ways that data interact across areas of study. These studies also frequently cite the benefit of data integration’s ability to comprehensively analyze complex systems and offer an avenue to shared methodologies and data. The previously mentioned interest in merging epistemologies ties directly into this concept of using data integration as a means by which to attain a shared methodology and an authentically collaborative academic effort. The more cohesive, or at least noncontradictory, data from across disciplines can be, the more likely it is to yield meaningful and realistic results.

Some of the benefits found in Fig. 1 also speak to the importance of epistemological pluralism in data collection and analysis. The incorporation of various epistemologies “offers knowledge left otherwise undetected” and suggests “increased reliability of convergent data.” Oftentimes data is incapable of fully explaining an event and requires another discipline’s perspective to fully understand the interplay of objects in a system. Conversely, if the same phenomena can be described via multiple theories or disciplines, this serves to prove a lack of bias in results.

Other stated benefits include “preservation of historical data,” “increased sample sizes for analysis,” and “less taxing on stakeholders.” These benefits are all associated with the centralizing of data in the form of databases or illustrative models. Creating this centralized data source allows for archival storage of previous data sets; increases sample sizes for improved reliability; and provides stakeholders with a single, comprehensive source for information.

In the discussion of how data ought to be integrated, there are unsurprisingly different schools of thought. Within a reductionist framework, by linking more complex attributes of a system to simpler or more easily modeled elements of that system, researchers can develop constituent models from only the most simplified elements and their assigned links (Silberstein 2002). This method offers increased certainty; however, it falls under scrutiny due to its susceptibility to disciplinary imperialism in establishing what matters in data and what links can reasonably be assumed (Eigenbrode et al. 2007). In one example of a reductionist framework, Lafuerza et al. (2016) suggest the use of a chain of simple models in order to generate near identical results to a complex model. They do this by implementing retroductive validation, also commonly referred to as hindcasting (Fig. 2), in which theories and models are developed that explain the outcomes of prior events, thus demonstrating the validity of the model (Nastar et al. 2018). Lafuerza’s technique assumes linearity as does any reductionist framework, which may not be accurate based on the nature of the complex model and the concerns present within complexity theory (Kallemeyn et al. 2020). The opposite of this reductionist method is the holistic method. Holistic science is interested in systems and their irreducible interactions. Representing a holistic approach, Nastar et al. (2018) doubt that integration is possible without “distortions, gaps, or inconsistencies,” suggesting instead that interdisciplinary knowledge is complementary but impossible to reduce or integrate. However, they also recommend that the goal ought to be to cultivate a comprehensive understanding of a research problem rather than break it down into its simpler elements, suggesting the data then should not be integrated but kept separate in an effort to create a conceptually complete picture. These two approaches have limitations and strengths that complement each other. Where reductionism is seen as oversimplified and failing to capture the intricacies of the system interactions, holism can sometimes become too complex to be modeled with adequate certainty (Fang and Casadevall 2011). Thus, it is often most advantageous to consider which framework offers more benefits for a given modeling scenario or if they can possibly be used in combination iteratively to achieve the most complete picture possible without sacrificing too much certainty.

Because some of the methods presented in Fig. 2 may not be commonly encountered, they will each be explained in order to provide adequate contextualization regarding the variety of methods formulated to combat the challenges of data integration. The largest group represented in this review are those studies that aim at data compilation, creating databases, generating illustrative models, or indexing of sources for data retrieval. These have been grouped together because they are useful sources of information, but they do not have predictive power on their own. Also noted in Fig. 2 are the “top-down approaches” and “bottom-up approaches.” Top-down models use a variety of statistical methods to represent systems-level behavior based on complex rules governing interactions. Bottom-up techniques rely on a method of understanding a complex system via the formation of isolated entities that are then connected through documented interactions. The bottom-up techniques presented in this review include matrices, network-based integration, modularity, and agent-based-modeling. Matrices compile information of disparate types and sources to determine correlations (Weckwerth 2011). As a more visual representation of this, network-based integration involves individual objects of a system interacting with each other along “edges.” This is best understood as nodes connected via lines in which lines are edges and nodes are objects in the system (Weighill et al. 2019). As an extension of this concept of networks, modularity typically operates such that a given discipline, focus area, or subtask constitutes a module, and that module can be connected to other modules via any number of realistic interdependencies in the systems of interest. As defined by Hinkel and Klein (2009), modularity is “the idea of encapsulating expert knowledge in the form of self-contained modules and making them available to others via well-defined interfaces.” A common concern with modularity is the chance of contradicting methodologies across modules. A clear understanding of roles, acceptable inputs, and goals helps combat the creation of contradictions within the system. The last of these bottom-up approaches is agent-based modeling. This modeling creates agents that have simple rules of interaction with other agents and the environment around them. These rules then play out in iterations creating a complex system. Agent-based modeling has been used in the field of ecology and anthropology for many years and has more recently come into use in disaster research (Chen et al. 2006). The potential of this modeling technique has not yet been fully explored within the field of disaster research.

The remainder of the techniques presented in Fig. 2 have primarily been used in combination with one of the three techniques already discussed as means of validating, simplifying, or further contextualizing the data. The process of “quantifying qualitative data” involves somehow codifying or classifying data that cannot otherwise be numerically modeled or visualized. In this review for instance, we classify the features of the papers being reviewed into a series of somewhat generalized descriptions and then assign values of 0 and 1 to indicate whether that feature is present in the study. The usefulness of this method is inherent in that patterns become markedly more recognizable; however, this does work to flatten the information, not expressing the extent to which a given feature is present in that paper. Hence it is necessary to also conduct more in-depth reviews of the work presented here and speak to unique elements not fully captured in the graphs presented. In this way, this review also demonstrates the use of “case analysis/narrative” as means of exploring the nuances presented in these studies. Some papers reviewed also utilized “scenarios.” In addition to facilitating conversation as discussed previously in this paper, scenarios can also represent a nearly exhaustive list of plausible outcomes. Thus, given a series of logically cohesive inputs (the scenario), the model can generate a more certain outcome than when no knowns are input. For example in Gomes et al. (2020), his concept of scenario development allows for social variables that often present sizable uncertainty to be manipulated to create plausible scenarios labeled Sustainability (Green Road), Regional Rivalry (Rocky Road), Inequality, Fossil-Fueled Development, and Middle of the Road (Gomes et al. 2020). This land use example may be adaptable to disaster research in which there could be similar plausible scenarios implemented surrounding a disaster event, and it could similarly serve to alleviate some of the large uncertainties that have as of yet been irreconcilable across disciplines.

Other methods noted in this analysis include “triangulation or convergent validation,” “signal-based integration,” “portfolio attribute development,” and “diffraction.” Triangulation is a validation method that validates a model by arriving at the same answer in multiple ways, often called convergent validation (Fielding 2012). The second method, “signal-based integration,” involves the tracking of a variable as it correlates to changing inputs (Weighill et al. 2019). For the third, Tobi (2014) proposed the formation of a single complex attribute based on the implementation of portfolio representation of measurement, noted in Fig. 2 as “portfolio attribute development.” Note, this technique has data level requirements to operate effectively and thus may not be universally applicable. Lastly, Uprichard and Dawney (2019) warn against forcing data to integrate, suggesting instead the concept of diffraction, which allows integration only when apt and meaningful. The methods presented in Fig. 2 serve to both show the variety of attempts at data integration and orient the reader for the following sections in which challenges and shortcomings will be presented. Although there is a wide range of methods, the issues they encounter are commonly shared.

Holistic frameworks suggest that systems are irreducible but do not speak to the issue of how this suggestion would function within the context of modeling, thus ignoring the emergent technologies that could offer new solutions for disaster research (National Academy of Sciences et al. 2005). Similarly, Miller et al. (2008) recommend that any attempts at data integration are not generalizable and must be formed anew in every context: “a reorganization of multiple, potentially equally valid ways of knowing requires a negotiation governed by the specifics of the question and the composition of the research team.” However, there is an interest in defining an agreed upon interdisciplinary methodology for research (Tobi and Kampen 2018), and there is presently a desire for actionable items and reduced abstraction of solutions as the result of research (Rodela and Alašević 2017). These two goals are linked in that by accomplishing the first, the researcher is notably closer to achieving the second. Some work would also suggest that data integration should only be done when to do so is useful or illustrative (Lynam et al. 2019) or when it does not introduce an unacceptable amount of uncertainty (Nateghi et al. 2019). Another point of concern is reliance on potentially biased or inadequate expert opinion. The concept of who qualifies as an expert is unregulated, and thus the results of expert opinions will likely not offer sufficient standardization for consistent interdisciplinary efforts (Brink et al. 2020). Additionally, an inability to discern between established fact and a previous conjecture based on an expert opinion also poses concern in that different disciplines may interrogate the validity of data differently (Newgreen et al. 2019). These are all necessary warnings to heed when attempting data integration, and it is within this context that this review attempts to evaluate the challenges faced with different integration methods.

As shown in Fig. 3, the most commonly occurring challenge is indeed the differences in syntax, data format, data type, and data granularity. This category is admittedly large; however, to divide these elements into separate categories did not seem appropriate because they are all approaching the common difficulty of data heterogeneity. This issue consistently plagues data integration efforts, and as of yet, no universally accepted solution has been found. The next challenge is that of different available input and output requirements. Some of the research reviewed noted this challenge in the context of the scale, focus, and purpose of measured data versus desired outcome. Often data are being pulled from sources that collected data for a different purpose than the research being conducted, and so the data does not offer adequate scale or appropriate focus to conclusively answer the research question. Considerable efforts have been made to address these concerns of data generalizability through semantic technologies. As briefly mentioned in the discussion of dialects and semantics, semantic technologies allow machines to not only combine data but engage in contextualization from different sources and interpretation for different uses. This would in turn allow machines to assist in, if not automate, the process of integrating data in meaningful and consistent ways through the establishment of a linked data infrastructure. Within ontology, this linked data infrastructure is often referred to as the Semantic Web, a web of data that removes the barriers between knowledge sources. To pursue this as a possible solution for the barriers faced in disaster research, there will need to be improvements in data sharing and agreed upon data structures (Shvaiko and Euzenat 2013).

Most of the other challenges presented here are somewhat self-explanatory or are alluded to elsewhere. However, a few points of contextualization and elaboration may be helpful. First, the challenge listed as “data integration must have robust rationale” in Fig. 3 has two contexts: avoiding contradictions in the underpinnings of merged epistemologies and knowing when integration may be unnecessary. Second, it is worth noting that, when top-down approaches are isolated in this analysis, the most frequently referenced challenges are “the unpredictability of human behavior and the different input/output requirements.” These challenges speak to the rigidity of the modeling requirements for top-down approaches and their inability to adequately represent the intricacies of social systems. As a last and lengthier point, Newgreen et al. (2019) urge practicality in interdisciplinary modeling, acknowledging the purpose as “to provide insight into relatively complex systems, not to produce facsimiles including every element.” They continue by explaining that this requires omission, simplification, and approximation in data when appropriate, necessary, and such that all essentials will still be considered. However, this manipulation of data could once again be susceptible to disciplinary imperialism in that the members of one epistemology would decide what matters in the data and what is most useful in modeling (Green and Anderson 2019). This may result in a less than comprehensive model, failing to reflect the perspectives of all disciplines involved. The concepts of omission, simplification, and approximation are further problematized in the papers analyzed, contributing to the challenges in Fig. 3 listed as “knowing when discrepancies are meaningful” and “integration must have robust rationale.” However, this frequently applied method of data handling also serves to alleviate some of the difficulties associated with system complexity and unpredictability of human behavior, noted in Fig. 3 as well. This example is not unique in its ability to facilitate one element of integration while simultaneously further complicating another, demonstrating the balance that must be struck in sound data integration.

In curating previous techniques or crafting new ones, it is useful to examine the points that these papers listed as shortcomings of their research, as well as their advice for future work. In Figs. 4 and 5, these features have been listed separately, primarily for the purpose of increased readability and more cohesive phrasing.

As seen in Fig. 4 the primary shortcoming noted in the studies reviewed was “model parameterization bias and substantial uncertainties.” These were grouped into a single category because they are frequently referenced together and contribute to difficulties in attaining accurate and actionable items from research. Sawada and Hanazaki (2020) connect these factors in stating, “The major limitation of socio-hydrological models is that they are often inaccurate due to the uncertainty in their input forcing, parameters, and descriptions of the processes.” Furthermore, situations are frequently encountered in data integration in which not all values are known for a given object in the system. The common solution for this is omission of the object. This parameterization bias is problematized by Le Sueur et al. (2020) who explain, “One limitation of this approach is the inevitable loss of information either due to differences in granularity or data capture; another is that the final patient group may suffer from selection biases.” These difficulties in the creation of models due to incongruencies and uncertainties in data are a common and not easily solved problem in data integration. They also further exacerbate issues of epistemological difference. In working to explain a system, different disciplinary perspectives will likely value and, consequently, prioritize different parameters in the explanation of a system. Thus, this shortcoming serves as yet another warning to cultivate collaborative interdisciplinary teams grounded in academic trust.

The next noted shortcoming is “difficulty in capturing the intricacies in systems.” This is typically an issue associated with the frequently insurmountable limitations of data acquisition and analysis. Oftentimes this shortcoming occurs within the context of dynamic systems, in which case models may no longer be valid as the situation progresses, which could be the case for some sea-level rise modeling (Kulp and Strauss 2019). This leads to the next shortcoming of “insufficient data to ensure the nature of relationships.” This speaks to the challenges in determining the rules of interaction in a system. This difficulty has two sources. The first is once again epistemological biases in which different perspectives offer competing theories of what matters in a system. The second difficulty is simply a matter of not being able to accurately model some elements in systems nor to fully extricate a single element in order to determine its interactions with the system. To the first, the solution is once again epistemological pluralism. To the second, there is no generalizable remedy, although a thorough review of systems science and how it ought to be applied may be helpful. This is of course the nature of the shortcoming listed as “no automated or generalizable solution.” It explains the frequent suggestion of best practices because there are situations in which there are no other means of explaining how to combat the intricate systems involved in data integration and interdisciplinarity. However, some focus area specific solutions can and have been offered, which explains why that shortcoming was only noted five times in the review. And finally, “sequential analysis” references those studies which problematize the use of procedures that act sequentially, not allowing for the realistic interaction of system elements. This can create skewed data based on the order in which analysis is conducted. This shortcoming is avoided through the implementation of interdependencies and iteration in the model.

Fig. 5 notes the suggested future work from the studies reviewed. These papers are much more consistent in their suggestions. The first of which is “more standardization of data and methods.” This would undoubtedly aid in the more seamless processing of data and in the increase in data availability through the formation of more comprehensive databases. Then, there is the suggestion for “further integration with other networks.” This would create increasingly comprehensive and realistic models, thus bolstering their predictive power. “More quality shared data and models” speaks to the need for more reliable data. Models frequently require an extensive amount of data to function and even more data to be validated. The more readily quality data can be acquired at little to no cost, the more powerful models can become.

There is also the matter of people in data integration staying mindful to prevent inadequately capturing the social dynamics of the problem (Lynam et al. 2019). This recommendation is included in Fig. 5 as “incorporation of more social dimensions.” These dimensions are sometimes, consciously or not, omitted due to the difficulties in modeling human behavior (noted in Fig. 3). However, doing so introduces epistemological bias and fails to adequately represent the system. The final piece of advice offered is “more multiscale and scalable analysis.” Frequently the different scale can modify the results of the study. By creating a scalable model, the results at different levels of granularity can be analyzed to provide more insight into the interactions within the system. In addition to these recommendations, some papers also more generally recommended continued efforts toward integration. This is not included in Fig. 5 due to the lack of specificity and thus the lack of explanatory power. All these recommendations should be kept in mind when attempting data integration projects.