Guiding Geospatial Analysis Processes in Dealing with Modifiable Areal Unit Problems

Geospatial analysis plays a critical role in a range of domains [30]. For example, public health professionals used geospatial analysis to track disease outbreaks and plan interventions. During the COVID-19 pandemic, analysts used GISystems to map infection hotspots, model transmission patterns, and allocate healthcare resources efficiently [38]. Practical applications of geospatial analysis in these professional domains involve complicated processes of managing multiple datasets, selecting appropriate spatial scales and methods for analysis, and interpreting geographic patterns. This can be extremely challenging for people without adequate GIS expertise [54] and spatial thinking skills [31, 36, 34].

Due to the unique nature of geographical data, spatial analysis results often suffer from uncertainties in data accuracies, measurement frameworks, transformation artifacts, and spatial heterogeneity [40]. Addressing these uncertainties is essential for ensuring reliable conclusions and decisions. In particular, the Modifiable Areal Unit Problem (MAUP) [20, 51] is a well-known issue that often makes the results of a spatial analysis unreliable. Although the concept of MAUP and related factors is well documented, most analysts choose to ignore MAUP effects in practice due to lack of expertise, high cognitive loads, and resource constraints [50, 26]. Even if analysts are committed to addressing the effects of MAUP, there is very little help and guidance on how to decide the proper strategies and methods in a specific problem-solving context.

To bridge this skill gap for addressing modifiable area unit problem in spatial analysis, we propose a machine guidance approach that captures the knowledge and experience necessary for dealing with MAUP into an intelligence agent. While human analysts conduct spatial analysis, a machine guidance agent is capable of monitoring the progression of the spatial analysis process and volunteers help and guide in two ways: (1) detect situations where MAUP takes effect and (2) direct users to take proactively measures to mitigate its impact on analytical results. Designing such a machine guidance agent requires that we answer a number of research questions:

1. 1.

   Why do analysts tend to ignore MAUP in spatial analysis? We identified seven (7) reasons why people failed to address MAUP effectively (see Section 3.3). This analysis provides us insights on opportunities for machine guidance.

2. 2.

   What factors contribute to the level of MAUP effects? The effects of MAUP on analytical results could range from negligible to serious depending on the degree of spatial autocorrelation and spatial heterogeneity, data aggregation methods, and the choices of scale and area units (see Section 3.1). Understanding these causal factors leads to ideas and methods to mitigate MAUP effects.

3. 3.

   What are the methods and tools available to address the effects of MAUP? We synthesize the scattered literature and identify eight methods that are used to help analysts understand the nature and extent of MAUP effects and minimize the effects on the analysis (see Section 3.2). Using these methods requires a significant level of GIS expertise and is cognitively challenging.

4. 4.

   What are the opportunities and strategies of machine guidance in addressing MAUP? Machine guidance exhibits helpful behavior that should be offered only when MAUP arises and when users need help mitigating the effects of MAUP. We identify seven recognizable opportunities and prescribe guidance strategies for them (see Section 4).

5. 5.

   How would users (analysts) experience machine guidance? We demonstrate how users experience guidance by presenting a scenario of use in the context of crime hotspot analysis where the machine guidance agent helps the analyst in dealing with the MAUP. Through the scenario, we gain insight into the expected behaviors of machine guidance.

6. 6.

   How can we enable machine guidance computationally? We show how machine guidance can be enabled computationally by a software agent that can engage with users in collaborative problem solving. Our computational framework was inspired by guidance research in visual analytics [10, 11], advances in mixed-initiative interfaces [53], and intention-based interactions with GIS [9].

By answering the above research questions systematically, this paper contributes to a theoretical foundation of machine guidance in GIScience research. Developing machine guidance tools for geospatial analysis is our long-term goal, and we provide here an initial framework for exploring the design challenges in both conceptual and computational levels.

Machine guidance is an active process of addressing the cognitive challenges and expertise gaps of users that hinder their analytical progress [11]. This approach argues for solving complex and difficult problems by bringing human and computer into a collaborative work relationship [61]. Collaboration is a process in which two or more agents work together to achieve a shared goal. In our case, we introduce a machine guidance agent to partner with a human agent in spatial analytic activities.

Figure 1 shows the collaborative relationship between human analysts and the guidance agent. A machine guidance agent is an intelligent computational agent that actively assists users during analytical processes by offering contextual guidance, recommendations, and feedback [12]. It can recognize when the analyst encounters difficulties and how to help [11] by integrating reasoning, planning, and communication.

A key capability of a machine guidance agent is to monitor the progression of a spatial analysis process and to volunteer help and guide when needed. As illustrated in Figure 2, the process of solving a domain-specific problem using geospatial analysis generally starts with developing a spatial representation of the problem, followed by the formulation of spatial questions and the assembly of analytical GIS workflows. Throughout this process, a machine guidance agent works alongside to assert necessary guidance when the human analyst gets lost in navigating the problem and solution space and to steer users away from any dangers and risks under uncertainties.

The task of guiding analysts in dealing with MAUP effects is the responsibility of the box labeled Guiding Analytical Strategies and Methods. Zooming into this box, our current work aims at the following two objectives (also summarized in Figure 3):

1. Obj-1

   Building awareness of MAUP effects. The guidance agent actively monitors the analytical process to identify indicators of MAUP occurrences (such as the use of area-aggregated data for analysis). When an MAUP issue occurs, the guidance agent should alert its dangers and potential effects. If the analyst ignores it or is reluctant to address it, the guidance agent plays a role in convincing the analyst to do more exploration to understand the effects on the analytical conclusions.

2. Obj-2

   Bridging the gaps of expertise in mitigating MAUP effects. If the analyst is committed to addressing the MAUP effects, the guidance agent will direct or assist the process of experimenting with multiple spatial units and scales, applying various methods to verify and confirm the choices of area units, and prescribing GIS workflows for proper implementation. Machine guidance simplifies this process by automating repetitive tasks, providing statistical references, and offering immediate feedback on potential solutions. This allows analysts to focus on steering the analysis to achieve confident results.

Given the above objectives, it is important to establish a deep understanding of how MAUP arises in spatial analysis, what factors contribute to the serenity of MAUP effects, and what methods and tools are available to explore and mitigate MAUP effects. We will answer the above questions through synthesizing the literature.

Many applications of geospatial analysis use area-aggregated data as the primary unit of analysis [64, 33]. Data aggregation by area units smooths out local variations, potentially masking important spatial patterns and heterogeneity within areal units. Spatial analysis using area-aggregated data often relies on the assumption of internal uniformity within each area unit. This assumption is rarely true in real world contexts, where factors such as population density, land use, and environmental conditions can vary considerably within a single region. A key issue stems from the wide variety of potential spatial units available for analysis, including administrative boundaries, census tracts, natural zones, and regular grids. The results of spatial analyses can differ markedly depending on which of these areal units is selected. Openshaw [49] demonstrated this phenomenon by showing how correlation coefficients changed when smaller spatial units were aggregated into larger ones. His findings revealed that correlation values can fluctuate between different spatial scales. This effect, known as the Modifiable Areal Unit Problem (MAUP), undermines the credibility of analyses based on arbitrarily chosen spatial units, casting doubt on the reliability and validity of the resulting conclusions.

The effects of MAUP on analytical results could range from negligible to very serious. This variability of MAUP effects can be explored by comparing analytical results on different spatial scales (thus the scale effect [27]) or using different zoning schemes (thus zoning effects [18]). Fotheringham and Wong [25] demonstrated that spatial aggregation introduces biases that vary depending on the chosen scale. This highlights how the choice of spatial scale significantly impacts analytical outcomes, emphasizing the importance of selecting an appropriate scale for an analysis. The zoning effect, on the other hand, arises from the specific configuration of spatial units. Even with the same number of zones, different boundary arrangements can produce drastically different statistical outcomes. Openshaw and Taylor [50] experimented with the use of alternative configurations of counties to compute the strength of correlation and they showed that the results of correlation coefficients ranging from 0.265 to 0.862, highlighting the inherent instability in spatial analysis.

Given the inherent complexities and challenges of addressing MAUP, there are critical moments where machine guidance can effectively assist analysts. In this section, we use the seven key challenges in addressing MAUP (as outlined in Section 3.3) to pinpoint critical moments when guidance is needed. Table 1 characterizes the possible guidance opportunities corresponding to the seven user challenges. For example, guidance can be inserted when the system detects that the analysis involves the use of area-aggregated data in geostatistical analysis (G1).

To take advantage of the guidance opportunities identified in Table 1, the guidance agent must form intention to volunteer guidance and formulate a strategy to generate guidance messages. Table 2 describes the guidance strategies we use as design rationales for our guidance agent. For each strategy, we specify the goals that can be achieved and prescribe a recipe for action. These guidance strategies are consistent with the guidance objectives described in Figure 3. It is important to note that the guidance agent does not dictate how the analyst deals with the MAUP issue. If the agent believes that the effect of MAUP should be handled, the guidance agent will convince the analyst to do more explorations and suggest suitable methods and operations to mitigate the MAUP effects according to the prescribed action recipes.

It is important to emphasize that the set of strategies prescribed in Table 2 is a significant finding of this paper. It fills a knowledge gap between mitigation goals (Table 1) and GIS methods (described in Section 3.2). Despite the abundance of methods available to address MAUP, there has been little understanding of how to effectively match and apply these methods to specific mitigation goals. For example, multi-scale analysis (M1) and sensitivity analysis (M2) are frequently cited as methods useful for dealing with MAUP, but exactly how to apply them is a knowledge inaccessible to most analysts.

To illustrate how a human analyst experiences interacting with the guidance system, we present a hypothetical scenario in which a public safety analyst uses geospatial analysis of crime hotspots to inform police actions.

Danny, a public safety analyst at the Baltimore City Police Department, is responsible for planning crime prevention strategies. He is charged with developing a police petrol plan on how to dispatch officers to neighborhoods based on crime hotspot patterns. Since the department has a limited number of police force to dispatch, it must ensure that the dispatch plan generates a measurable reduction in crime rates. It is very important that Danny derives reliable and trustworthy results from his analysis. He has access to ArcGIS Desktop and crime data from the last few months.

Danny is familiar with basic concepts and methods of GIS analysis, but he is not an expertise in GIS tools and algorithms. Danny is representative of a class of analysts who are experts in their fields but have limited or no knowledge of geospatial analysis methods and tools [48, 62]. These analysts lack specialized training in GIScience or have only surface knowledge of MAUP.

Danny has access to a crime incident dataset that contains ten types of crime (see the picture of MG 1 in Table 3). Each type of crime has different underlying mechanisms and processes that produce the crime patterns. Criminogenic situations can vary in scale, duration and impact, affecting entire regions or specific groups [23]. This raises challenges with respect to the selection of an appropriate spatial unit to identify hotspot areas [44].

Based on the narrative of the scenario above, we present a hypothetical sequence of interactions between the User (Danny) and the guidance agent (MG) in Table 3. This hypothetical dialogue showcases how machine guidance can systematically address MAUP by raising awareness, recommending alternative methods, and providing statistical support to improve decision making under practical constraints.

As Danny analyzes crime hotspots in neighborhoods, the guidance agent actively monitors the process and detects potential MAUP effects due to spatial aggregation. At step (User 3), the system sensed that the user is unaware of the MAUP problem, an alert guidance is initiated to warn the user about the danger. To help the user understand the risk of MAUP effects, the guidance agent repeated the analysis using an alternative spatial unit, census tracts, and showed the user that the result is significantly different(MG 3-a). To further convince the user to take steps to mitigate risk, the agent computed LISA (as a prescribed strategy S8), suggesting strong MAUP effects (MG 3-b). These maps and messages convinced Danny to commit serious effort to mitigate the MAUP effect.

At step (User 4), Danny acknowledges the rise of a MAUP issue and decides to explore the likely effects. However, Danny does not know how to proceed. He asks for help directly and the guidance agent suggests using disaggregated data where available (S4) and offers KDE (S5) as an alternative method for density calculations, mitigating the distortions introduced by arbitrary spatial units. In this stage, Danny was guided to choose mitigation methods. He was also assisted in executing a proper GIS workflow for exploratory analysis. For practical reasons, Danny is not free to choose any area units other than neighborhood boundaries. The guidance agent adapted a strategy to verify the hotspots using kernal density surface representation (MG 4). MG 5 was generated using the Incremental Spatial Autocorrelation tool ¹¹1https://desktop.arcgis.com/en/arcmap/latest/tools/spatial-statistics-toolbox/incremental-spatial-autocorrelation.htm to determine an appropriate spatial scale (M8), which is then applied as the distance banding parameter for Hotspot Analysis ²²2https://desktop.arcgis.com/en/arcmap/latest/tools/spatial-statistics-toolbox/hot-spot-analysis.htm with Gi* statistics (M8). Such statistical validation is used as additional evidence to convince Danny that he should take measures to minimize uncertainty and improve the reliability of their conclusions.

Our approach would not be complete without discussing the feasibility of achieving our design goals through machine intelligence. To demonstrate the feasibility of machine guidance, we are developing a prototype design that supports the guidance behavior demonstrated in the scenario of Table 3. A full discussion on that prototype implementation is beyond the scope of this paper. However, we do want to briefly describe the computational frameworks employed and shed light on the practicality of implementing machine guidance.

The Modifiable Areal Unit Problem (MAUP) continues to pose a significant challenge in GIScience, yet discussions surrounding its causes, consequences, and solutions remain fragmented. Although existing research has primarily emphasized the scientific implications of MAUP, practical strategies for addressing it in real-world applications are still limited and underdeveloped [50, 25]. Our analysis reveals that many analysts tend to overlook MAUP or underestimate its impact, underscoring a critical disconnect between theoretical understanding and practical implementation.

Our work contributes to a practical approach to address MAUP in geospatial analysis. We proposed to introduce an intelligent agent to guide analysts in mitigating the effect of MAUP. As the first step toward this long-term goal, this paper established a preliminary theory of machine guidance by answering a number of fundamental research questions. We identified multiple opportunities for the machine to guide the analysts by alerting to the rise of MAUP, assessing the impact of MAUP, choosing mitigation methods, and generating visual guidance messages using GIS functions and tools. In terms of choosing what guidance features to be designed, we set two sets of objectives machine guidance in MAUP: (1) building awareness (2) supplement user’s expertise in mitigating MAUP effects. This level of understanding allows for further refinement and formalization of the related expertise in computational systems.

MAUP in geospatial analysis poses challenges in identifying its causes, selecting mitigation strategies, and interpreting scale-dependent results [63, 50]. Machine guidance has the potential to provide a proactive solution for addressing MAUP by alerting analysts to potential consequences, offering suitable methods, and facilitating executions in the GISystem. Given the resolution-dependent nature of geographic data [32], the selection of appropriate methods is crucial. Visual guidance, such as standardized map comparisons (Table 3), helps analysts interpret MAUP effects more effectively [57, 13], reducing the likelihood of overlooking its impact [21, 59, 1]. Addressing MAUP through machine guidance demonstrates its potential to enhance geospatial analysis in various domains by expanding its knowledge base and integrating domain-specific solutions [45, 43, 56].

The work presented in this paper is the first step towards the goal of active machine guidance when analysts encounter MAUP during geospatial analysis. Although we made a convincing argument for the feasibility of machine guidance and its capacity to address MAUP, the scientific merit of this approach needs to be assessed by the usefulness of the tool (machine guidance agent) when it is implemented, refined, and tested. Our ongoing research focuses on evaluating and refining the proposed strategies to ensure practical applicability. Observing how participants interact with the system, our aim is to gain a deeper understanding of when and how the guidance should be introduced when addressing the MAUP. We are collecting data on user experience and feedback and identify areas for improvement. We apply a human-centered approach to further refine both the conceptual and computational components. The findings of the study of machine guidance are likely to inspire and inform researchers in both GIS and Human-Computer Interaction (HCI) regarding the design of interactive components in GISystems.