Understanding Reader Takeaways in Thematic Maps Under Varying Text, Detail, and Spatial Autocorrelation

The role of text elements in visualization remains a polarizing topic, with some work concluding that visual elements are more influential than text (e.g. [24, 34]) while other work suggests the opposite (e.g. [4, 25]). For example, Kim et al. [24] showed that when both chart and caption emphasized a high-prominence feature, it was predominantly considered the main takeaway by readers i.e., readers were found less likely to use information from the text to form their takeaways. Similarly, O’Brien and Lauer [34] found that deceptive techniques (such as truncated or inverted axes) caused readers to misinterpret information even when paired with accurate explanatory text. These works suggest that a reader’s attention is more naturally influenced by the visualization rather than textual attributes.

In contrast, several studies support the view that text annotations can have a strong influence on a reader’s focus, inferred takeaways, and preferences. Borkin et al.’s eye-tracking study [3] found that titles and text are key elements in a visualization that help with recall, with readers fixating on text even if it did not appear in the title. Kong et al. [25] found that the slant of the title in a visualization influenced the perceived main idea, with readers arriving at divergent takeaways from the same visualization. Ottley et al. [36], also through an eye-tracking study, found that visualizations primarily facilitated the identification of main topics. Once the main topics were identified, however, it was found that readers extracted information from the text annotations. Additionally, participants with a preference for visual information over text information still preferred charts with a higher number of annotations. Finally, Hearst and Tory [21] examined individual preferences for visualizations when interacting with chatbots, finding that more than half of surveyed participants preferred seeing text annotations without visualizations.

A recent work closest in spirit to our study is Stokes et al. [48], which examined the interplay between text and visual elements in line charts. Their work captured reader preferences and takeaways when using these two modalities and distilled major findings into design guidelines for integrating text annotations into line charts. However, this study only examined annotations in univariate line charts and used synthetic datasets with abrupt peaks and trends. In contrast, our study expands on Stokes et al.’s initial findings by addressing several limitations: (1) we explore a more complex visualization technique (maps) and its interplay with text annotations; (2) we use real-world datasets and apply accepted methods in thematic cartography to generate base maps. Through preliminary discussions with a cartography expert, it is hypothesized that the guidelines for line charts as recommended by Stokes et al. may not naturally extend to thematic maps, due to the nuanced complexities of geospatial datasets and related visualization techniques. Although recent research by Stokes et al. [47] supports using text-only options when presenting data information, we excluded this option due to the intricacy of maps and possible biases arising from converting geospatial data to text (this limitation is further discussed in Section 6).

A thematic map is a type of map that emphasizes the spatial distribution of a particular theme, topic, or attribute within a specific geographic area. Instead of showing traditional geographical features like rivers, mountains, and political boundaries, thematic maps use various visual elements such as colors, symbols, and patterns to represent data related to a specific subject (a theme) [14]. For example, a population density thematic map might use different shades of color to depict the concentration of people in different regions, with darker colors indicating higher population densities. Thematic maps are valuable tools for visualizing and analyzing spatial data for a wide range of applications, including urban planning, environmental studies, and social sciences [43].

Spatial Autocorrelation: A critical visual task when using thematic maps is the identification of spatial clusters. Typically, color changes have been used to provide visible clustering among regions that have similarities or differences. The measurement of spatial autocorrelation with respect to an attribute of interest has been the focus of previous work on clustering map components. Some popular metrics proposed include: Join count statistic [10] developed for binary variables based on the probability that a unit area belongs to the same class as its adjacent areas; Moran’s I [32] which considers pairwise products of deviations; and Geary’s C [17] which uses the sum of squared distances. In this study, Moran’s I was selected due to its emphasis on the global detection of similarities among regions in contrast with Geary’s C, which is more appropriate for detecting spatial heterogeneity.Classification Scheme: After identifying and calculating appropriate spatial autocorrelation metrics, the analysis naturally segues into choosing a classification scheme, a key aspect of thematic map design. While spatial autocorrelation measures the extent of spatial association among individual observations, classification schemes provide a set of rules for assigning each map region to a class. Selection of the classification scheme plays a major role in the appearance of spatial clusters, which could reveal key trends or patterns that would otherwise not be readily apparent [28]. For univariate data, the simplest class selection methods rely on quantiles and standard deviations [30]. More complex methods include Jenks’ natural breaks [23], which maximizes interclass variance and minimizes intraclass variance; Cromley’s [11] minimum boundary error, which seeks to maximize spatial similarity among adjacent areas in the same class; and Jenks Caspall [23], which minimizes the deviation around class means using a heuristic process. Each method results in a different map appearance, which underlines the importance of the judicious selection of a classification scheme based on the statistical properties of the dataset. After careful consideration of the attributes of the dataset used in this study, we used the 7-class Fisher-Jenks method as our primary classification scheme for all generated maps.

Text in Maps: In addition to graphical components, it is essential to highlight the distinct roles and importance of textual elements in maps, a notion supported by previous research [15]. The effectiveness of textual elements largely hinges on their presentation. It is important to differentiate between text annotations or narratives, which supply pivotal additional content and explanations, and labels, which identify and delineate the elements they are attached. Traditionally, cartographic research has primarily centered on the study of text labels, focusing on enhancing their legibility and facilitating the correct association with the respective features they denote; a guidance substantiated in the detailed categorization by Dent et al. [13].

Data Storytelling and Data Journalism: In recent years, accessible and interactive representations of geographic information called story maps have also risen in popularity [42]. Story maps integrate maps and text, organized in the form of focused narratives. They have been used as an engaging method for showing compelling evidence of the rise in global sea levels [46] and the spread of COVID-19 [38]. In the context of data journalism, Song et al. [46] studied whether themes (US presidential campaign donations, US coastal sea-level rise), genres (longform infographic, dynamic slideshow), or tropes (color highlighting, leader lines), would influence reader retention or comprehension. They found that the story theme had no influence, while participants performed better using long-form infographics and leader lines. Individual audience differences by expertise and prior beliefs also impacted participant response. Our work continues to build on these findings by closely examining the interplay between visual and textual elements in the comprehension process from a data visualization and cartography standpoint. Our examination will highlight the importance of proper text annotations in conveying clear and concise information and in balancing visual and textual elements for an optimal reader experience.

We considered best practices in map design and analyzed existing gaps to identify map design factors for evaluation. To inform the map design factors tested in this study, we consulted Slocum et al.’s [43] textbook on thematic cartography and selected elements that could yield measurable variations in reader takeaways. Additionally, we drew insights from various other sources, including Stokes et al.’s [48] work to further inform the design of the study. Ultimately, we identified these five elements: Map Type, Map Detail, Semantic Level, Spatial Autocorrelation, and Text-Map Detail Alignment.

Map Type: Map types, as a fundamental design factor, differ with respect to their purpose and distinct representations  [33]. In this study, we limit our scope to common map types used for quantitative thematic mapping [14].

Slocum et al.’s textbook [46] on thematic cartography categorizes three primary map types suitable for visualizing a univariate measure in a geographic context, each accompanied by relevant examples:

Map Detail: A map’s geographic detail has been known to influence its legibility and a reader’s cognitive load for processing the presented information [1]. While this factor [6] has been theoretically considered in past studies, its effect on reader takeaways has not been studied, specifically in the presence of changing text elements. In this study, we chose to use U.S. maps at the state or county level, with Alaska and Hawaii omitted. Semantic Level: We apply Lundgard and Satyanaryan’s framework [31], which classifies accompanying text for visualizations into four levels. Originally intended to improve text accessibility for the visually impaired, Stokes et al. [48] utilized this framework to examine different aspects of a visualization’s content.

Spatial Autocorrelation: Changes in spatial autocorrelation result in changes in map appearance, as well as information regarding associations among adjacent locations [9, 27]. Examining spatial autocorrelation as a factor in map perception allows for the examination of how readers perceive spatial patterns and clusters, which is crucial for understanding spatial cognition. A pioneering study by Olson [35] examined the relationship between people’s perception of complexity and autocorrelation and clustering. The experiment found that highly clustered visualizations were generally rated as less complex, while dispersed visualizations were seen as more complex. Participants’ open-ended feedback supported these findings. Later, Bunch and Lloyd [6] explored various types of cognitive load in the context of geographic information, distinguishing between intrinsic load (associated with inherent complexity), extraneous load (related to presentation format), and germane load (linked to processing and schema automation). Given these insights, we hypothesize that spatial autocorrelation influences reader understanding.

In this paper, we use the metric Moran’s I [32], which calculates the degree of similarity between neighboring observations to measure a map’s global spatial autocorrelation. Moran’s I is sensitive to broad patterns and is ideal for identifying overarching trends in spatial datasets, making it a suitable measure of spatial autocorrelation for the purposes of our study.

Text-Map Detail Alignment: This variable focuses on the relationship between the granularity of text annotations and the corresponding map detail. When the level of detail in text annotations matches that of the map, such as county-level text on a county-level map or state-level text on a state-level map, it is considered aligned. Conversely, misalignment occurs when the granularity of text annotations does not correspond with the map detail. For instance, county-level text annotations on a state-level map represent a misalignment. Our interest in this variable stems from a hypothesis that the alignment (or misalignment) of these elements is not merely a design choice, but a factor that can lead to interesting and measurable differences in how readers process and understand thematic maps. We included this variable to primarily explore cognitive processing in map reading, as misaligned text and map detail may pose cognitive challenges, requiring readers to integrate separate pieces of information.

To effectively author annotations, we consulted Fairbairn’s [15] paper, which lists a comprehensive taxonomy of text symbols that can appear on maps. They can be (1) descriptive, which reflects features that are symbolized on the map; (2) analytical, which links the reader with feature attributes; and (3) positional, which describes or confirmations location. These text types have further subcategories. However, as our study only utilizes univariate thematic maps, many of these text types are not applicable. For instance, it would be unsuitable to include warning text or longitude and latitude markings on choropleth maps. Consequently, we will use descriptive (L4), determinative (L2), interpretative (L3-L4), and temporal/positional (L4) text in the study.

While past work [19] has focused on exploring the effects of text and map design factors on a reader’s comprehension, our work differs in that we considered the interdependence of these factors on novel measures of design effectiveness via the takeaways provided. First, the participants were asked to formulate a minimum of 1 to a maximum of 5 takeaways, where each takeaway is a set of sentences describing the most important information gleaned from the annotated map. Then we asked participants to rate the perceived influence of map and text in their process of obtaining information from an annotated map. The takeaways were then rated by the project team with respect to the level of geographic detail of information provided (county, state, or region) and semantic level. Thus, the following research questions and related hypotheses focused on the effects of the design variables on these three performance metrics: source, detail, and semantic level of takeaways.

Research Question 1 (RQ1): How does the design of a map influence a reader’s primary source for information (map or text annotations)?

Research Question 2 (RQ2): How does map design affect the granularity of takeaways considering the map type, geographic detail, and text content?

Research Question 3 (RQ3): How does map design impact the semantic level of takeaways in relation to geographic detail and text content?

All hypotheses and analyses scripts are preregistered on OSF ¹.

Participants for our study were recruited via Prolific.co, with eligibility criteria including U.S. residency, desktop computer usage, and a minimum 95% task acceptance rate on Prolific.co. The survey, designed for a 20-minute completion time, offered a $4 compensation upon successful completion, equating to an hourly rate of $12. We considered 103 participants in our results (65 male, 36 female, 2 non-binary, M_age = 37.2, SD_age = 23.3). The majority (61) had at least a 4-year degree.

The experimental design, as depicted in Fig. 2, was developed using the JMP® statistical analysis tool, where Map Type, Map Detail, Spatial Autocorrelation, Semantic Level, and Text-Map Detail Alignment were considered to construct an efficient experimental design. We began by assembling six real-world datasets at both the state and county levels.

Generating the stimuli followed a structured pipeline, visually represented in Fig. 3. The pipeline consisted of several key stages: encoding the assembled data into GeoJSON objects, applying thematic map design principles, and introducing variations or constants in map design factors as dictated by our experimental design (Fig. 2). An important consideration was the spatial autocorrelation metric, Moran’s I, which is determined by the dataset properties. We documented the values of Moran’s I for both the state and county levels as indicators of spatial autocorrelation for each dataset (Fig. 4).

The full survey pipeline is detailed in Fig. 6. Before the study starts, the participants are either assigned to Group A or B. The exact stimuli seen for each group are in Fig. 2. The first section was an attention check and terminology introduction. Participants were shown a map with four basic components of thematic maps: legend, title, map, and annotations. After seeing this introduction page, in the training section, they were asked to identify two of the four components at random. Participants must answer both correctly to continue.

As part of our study, we focused on gathering data on specific output variables to understand how participants interact with and interpret the map stimuli. Our primary output variables are:

Source: An ordinal measure of where the reader obtained the takeaway. This variable can assume five different values, representing the source of the reader’s takeaway; a value of 1 denotes deriving insight solely from the text, while a value of 5 signifies solely relying on the map. Values between 1 and 5 indicate varying degrees of combined influence from both the text and the map.

Takeaway Granularity: This variable captures the level of detail in the participants’ takeaways, categorized into county, state, or region.

Takeaway Semantic Level: Assessing the semantic depth of the takeaways, classified according to the framework outlined by Lundgard and Satyanaryan [31].

The third section was a continuation of the training, primarily designed to inform the participant what the takeaways are and how to form them after reading a visualization. We defined a takeaway as “a key fact, point, or idea to be remembered after viewing the map”, adapted from Stokes et al.’s [48] study. We provide a list of examples of takeaways and a brief explanation of how they were formed. We clarified that a reader can form a takeaway from reading the annotations about the map, reading the map only, or a combination of both. After this training, participants were asked two multiple-choice questions. There were four answer choices and three of them were correct. The incorrect answer was a direct contradiction to one of the annotations.

The fourth section of the study consisted of the stimuli in Experiment 1. Each participant saw six maps and was asked to write their takeaways for each map. They were allowed a maximum of five takeaways and a minimum of one takeaway per map. For each takeaway, participants reported how they formed their takeaways on a 5-point scale (1- all from text, 2- more text, 3- same, 4- more map, 5- all from the map), which constitutes the Source variable.

The fifth section of the study was an abridged version of the Visual Literacy Assessment Test [26], a validated way of assessing critical reading of various data visualizations such as line, pie, bar, stacked bar, etc. Since our study is primarily concerned with spatial data, we used questions pertaining to scatterplots, bubble charts, and choropleth maps, for a total of 17 questions. The purpose of this section was to reduce stimuli fatigue by adding variance to the tasks. It is also to properly test participants on their ability to critically read spatial data visualizations in lieu of relying on a self-reported metric. The sixth section of the study consisted of the stimuli in Experiment 2. Mechanics from Experiment 1 were repeated. Finally, the seventh and final section was a demographic survey.

Statistical analysis was conducted using the SAS®statistical software package. Due to the non-continuous nature of some of the variables, such as Map Detail, we used a flexible family of models called generalized linear mixed models (GLMM) for both estimation and tests of significance [49]. The SAS®software’s GLIMMIX package was used to estimate and test different GLMM’s. We initially fitted models that had active main effects and two-factor interactions (2FI); then, we used F-tests to determine and eliminate effects that had no significant impact on the response (Source, Takeaway Granularity, Takeaway Semantic Level) via the p-value criterion. Finally, the final GLMM used for analysis only consisted of main effects and/or 2FI that had p-values less than the significance level of 0.05. We believe that these model forms, specifically with the addition of the 2FI, were parsimonious enough to be interpretable but complex enough to capture interesting joint effects of the variables of interest.

Post-hoc analysis of the difference in contrasts between groups (i.e., categories of the independent variables) yielded estimates of odds ratios. For binary responses, odds are the ratios of the probabilities of success vs. failure; in the case of ordered data, the odds are proportional odds which are equal odds (with respect to any two categories) of the outcome being a higher-level vs. a lower level category. Odds ratios, on the other hand, are calculated as the ratios of the odds of two groups. For example, when examining the effect of a specific map design feature such as Map Type (independent variable) on the likelihood of achieving a higher Source value (outcome), an odds ratio for Chloropleth (Group 1) and Isarithmic (Group 2) that is greater than 1 would suggest that the Group 1 odds of yielding a high Source value is greater than the Group 2 odds. If the odds ratio is not significantly different from 1, this indicates that there is no difference between the two groups with respect to their impact on the Source value.

Hypothesis 1a (H1a): Map type influences a reader’s reliance on text annotations.

Due to the ordinal nature of the Source variable, where higher values indicate more reliance on maps, while lower values show more reliance on text, we used a proportional odds model to analyze the impact of the map and text variables on the comparative odds of being more reliant on map over text (odds ratios). Results from the fitted proportional odds model show that the effect of Map Type on Source was found to be significantly dependent on the semantic level of text ² i.e., the Map Type × Semantic Level interaction effect was significant (p = 0.0004), but not on Text-Map Detail Alignment (p > 0.05). Estimates of the multiplicative effect on the odds ratios (Fig. 9) show that for the choropleth map, perceiver-independent annotations increase the chance of using the map as the primary source of takeaways by a factor of 2.12 × (1/0.47 = 2.12) when compared to perceiver dependent annotations. This implies that for every 100 people, 32 will report relying more on the map when reading a choropleth map with perceiver-dependent annotations and 68 will report relying more on the map when reading a choropleth map with perceiver-independent annotations. To further clarify, consider a sample of 100 map readers. The proportion of readers that are in the perceiver-dependent group versus perceiver-independent group can be determined by the ratio of 1 to 2.12. When these ratios are normalized (i.e., \(\frac{1}{1+2.12}\) for perceiver-dependent and \(\frac{2.12}{1+2.12}\)) for perceiver-independent annotations), they correspond to approximately 32 and 68 readers. For participants reading an isarithmic map, utilizing perceiver-independent annotations as opposed to perceiver-dependent annotations results in a 1.27 × (p = 0.085) increase in the odds of predominantly relying on the map. In the scenario of the hexbin map interaction, perceiver-independent annotations result in a 2.86 × (p < 0.001) increase in the odds of a reader focusing more on the map compared to when perceiver-dependent annotations are present. Annotations at the dependent level generally increase reliance on text annotations, aligning with the hypothesis’s predicted directionality.

These results indicate that the semantic level’s effect (dependent vs. independent) on takeaways is the strongest for isarithmic maps, followed by choropleth maps, and is least apparent for hexbins. This indicates that a higher semantic level tends to foster more reliance on text annotations, with this tendency being most prominent in isarithmic maps and least so in hexbins.Hypothesis 1b (H1b): Increased geographic detail in maps leads to greater reliance on text annotations.

We used a proportional odds model to analyze the impact of the map and text variables on a higher Source (i.e. reader being more reliant on map). The Fixed Effect tests (Fig. 8) indicate that Map Detail is significant on Source (p = 0.0164). Readers who are shown county-level maps tend to display a higher reliance on text than on the map itself. The odds of a reader relying more on the map for insights are 1.22 × higher (p = 0.016) when the map detail is at the state level compared to the county level (Fig. 10) supporting this hypothesis.

Hypothesis 1c (H1c): Higher spatial autocorrelation in maps might reduce reliance on text annotations.

This hypothesis is rejected. The results reveal that Moran’s I is not significant on Source for both experiments (Fig. 8). Similarly, text attributes exhibit no interaction with Moran’s I. This leads to the conclusion that differences in geographic correlations within the phenomenon mapped do not influence a reader’s reliance on annotations.

Hypothesis 2a (H2a): Map type affects the granularity of a reader’s takeaways.

The Takeaway Granularity output variable, which takes on the values county, state, region is an ordinal data type. We used a proportional odds model to analyze the impact of the input variables on the comparative odds of having a coarser (regional) takeaway.

We examine the effect of Map Type on Takeaway Granularity, to which there is a significant effect (p = 0.0003). According to the odds ratio model which is summarized in Fig. 11, readers are 10.70 × (p < 0.001) more likely to produce takeaways at a coarser granularity when reading a choropleth map vs. an isarithmic map. Readers are also 12.5 × more likely to have coarser takeaways when reading a hexbin vs. an isarithmic map. When comparing hexbins to choropleth maps, hexbins were 1.11 × (p = 0.756) more likely to create coarser takeaways, although this result is not significant. Thus, we conclude that people are least likely to have a coarser-detailed takeaway when reading isarithmic maps.

The immediate implications of these results suggest that choropleth and hexbin maps lead to coarser granularity takeaways, while isarithmic maps help readers focus on finer details. These findings underscore the importance of the contextual use of each map type. For example, in teaching geographic patterns, choropleth maps may be more effective in highlighting trends at the regional level, while isarithmic maps may be better suited for showing phenomena at a finer level.Hypothesis 2b (H2b): The granularity of takeaways varies with map detail.

The test of Fixed Effects (Fig. 8) suggests that both Text-Map Detail Alignment (p < 0.0001) and Text-Map Detail Alignment × Semantic Level (p < 0.0001) have a significant influence on the granularity of a reader’s takeaway. When viewing a county-level map with text-map detail misalignment, the odds of a reader offering a coarser takeaway granularity are 23.03 × (p < 0.001) greater (Fig. 12) than when reading a map with an aligned text-map detail. This result suggests that detailed maps paired with coarser text detail lead to more generalized takeaways. For less detailed (state-level) maps, text detail, regardless of whether it is coarser or finer, tends to elicit more precise (state or county-level) takeaways (Fig. 12). It is noteworthy that regardless of the level of text detail, county-level maps yield same-level (county) coarser takeaways (at the state or regional level) compared to state-level maps, since there is no finer level of detail beyond the county level. This suggests that the production of finer-grained takeaways is not caused by higher-detailed maps.Hypothesis 2c (H2c): Higher spatial autocorrelation leads to takeaways with coarser detail.

Due to the categorical nature of the dependent variable such as Takeaway Granularity and the continuous nature of the independent variable (Moran’s I for Spatial Autocorrelation), we use nominal logistic regression to model the odds of having a coarser takeaway. Recall that the ordinal ordering from fine to coarse is county, state, region.

The effect of Spatial Autocorrelation depends only on Map Type, but not on text elements (Fig. 8). The results indicate that higher levels of spatial autocorrelation produce finer details for both choropleth and hexbin but higher levels of Moran’s I produce coarser level of detail for the isarithmic maps. This is seen by the negative coefficient values for choropleth (− 1.8869) and hexbin (− 2.2194) maps, but a positive coefficient (3.3453) for the isarithmic map (see Fig. 13). Thus, this hypothesis is only supported for isarithmic maps.

Hypothesis 3a (H3a): Map type influences the semantic level of a reader’s takeaways.

To better analyze the semantic level of reader takeaways, which is categorized into ordinal levels (L2, L3, or L4), we again employed a proportional odds model. This model is used to assess the influence of various map and text factors on the likelihood of a reader deriving a takeaway at a more advanced semantic level.

Our model (Fig. 14) that estimates the probability that the takeaway will be at a higher semantic level. Fixed Effect tests (Fig. 8) show that Map Type (p < 0.0001) has an effect on Takeaway Semantic Level, but the direction and magnitude of the effect is dependent on the Text-Map Detail Alignment (i.e. Text-Map Detail Alignment × Map Type is also significant at p = 0.0001). The interpretations of the odds ratios are as follows:

When text and map detail were unaligned, the odds that the reader has a higher semantic level takeaway is 8.33 × (p = 0.214) when reading an isarithmic map vs. a choropleth map. This difference is more pronounced when text and map detail were aligned, here the odds increased to 100 × (p = 0.012) for the same comparison - isarithmic map vs. a choropleth map.

When text and map detail were unaligned, the odds that a reader provides a higher semantic level takeaway is 3.43 × (p = 0.241) greater for a choropleth map vs. a hexbin. When text and map detail were aligned, the odds of readers providing a higher semantic level takeaway is 1.20 × (p = 0.525) greater for a hexbin vs. a choropleth.

When text and map detail were unaligned, the odds that a reader provides a higher semantic level takeaway is 28.07 × (p < 0.001) for an isarithmic map vs. a hexbin. When text and map detail were aligned, the odds increased to 69.66 × (p < 0.001) for the same comparison - an isarithmic map vs. a hexbin.

These quantitative results show that regardless of the level of text-map detail alignment, isarithmic maps produce significantly higher semantics than choropleths, with the effect stronger when the text detail is the aligned. Thus, this hypothesis is supported.Hypothesis 3b (H3b): Coarser map details lead to higher-level semantic takeaways. We build an odds model to analyze the impact of map and text factors on the comparative odds of having a higher semantic level takeaway. We see that test of Fixed Effects indicates that Map Detail (p < 0.0001) has a significant effect on the semantic level of the takeaways (Fig. 8). More specifically, county-level maps produce higher semantic level takeaways over state-level maps by a factor of 1.64 × (p < 0.001) (Fig. 15 B). This implies that finer map details are more likely to lead to takeaways at a higher semantic level. The opposite of our hypothesis is true, so it is rejected.

When examining this hypothesis, we also discovered that Semantic Level has a highly significant effect on the reader’s Takeaway Semantic Level (p < 0.0001). In experiment 2 (Fig. 15 A), the pairwise comparisons in our odds ratios model yielded: Readers have 4.177 × (p < 0.001) higher odds of having a higher level semantic takeaway when reading L3 annotations compared to L2 annotations. When reading L4 annotations compared to L2, the odds increase significantly to 25 × (p < 0.001) higher. Moreover, readers have 6.25 × (p < 0.001) higher odds of experiencing a higher level semantic takeaway when comparing L4 to L3 annotations. From these findings, we conclude that higher semantics in the text annotations result in higher semantics in the takeaways.Hypothesis 3c (H3c): Spatial autocorrelation within a map dataset influences the semantic level of takeaways.

For this hypothesis, we use nominal logistic regression (Fig. 16) to model how Moran’s I affects the semantic level of a reader’s takeaway. Spatial Autocorrelation × Map Type (p < 0.0001) has a significant effect on Takeaway Semantic Level. For isarithmic maps, higher Moran’s I results in lower semantic levels, as denoted by a significant p-value and negative coefficient(coefficient = −6.8788, p = 0.0410). Hexbins produce takeaways at a higher semantic level with higher spatial autocorrelation (coefficient = 4.7801, p = 0.0019). Choropleth maps also are more likely to produce takeaways at a higher semantic level with an increasing Moran’s I (coefficient = 1.9961, p = 0.1305), but this result was not significant.

The preceding subsections focused on examining the influence of map variables on participant takeaways. This subsection aims to concisely present insights gained from analyzing participants’ VLAT scores. It’s important to note that VLAT scores are observed outcomes; hence, any causal relationships cannot be established based on these scores.

The Pearson correlation coefficient between VLAT scores and the variable Source is approximately 0.085 (p < 0.001), signifying a significant correlation, albeit a slight correlation. This suggests that participants with higher visual literacy are more inclined to extract information directly from the map. However, we reiterate this relationship is correlative and does not imply causation. Furthermore, the nature of the Source as ordinal data presents limitations when correlating with VLAT scores, which are cardinal in nature.

Similarly, the Pearson correlation coefficient of -0.033 (p < 0.023) shows a weak, yet statistically significant negative relationship between VLAT scores and the Takeaway Semantic Level. This indicates that higher VLAT scores correlate with lower levels of takeaway semantics. It further suggests that individuals with better visual literacy are inclined to rely less on personal interpretation, favoring direct data extraction (e.g., reading specific values from a map) rather than making abstract or trend-based inferences.

Regarding the volume of takeaways, the mean number recorded per dataset is 774, with a standard deviation of approximately 11.85. This statistic reflects that each dataset in the study, on average, elicited about 774 takeaways. On an individual level, participants contributed an average of 46 takeaways each, with a standard deviation of 12.48.

Effects of Geographic Detail: As per H1b and H3b, readers (without specialized training in map reading and use) who saw maps with higher geographic detail (i.e. county-level maps) tend to have a higher reliance on text and produce higher semantic level takeaways. This could be due to the variation in the expertise of the map reader. Counterintuitive, our results suggest that is the aim is to impart fine-grained information to the reader, less detailed maps are more effective, particularly when combined with text annotations at the same detail as the map. A focused approach that emphasizes the regions of interest while potentially reducing the prominence of irrelevant areas can enhance the reader’s comprehension. This strategy is often adopted in census maps, where a targeted representation not only prevents information overload but also facilitates a more insightful reading experience.

Effects of Semantic Level of Text Annotations: Based on the outcomes from H2b and H3b, the semantic level of text annotations significantly impacts the granularity and semantic level of takeaways. Annotations with higher semantic levels (dependent) are more likely to produce spatially coarser takeaways and higher semantic level takeaways. Therefore, when the intention is to elicit broad or high-level understanding, using dependent level annotations can be effective. To further contextualize, it is worth noting that annotations that succinctly encapsulate trends and delineate regional patterns stand as powerful tools in encouraging a high-level comprehension. This approach is already employed in various real-world scenarios including, but not limited to, weather forecasting and temperature mapping, where conveying overarching trends and patterns to news viewers takes precedence.

Spatial Autocorrelation and Map Type: Results from H1c and H2c show that spatial autocorrelation does not significantly influence the source of takeaways or reliance on text, but it does influence the detail of takeaways depending on the map type. Higher levels of spatial autocorrelation produce finer details for choropleth and hexbin maps but coarser detail for isarithmic maps. Therefore, when designing maps, the author should be aware that the type of map chosen will impact the granularity of the takeaways. Being aware of the spatial autocorrelation existing in the dataset can guide the optimal choice of map type to communicate details more effectively. For instance, in scenarios where fine detail is important, leveraging the strengths of choropleth or hexbin maps can be more productive. Conversely, if the intention is to present a more generalized overview, isarithmic maps are a better choice.

Balancing Visual Complexity with Semantic Detail: The results from H1a and H3a reveal that map type impacts both the reliance on text and the depth of understanding derived from the map. Isarithmic maps, with their de-emphasis of the administrative units for which data were collected, often lead to seek clarity from text annotations, resulting in a higher level of understanding. Conversely, map types that visually emphasize the administrative units for which data were collected, such as choropleth maps, encourage readers to derive information directly from the map, particularly when paired with more accessible annotations. Regardless of the text-map detail alignment, hexbins and choropleths do not differ significantly with respect to their effect on the semantic level of takeaways. Additionally and as expected, dependent level annotations produce higher level semantics in the takeaways, regardless of the Map Type. Therefore, map creators should strategically choose the map type based on the desired depth of understanding and the reliance on the map itself versus the accompanying text. Hexbin or choropleth maps are preferable when the map should be the primary source of information, especially in cases where the complexity of information can’t be altered. Conversely, isarithmic maps are more suited when a higher depth of understanding is the goal, but they may risk overwhelming the reader. Given this, it is evident why most online news sources leverage choropleth maps or cartograms — these maps are extensively used in representing census, socioeconomic, or demographic data where delineating administrative units clearly is pivotal. This style accommodates a direct, unambiguous visualization of data corresponding to specific regions. In contrast, isarithmic maps naturally fit scenarios requiring the representation of continuous spatial data distributions, like weather forecasts or temperature maps. This is because they effectively convey gradients and variations, offering a holistic view of data trends across different geographical expanses. Therefore, understanding the inherent strengths of each map type can guide designers in choosing the most effective way to communicate data publicly.

To synthesize, the study illustrates the complex interplay between map design elements and their impact on reader comprehension and information takeaway. Notably, this research highlights the importance of considering the interaction between different map elements, such as type, detail, and spatial autocorrelation, along with the semantic level and detail of accompanying text when designing maps. By strategically utilizing these elements, cartographers can better guide reader comprehension and the granularity and depth of their takeaways.

In this subsection, we refer to participants by their ID from P1-P103.

Subjectivity of Takeaways: The interpretation of maps can be influenced by personal context and knowledge. For example, P30 writes “Florida does achieve more employment than Ohio, where I am located. The population might be more homogenous [sic],” while P47 drew on their knowledge about the U.S. agricultural industry: “The agricultural sector is highly dependent on government policies and subsidies, which have a significant impact on the prices of agricultural commodities.” These insights can provide unique perspectives on the data, but also highlight the potential for subjectivity in interpretation.

Some responses provided by participants were not easily verifiable. For example, P2 wrote about the map in Fig. 1 C, “less vacancy in large metro areas.” Although this may be true, we surmise that different readers may disagree on the validity of this statement. This example demonstrates how readers may bring their own assumptions to the interpretation of maps, which could lead to unobjective conclusions.

These results indicate that participants bring their own unique perspectives and experiences to the interpretation of visual information. This subjectivity can lead to diverse and creative insights, but can also introduce biases and inaccuracies in interpretation, which can hinder effective decision-making. Therefore, it is important to acknowledge and account for the subjective nature of map reading in research and in practical applications, while also valuing the potential for diverse perspectives and insights.

VLAT Scores: Participant scores on the VLAT were not considered in the results because the inclusion of these results did not yield any meaningful conclusions. We believed that VLAT scores would correlate positively with the number of takeaways per person. However, it was found that participants would provide the same number of takeaways for each map stimulus throughout the entire duration of the study. Thus, the VLAT section in the study was primarily a mechanism for preventing stimuli fatigue.

Repeating content in annotations: Our study observed instances where participants echoed the content of annotations in their responses, reflecting the impact of annotations on the interpretation process. This aligns with findings from Hullman et al., highlighting how the presence or absence of annotations can steer viewers’ attention and shape their understanding of the visualization [22]. However, this effect was small, as only 44 takeaways copied the text annotations from the maps, with 2 takeaways copying Semantic Level 2 annotations, 28 copying Level 3, and 14 copying Level 4. This distribution suggests that the majority of these repetitive takeaways are categorized under semantic levels 3 and 4. The higher occurrence in these levels indicates that while some participants directly mirrored the annotations, they predominantly did so in contexts requiring a more complex understanding (levels 3 and 4), rather than merely restating basic facts (level 2).