1 Embedding space

Location is not just another column in a dataset. Once observations happen somewhere, they become part of a spatial structure shaped by distance, adjacency, containment, connectivity and scale. That is why spatial data are analytically different from non-spatial data: geography changes both what we can observe and how safely we can interpret relationships (Lovelace, Nowosad, and Muenchow 2019; Bivand, Pebesma, and Gómez-Rubio 2013).

This matters even more in the current data-rich environment. Administrative records, mobile devices, transport platforms, satellite systems, environmental sensors and public-service databases now generate large volumes of data with an explicit geographic reference. In practice, many datasets that social scientists work with now contain coordinates, area identifiers, routes, catchments or flows between places. Spatial thinking is thus a core component of applied data work (Lovelace, Nowosad, and Muenchow 2019; Brunsdon and Comber 2015).

For this course, the main question is practical: how do we make sense of spatial data? Across chapters, we will move from describing different spatial data types to handling them in R, exploring point events, modelling flows, and then asking how spatial dependence and heterogeneity shape the results we obtain.

1.1 Why spatial data are special

Spatial data are special because the location of an observation provides social and geographical context. The same value attached to two observations may imply something very different depending on whether those observations are adjacent, far apart, embedded in the same administrative hierarchy or connected by a transport corridor. Spatial context therefore affects description, explanation and policy interpretation.

Geometry is the first reason. Spatial data usually come with a geometry such as a point, line, polygon or raster cell. That geometry allows us to compute distance, overlap, contiguity, travel cost, exposure and accessibility. In non-spatial tabular data, records can often be reordered without changing the analysis; in spatial data, relative position is part of the information itself.

Dependence is the second reason. Nearby places often share infrastructure, labour markets, environmental risks, housing systems, institutional arrangements and social composition. Values observed in one place are frequently related to values observed nearby (Anselin 1988). This violates the simplifying assumption that observations are independent draws and forces us to think carefully about clustering, spillovers and neighbourhood effects.

Scale is the third reason. A pattern that looks smooth at the department level can become highly uneven when measured for smaller local areas, and a local hotspot can disappear once data are aggregated upward. Spatial data therefore require decisions about the geographic unit of analysis, and those decisions are not neutral because they shape the patterns we see.

Spatial data are often nested. Individual events can be grouped into blocks, neighbourhoods, municipalities, departments or functional regions. Data collected at one level are often used to inform decisions at another. That hierarchical structure is part of what makes spatial analysis powerful. It can also be a source of ambiguity if we do not match the question, the scale and the data structure carefully.

1.2 The data revolution and the rise of spatial data

One useful way to understand the current importance of spatial data is to think about how many data systems have become spatially enabled. A census table may contain a small-area code. A health record may include the service area of the nearest hospital. A transport feed may record origin and destination, and a social-media or mobile dataset may include coordinates or inferred locations. Even when the geography is coarse, location has become a key attribute in a growing share of the data used in business, applied research and policy analysis.

This expansion has produced a surge of spatial data. Data in which location is directly observed, inferred or embedded through an area identifier. The shift matters because it changes the types of questions we can ask. Instead of asking only whether unemployment, injuries or access are high or low, we can ask where they are concentrated, how uneven they are, whether nearby places resemble one another, and whether relationships vary across space.

The growth of spatial data has not removed the need for judgement. It has increased it. More coordinates and maps can create a false sense of precision unless we also understand projections, geocoding errors, confidentiality constraints, aggregation, missing coverage and scale sensitivity (Lovelace, Nowosad, and Muenchow 2019; Bivand, Pebesma, and Gómez-Rubio 2013). Spatial data skills combine technical handling with interpretation. That is reading geometry correctly, choosing defensible units of analysis, and connecting outputs to the social or policy process that produced them.

Most contemporary data science and machine learning workflows do not treat space as a core modelling problem. They may include coordinates as features or maps, but they often assume that observations are exchangeable and that relationships are invariant across study areas. This course is about what changes once we take space seriously. It introduces the main data structures and modelling questions needed to embed space in a principled way, rather than adding geography as an afterthought.

For this Uruguay course, that means learning to work with several distinct but related data structures:

area or surface data attached to administrative or analytical units
point events such as Montevideo traffic injuries
connections between places, especially origin-destination flows
outcomes and relationships that may cluster or vary across space

1.3 Main types of spatial data

There are several valid ways to classify spatial data. A helpful starting distinction is between vector data, which represent discrete features using points, lines and polygons, and raster data, which represent space as a regular grid of cells (Lovelace, Nowosad, and Muenchow 2019; Brunsdon and Comber 2015). For course purposes, however, it is often clearer to organise the material around the kinds of analytical and modelling strategies the data call for.

Data types used to introduce the main families of spatial data covered in the course. Area / lattice data source: Önnerfors et al. (2019). Point data source: Tao et al. (2018). Flow data source: Rowe and Patias (2020). Trajectory data source: Kwan and Lee (2004).

Table 1.1 summarises the main families of spatial data that matter for this course. The goal is not to create a dense taxonomy for its own sake, but to identify when the structure of the data changes the analytical logic. If the same broad logic applies, it is useful to signal internal variation without multiplying categories unnecessarily.

Table 1.1: Main families of spatial data and the kinds of analytical questions they tend to raise.

Type	Typical geometry	What a record represents	Examples and applications	Why it matters analytically
Areal or surface data	Polygons or regular grid cells	An attribute measured for an area or surface unit	accessibility indicators, census rates, deprivation by local area, land cover, temperature surfaces	Requires attention to boundaries, aggregation, adjacency, resolution and MAUP
Point data	Points	An event, object or observation at a location	Traffic injuries, schools, hospitals, incidents	Supports density estimation, hotspot analysis and distance-based reasoning
Connection data	Lines, edges, routes, origin-destination links	A physical connection, route or interaction between places	roads, bus routes, rivers, utility lines, commuting, migration, freight	Emphasises connectivity, topology, direction and separation rather than only point locations
Trajectory data	Ordered coordinates through time	A path followed by a moving person or object	GPS traces, mobile phone tracks, vehicle logs	Combines space and time, making sequence and temporality central

1.3.1 Areal or surface data

Areal and gridded data can be treated together in this course because they usually support the same broad analytical logic. Attributes are attached to spatial units that partition a study area. Those units may be irregular polygons such as neighbourhoods, census tracts, municipalities or departments. Polygon-based area data depend strongly on administrative boundaries and zoning systems. Alternatively they can be regular cells such as grids, hexagons, pixels and raster surfaces. Raster and gridded data depend more strongly on resolution, cell size and resampling choices.

They two types of data are often considered separetely because these features. However, we take a different approach here grouping them. They consider these data as part of the same family as we consider them from a modelling perspective. The same modelling approaches can be applied for two data types. In both cases, the data tell us how an outcome, exposure or condition varies across spatial units, rather than across individual events or explicit origin-destination links. We can also use them to address similar questions about spatial variation, neighbouring units, aggregation, scale and heterogeneity.

Areal or surface data are central to later chapters in the course. When we look at spatial dependence and spatial heterogeneity, we will work mainly with polygon-based indicators linked to areas and ask how neighbouring places are related, and whether a single model can describe the whole geography equally well. The key lesson is the spatial partition of a territory is part of the analytical problem.

1.3.2 Point data

Point data represent discrete locations. A point may indicate where a crash happened, where a school is located, where a household was sampled, or where an environmental reading was recorded. Because point data preserve a relatively fine spatial resolution, they are often the closest available representation of where an observed event or object is actually situated (Tao et al. 2018; Brunsdon and Comber 2015).

Point data are especially valuable for exploratory analysis. They allow us to ask whether events are concentrated, dispersed, clustered around infrastructure, or unevenly distributed relative to population and services. In this course, traffic injuries data provide a concrete example of how point records can be turned into density surfaces and interpreted as patterns of concentration rather than as isolated observations.

1.3.3 Connection data

In this course, we consider line, network and flow representations as one broader family of connection data. They all describe relationships in space rather than isolated locations. A line represents fixed infrastructure such as a road, rail corridor or river. It can also represent an interaction between places, such as commuting, migration, freight, or an activity-space link between an origin and a destination (Rowe and Patias 2020). What these representations share is an emphasis on connectivity, separation and path structure. Two places may be close in straight-line distance but weakly connected in practice, while two more distant places may be strongly linked through transport infrastructure, migration systems or daily mobility.

Within this broader category, the course focuses mainly on flows. These are origin-destination relationships with a magnitude and often a direction, so the analytical logic differs from point and area data. Distance, travel impedance, origin characteristics and destination attractiveness may all matter at the same time. In the course, we use movement data to show how these relational data can be visualised, summarised and modelled.

1.3.4 Trajectory data

Trajectory data record the path of a moving person or object through space and time. They usually come from GPS devices, mobile-phone traces, logistics systems, vehicles or app-based mobility platforms. They are increasingly common in research on travel behaviour, exposure, segregation and accessibility (Kwan and Lee 2004; Kong et al. 2018). Unlike static point data, trajectory data preserve sequence, timing and movement between observations.

We do not analyse trajectory data directly in this course, but it is worth identifying them as part of the wider spatial-data landscape. They show clearly that spatial analysis is not only about “where” an event occurs, but also about how movement unfolds, how routes are constrained, and how exposure changes over time and across places.

1.4 Hierarchies, scale and context

An important property of spatial data is that observations are often embedded in multiple geographic levels. A crash happens at a point, but it can also be counted inside a street segment, a neighbourhood, a municipality and a department. A hospital serves a catchment area, but it is also located in an administrative unit and connected to a transport network. The same empirical process can therefore be represented in several ways depending on the question we ask.

This matters because each representation may reveals or hide patterns. Point data preserve local detail, but they can be noisy and difficult to compare across areas with different population sizes. Area data make comparison easier, but they summarise away within-area variation. Flow data reveal connections across places, but they can conceal what is happening inside each origin or destination. Good spatial analysis requires students to move deliberately between these levels rather than assuming one geography is naturally correct.

Hierarchy is not only geographical. It can also be temporal and spatio-temporal. Daily observations sit inside weeks and months. Hourly mobility traces sit inside daily or weekly trajectories. Local events are embedded in both wider territories and longer time frames. In practice, this means that “embedding space” often also means deciding how a process is nested in time and whether the unit we observe is the same unit at which the process operates (Kwan and Lee 2004; Kong et al. 2018).

flowchart TB
  subgraph L1["Small statistical areas"]
    A1["Blocks / output areas"]
    A2["Blocks / output areas"]
    A3["Blocks / output areas"]
    A4["Blocks / output areas"]
  end

  subgraph L2["Intermediate local areas"]
    B1["Neighbourhoods / census tracts"]
    B2["Neighbourhoods / census tracts"]
  end

  subgraph L3["Administrative areas"]
    C1["Municipalities / districts"]
    C2["Departments / provinces / states"]
  end

  subgraph L4["National territory"]
    D1["Country"]
  end

  subgraph L5["Supranational areas"]
    E1["Macro-region / international region"]
  end

  A1 --> B1
  A2 --> B1
  A3 --> B2
  A4 --> B2
  B1 --> C1
  B2 --> C1
  C1 --> C2
  C2 --> D1
  D1 --> E1

The diagram above provides a hierarchical representation of geographies. It illustrates nesting structure of places and spatial information. Small statistical areas are grouped into larger local areas, those into wider administrative units, then into countries, and in some cases into supranational regions. This is exactly why spatial data can be summarised at multiple levels and why a pattern that is visible at one level may look different at another.

1.5 Space as proxy and model choice

A core idea to carry through the course is thinking of space being a proxy for something else. A cluster on a map may point to labour-market sorting, service inequality, environmental exposure, historical segregation, administrative practice, network accessibility or some mixture of these. The coordinates or boundaries matter, but they are rarely the substantive mechanism in themselves (Anselin 1988; Lovelace, Nowosad, and Muenchow 2019).

This is why modelling space requires thinking beyond geometry. The aim is not to “add space” to a model in the abstract. The aim is to decide what aspects of the world a spatial representation is standing in for, and whether that representation preserves the part of reality we care about most. That same logic explains why there is always an element of simplification in spatial modelling. We never model the full complexity of the real world. We create a stylised version that retains the relationships, scales and spatial structures most relevant to the problem at hand (Bivand, Pebesma, and Gómez-Rubio 2013; Brunsdon and Comber 2015).

Those choices shape methodology. If the key issue is local concentration, point-pattern thinking may be appropriate. If the question is about interaction between origins and destinations, we need a flow model. If the problem is unevenness across territorial units, areal or surface data may be the right abstraction. If the concern is whether nearby places influence one another or whether relationships vary across territory, we need to think in terms of spatial dependence and heterogeneity. The modelling strategy follows from the process we think space is proxying.

1.6 Key complexities and challenges

Working with spatial data is rewarding because geography helps us explain social and environmental processes more realistically. It is also demanding because spatial data come with extra layers of complexity that affect validity, interpretation and reproducibility. Next we identify and discuss these issues as part of ordinary good practice in working and modelling spatial data.

1.6.1 Geometry, topology and spatial relations

Spatial records are not only rows in a table. They also include geometric information that determines whether features touch, overlap, contain one another, intersect or remain disconnected. Small errors in geometry, such as invalid polygons, duplicated vertices or mismatched boundaries, can produce misleading joins, wrong neighbour structures or failed overlays.

This is one reason the next chapter spends time on inspecting and validating spatial objects. In spatial work, data engineering includes checking whether the geometry makes sense for the question at hand and whether the spatial relations implied by the data match the reality the analysis seeks to represent.

1.6.2 Coordinate reference systems and measurement

Distances, areas and directions only make sense relative to a coordinate reference system (CRS). A dataset stored in geographic coordinates can be perfectly suitable for display while being inappropriate for measuring distances or areas directly. Likewise, combining layers with different CRSs can silently distort overlays if we are not careful about transformation and alignment.

This is a technical issue, but it has substantive consequences. A travel-accessibility indicator, a kernel density surface or a buffer around a health facility can all be inaccurate if the projection is inappropriate. Spatial analysis therefore requires students to understand how spatial measurement is encoded mathematically in the data structure itself.

1.6.3 Modifiable Areal Unit Problem (MAUP)

The Modifiable Areal Unit Problem (MAUP) is one of the classic challenges in spatial analysis (Openshaw 1981). It refers to the fact that our results can change when we alter the size of the spatial units used in analysis, or when we redraw boundaries while keeping the same underlying observations. Because many social datasets are analysed through administrative or analytical zones, MAUP is a routine problem rather than a rare technical curiosity.

Two related mechanisms sit inside the MAUP. The scale effect appears when we move from smaller units to larger units and patterns become smoother or sharper simply because of aggregation. The zonation effect appears when the number of zones stays similar but the boundaries are arranged differently, producing different averages, correlations or apparent clusters (Fotheringham and Wong 1991).

Figure 1.1: Graphical illustration of Modifiable Areal Unit Problem (MAUP). The same 40 individuals are shown on the left, then aggregated into alternative zone systems. The top row shows the scale effect. The bottom row shows the zonation effect.

Figure 1.1 illustrates the MAUP effects. The left panel shows the origiinal individual observations. The top row shows the scale effect by comparing coarser and finer aggregations of the same points. The bottom row shows the zonation effect by comparing two alternative boundary systems applied to the same observations. The key message is that the underlying people or events do not change, yet the area-level pattern does.

This has direct implications for applied work. Accessibility, deprivation, educational outcomes or public-health indicators may tell a different story when analysed for broad administrative units than when analysed for smaller local areas. Practical responses include checking sensitivity across scales, using the smallest defensible geography available, and considering functional geographies that better reflect the process under study (Casado-Díaz, Martínez-Bernabéu, and Rowe 2017; Arribas-Bel, Garcia-López, and Viladecans-Marsal 2021; Singleton and Spielman 2013; Patias, Rowe, and Cavazzi 2019). None of these strategies eliminates MAUP, but they make its consequences visible and easier to discuss.

1.6.4 Ecological fallacy

Ecological fallacy occurs when we infer something about individuals from patterns that have only been observed for groups or areas. An area-level relationship can be informative and substantively important, but it does not automatically describe the behaviour or characteristics of the individuals living in those areas. The problem arises when we move from statements about places to statements about people without changing the evidence base.

Figure 1.2: Graphical illustration of ecological fallacy. The left panel shows individual observations with a negative within-area relationship. The right panel shows the same data aggregated to area means, which now suggest a positive area-level relationship.

Figure 1.2 is easiest to read as a shift in perspective. Start with the circles, which represent individual observations grouped into three areas. Within each area, the pattern slopes downward: as the predictor increases, the outcome tends to fall. If we stopped there, we would conclude that the relationship is negative. But then each area is reduced to a single summary point, its average predictor and average outcome. Once we look only at those area means, the pattern appears to slope upward instead. The underlying individuals have not changed at all; only the level of aggregation has changed. That is the intuition behind ecological fallacy.

This issue is especially relevant in policy-oriented work because many spatial datasets are aggregated by design. A department with low average accessibility or high deprivation may still contain very different neighbourhoods, households and individual experiences within it. The same problem appears in public health. An area with high average deprivation may show poor average health outcomes, but that does not mean every deprived individual in that area experiences the same level of health risk, nor that every healthier individual lives in a less deprived context.

Robinson’s classic example makes this especially clear (Robinson 1950). Using data for U.S. states in 1930, he compared the share of the population that was foreign-born with the share that was literate. At the state level, the aggregate correlation was positive, which could be read as suggesting that places with more foreign-born residents were also more literate. But when the same relationship was examined at the individual level, the association was negative. The aggregate pattern and the individual-level pattern pointed in different directions. The practical lesson is to keep interpretations at the scale of the evidence unless individual-level data are also available.

1.6.5 Spatial dependence

Spatial dependence means that values observed in one location are related to values observed in other locations, and closer locations may be more associated (Anselin 1988). This can arise because of spillovers, diffusion, shared context, common infrastructure, environmental exposure or market linkages. In social-science terms, places are often not isolated cases. They are interdependent settings.

This matters because many standard statistical tools assume independence across observations. If neighbouring values are related, model residuals can cluster. Standard errors can be misleading and coefficient interpretation can become less reliable. Later in the course, we return to this idea to show how to diagnose dependence and think about neighbour structures more explicitly.

1.6.6 Spatial heterogeneity and nonstationarity

Spatial heterogeneity refers to unevenness across space. Some places systematically display higher values, lower values or different combinations of characteristics than others. These contrasts are often visible long before any formal model is fitted. In many applied settings, this is the first substantive reason to map data. We are interested in assessing if the process of interest is distributed evenly or unevenly across space.

Nonstationarity takes the idea a step further. It suggests that the relationship between an outcome and its predictors may vary across space. A factor associated with hospital accessibility or population ageing in one part of Uruguay may be weakly related, or differently related, in another. In later chapters, we distinguish between detecting spatial structure or whether a single global coefficient hides meaningful local variation.

1.6.7 Coverage, quality and confidentiality

A final challenge is that spatial data are not automatically better because they are spatially detailed. Coordinates may be missing or imprecise. Area codes may change over time. Platform data may overrepresent certain groups. Administrative systems may record the place of reporting rather than the place where an event actually occurred. The more spatially precise the data become, the more confidentiality concerns can restrict access, aggregation choices and reproducibility.

Course prompt

Before moving to the next chapter, identify the spatial structure of a dataset you already know. Ask yourself four questions: What is the observational unit? What geometry is attached to it? At what scale is it reported? And what might change if the geography were redrawn or aggregated differently? These questions are often enough to reveal whether a problem is mainly about points, areas, networks, flows, or a combination of several data structures.

1.7 What to carry forward

Three ideas from this chapter should stay with us through the rest of the course.

Spatial data are different because geometry, distance, adjacency, hierarchy and scale are part of the information, not just contextual detail.
Different spatial data types invite different descriptive and modelling strategies, so methods should be matched to the structure of the data rather than applied mechanically.
Good spatial analysis is about choosing appropriate geographies, recognising common pitfalls such as MAUP and ecological fallacy, and building models that respect spatial structure.

The next chapter turns to the practical toolkit required to do that work in R: reading spatial data, inspecting geometry, transforming coordinate systems, joining attributes, validating objects and producing clear maps.

Anselin, Luc. 1988. Spatial Econometrics: Methods and Models. Vol. 4. Springer Science & Business Media.

Arribas-Bel, Daniel, M.-À. Garcia-López, and Elisabet Viladecans-Marsal. 2021. “Building(s and) Cities: Delineating Urban Areas with a Machine Learning Algorithm.” Journal of Urban Economics 125 (September): 103217. https://doi.org/10.1016/j.jue.2019.103217.

Bivand, Roger S., Edzer Pebesma, and Virgilio Gómez-Rubio. 2013. Applied Spatial Data Analysis with r. Springer New York. https://doi.org/10.1007/978-1-4614-7618-4.

Brunsdon, Chris, and Lex Comber. 2015. An Introduction to r for Spatial Analysis & Mapping. Sage.

Casado-Díaz, José Manuel, Lucas Martínez-Bernabéu, and Francisco Rowe. 2017. “An Evolutionary Approach to the Delimitation of Labour Market Areas: An Empirical Application for Chile.” Spatial Economic Analysis 12 (4): 379–403. https://doi.org/10.1080/17421772.2017.1273541.

Fotheringham, A Stewart, and David WS Wong. 1991. “The Modifiable Areal Unit Problem in Multivariate Statistical Analysis.” Environment and Planning A 23 (7): 1025–44.

Kong, Xiangjie, Menglin Li, Kai Ma, Kaiqi Tian, Mengyuan Wang, Zhaolong Ning, and Feng Xia. 2018. “Big Trajectory Data: A Survey of Applications and Services.” IEEE Access 6: 58295–306. https://doi.org/10.1109/access.2018.2873779.

Kwan, Mei-Po, and Jiyeong Lee. 2004. “Geovisualization of Human Activity Patterns Using 3D GIS: A Time-Geographic Approach.” Spatially Integrated Social Science 27: 721–44.

Lovelace, Robin, Jakub Nowosad, and Jannes Muenchow. 2019. Geocomputation with r. Chapman; Hall/CRC. https://doi.org/10.1201/9780203730058.

Önnerfors, Åsa, Mariana Kotzeva, Teodóra Brandmüller, et al. 2019. “Eurostat Regional Yearbook 2019 Edition.”

Openshaw, Stan. 1981. “The Modifiable Areal Unit Problem.” Quantitative Geography: A British View, 60–69.

Patias, Nikos, Francisco Rowe, and Stefano Cavazzi. 2019. “A Scalable Analytical Framework for Spatio-Temporal Analysis of Neighborhood Change: A Sequence Analysis Approach.” In, 223–41. Springer International Publishing. https://doi.org/10.1007/978-3-030-14745-7_13.

Robinson, WS. 1950. “Ecological Correlations and Individual Behavior.” American Sociological Review 15 (195): 351–57.

Rowe, Francisco, and Nikos Patias. 2020. “Mapping the Spatial Patterns of Internal Migration in Europe.” Regional Studies, Regional Science 7 (1): 390–93. https://doi.org/10.1080/21681376.2020.1811139.

Singleton, Alexander D., and Seth E. Spielman. 2013. “The Past, Present, and Future of Geodemographic Research in the United States and United Kingdom.” The Professional Geographer 66 (4): 558–67. https://doi.org/10.1080/00330124.2013.848764.

Tao, Sui, Jonathan Corcoran, Francisco Rowe, and Mark Hickman. 2018. “To Travel or Not to Travel: ‘Weather’ Is the Question. Modelling the Effect of Local Weather Conditions on Bus Ridership.” Transportation Research Part C: Emerging Technologies 86 (January): 147–67. https://doi.org/10.1016/j.trc.2017.11.005.

# Embedding space Location is not just another column in a dataset. Once observations happen somewhere, they become part of a spatial structure shaped by distance, adjacency, containment, connectivity and scale. That is why spatial data are analytically different from non-spatial data: geography changes both what we can observe and how safely we can interpret relationships [@lovelace2019; @bivand2013]. This matters even more in the current data-rich environment. Administrative records, mobile devices, transport platforms, satellite systems, environmental sensors and public-service databases now generate large volumes of data with an explicit geographic reference. In practice, many datasets that social scientists work with now contain coordinates, area identifiers, routes, catchments or flows between places. Spatial thinking is thus a core component of applied data work [@lovelace2019; @comber2015]. For this course, the main question is practical: how do we make sense of spatial data? Across chapters, we will move from describing different spatial data types to handling them in `R`, exploring point events, modelling flows, and then asking how spatial dependence and heterogeneity shape the results we obtain. ## Why spatial data are special Spatial data are special because the location of an observation provides social and geographical context. The same value attached to two observations may imply something very different depending on whether those observations are adjacent, far apart, embedded in the same administrative hierarchy or connected by a transport corridor. Spatial context therefore affects description, explanation and policy interpretation. Geometry is the first reason. Spatial data usually come with a geometry such as a point, line, polygon or raster cell. That geometry allows us to compute distance, overlap, contiguity, travel cost, exposure and accessibility. In non-spatial tabular data, records can often be reordered without changing the analysis; in spatial data, relative position is part of the information itself. Dependence is the second reason. Nearby places often share infrastructure, labour markets, environmental risks, housing systems, institutional arrangements and social composition. Values observed in one place are frequently related to values observed nearby [@anselin1988spatial]. This violates the simplifying assumption that observations are independent draws and forces us to think carefully about clustering, spillovers and neighbourhood effects. Scale is the third reason. A pattern that looks smooth at the department level can become highly uneven when measured for smaller local areas, and a local hotspot can disappear once data are aggregated upward. Spatial data therefore require decisions about the geographic unit of analysis, and those decisions are not neutral because they shape the patterns we see. Spatial data are often nested. Individual events can be grouped into blocks, neighbourhoods, municipalities, departments or functional regions. Data collected at one level are often used to inform decisions at another. That hierarchical structure is part of what makes spatial analysis powerful. It can also be a source of ambiguity if we do not match the question, the scale and the data structure carefully. ## The data revolution and the rise of spatial data One useful way to understand the current importance of spatial data is to think about how many data systems have become spatially enabled. A census table may contain a small-area code. A health record may include the service area of the nearest hospital. A transport feed may record origin and destination, and a social-media or mobile dataset may include coordinates or inferred locations. Even when the geography is coarse, location has become a key attribute in a growing share of the data used in business, applied research and policy analysis. This expansion has produced a surge of spatial data. Data in which location is directly observed, inferred or embedded through an area identifier. The shift matters because it changes the types of questions we can ask. Instead of asking only whether unemployment, injuries or access are high or low, we can ask where they are concentrated, how uneven they are, whether nearby places resemble one another, and whether relationships vary across space. The growth of spatial data has not removed the need for judgement. It has increased it. More coordinates and maps can create a false sense of precision unless we also understand projections, geocoding errors, confidentiality constraints, aggregation, missing coverage and scale sensitivity [@lovelace2019; @bivand2013]. Spatial data skills combine technical handling with interpretation. That is reading geometry correctly, choosing defensible units of analysis, and connecting outputs to the social or policy process that produced them. Most contemporary data science and machine learning workflows do not treat space as a core modelling problem. They may include coordinates as features or maps, but they often assume that observations are exchangeable and that relationships are invariant across study areas. This course is about what changes once we take space seriously. It introduces the main data structures and modelling questions needed to embed space in a principled way, rather than adding geography as an afterthought. For this Uruguay course, that means learning to work with several distinct but related data structures: - area or surface data attached to administrative or analytical units - point events such as Montevideo traffic injuries - connections between places, especially origin-destination flows - outcomes and relationships that may cluster or vary across space ## Main types of spatial data There are several valid ways to classify spatial data. A helpful starting distinction is between **vector data**, which represent discrete features using points, lines and polygons, and **raster data**, which represent space as a regular grid of cells [@lovelace2019; @comber2015]. For course purposes, however, it is often clearer to organise the material around the kinds of analytical and modelling strategies the data call for. ![Data types used to introduce the main families of spatial data covered in the course. Area / lattice data source: @onnerfors2019eurostat. Point data source: @tao2018. Flow data source: @rowe2020. Trajectory data source: @kwan2004geovisualization.](figs/ch1/datatypes.png) @tbl-spatial-data-types summarises the main families of spatial data that matter for this course. The goal is not to create a dense taxonomy for its own sake, but to identify when the structure of the data changes the analytical logic. If the same broad logic applies, it is useful to signal internal variation without multiplying categories unnecessarily. | Type | Typical geometry | What a record represents | Examples and applications | Why it matters analytically | |---------------|---------------|---------------|---------------|---------------| | Areal or surface data | Polygons or regular grid cells | An attribute measured for an area or surface unit | accessibility indicators, census rates, deprivation by local area, land cover, temperature surfaces | Requires attention to boundaries, aggregation, adjacency, resolution and MAUP | | Point data | Points | An event, object or observation at a location | Traffic injuries, schools, hospitals, incidents | Supports density estimation, hotspot analysis and distance-based reasoning | | Connection data | Lines, edges, routes, origin-destination links | A physical connection, route or interaction between places | roads, bus routes, rivers, utility lines, commuting, migration, freight | Emphasises connectivity, topology, direction and separation rather than only point locations | | Trajectory data | Ordered coordinates through time | A path followed by a moving person or object | GPS traces, mobile phone tracks, vehicle logs | Combines space and time, making sequence and temporality central | : Main families of spatial data and the kinds of analytical questions they tend to raise. {#tbl-spatial-data-types} ### Areal or surface data Areal and gridded data can be treated together in this course because they usually support the same broad analytical logic. Attributes are attached to spatial units that partition a study area. Those units may be irregular polygons such as neighbourhoods, census tracts, municipalities or departments. Polygon-based area data depend strongly on administrative boundaries and zoning systems. Alternatively they can be regular cells such as grids, hexagons, pixels and raster surfaces. Raster and gridded data depend more strongly on resolution, cell size and resampling choices. They two types of data are often considered separetely because these features. However, we take a different approach here grouping them. They consider these data as part of the same family as we consider them from a modelling perspective. The same modelling approaches can be applied for two data types. In both cases, the data tell us how an outcome, exposure or condition varies across spatial units, rather than across individual events or explicit origin-destination links. We can also use them to address similar questions about spatial variation, neighbouring units, aggregation, scale and heterogeneity. Areal or surface data are central to later chapters in the course. When we look at spatial dependence and spatial heterogeneity, we will work mainly with polygon-based indicators linked to areas and ask how neighbouring places are related, and whether a single model can describe the whole geography equally well. The key lesson is the spatial partition of a territory is part of the analytical problem. ### Point data Point data represent discrete locations. A point may indicate where a crash happened, where a school is located, where a household was sampled, or where an environmental reading was recorded. Because point data preserve a relatively fine spatial resolution, they are often the closest available representation of where an observed event or object is actually situated [@tao2018; @comber2015]. Point data are especially valuable for exploratory analysis. They allow us to ask whether events are concentrated, dispersed, clustered around infrastructure, or unevenly distributed relative to population and services. In this course, traffic injuries data provide a concrete example of how point records can be turned into density surfaces and interpreted as patterns of concentration rather than as isolated observations. ### Connection data In this course, we consider line, network and flow representations as one broader family of **connection data**. They all describe relationships in space rather than isolated locations. A line represents fixed infrastructure such as a road, rail corridor or river. It can also represent an interaction between places, such as commuting, migration, freight, or an activity-space link between an origin and a destination [@rowe2020]. What these representations share is an emphasis on connectivity, separation and path structure. Two places may be close in straight-line distance but weakly connected in practice, while two more distant places may be strongly linked through transport infrastructure, migration systems or daily mobility. Within this broader category, the course focuses mainly on **flows**. These are origin-destination relationships with a magnitude and often a direction, so the analytical logic differs from point and area data. Distance, travel impedance, origin characteristics and destination attractiveness may all matter at the same time. In the course, we use movement data to show how these relational data can be visualised, summarised and modelled. ### Trajectory data Trajectory data record the path of a moving person or object through space and time. They usually come from GPS devices, mobile-phone traces, logistics systems, vehicles or app-based mobility platforms. They are increasingly common in research on travel behaviour, exposure, segregation and accessibility [@kwan2004geovisualization; @kong2018]. Unlike static point data, trajectory data preserve sequence, timing and movement between observations. We do not analyse trajectory data directly in this course, but it is worth identifying them as part of the wider spatial-data landscape. They show clearly that spatial analysis is not only about "where" an event occurs, but also about how movement unfolds, how routes are constrained, and how exposure changes over time and across places. ## Hierarchies, scale and context An important property of spatial data is that observations are often embedded in multiple geographic levels. A crash happens at a point, but it can also be counted inside a street segment, a neighbourhood, a municipality and a department. A hospital serves a catchment area, but it is also located in an administrative unit and connected to a transport network. The same empirical process can therefore be represented in several ways depending on the question we ask. This matters because each representation may reveals or hide patterns. Point data preserve local detail, but they can be noisy and difficult to compare across areas with different population sizes. Area data make comparison easier, but they summarise away within-area variation. Flow data reveal connections across places, but they can conceal what is happening inside each origin or destination. Good spatial analysis requires students to move deliberately between these levels rather than assuming one geography is naturally correct. Hierarchy is not only geographical. It can also be temporal and spatio-temporal. Daily observations sit inside weeks and months. Hourly mobility traces sit inside daily or weekly trajectories. Local events are embedded in both wider territories and longer time frames. In practice, this means that "embedding space" often also means deciding how a process is nested in time and whether the unit we observe is the same unit at which the process operates [@kwan2004geovisualization; @kong2018]. ```{mermaid} flowchart TB subgraph L1["Small statistical areas"] A1["Blocks / output areas"] A2["Blocks / output areas"] A3["Blocks / output areas"] A4["Blocks / output areas"] end subgraph L2["Intermediate local areas"] B1["Neighbourhoods / census tracts"] B2["Neighbourhoods / census tracts"] end subgraph L3["Administrative areas"] C1["Municipalities / districts"] C2["Departments / provinces / states"] end subgraph L4["National territory"] D1["Country"] end subgraph L5["Supranational areas"] E1["Macro-region / international region"] end A1 --> B1 A2 --> B1 A3 --> B2 A4 --> B2 B1 --> C1 B2 --> C1 C1 --> C2 C2 --> D1 D1 --> E1 ``` The diagram above provides a hierarchical representation of geographies. It illustrates nesting structure of places and spatial information. Small statistical areas are grouped into larger local areas, those into wider administrative units, then into countries, and in some cases into supranational regions. This is exactly why spatial data can be summarised at multiple levels and why a pattern that is visible at one level may look different at another. ## Space as proxy and model choice A core idea to carry through the course is thinking of space being a proxy for something else. A cluster on a map may point to labour-market sorting, service inequality, environmental exposure, historical segregation, administrative practice, network accessibility or some mixture of these. The coordinates or boundaries matter, but they are rarely the substantive mechanism in themselves [@anselin1988spatial; @lovelace2019]. This is why modelling space requires thinking beyond geometry. The aim is not to "add space" to a model in the abstract. The aim is to decide what aspects of the world a spatial representation is standing in for, and whether that representation preserves the part of reality we care about most. That same logic explains why there is always an element of simplification in spatial modelling. We never model the full complexity of the real world. We create a stylised version that retains the relationships, scales and spatial structures most relevant to the problem at hand [@bivand2013; @comber2015]. Those choices shape methodology. If the key issue is local concentration, point-pattern thinking may be appropriate. If the question is about interaction between origins and destinations, we need a flow model. If the problem is unevenness across territorial units, areal or surface data may be the right abstraction. If the concern is whether nearby places influence one another or whether relationships vary across territory, we need to think in terms of spatial dependence and heterogeneity. The modelling strategy follows from the process we think space is proxying. ## Key complexities and challenges Working with spatial data is rewarding because geography helps us explain social and environmental processes more realistically. It is also demanding because spatial data come with extra layers of complexity that affect validity, interpretation and reproducibility. Next we identify and discuss these issues as part of ordinary good practice in working and modelling spatial data. ### Geometry, topology and spatial relations Spatial records are not only rows in a table. They also include geometric information that determines whether features touch, overlap, contain one another, intersect or remain disconnected. Small errors in geometry, such as invalid polygons, duplicated vertices or mismatched boundaries, can produce misleading joins, wrong neighbour structures or failed overlays. This is one reason the next chapter spends time on inspecting and validating spatial objects. In spatial work, data engineering includes checking whether the geometry makes sense for the question at hand and whether the spatial relations implied by the data match the reality the analysis seeks to represent. ### Coordinate reference systems and measurement Distances, areas and directions only make sense relative to a coordinate reference system (CRS). A dataset stored in geographic coordinates can be perfectly suitable for display while being inappropriate for measuring distances or areas directly. Likewise, combining layers with different CRSs can silently distort overlays if we are not careful about transformation and alignment. This is a technical issue, but it has substantive consequences. A travel-accessibility indicator, a kernel density surface or a buffer around a health facility can all be inaccurate if the projection is inappropriate. Spatial analysis therefore requires students to understand how spatial measurement is encoded mathematically in the data structure itself. ### Modifiable Areal Unit Problem (MAUP) The Modifiable Areal Unit Problem (MAUP) is one of the classic challenges in spatial analysis [@openshaw1981modifiable]. It refers to the fact that our results can change when we alter the size of the spatial units used in analysis, or when we redraw boundaries while keeping the same underlying observations. Because many social datasets are analysed through administrative or analytical zones, MAUP is a routine problem rather than a rare technical curiosity. Two related mechanisms sit inside the MAUP. The **scale effect** appears when we move from smaller units to larger units and patterns become smoother or sharper simply because of aggregation. The **zonation effect** appears when the number of zones stays similar but the boundaries are arranged differently, producing different averages, correlations or apparent clusters [@fotheringham1991modifiable]. ```{r} #| label: fig-maup #| fig-cap: "Graphical illustration of Modifiable Areal Unit Problem (MAUP). The same 40 individuals are shown on the left, then aggregated into alternative zone systems. The top row shows the scale effect. The bottom row shows the zonation effect." #| echo: false #| warning: false #| message: false library(ggplot2) set.seed(42) push_from_breaks <- function(v, breaks, gap = 0.035) { for (b in breaks) { close <- abs(v - b) < gap v[close & v < b] <- b - gap v[close & v >= b] <- b + gap } v } pts <- data.frame( x = runif(40, min = 0.06, max = 0.94), y = runif(40, min = 0.06, max = 0.94) ) pts$x <- push_from_breaks(pts$x, breaks = c(0.25, 0.5, 0.75)) pts$y <- push_from_breaks(pts$y, breaks = c(0.5)) score <- 0.9 * pts$x + 0.3 * pts$y pts$value <- 0L pts$value[order(score, decreasing = TRUE)[1:24]] <- 1L pts$group <- ifelse(pts$value == 1L, "Blue", "Red") panel_boxes <- data.frame( panel = c("Source", "Scale coarse", "Scale fine", "Zone horizontal", "Zone vertical"), x0 = c(1.8, 7.2, 12.1, 7.2, 12.1), y0 = c(2.25, 5.05, 5.05, 0.10, 0.10), w = c(4.4, 4.4, 4.4, 4.4, 4.4), h = c(4.4, 4.4, 4.4, 4.4, 4.4) ) make_pts_panel <- function(panel_name) { box <- panel_boxes[panel_boxes$panel == panel_name, ] transform( pts, panel = panel_name, xp = box$x0 + x * box$w, yp = box$y0 + y * box$h ) } make_zone_panel <- function(panel_name, zones) { box <- panel_boxes[panel_boxes$panel == panel_name, ] zones$panel <- panel_name zones$xmin_p <- box$x0 + zones$xmin * box$w zones$xmax_p <- box$x0 + zones$xmax * box$w zones$ymin_p <- box$y0 + zones$ymin * box$h zones$ymax_p <- box$y0 + zones$ymax * box$h zones$share_blue <- mapply(function(xmin, xmax, ymin, ymax) { inside <- pts$x >= xmin & pts$x < xmax & pts$y >= ymin & pts$y < ymax if (any(inside)) mean(pts$value[inside]) else NA_real_ }, zones$xmin, zones$xmax, zones$ymin, zones$ymax) zones$fill_group <- ifelse(is.na(zones$share_blue), NA_character_, ifelse(zones$share_blue >= 0.5, "Blue zone", "Red zone") ) zones } zones_scale_coarse <- data.frame( xmin = c(0, 0.5, 0, 0.5), xmax = c(0.5, 1, 0.5, 1), ymin = c(0.5, 0.5, 0, 0), ymax = c(1, 1, 0.5, 0.5) ) zones_scale_fine <- expand.grid( xmin = c(0, 0.25, 0.5, 0.75), ymin = c(0, 0.5) ) zones_scale_fine$xmax <- zones_scale_fine$xmin + 0.25 zones_scale_fine$ymax <- zones_scale_fine$ymin + 0.5 zones_horizontal <- data.frame( xmin = c(0, 0), xmax = c(1, 1), ymin = c(0.5, 0), ymax = c(1, 0.5) ) zones_vertical <- data.frame( xmin = c(0, 0.5), xmax = c(0.5, 1), ymin = c(0, 0), ymax = c(1, 1) ) zones_all <- rbind( make_zone_panel("Scale coarse", zones_scale_coarse), make_zone_panel("Scale fine", zones_scale_fine), make_zone_panel("Zone horizontal", zones_horizontal), make_zone_panel("Zone vertical", zones_vertical) ) pts_all <- rbind( make_pts_panel("Source"), make_pts_panel("Scale coarse"), make_pts_panel("Scale fine"), make_pts_panel("Zone horizontal"), make_pts_panel("Zone vertical") ) box_outline <- transform( panel_boxes, xmin = x0, xmax = x0 + w, ymin = y0, ymax = y0 + h ) arrow_df <- data.frame( x = c(6.45, 6.45), y = c(6.95, 3.15), xend = c(7.02, 7.02), yend = c(6.95, 3.15), label = c("Scale effect (aggregate)", "Zoning effect (aggregate)"), lx = c(4.0, 4.0), ly = c(7.75, 1.55) ) ggplot() + geom_rect( data = zones_all, aes( xmin = xmin_p, xmax = xmax_p, ymin = ymin_p, ymax = ymax_p, fill = fill_group ), colour = "black", linewidth = 1.1, alpha = 0.72 ) + geom_rect( data = box_outline, aes(xmin = xmin, xmax = xmax, ymin = ymin, ymax = ymax), fill = NA, colour = "#666666", linewidth = 1 ) + geom_point( data = pts_all, aes(x = xp, y = yp, fill = group), shape = 21, size = 2.3, stroke = 0.45, colour = "#1f1f1f" ) + geom_segment( data = arrow_df, aes(x = x, y = y, xend = xend, yend = yend), arrow = arrow(length = grid::unit(0.18, "inches")), linewidth = 0.9, colour = "#666666" ) + geom_text( data = arrow_df, aes(x = lx, y = ly, label = label), hjust = 0.5, size = 3.6, colour = "#555555", fontface = "bold" ) + annotate( "text", x = -0.2, y = 5.45, hjust = 1, vjust = 0.5, label = "40 individuals:", size = 4.3, colour = "#375a7f", fontface = "bold" ) + annotate( "text", x = -0.2, y = 4.6, hjust = 1, vjust = 0.5, label = "24 blues (60%)", size = 3.6, colour = "#375a7f", fontface = "bold" ) + annotate( "text", x = -0.2, y = 3.75, hjust = 1, vjust = 0.5, label = "16 red (40%)", size = 3.6, colour = "#d95f02", fontface = "bold" ) + scale_fill_manual( values = c( "Blue" = "#4c6f7a", "Red" = "#e76f51", "Blue zone" = "#b8c8cf", "Red zone" = "#f2b8aa" ), guide = "none" ) + coord_equal(xlim = c(-4.2, 17.1), ylim = c(-0.3, 10.0), expand = FALSE) + labs(x = NULL, y = NULL) + theme_void() + theme(plot.margin = margin(10, 10, 10, 40)) ``` @fig-maup illustrates the MAUP effects. The left panel shows the origiinal individual observations. The top row shows the **scale effect** by comparing coarser and finer aggregations of the same points. The bottom row shows the **zonation effect** by comparing two alternative boundary systems applied to the same observations. The key message is that the underlying people or events do not change, yet the area-level pattern does. This has direct implications for applied work. Accessibility, deprivation, educational outcomes or public-health indicators may tell a different story when analysed for broad administrative units than when analysed for smaller local areas. Practical responses include checking sensitivity across scales, using the smallest defensible geography available, and considering functional geographies that better reflect the process under study [@casado-díaz2017; @arribas-bel2021; @singleton2013; @patias2019]. None of these strategies eliminates MAUP, but they make its consequences visible and easier to discuss. ### Ecological fallacy Ecological fallacy occurs when we infer something about individuals from patterns that have only been observed for groups or areas. An area-level relationship can be informative and substantively important, but it does not automatically describe the behaviour or characteristics of the individuals living in those areas. The problem arises when we move from statements about places to statements about people without changing the evidence base. ```{r} #| label: fig-ecological-fallacy #| fig-cap: "Graphical illustration of ecological fallacy. The left panel shows individual observations with a negative within-area relationship. The right panel shows the same data aggregated to area means, which now suggest a positive area-level relationship." #| echo: false #| warning: false #| message: false set.seed(7) areas <- data.frame( area = c("Area A", "Area B", "Area C"), x_center = c(2.5, 5.5, 8.5), y_center = c(2.4, 5.1, 7.8) ) make_area_points <- function(area, x_center, y_center, n = 18) { x <- rnorm(n, mean = x_center, sd = 0.55) y <- y_center - 0.55 * (x - x_center) + rnorm(n, sd = 0.28) data.frame(area = area, x = x, y = y) } individuals <- do.call( rbind, Map(make_area_points, areas$area, areas$x_center, areas$y_center) ) area_means <- aggregate(cbind(x, y) ~ area, data = individuals, FUN = mean) ggplot() + geom_smooth( data = individuals, aes(x = x, y = y), method = "lm", se = FALSE, linewidth = 0.9, colour = "#6b7280", linetype = "dashed" ) + geom_point( data = individuals, aes(x = x, y = y, fill = area), shape = 21, size = 2.4, stroke = 0.35, alpha = 0.9, colour = "#1f1f1f" ) + geom_smooth( data = area_means, aes(x = x, y = y), method = "lm", se = FALSE, linewidth = 1, colour = "#b45309" ) + geom_point( data = area_means, aes(x = x, y = y, fill = area), shape = 23, size = 4, stroke = 0.6, colour = "#1f1f1f" ) + annotate( "text", x = 3.1, y = 8.85, label = "Individual-level pattern:\nnegative relationship within areas", size = 3.5, colour = "#4b5563", hjust = 0 ) + annotate( "text", x = 6.2, y = 3.0, label = "Area means:\napparently positive relationship", size = 3.5, colour = "#92400e", hjust = 0 ) + scale_fill_manual( values = c( "Area A" = "#9ecae1", "Area B" = "#fdd0a2", "Area C" = "#c7e9c0" ), guide = "none" ) + coord_cartesian(xlim = c(1, 10), ylim = c(1.5, 9.2)) + labs( x = "Predictor", y = "Outcome" ) + theme_minimal(base_size = 11) + theme( panel.grid.minor = element_blank() ) ``` @fig-ecological-fallacy is easiest to read as a shift in perspective. Start with the circles, which represent individual observations grouped into three areas. Within each area, the pattern slopes downward: as the predictor increases, the outcome tends to fall. If we stopped there, we would conclude that the relationship is negative. But then each area is reduced to a single summary point, its average predictor and average outcome. Once we look only at those area means, the pattern appears to slope upward instead. The underlying individuals have not changed at all; only the level of aggregation has changed. That is the intuition behind ecological fallacy. This issue is especially relevant in policy-oriented work because many spatial datasets are aggregated by design. A department with low average accessibility or high deprivation may still contain very different neighbourhoods, households and individual experiences within it. The same problem appears in public health. An area with high average deprivation may show poor average health outcomes, but that does not mean every deprived individual in that area experiences the same level of health risk, nor that every healthier individual lives in a less deprived context. Robinson's classic example makes this especially clear [@robinson1950ecological]. Using data for U.S. states in 1930, he compared the share of the population that was foreign-born with the share that was literate. At the state level, the aggregate correlation was positive, which could be read as suggesting that places with more foreign-born residents were also more literate. But when the same relationship was examined at the individual level, the association was negative. The aggregate pattern and the individual-level pattern pointed in different directions. The practical lesson is to keep interpretations at the scale of the evidence unless individual-level data are also available. ### Spatial dependence Spatial dependence means that values observed in one location are related to values observed in other locations, and closer locations may be more associated [@anselin1988spatial]. This can arise because of spillovers, diffusion, shared context, common infrastructure, environmental exposure or market linkages. In social-science terms, places are often not isolated cases. They are interdependent settings. This matters because many standard statistical tools assume independence across observations. If neighbouring values are related, model residuals can cluster. Standard errors can be misleading and coefficient interpretation can become less reliable. Later in the course, we return to this idea to show how to diagnose dependence and think about neighbour structures more explicitly. ### Spatial heterogeneity and nonstationarity Spatial heterogeneity refers to unevenness across space. Some places systematically display higher values, lower values or different combinations of characteristics than others. These contrasts are often visible long before any formal model is fitted. In many applied settings, this is the first substantive reason to map data. We are interested in assessing if the process of interest is distributed evenly or unevenly across space. Nonstationarity takes the idea a step further. It suggests that the relationship between an outcome and its predictors may vary across space. A factor associated with hospital accessibility or population ageing in one part of Uruguay may be weakly related, or differently related, in another. In later chapters, we distinguish between detecting spatial structure or whether a single global coefficient hides meaningful local variation. ### Coverage, quality and confidentiality A final challenge is that spatial data are not automatically better because they are spatially detailed. Coordinates may be missing or imprecise. Area codes may change over time. Platform data may overrepresent certain groups. Administrative systems may record the place of reporting rather than the place where an event actually occurred. The more spatially precise the data become, the more confidentiality concerns can restrict access, aggregation choices and reproducibility. ::: callout-note ## Course prompt Before moving to the next chapter, identify the spatial structure of a dataset you already know. Ask yourself four questions: What is the observational unit? What geometry is attached to it? At what scale is it reported? And what might change if the geography were redrawn or aggregated differently? These questions are often enough to reveal whether a problem is mainly about points, areas, networks, flows, or a combination of several data structures. ::: ## What to carry forward Three ideas from this chapter should stay with us through the rest of the course. 1. Spatial data are different because geometry, distance, adjacency, hierarchy and scale are part of the information, not just contextual detail. 2. Different spatial data types invite different descriptive and modelling strategies, so methods should be matched to the structure of the data rather than applied mechanically. 3. Good spatial analysis is about choosing appropriate geographies, recognising common pitfalls such as MAUP and ecological fallacy, and building models that respect spatial structure. The next chapter turns to the practical toolkit required to do that work in `R`: reading spatial data, inspecting geometry, transforming coordinate systems, joining attributes, validating objects and producing clear maps.