Science Education Systems · Article 92. Maya, Jia Jun, Hana and Ethan are fictional learners used to make scientific reasoning visible. This article owns one distinct scientific job: spatial analysis—understanding how location, distance, neighbourhood, direction, scale and spatial dependence shape scientific patterns. It does not replace fieldwork, sampling or general data interpretation. Its canonical question is: what does where something happens tell us about why and how it happens?
The 50-second parent route
A map is not decoration. Location can be a variable, a source of dependence, a clue to mechanism and a boundary on generalisation.
The route is:
question → spatial unit → coordinates → scale → neighbourhood → distance → pattern → spatial dependence → model → uncertainty → mechanism → validation → bounded conclusion
The fastest diagnostic is to ask: If I shuffled the locations while keeping the measurements the same, would the scientific meaning change? If yes, space is part of the evidence and should be analysed explicitly.
This article extends How Scientific Fieldwork Works, How Scientific Sampling Works, How Scientific Scale Works and How Scientific External Validity Works.
1. Spatial analysis begins when location matters
Temperature varies across a city. Species cluster near water. Disease cases concentrate around transport routes. Soil properties change down a slope. Student travel time depends on network distance, not straight-line distance. The location of each observation contributes scientific information.
2. Space is not merely another column in a spreadsheet
Nearby observations often share environmental conditions, infrastructure, ancestry, exposure and measurement systems. This creates dependence that ordinary methods assuming independence may not handle correctly.
3. Maya’s first error is map-reading without measurement
She sees two clusters on a map and declares a causal hotspot. Her repair is to define the population at risk, sampling effort and statistical expectation before interpreting visual concentration.
4. Jia Jun’s first error is using straight-line distance automatically
Two locations are one kilometre apart across a river but ten kilometres apart by road. His repair is to choose the distance metric matching the mechanism.
5. Hana’s first error is treating every point as independent
Fifty sensors within one small park are analysed like fifty unrelated sites. Her repair is to model spatial autocorrelation and true sampling units.
6. Ethan’s first error is changing map boundaries until the story looks strong
His repair is to recognise that aggregation units can change apparent correlations. Spatial conclusions should be robust to reasonable boundary choices.
7. Coordinate systems define where observations are
Latitude-longitude, projected coordinates and local grids represent position differently. Mixing systems without conversion can move points dramatically.
8. Projection changes shape, area or distance
No flat map preserves every property of the Earth perfectly. The projection should suit the analysis: some preserve area better, others direction or local distance.
9. Spatial accuracy matters
A GPS coordinate may have metre-scale uncertainty. If the scientific boundary is only a metre wide, location uncertainty becomes part of classification uncertainty.
10. Spatial resolution sets what can be detected
A satellite pixel covering hundreds of metres cannot resolve a one-metre habitat patch. Fine resolution reveals local detail but may increase noise, storage and computational cost.
11. Grain and extent are different
Grain is the size of the smallest spatial unit measured. Extent is the total area covered. A study can have fine grain over a small extent or coarse grain over an entire region.
12. Scientific conclusions depend on both grain and extent
A neighbourhood pattern may disappear at national scale. A regional gradient may be invisible in one small plot. Scale changes what relationships become visible.
13. Primary Science can learn spatial thinking through simple maps
Where are plants growing?
Where is the shade?
Where is the ground wet?
Overlaying simple observations reveals that location can connect variables.
14. Primary 3 can use position and direction
Record where objects or organisms occur relative to a landmark. Build the habit of connecting observation to place.
15. Primary 4 can compare zones
Near drain versus far from drain.
shade versus open area.
edge versus centre.
Students learn that space can structure comparisons.
16. Primary 5 can measure distance
Does temperature change with distance from a wall? Does plant abundance change with distance from a path? A transect turns spatial position into a quantitative variable.
17. Primary 6 can detect clusters and gradients
Students can distinguish a cluster of similar values from a smooth directional change and suggest different mechanisms for each.
18. Secondary Science can formalise spatial dependence
Distance matrices.
neighbourhood weights.
spatial autocorrelation.
interpolation.
point processes.
network distance.
Students can see maps as quantitative models.
19. Distance should reflect the mechanism
Euclidean distance fits movement across open space. Network distance fits roads, rivers or pipes. Travel time fits accessibility. Resistance distance may fit ecological movement through difficult landscapes.
20. Direction can matter as much as distance
Downwind is different from upwind.
downstream is different from upstream.
downslope is different from upslope.
Anisotropic processes have directional structure.
21. Neighbourhoods need definitions
Nearest five points?
within 500 metres?
sharing a boundary?
connected by the same road?
Spatial models depend on what counts as a neighbour.
22. Different neighbourhood definitions can change results
If the conclusion exists only under one arbitrary neighbourhood choice, it may be fragile. Sensitivity analysis should compare plausible definitions.
23. Spatial autocorrelation means nearby values are related
Positive autocorrelation means nearby values tend to resemble one another. Negative autocorrelation means neighbours tend to differ. Zero autocorrelation means location gives little information about similarity at that scale.
24. Positive spatial autocorrelation is common
Nearby places share climate, soil, infrastructure and exposure. This creates smooth patches or clusters rather than random scatter.
25. Spatial autocorrelation violates ordinary independence assumptions
If neighbouring observations share information, effective sample size is smaller than the raw count. Standard errors based on independence can become too narrow.
26. Moran-like statistics summarise global spatial autocorrelation
They compare value similarity with spatial proximity under a defined weight structure. A global statistic tells whether spatial pattern exists overall, not necessarily where it occurs.
27. Local indicators identify spatial pockets
High values surrounded by high values.
low surrounded by low.
high outlier among low.
low outlier among high.
Local statistics reveal where unusual spatial relationships occur.
28. Multiple testing applies to hotspot maps
Testing many locations creates false-positive opportunities. Spatial dependence also complicates ordinary corrections. Hotspots should not be treated as unquestionable simply because software colours them red.
29. Density and count are different
Ten cases in a small population can represent higher rate than fifty cases in a huge population. Maps of raw counts often reflect where more people or organisms exist.
30. Denominators matter spatially
Disease cases per population.
accidents per traffic volume.
species detections per survey effort.
Spatial comparisons require relevant exposure denominators.
31. Population-at-risk maps change hotspot interpretation
A city centre may show many events because many people pass through it. Rate maps can reveal whether risk per person is actually unusual.
32. Small-area rates can be unstable
One event in a tiny population produces a large rate. Empirical Bayes smoothing or hierarchical models can reduce extreme noise while preserving real differences cautiously.
33. Smoothing trades local detail for stability
Too much smoothing hides real hotspots. Too little leaves noise. The smoothing scale should match the phenomenon and decision.
34. Point patterns ask where individual events occur
Trees.
nests.
earthquakes.
incidents.
shops.
The pattern can be clustered, regular or compatible with spatial randomness.
35. Clustering can arise from shared habitat
Plants cluster where soil is wet. Disease clusters where exposure is common. Shops cluster near customers. The cluster itself does not identify the mechanism.
36. Clustering can arise from contagion or interaction
One event changes the probability of nearby events. Infection, fire spread or ecological facilitation can produce self-reinforcing patterns.
37. Clustering can arise from observation bias
More observers search near roads. More sensors are installed in cities. Apparent clusters can reflect sampling effort rather than underlying phenomena.
38. Complete spatial randomness is a reference model
It asks what the point pattern would look like if events occurred independently and uniformly over the study area. Real systems rarely satisfy this exactly, but the model provides a baseline.
39. Inhomogeneous patterns require better reference models
If population density varies, uniform randomness is inappropriate. A fair null model should account for known background intensity.
40. Edge effects matter
Points near the study boundary have fewer observable neighbours because part of their surrounding area lies outside the map. Spatial statistics often need edge correction.
41. Study boundaries can create artificial patterns
A cluster appears cut in half because the map ends at an administrative border. Science should distinguish the phenomenon boundary from the data boundary.
42. Administrative boundaries are often convenient, not causal
Neighbourhoods, districts and countries matter socially, but physical phenomena may cross them freely. The spatial unit should fit the mechanism where possible.
43. The modifiable areal unit problem is fundamental
Aggregate the same point data into different zone sizes or boundaries and correlations can change. Spatial relationships may be properties of aggregation rather than individuals.
44. Scale effect changes results with unit size
Fine zones show local variation. Large zones average differences away. A relationship can strengthen or weaken depending on aggregation scale.
45. Zoning effect changes results with boundary placement
Even when zone size remains similar, redrawing boundaries can change which observations are grouped together and therefore change statistics.
46. Ecological fallacy confuses group and individual relationships
A region with higher average income may have lower disease rate. That does not prove richer individuals within each region have lower risk. Group-level and individual-level effects differ.
47. Atomistic fallacy runs the other direction
An individual-level relationship may not describe group-level processes. Scale of inference should match scale of data and mechanism.
48. Interpolation estimates between measured locations
Field measurements are discrete. Interpolation creates a continuous surface estimate. Every interpolated value is inferred, not directly measured.
49. Nearest-neighbour interpolation is simple
Assign each unsampled location the value of the nearest sampled point. It preserves measured values but creates abrupt boundaries.
50. Inverse-distance weighting assumes closer points matter more
Nearby observations receive larger weights. The power parameter controls how quickly influence decays with distance.
51. Kriging models spatial covariance
Geostatistical kriging uses an estimated spatial dependence structure to produce predictions and uncertainty. It requires stronger modelling assumptions but can use spatial correlation efficiently.
52. Variograms describe spatial dependence by distance
How does expected difference between observations change as separation increases? The variogram reveals range, sill and short-distance behaviour under the model.
53. Range is a spatial dependence distance
Beyond a certain separation, observations may become effectively uncorrelated under the fitted model. That distance has scientific meaning about process scale.
54. Nugget captures short-scale variation or measurement error
A non-zero jump near zero distance can represent microscale heterogeneity, instrument noise or both.
55. Sill represents total variance under the model
The variogram approaches a plateau as spatial dependence fades. Interpretation depends on model form and data support.
56. Interpolation uncertainty should be mapped
Predictions far from sampled locations are usually less certain. A smooth colour surface without uncertainty can create false confidence.
57. Extrapolation beyond the sampled region is riskier
Interpolation operates inside spatial support. Predicting outside the observed domain adds boundary-condition uncertainty.
58. Spatial regression accounts for location-linked structure
Ordinary regression residuals may remain spatially correlated, indicating missing spatial processes. Spatial models can represent dependence explicitly.
59. Spatial lag models represent neighbour influence
The outcome in one place depends partly on outcomes in neighbouring places. This can model diffusion, imitation or interaction, but causal interpretation needs mechanism.
60. Spatial error models represent omitted spatial structure
Unmeasured spatial variables create correlated residuals. Modelling that structure improves uncertainty estimates but does not identify the missing variables automatically.
61. Spatial confounding can distort exposure effects
An exposure varies geographically along with socioeconomic, environmental or infrastructural variables. Apparent causal effects may reflect shared spatial pattern.
62. Fine spatial adjustment can over-control
If location is adjusted so aggressively that real exposure variation disappears, causal estimates can become unstable or target a different question.
63. Spatial fixed effects compare within zones
They can control stable differences between regions but cannot remove confounding that varies within zones or over time.
64. Spatial random effects model residual heterogeneity
Hierarchical models can share information among neighbouring regions while allowing each area its own estimate.
65. Networks create non-Euclidean space
Roads, rivers, power lines, social connections and neural pathways constrain movement. Distance should follow the network where mechanism travels along edges.
66. Shortest-path distance can be more meaningful than straight-line distance
Two stations close geographically may require a long route through the network. Accessibility follows connectivity.
67. Travel time can be a better metric than distance
Congestion, speed limits, transfers and terrain change effective separation. For service access, minutes may matter more than kilometres.
68. Resistance surfaces model difficult movement
Animals, pollutants or people move more easily through some landscapes than others. Assigning movement cost to terrain creates effective distances that reflect barriers.
69. Directional flow changes neighbourhood
Upstream sites can influence downstream sites more than the reverse. Spatial weights should respect process direction.
70. Wind creates moving neighbourhoods
Downwind exposure depends on meteorological conditions. The relevant spatial relation changes over time.
71. Space and time often interact
A disease cluster moves. A heat plume changes by hour. A species expands its range. Spatiotemporal analysis models location and time together.
72. Static maps can hide dynamic processes
Two maps at different times may look similar while individual events moved. Animation, trajectories or change maps can reveal dynamics.
73. Space-time sampling needs balance
Many locations on one day versus one location every day answer different questions. Design should match whether spatial variation or temporal variation is the priority.
74. Worked case: schoolyard surface temperature
Students record temperature at grass, concrete and shaded points. A strong spatial design spreads measurements across the grounds, uses consistent sensor height and time windows, and records shade rather than selecting only convenient points.
75. The schoolyard case teaches clustering
Concrete points cluster near buildings. Grass points cluster near fields. Surface type and location are confounded, so a simple grass-versus-concrete average may mix material and microclimate.
76. The schoolyard case teaches matched locations
Compare grass and concrete under similar shade and time where possible. Spatial matching improves the causal contrast.
77. Worked case: stream pollution
Measure a water-quality indicator upstream and at several downstream distances from a discharge point. The expected mechanism predicts a directional gradient modified by flow and tributaries.
78. The stream case teaches network distance
Two sites physically close across a bend may be far apart along water flow. Distance should follow the river network.
79. The stream case teaches dilution and new inputs
A decreasing downstream concentration may reflect dilution. A sudden increase may indicate another source. Spatial pattern generates mechanistic hypotheses.
80. Worked case: disease mapping
Raw cases concentrate in densely populated areas. Rates reveal a different pattern. Age structure or access to testing may explain residual spatial variation. Mapping is the beginning, not the causal conclusion.
81. Disease maps need privacy protection
Fine geolocation can identify individuals, especially for rare conditions. Spatial resolution should balance scientific usefulness and confidentiality.
82. Spatial aggregation can protect privacy but reduce detail
Larger zones lower re-identification risk while hiding local clusters. The trade-off should be explicit.
83. Worked case: species distribution
Observations cluster near hiking trails. Is the species trail-associated, or are observers simply more likely to search there? Sampling effort is a spatial confounder.
84. Presence-only data need background modelling
Observed locations show where species were detected, not where they were absent. Models require assumptions about availability and observation effort.
85. Remote sensing creates wall-to-wall spatial data
Satellite imagery covers large regions consistently, but pixels still represent sensor signals requiring classification or physical interpretation.
86. Ground truth anchors remote spatial models
Field measurements at known coordinates validate whether spectral patterns really represent vegetation, temperature, moisture or another target.
87. Mixed pixels are a scale problem
One pixel may contain road, tree and water. The measured signal combines them. Fine-scale boundaries are blurred by spatial resolution.
88. Spatial uncertainty propagates
Location error, classification error, interpolation error and model error combine. Final maps should not imply sharper boundaries than the evidence supports.
89. Cartographic choices influence perception
Colour breaks, classification bins, symbol sizes and map projection can make patterns appear stronger or weaker. Visualisation should support evidence, not manufacture drama.
90. Choropleth maps require normalised variables
Mapping raw population counts by area often produces misleading darkness in populous regions. Rates or densities may be more appropriate depending on the question.
91. Colour scales need meaningful ordering
Sequential scales suit low-to-high quantities. Diverging scales suit deviations around a meaningful centre. Category colours should not imply numeric order accidentally.
92. Legends are part of scientific communication
Units, dates, source, classification method and missing-data symbols should be clear enough for readers to interpret the map without guessing.
93. Maps should show missingness
Blank areas can mean no data, zero events, outside study area or excluded region. These meanings must be distinguished.
94. Spatial analysis and fieldwork are different
Fieldwork collects evidence in real environments. Spatial analysis studies the location structure of that evidence. One can use field data, remote data or administrative data.
95. Spatial analysis and sampling are different
Sampling determines where observations are collected. Spatial analysis asks what the resulting arrangement means, including dependence and coverage gaps.
96. Spatial analysis and scale are inseparable
Patterns can reverse or disappear when grain and extent change. Scale should be treated as part of the scientific hypothesis.
97. Spatial analysis and external validity are connected
A result found in one region may depend on local spatial structure. Transporting it elsewhere requires comparing neighbourhood, infrastructure, climate and population context.
98. Spatial analysis and counterfactual reasoning are connected
Policy interventions at one location often need spatial controls or synthetic comparison regions. Spillovers across borders complicate the counterfactual.
99. Spatial analysis and boundary conditions are connected
A model calibrated in one environmental range may fail in another region. Spatial maps can reveal where the system enters a new regime.
100. AI can analyse large spatial datasets
Satellite imagery, street-level images, sensor networks and geotagged records create volumes too large for manual inspection. AI can classify, segment and detect spatial patterns.
101. AI inherits spatial sampling bias
If training images come mostly from cities, well-mapped countries or popular roads, performance may be weaker elsewhere.
102. Geolocation models can exploit shortcuts
Watermarks, camera artefacts, road signs or dataset-specific cues may drive apparent geographic skill. Construct validity matters.
103. AI can overstate location precision
A model may return one exact coordinate even when evidence supports only a broad region. Spatial uncertainty should be represented honestly.
104. AI can help learners practise spatial reasoning
Useful prompts include: “Create two maps where raw counts and rates tell different stories,” “Give me a clustering pattern caused by sampling effort,” “Show why network distance differs from straight-line distance,” and “Ask me to choose a spatial scale for this mechanism.”
105. Parents can teach spatial reasoning through familiar routes
Why can a nearby place take longer to reach? Why are some streets hotter? Why do certain plants grow along one side of a path? Everyday geography makes distance, exposure and neighbourhood concrete.
106. Small-group tuition can use map diagnostics
Give learners a map with counts, population and sampling locations. Ask what they can conclude before calculating anything. Strong students look for denominators, effort, spatial coverage and alternative explanations.
107. Examination questions often hide spatial dependence
Samples taken close together, along a transect or upstream/downstream are not interchangeable. Diagram location can be part of the scientific variable.
108. Independent-attempt task 1: choose the right distance
For a road-access problem, river-pollution problem and airborne-pollution problem, choose straight-line, network, flow or wind-informed distance and explain why.
109. Independent-attempt task 2: expose the denominator
Create a map where Area A has more cases but Area B has higher rate. Explain why raw count and risk answer different questions.
110. Independent-attempt task 3: change the zones
Group the same points into large and small regions. Compare the resulting averages. Explain how the modifiable areal unit problem changes interpretation.
111. Independent-attempt task 4: map uncertainty
Place five measured points on a blank grid. Predict where interpolation will be most and least certain. Then explain why a smooth surface should include uncertainty.
112. Diagnostic error: map equals mechanism
A cluster is observed and a cause is declared. Repair by proposing competing mechanisms and checking exposure, sampling and denominators.
113. Diagnostic error: proximity equals causality
Two things are nearby, so one is said to cause the other. Repair by asking whether the mechanism travels through the relevant spatial pathway.
114. Diagnostic error: raw counts mapped without denominator
Population density drives apparent hotspot. Repair by calculating a relevant rate or exposure-adjusted measure.
115. Diagnostic error: straight-line distance everywhere
Mechanism follows road, river or network. Repair by choosing a meaningful distance metric.
116. Diagnostic error: independence assumed
Dense nearby samples are treated like independent replicates. Repair by modelling spatial dependence or changing sampling design.
117. Diagnostic error: boundary treated as natural
Administrative zones are interpreted as if the phenomenon stops at the border. Repair by distinguishing data boundaries from process boundaries.
118. Diagnostic error: smooth map hides sparse data
Interpolation creates confident-looking colour in areas far from measurements. Repair by showing sampling locations and prediction uncertainty.
119. Diagnostic error: aggregation used to infer individuals
Area-level correlation becomes a claim about every person. Repair by matching inference level to data level.
120. The independence test
Give a learner a new map without instructions. Can they identify spatial units, denominators, sampling effort, distance mechanism, dependence and uncertainty before jumping to a story? That is transferable spatial reasoning.
121. The evidence boundary
Spatial patterns are descriptive until mechanism, comparison and sampling support stronger inference. Maps reveal where to ask questions; they do not automatically answer why a pattern exists.
122. A compact spatial-analysis checklist
- What spatial question is being asked?
- What is the unit: point, line, area, raster cell or network node?
- What coordinate system is used?
- What spatial accuracy and resolution apply?
- What grain and extent define the study?
- What distance metric matches the mechanism?
- What counts as a neighbour?
- Is spatial autocorrelation expected?
- Are observations truly independent?
- What denominator or exposure is needed?
- Could sampling effort create apparent clusters?
- Could zone boundaries change the result?
- Is interpolation supported by nearby data?
- How is spatial uncertainty shown?
- Does the map support description, prediction or causality?
- What mechanism could generate the observed pattern?
123. Frequently asked questions
What is scientific spatial analysis?
Scientific spatial analysis is the quantitative study of how location, distance, neighbourhood, direction and spatial dependence structure observations and scientific relationships.
Why can’t nearby observations always be treated as independent?
Because neighbouring locations often share environment, exposure, infrastructure or other conditions, making their values correlated.
What is spatial autocorrelation?
It is the tendency for values at nearby locations to be more similar or more different than expected under a chosen spatial reference model.
What is the modifiable areal unit problem?
It is the fact that statistical relationships can change when the same underlying data are aggregated into different spatial scales or zone boundaries.
How does spatial reasoning help PSLE Science?
It strengthens interpretation of diagrams, transects, sampling locations, gradients and fair comparisons across places.
How does it deepen in Secondary Science?
Students can analyse autocorrelation, interpolation, spatial confounding, networks, scale effects, geostatistics and spatiotemporal patterns more formally.
124. Continue the Science Education Systems series
Conclusion: Where something happens can be part of why it happens
Maya sees the map.
Jia Jun asks what distance means.
Hana checks whether nearby observations are truly independent.
Ethan asks which mechanism could have drawn the pattern onto space.
Science needs all four.
Define the spatial unit.
choose the scale.
respect the coordinate system.
measure the right distance.
model neighbourhood and dependence.
show the denominator.
map uncertainty.
Then let location become evidence without allowing the map to become the explanation.
