Geospatial analysis

A geospatial analysis of population health keeps geography as a variable instead of averaging it out, so it can estimate where a problem is concentrated with more geographical precision than an aggregate rate. It applies when you have georeferenced health data, such as incidence, mortality, risk factor or resource data, and the question is where to act, not just how much happens overall. When an indicator is aggregated at regional level, the internal differences disappear into the average: a region with a reasonable mean rate can hide areas of high concentration next to areas with hardly any cases. Spatial analysis locates those concentrations, assesses whether access to services is equitable across areas and provides the geographical basis for directing resources where the impact is greatest, instead of spreading them evenly over a map that is not even.

For a geospatial health analysis, we need to be able to place each record reliably in space, using coordinates or an area such as a census tract, municipality or health district. Calculating rates also requires population denominators and boundary data for the areas being analysed. The data do not need to arrive clean: we check the geocoding and the stability of the rates, and define the spatial structure, that is, which areas are treated as neighbours by contiguity, distance or another justified rule. Scale affects the answer: a pattern visible at census-tract level may disappear when data are aggregated by municipality, while very small areas increase instability and the risk of re-identification. For sensitive data, we can work with aggregated areas and explain the resulting loss of resolution. A spatial analysis cannot be more precise than the location data on which it is built.

In areas with small populations, spatial smoothing is applied where appropriate, using Bayesian approaches or other rate stabilisation methods, so that the chance occurrence of one or two cases does not produce a map that seems to flag a problem where there is only noise. It applies whenever rates are mapped over small geographical units, where the small denominator makes the rate highly unstable. One or two deaths in an area with few inhabitants send the crude rate soaring and produce striking but misleading maps: the intense colour reflects the small size of the denominator, not a real epidemiological signal. Stabilisation borrows information from neighbouring areas to produce a more stable estimate, so that the map better approximates the underlying pattern rather than the artefact of the denominator. That way, what is flagged as a risk area is a plausible concentration, not a chance result of working with small populations.

A cluster is assessed with spatial autocorrelation statistics and their tests, not by eye on the map. We first assess global spatial dependence, for example with Moran's I: whether neighbouring areas show similar values more often than expected under the null hypothesis. We then use local statistics such as LISA or Getis-Ord to locate potential hot and cold spots. Local significance is not read in isolation: testing many areas increases the chance of apparent clusters, so we take multiple comparisons and the stability of the rates into account. The result also depends on how areas and their neighbourhood relationships were defined. The report therefore distinguishes a pattern that merely catches the eye from a concentration supported by the analysis, with its assumptions and limitations made clear.

Straight-line distance does not reflect the real barrier to access; accessibility analysis uses isochrones to estimate travel time over the available network, whether roads, public transport or another mode relevant to the context. It applies when you want to assess equity of access to health services or decide where coverage is lacking. Two points at the same distance in kilometres can have very different travel times depending on the roads, the terrain, transport or rurality, so straight-line distance can show a population as covered when in practice it is far away. Isochrones calculate which area can be reached from each facility within a time threshold set on clinical, operational or planning criteria, and identify the areas left without coverage. This turns an equity question into concrete geographical evidence: which population lies beyond a reasonable travel time, and where a coverage gap justifies an intervention or a new resource.

Georeferenced data can violate the independence assumption of classical regression, since nearby areas resemble each other, so when that dependence is present we use spatial regression models, which address it instead of ignoring it. It applies to analysing which geographical, environmental or socioeconomic factors are associated with the distribution of a disease. A conventional regression assumes that each observation is independent of the others. In space, neighbouring areas tend to resemble each other, and applying the classical model without checking can underestimate uncertainty and produce false precision. That is why the first step is to check for relevant spatial autocorrelation and, if it is present, spatial regression and CAR models incorporate the neighbourhood structure, so that the uncertainty around the associated factors is estimated more appropriately. The scope needs to be clear: these models identify spatial associations and factors associated with a pattern, not causal relationships; establishing causality requires a different design.

Geospatial analysis of population health describes and explains the geographical distribution of a health problem: where it is concentrated, what spatial pattern it follows and which factors are associated with it. Ex post evaluation of interventions estimates the attributable effect of a specific intervention or policy that has already been implemented, against a counterfactual. They do not answer the same question: geospatial analysis answers "where is it and what is it associated with?", from a descriptive, associative or explanatory perspective, but not a causal one in itself; ex post evaluation answers "what effect did this intervention have?", in causal terms. The two can be chained, since a geospatial analysis can reveal where to target an intervention that is later evaluated with an ex post design, but one maps and explains an associated pattern and the other attributes an effect to a specific action. If the question is where the problem is and how it is distributed, it is geospatial analysis; if it is what effect an intervention already rolled out has produced, it is ex post evaluation.

We deliver the thematic maps and the interpreted results of the spatial analysis: maps of smoothed rates and relative risk, the identification of hot spots and cold spots with their statistical tests, accessibility analyses where relevant and, if the question calls for it, spatial regression models with the factors associated with the distribution. It is for those who need geographical evidence to plan: where to target a programme, where to site a resource, where access is not equitable. The deliverable separates a statistically supported pattern from a visual reading, so that decisions rest on the signal and not on the artefact. The client provides the records and whatever geographical or georeferenced information is available, and makes the planning decisions; the service estimates, models and maps, and provides the geographical interpretation with its limits made explicit.