Mobile Clinic Siting with Raster Service Areas and Census Tracts
A county health department must rank 18 census tracts for a mobile clinic. Analysts have a 100 m population raster, a 30 m land-cover raster, a 10 m DEM, and five candidate clinic points. They compute 15-minute drive-time service areas in Network Analyst, convert them to raster, and overlay population, slope, and land cover. Cells with slope above 8° or land cover other than developed are excluded. Results show candidate A serves 4,200 residents, B 3,900, and C 5,100. Candidate C's service area crosses a river with one bridge; one high-population tract has its centroid inside C's area, but most of its residents live outside. Two tracts are split by the service-area edge. The department wants a defensible ranking in two weeks and has one analyst. Staff propose accepting the raster overlay as final because it is 'objective and repeatable.'
1. Identify the main methodological problems in the proposed ranking, including MAUP, edge effects, and network assumptions.
Read the worked answer
Main problems:
- Centroid fallacy and MAUP: Assigning a tract to a candidate because its centroid falls inside the service area can invert the result. A tract can be large, irregularly shaped, and mostly outside the 15-minute area. Candidate C's high-population tract is exactly this case. The modifiable areal unit problem also means that aggregating to tracts after raster overlay can change the ranking depending on boundaries and zone size.
- Raster conversion of a network service area: Converting a drive-time polygon to raster imposes an arbitrary cell grid and can include or exclude edge cells. The 100 m population raster may itself be allocated from block centroids, so the overlay inherits a population-distribution assumption.
- Edge effects: Tracts split by the service-area boundary are counted all-or-nothing. Population near the boundary may be within 15 minutes in one direction and outside in another; the analysis should allocate population by area or, better, by a finer dasymetric surface.
- Network assumptions: The one-bridge river crossing is a bottleneck. Drive times may assume free-flow speeds, ignore turn restrictions, or understate congestion and bridge delays. A 15-minute area that depends on a single bridge is less reliable than one with multiple routes.
- Threshold sensitivity: The 8° slope exclusion and developed-land-cover filter are binary cutoffs. Small changes near the threshold can change served population.
Conclusion: The overlay is repeatable but not automatically defensible. It answers a raster question, not necessarily the clinic-access question the department is asking.
2. Recommend a validation and sensitivity approach the analyst can complete in two weeks. What should be compared, and what should be reported?
Read the worked answer
Recommended approach:
- Population allocation: Replace tract-centroid assignment with population-weighted areal interpolation or a dasymetric method using block or parcel data. Allocate split-tract population in proportion to where people actually live, not to tract area alone. If block data are unavailable, at least clip the population raster to the service area and compare with the tract-level result.
- Network validation: Check the drive-time network for bridge penalties, turn restrictions, and peak-hour speeds. Re-run C with a 20% bridge delay and with a 10-minute and 20-minute threshold. If C's rank collapses under plausible delay, report it as sensitive.
- Sensitivity tests: Vary the slope cutoff (6°, 8°, 10°), land-cover reclassification, and service-area threshold. Produce a rank table showing whether A, B, and C keep their order. Use a simple decision rule: report a candidate as robust only if it stays in the top two across the tested assumptions.
- Field or external check: Validate the top two candidates with drive checks during peak and off-peak periods, and compare against existing clinic use or transit access.
- Report: Present a range, not a single number: for example, 'C serves 4,600-5,300 under tested assumptions, but depends on one bridge; A is stable at 4,000-4,400.' Include maps of the split tracts and the bridge bottleneck. State that the raster overlay was a screening step, not the final evidence.
3. How should the analyst present uncertainty to decision-makers without paralyzing the choice?
Read the worked answer
Presentation strategy: Separate the decision into a screening result, a robustness check, and a recommendation. Start with the finding that all three candidates are plausible and that the initial ranking is not stable enough to justify a single winner. Show a simple table with candidate, base population served, tested range, and key risk. For example, C has the highest base number but the widest range because of the single bridge and split tract; A is lower but stable.
Use plain language: 'C looks strongest on the map, but its advantage depends on one river crossing and on counting a tract whose residents mostly live outside the service area. A is slightly lower but remains in the top group under every test.' Avoid false precision such as 5,100 versus 4,200 residents when the allocation method cannot support that difference.
Recommend a two-stage decision: choose A and C for further field verification, or choose A if the department prioritizes reliability and C if it can accept bridge risk and verify access. Include a map showing the split tract and the bridge, plus a short methods note. This gives decision-makers a defensible choice without requiring them to understand every GIS step.