Understanding the Geography of Healthcare Access in South Africa
Evaluating Pharmacy Access in Gauteng and KwaZulu-Natal Under the National Health Insurance Act
MUSA Smart Cities Practicum | Spring 2026
Team: Tess Anh Thu Vu, Joey Cahill, Jillian Kalman, Alex Stauffer
Instructors: Michael Fichman & Matthew Harris
Client: Distributed AI Research Institute (DAIR)
Raesetje Sefala (Researcher) | Nyalleng Moorosi (Research Fellow)
Background
The National Health Insurance Act and Healthcare Access
The catalyst of this project starts with a new healthcare policy passed by the South African parliament that intends to expand public healthcare services, fixing what is currently an imbalanced system. Right now, public and private medicine funding is split 50/50 by the government. However, the public health system serves over 80% of the population, whereas private pharmacies serve about 20% (Mkhize N.I., 2026). The inefficiency of this system on the supply and demand of medicines has led to a number of social repercussions like lower than average vaccination rates and higher than average rates of communicable diseases (Aweeka 2025). This new bill seeks to bridge the gap between private and public disparities by making public health insurance acceptable in private pharmacies where supply is at a surplus.
The National Health Insurance (NHI) Act of 2024 represents a landmark transformation in South African healthcare policy, establishing a public fund to subsidize the provision of care and medicines across the nation. This legislation aims to address the profound inequalities in healthcare access that persist nearly three decades after the end of apartheid, creating a universal healthcare system that theoretically enables all South Africans to receive quality healthcare services regardless of their socioeconomic status (Ramaphosa, 2024).
The passage of the NHI Act has prompted critical questions about implementation readiness, particularly regarding the spatial distribution of pharmaceutical services. While the NHI creates funding mechanisms for healthcare provision, its success depends fundamentally on whether populations can physically access healthcare infrastructure, including pharmacies where they can obtain prescribed medications. This spatial dimension of healthcare access is not uniformly distributed across South Africa and reflects historical patterns of development, urbanization, and the enduring legacy of apartheid-era spatial planning.
The Legacy of Apartheid Geography
The precursor of this pharmasocial imbalance is the persistent legacy of Apartheid in South Africa that was officially stripped of use in the 1990s. During this relatively short period of time, political power was consolidated to the white minority settlers who were motivated by racism and hierarchy in how they ran the country. Native Black African populations were stripped of voting rights, land ownership rights, and the ability to hold office. While segregation is no longer written in the South African constitution, the forced restructuring and separation that Apartheid instilled has continued to influence infrastructural development, the movement of people, and ideological beliefs to this day. The activists and authors behind this new healthcare policy are interested in expanding pharmaceutical access to the average South African to improve both individual health outcomes and societal public health outcomes.
With this, understanding healthcare access in contemporary South Africa requires grappling with the historical geography of apartheid. Crucially, the Group Areas Act of 1950 and subsequent legislation systematically segregated residential areas by race, concentrating Black African and Coloured populations in townships and homelands while reserving well-serviced urban areas for white residents. As an intended consequence, this spatial engineering created patterns of uneven development that persist today as former township areas often lack the commercial infrastructure, including pharmacies, that developed in historically white areas (Johnson-Castle, 2014).
The spatial mismatch between where people live and where services are located represents a fundamental challenge for NHI’s smooth implementation. Populations in former township areas and informal settlements may need to travel considerable distances to access pharmacy services that creates barriers which are particularly acute for the elderly, disabled, and those without private transportation. The informal economy possesses a substantial portion of South Africa’s workforce, further compounds and complicates access patterns, as workers may have limited time flexibility for healthcare visits (McLaren, Ardington & Leibbrandt, 2014).
Research Motivation and Client Context
The Distributed AI Research Institute’s (DAIR’s) interest stems from their broader research agenda examining how data-driven technologies can be deployed to understand and address social inequalities rooted in social ethics. The organization has previously conducted influential research on spatial apartheid, developing datasets and analytical frameworks for studying the persistent spatial inequalities in South African settlements (Sefala et al., 2021).
This project builds on DAIR’s established expertise while extending their analytical framework to healthcare infrastructure. By creating a well-documented, replicable methodology for two pilot provinces, the research provides a foundation that DAIR, should they desire, could expand and generalize to evaluate pharmacy access across all nine South African provinces.
Research Questions
The historical significance of South African society is what will be used to inform the analysis and the features chosen for the final product. The problem in its purest form is an inequitable access to essential medicines asserted by unorthodox geospatial design and privatized healthcare. This project measures and quantifies accessibility to pharmacies to answer the questions:
- What is the spatial distribution of pharmacies in Gauteng and KwaZulu-Natal?
- What populations have adequate access to pharmacy services, and what populations face access barriers?
- To what extent do patterns of pharmacy access correlate with historical geographies of apartheid?
The following report will detail how access to pharmacies are defined, the methodology behind the metrics, the instruments used to deploy the application, and the data that was collected to illustrate the findings.
Study Area
Gauteng Province
Gauteng, South Africa’s smallest province by area but most populous contains approximately 16 million residents representing roughly 26% of the national population. The province encompasses the cities of Johannesburg, Pretoria (Tshwane), and the East Rand (Ekurhuleni), which forms the economic beating heart of South Africa and the broader Southern African region.
The province’s spatial structure reflects its mining and industrial heritage as Johannesburg developed around gold mining operations with segregated residential areas creating the townships of Soweto, Alexandra, and numerous smaller settlements. Pretoria, which was established as an administrative center, displays similar patterns of racial residential segregation. The East Rand industrial corridor connects these major centers while incorporating additional township complexes.
Gauteng’s economic centrality creates complex healthcare access dynamics because while the province has relatively high pharmacy density overall, the distribution is uneven. Historically white suburbs contain concentrations of commercial pharmacy services, while township areas and informal settlements often have limited pharmacy presence, requiring residents to travel to commercial centers for pharmaceutical services.
KwaZulu-Natal Province
KwaZulu-Natal, located on South Africa’s eastern coast has approximately 12 million residents, making it the second most populous province. In juxtaposition to Gauteng, this province displays greater geographic diversity, encompassing coastal urban areas around Durban (eThekwini), extensive rural areas, and former homeland territories that were designated as independent states during apartheid (Lawal, 2025).
The city of Durban serves as the provincial economic center and contains South Africa’s largest port. However, unlike Gauteng’s concentrated urbanization, KwaZulu-Natal has substantial rural populations, including communities in former KwaZulu homeland areas where they face healthcare access challenges that are distinct from urban townships, including greater distances to healthcare infrastructure and limited transportation options.
The contrast between Gauteng and KwaZulu-Natal provides methodological advantages for this research. By developing approaches that work across both a highly urbanized province and one with significant rural populations, the methodology of measuring access to both of them demonstrates generalized applicability, coupled with human decision-making, to the diverse geographic contexts present across South Africa.
Administrative Geography
South Africa’s administrative hierarchy structures the geographic analysis. The nine provinces divide into either metropolitan municipalities (for major urban areas) or district municipalities (for other areas). District municipalities further subdivide into local municipalities. For census purposes these administrative units contain wards that serve as the smallest geographic unit for which 2023 Census data are available.
Wards are population-scaled units, meaning that rural wards cover much larger geographic areas than urban wards while containing similar population counts. This creates analytical challenges in that rural wards may encompass multiple dispersed settlements with varying access to pharmacy services, while urban wards represent more spatially concentrated populations.
Below the ward level, Small Area Layers (SALs) provide finer geographic resolution but are only available from the 2011 Census. SALs contain approximately 500 households each, providing a scale more appropriate and granular for local accessibility analysis but requiring methodological approaches to estimate current populations from decade-old data.
Literature Review
Spatial Accessibility in Healthcare Research
Clinicials and researchers have been measuring healthcare access for decades. Typically, researchers find that rural, low-income areas are the most vulnerable to being under-served in access to healthcare. This is because remoteness is working against them. Most research indicates a 20-30 minute commute for healthcare is considered reasonable, which is hard to acheive in rural areas. In countries where the cost of healthcare is increasing, this gap of access continues to increase, and is exaserbated by spatial context (Murphy and Rodis 2025). Healthcare accessibility is oftem simplified to ‘living near health services’, but this is been proven over time to be insufficient, especially when using Euclidean distance. Variables like supply and infastructure are overlooked in this method (McClaren et al. 2014). Overtime, researchers have developed a method called Enhanced Two-Step Floating Catchment Area method. This is a gravity-based model that measures acces through time-traveled, supply capacity, and the role of demand. This will be explored in the methodology. Accessibility to pharmacies is crucial for public health: it promotes medication adhernece, increases health education/awareness, and normalizes vaccination and testing behaviors (L Berenbrok 2022).
South African Spatial Research Context
Apartheid is rightfully at the center of spatial research in South Africa and the motiviation behind this research. Although near 40 years in the past, inequalities still persist on many fronts (Lowal 2025). With the work of DAIR, they seek to understand how apartheid has evolved over time into the mainstream culture and society of South Africa today. Because of the manufactured nature of apartheid, South Africa requires different classifications than just urban/subruban/rural that are often found around the world in some iteration. These concepts are present in SA, but are broken down into a more fine-grained subdivision based on their historical context that is not found anywhere else. Neighborhood types such as townships and informal residential have been introduced to this research using thousands of high resolution satellite imagery and integrated into detailed datasets by DAIR. They also collected spatial data on economic class (wealthy/non-wealthy), land use (industrial/commercial/residential), and building density. DAIR plans to continue this research to understand the standard of living and demographic make-up of South Africa as time goes forward (R Sefala 2024).
Pharmacy Regulation and Trade in South Africa
Pharmacies are regulated by the South African Pharmacy Council (SAPC), constituted of a 25-member collective of pharmaceutical professionals. SAPC is an independent, self-funded, statutory body mandated in terms of the Pharmacy Act, 53 of 1974 to regulate the pharmacy profession in the country with powers to register pharmacy professionals and pharmacies, control of pharmaceutical education, and ensuring good pharmacy practice. They maintain multiple committiees and allow you to search pharmacies or pharmacists that are recognized and approved by their regulations, including histories of inspections (pharmcouncil.co.za). Due to the gaps in the regulated pharmaceutical market, an illicit drug trade has emerged. Those who cannot access or are denied access to regulated medicine turn to illegal vendors and street markets to access things like antibiotics, contraceptives, and pain relievers. The consequences of widespread unregulated medicine include increased drug abuse/misuse, increased risk of adverse advents, and decreased rates of dosage adherance (Mutandiro 2025).
Methodology
Data Sources and Collection Strategy
The project assembles pharmacy location data from multiple sources, reflecting the fragmented nature of pharmacy information in South Africa, and unfortunately, no unified national database of pharmacies exists, requiring data integration across heterogenous sources.
Raw Data Inventory
Geographic Boundary Data:
| Dataset | Source | Date | Columns | Records |
|---|---|---|---|---|
| South Africa Wards | South African Census | 2023 | 12 | 4,468 |
| South Africa Small Area Layers (SAL) | South African Census | 2011 | 60 | 39,177 |
Reference Lookup Tables:
| Dataset | Description | Columns | Records |
|---|---|---|---|
CITY_PROVINCE_LOOKUP |
Mapping table linking cities/suburbs to provinces and municipality hierarchies for Gauteng and KwaZulu-Natal | 6 | 290 |
The CITY_PROVINCE_LOOKUP table was created to help with handling province assignment for pharmacy records lacking explicit province information, which contains mappings from city/suburb names to their corresponding PROVINCE, LOCAL_MUNICIPALITY, DISTRICT_MUNICIPALITY, METROPOLITAN_MUNICIPALITY, and CITY values. This lookup supports the multi-stage address matching process described in the data processing pipeline.
Network Data:
| Dataset | Source | Resolution/Format | Description |
|---|---|---|---|
| OSMnx Pedestrian Network | OpenStreetMap via OSMnx | Vector network (nodes/edges) | Walkable street network for pedestrian accessibility calculations |
| OSMnx Drive Network | OpenStreetMap via OSMnx | Vector network (nodes/edges) | Road network for driving accessibility calculations |
- The OSMnx pedestrian and drive networks are extracted from OpenStreetMap for both provinces, enabling network-based distance calculations to pharmacies as specified by the client.
Private Health Insurance Provider Networks:
Private health insurance providers publish lists of pharmacies in their preferred provider networks. These lists, while not comprehensive of all pharmacies, provide extensive coverage of commercially operating pharmacies:
Total Raw Private Pharmacy Records: 9,114
Government Hospital Data:
| Dataset | Source | Date | Records |
|---|---|---|---|
| Hospital Pharmacies | DAIR Institute (pre-cleaned) | 2026 | 272 |
Public hospitals contain pharmacies that dispense medications to patients, and location data from government sources confirmed by the DAIR Institute provides coordinates for these public pharmacy access points, though not detailed address information—private pharmacies, on the other hand, lack coordinates, but have detailed address information.
SAPC Registry (Scraped):
| Dataset | Source | Date | Records |
|---|---|---|---|
| SAPC Active Pharmacies — Gauteng | South African Pharmacy Council Registry | February 2026 | 1,961 |
| SAPC Active Pharmacies — KwaZulu-Natal | South African Pharmacy Council Registry | February 2026 | 894 |
The South African Pharmacy Council maintains a registry of all registered pharmacies at pharmcouncil.co.za. Unlike the insurance network lists, which only include pharmacies participating in specific provider agreements, the SAPC registry represents the full population of legally registered pharmacies. Records were collected programmatically via a trie-based web scraper that submitted search queries for every 3-character alphabetical prefix within each province and fetched detail pages for Active pharmacies to retrieve full address, owner, responsible pharmacist, licence number, and inspection history. Inactive and Erased records are retained with table-level data only. The process is documented in 01_sapc_scraper.ipynb.
SAPC Source Limitations: The registry reflects registration status at the time of query, so pharmacies registered when the insurance network data was collected may have subsequently been deregistered, or vice versa. Detail pages are only accessible within the same browser session that executed the search, meaning the scraper must fetch detail pages immediately during the search loop rather than in a separate pass. Pharmacies with Inactive or Erased status are retained in the raw output with table-level data only and are excluded during the master file build. Additionally, the registry search has a minimum of 3-character prefix enforcement, meaning pharmacies with very short names may require multiple prefix expansions to be captured.
Data Extraction Methods:
- PDF Conversion: GEMS, Wooltru, and Vitality pharmacy lists were converted from PDF to tabular format using Adobe Acrobat text extraction.
- HTML Table Copy: Momentum pharmacy data was copied directly from the provider website where it was already displayed in table format.
- CSV Download: SAMWUMED provided direct CSV download functionality from their website.
- Web Scrape: SAPC records were accessed via a python scraper which queried a search function hosted on their website.
Final Transformed Data
| Dataset | Description | Records |
|---|---|---|
PHARMACIES_COMBINED |
Unified pharmacy database for Gauteng and KwaZulu-Natal after cleaning, standardization, and province filtering | 5,436 |
geocoded_with_sapc.csv |
Full geocoded pharmacy list with SAPC regulatory fields appended where a confident fuzzy match was found | ~5,400 |
sapc_needs_geocoding.csv |
SAPC records with no confident match to the geocoded list, routed to a standalone geocoding pass | ~1,200 |
PHARMACIES_MASTER_FINAL.csv |
Authoritative master pharmacy file combining SAPC-registered and Places-only records after spatial validation, deduplication, and column consolidation. This is the input to all accessibility analysis. | 2,152 |
The reduction from 9,114 raw private pharmacy records plus 272 hospital records (9,386 total) to 5,436 combined records reflects the removal of pharmacies outside Gauteng and KwaZulu-Natal provinces, deduplication of pharmacies appearing across multiple insurance networks, and removal of records with insufficient address information for geocoding.
The PHARMACIES_MASTER_FINAL.csv file is the primary output of the Python geocoding pipeline and represents a different dataset from PHARMACIES_COMBINED. Where PHARMACIES_COMBINED consolidates the insurance network sources, the master file layers the SAPC registry on top of the geocoded insurance list, adds coordinate data from the Places API and hospital shapefile, removes records that fail spatial validation, and resolves duplicate records through both identifier-based and coordinate-proximity deduplication. SAPC rows take priority over Places-only rows in any collision on place_id. The master file includes a source field (sapc or places) and a spatial_check field (ok, no_coords, outside_province, outside_sa) that allow users to filter to any confidence tier. A sequential record_id field (PH00001 through PHnnnnn) provides a stable primary key for joining to accessibility outputs.
Data Source Limitations:
- Network Participation Bias: Pharmacies must participate in at least one of the insurance networks to appear in the dataset, meaning independent pharmacies serving primarily uninsured or cash-paying populations may be systematically underrepresented.
- Temporal Inconsistency: Source data spans from January 2024 to February 2026 with some sources undated, so pharmacies may have opened, closed, or relocated during this period, creating potential staleness.
- PDF Extraction Errors: Text extraction from PDF documents using Adobe Acrobat and Snowflake AI tools may introduce errors or formatting artifacts that could affect geocoding accuracy for example.
- Missing Identifiers: Momentum and Vitality sources lack
PRACTICE_NUMidentifiers, preventing deduplication for approximately 1,500 records. - Hospital Data Gaps: Hospital records contain coordinates but lack street addresses, limiting cross-validation capabilities and preventing address-based quality checks.
Validation Sources:
- SAPC Registration Database: The South African Pharmacy Council provides a searchable database for verifying pharmacy registration status (https://pharmcouncil.co.za/Pharmacies_Overview)
- Google Places API: Provides geocoding services and place verification for pharmacy locations
Data Processing Pipeline
Data processing occurs within Snowflake, a cloud data platform that serves as the central workspace for assembling, cleaning, and validating pharmacy data. The pipeline follows a structured workflow documented in SQL scripts.
Database Structure
The Snowflake database follows a three-schema architecture separating raw ingestion, intermediate processing, and final production tables:
MUSA_DAIR_DB
├── RAW (source data as ingested)
│ ├── PHARMACIES_GEMS_GAUTENG
│ ├── PHARMACIES_GEMS_KZN
│ ├── PHARMACIES_MOMENTUM_GAUTENG
│ ├── PHARMACIES_MOMENTUM_KZN
│ ├── PHARMACIES_WOOLTRU_GAUTENG
│ ├── PHARMACIES_WOOLTRU_KZN
│ ├── PHARMACIES_SAMWUMED
│ ├── PHARMACIES_VITALITY_WELLNESS
│ ├── PHARMACIES_HOSPITALS
│ ├── SAL_POLYGON_2011
│ ├── WARDS_POLYGON_2023
│ └── RAW_PHARMACIES_GEOJSON
├── INTERMEDIATE (cleaned and joined data)
│ ├── PHARMACIES_COMBINED
│ ├── CITY_PROVINCE_LOOKUP
│ └── WARD_JOINED_SAL (to be created)
└── MART (final production tables for Mapbox GL)
├── PHARMACIES_MASTER_FINAL_GEOJSON
├── CITY_PROVINCE_LOOKUP
├── PHARMACIES_MASTER_FINAL
├── KZN_BOUNDARIES
├── GAUTENG_BOUNDARIES
└── ALL_ACCESS_GEOJSONRAW Schema: Contains source data in its original structure with minimal transformation. Each pharmacy source maintains a separate table preserving the original column schema. Geographic boundary files (SAL polygons from 2011, Ward polygons from 2023) are stored here after initial import.
INTERMEDIATE Schema: Contains cleaned, standardized, and joined datasets used during analysis. The PHARMACIES_COMBINED table consolidates all pharmacy sources with harmonized schema. Lookup tables and spatial join outputs reside here:
WARD_JOINED_SAL: Areal-weighted intersection of 2023 Wards with 2011 SALs, containing weighted population allocations
MART Schema: Contains final, analysis-ready tables for the Mapbox GL web map application, so tables in this schema are formatted for efficient tile generation and frontend querying.
Processing Stages
Stage 1: Source Ingestion (01_combine_pharmacies.sql)
Raw pharmacy data from each source is ingested into Snowflake and combined into a unified table structure. Each source has distinct column schemas that were harmonized:
- GEMS data includes
PROVINCE,CITY,SUBURB,ADDRESS,PRACTICE_NAME,PRACTICE_NUM,PHONE - Momentum data includes
PRACTICE_NAME,ADDRESS,STREET,AREA,POSTAL,PHONE,FAX(noPRACTICE_NUM) - Wooltru data includes
PRACTICE_NUM,PRACTICE_NAME,ADDRESS,CITY,PHONE - SAMWUMED data includes
PRACTICE_NUM,PRACTICE_NAME,PHONE,EMAIL,ADDRESS,SUBURB,CITY - Vitality data includes
PROVINCE,PRACTICE_TYPE,PRACTICE_NAME,ADDRESS,SUBURB,CITY,PHONE(noPRACTICE_NUM)
The combination script standardizes naming conventions (INITCAP formatting), cleans phone numbers (removing non-numeric characters, standardizing to 10-digit format), and concatenates address components into complete geocodable addresses.
Stage 1 Limitations: Schema harmonization requires assumptions about field equivalence (e.g. treating AREA, SUBURB, and CITY as interchangeable geographic descriptors), address concatenation order affects geocoding performance and quality may vary by source, INITCAP formatting may incorrectly case acronyms or non-English names.
Stage 2: Province Population and Filtering (03_fill_gaps_pharmacies.sql)
Many source records lack explicit province information, so the CITY_PROVINCE_LOOKUP table (n = 290) containing known cities and suburbs within Gauteng and KwaZulu-Natal allows province inference from address components. The lookup table maps city/suburb names to their corresponding province and municipality hierarchy (local, district, or metropolitan municipality) where the script applies multiple matching strategies:
- Direct match on
CITYcolumn - Match on last comma-separated address segment
- Match on second-to-last address segment
- Contains match for city names within address strings
Records that cannot be matched to either Gauteng or KwaZulu-Natal are removed from the analysis scope. Province lookup updates (02_province_lookup_update.sql) address gaps identified during manual review of dropped records, ensuring that missing suburbs are added before final filtering.
Stage 2 Limitations: Province inference relies on the completeness of the lookup table, so missing suburb names result in false negatives (legitimate GP/KZN pharmacies incorrectly dropped). However, the script provides an opportunity for manual human checks, confirmation, and potential revision before dropping the observations from the table. The contains-match fallback may produce false positives for short city names appearing as substrings in unrelated addresses, and municipality hierarchy assignment depends on lookup table accuracy, which may not reflect recent boundary changes or naming conventions.
Stage 3: Deduplication (04_deduplicate_pharmacies.sql)
The same pharmacy may appear across multiple insurance provider networks with slightly varying information, so conservative deduplication uses PRACTICE_NUM as the authoritative identifier where available, aggregating within sources while preserving the most complete address information.
Records without PRACTICE_NUM (notably Momentum and Vitality sources) cannot be deduplicated using identifier-based methods and are retained for post-geocoding spatial deduplication based on coordinate proximity and fuzzy matching.
Stage 3 Limitations: Identifier-based deduplication assumes PRACTICE_NUM uniquely identifies a physical location within their respective pharmaceutical company, not across. However, a single PRACTICE_NUM may be associated with multiple branch locations, or different PRACTICE_NUMs may represent the same pharmacy under different registrations. Records lacking PRACTICE_NUM (~1,500) remain as potential duplicates until spatial deduplication, which inflates counts prior to final datasets. In addition, the longest address filter for selecting the record assumes address length correlates with geocoding quality.
Stage 4: Hospital Integration (05_add_hospitals.sql)
Public hospital records from government sources are appended to the pharmacy database. Hospital records include geographic coordinates but lack detailed address information, with PRACTICE_TYPE designated as “Hospital” and FUNDING as “Public” to distinguish from private pharmacies.
Stage 4 Limitations: Hospital coordinates may represent facility points within the campus and not pharmacy-specific locations within hospital campuses, it is also possible that not all hospitals contain pharmacies and some may only have dispensaries with limited hours or medication availability, and the dataset also may not capture clinic-based pharmacies or community health centers that provide pharmaceutical services.
Population Data and Downscaling: Step-Down Methodology and Daysymetric Mapping
This approach estimates small-area population counts for 2023 (SAL-level) using a combination of
2011 SAL population data, 2023 ward-level projections, and spatial weighting.
Using the Step Down Projection Method: It leverages population growth patterns and land weights to distribute ward-level counts to finer spatial units.
SAL Deduplication (Preprocessing)
The ArcGIS Pro spatial join that assigned SAL polygons to 2020 ward boundaries produced multiple rows per SAL where join artifacts created identical copies. Of 39,177 input features, 37,610 EA_CODEs appear once, 748 appear twice, and 22 appear three or more times. Five diagnostic checks confirm all 770 duplicate groups carry identical ward assignment, area, geometry, and all attribute values. Deduplication via drop_duplicates(subset="EA_CODE", keep="first") produces 38,380 unique SAL records with no analytical information lost. This deduplicated shapefile (sal_w_ward_dedup.shp, EPSG:32735) serves as the geometric foundation for all downstream notebooks.
Data Preparation
2011 SAL census shapefile (ea_sal_kzn_gp.shp)
Already filtered to just Gauteng and KZN
2023 Ward shapefile and population (SA_Wards2020.dbf and census_ward_2023_with_pop.csv)
These were joined on ‘WardId’
| Column | Type | Count | Unique |
|---|---|---|---|
| Province | object | 1430 | 2 |
| Municipali | object | 1430 | 53 |
| CAT_B | object | 1430 | 53 |
| WardNo | int64 | 1430 | 135 |
| District | object | 1430 | 16 |
| DistrictCo | object | 1430 | 16 |
| Date | datetime64[ms] | 1430 | 2 |
| WardID | object | 1430 | 1430 |
| WardLabel | object | 1430 | 1430 |
| geometry | geometry | 1430 | 1430 |
| Total | object | 1430 | 1430 |
Spatial Joining and Tabulation
Using ArcGis:
Tabulate Intersection–> Input Zone: 2011 SAL geometries
Input Class: 2023 Ward Geometries
This table identifies the ward(s) that each SAL encompasses, the percentage of the area of the ward that the SAL takes up, and the area (m sq.).
| EA_CODE | WardID |AREA | Percentage|
|—————-|—————-|—————-|———-| | 50310001 | 52103001 | 6967088.20 | 99.99| | 50310001 |52106004 | 28.807487153359173 |0.0004| | 50310001 | 52106014 | 43.38680092706605 |0.0003|
This is joined back to the SAL layer by EA_CODE –> Summarize Table –> AREA==Maximum
| EA_CODE | WardID | AREA | Province | district |
|---|---|---|---|---|
| 50310001 | 52103001 | 6967088 | province | district |
| 58110038 | 52606020 | 48373210 | province | district |
| 76410132 | 74805033 | 15722398 | province | district |
| 53810017 | 52606020 | 5720 | province | district |
Limitations: SALs where the ward share is evenly split between one or more wards lose some spatial meaning when it gets paired with the highest share ward since it does not reflect where people actually live inside the SAL/ward.
Density Calculations for weighting
As indicated by DAIR, South Africa census has a history up undercounting populations.
There were 2,084 SALs with a null population, so to avoid perpetuating further underestimation, housing counts (‘houses2011’) were used as a population proxy, which is then multiplied by three (average house size per SA census website). If the SAL had null/0 population and a house count of 0, their final population remained at 0.
Density= 2011 population count/ SAL area (km sq.)
Areal-Weighted Dasymetric Mapping
We estimate ward-level 2011 counts by grouping at ward-level and summing SAL population counts. This is used to calculate the share of population the SAL contains within the ward ‘share2011’.
share2011= SAL 2011 Population / Ward 2011 Total
An earlier iteration of the model implemented dasymetric mapping weights that combined population share with log-density:Dasym weight= share2011 * density_log
The log of density was used to calculate the weight for several reasons. The spatial data was extremely skewed, with some SALs being the size of one apartment building and some being entire farming communities. The log was used to capture relative density to avoid extreme over estimation.
Final implementation: In the cleaned pipeline (tess_newpred_clean.ipynb), the dasymetric weight is set equal to share2011 alone, without log-density re-weighting. This change was made because both share2011 and log_density derive from sal2011_pop, and combining them would double-count the 2011 population signal. The ward share sums are validated to equal exactly 1.000000 for every ward (pycnophylactic constraint), ensuring mass preservation.
Before calculating the final estimate with the weight, it is grouped by ward then normalized by the sum of SAL weights. This is to capture SAL population relative to its own Ward (our coarsest unit for which we have real counts). It avoids unrealistic overestimating in urban pockets and undue undercounting in rural areas.
Finally, 2023 Ward population counts were used with the dasymetric weights to estimate 2023 SAL level population projectionsSAL 2023 estimate= daysm weight * Ward 2023 population
| Variable | Value |
|---|---|
| WardID | 59500022 |
| EA_CODE | 59913668 |
| sal2011_pop | 1306.0 |
| PR_NAME | KwaZulu-Natal |
| ward2011_sum | 32132.0 |
| ward2023_pop | 21793.913535 |
| sal2023_est | 1225.25738 |
| EA_GTYPE | Urban |
| EA_TYPE | Township |
| econ_status | Non_Wealthy |
| houses2011 | 10.0 |
| area_km2 | 0.001971 |
| sal_dense | 662594.713465 |
| log_density | 13.40392 |
| share2011 | 0.040645 |
| dasym_weight | 0.05622 |
| growth_rate | -0.005304 |
This SAL is a pocket of land roughly the size of one small to mid-size apartment building with 1306 people, giving it a density of 662,594 people/ kmsq.: Impossibly dense, yet here it is in black and white. If we used the absolute density in the dasymetric calculation, the 2023 estimate would be ~26,000 for one apartment building. But when using the log density, the estimate sits at a modest 1,225: reflecting realistic counts and the ward level decline in growth, as well.
Estimation Justification and Reinforcement
If the 2023 SAL estimates were properly dissolved into SAL zones, the difference between the 2023 Census ward counts and our SAL estimates should be extremely minimal; as demonstrated below. Our output for this equation is 0.2023 Ward Population Sum- 2023 SAL Estimate Sum
Total estimated 2023 population across both provinces: 27,523,308. This exactly matches the sum of ward-level 2022 Census totals, confirming mass preservation. Mean SAL population: 717; median: 667; max: 13,852. There are 1,270 SALs with zero estimated population (SALs that were vacant with no dwellings in 2011).
Geocoding and Coordinate Assignment
Address strings assembled in PHARMACIES_COMBINED require geocoding to assign geographic coordinates. The project uses the Google Places Text Search API as its primary geocoding engine, submitting structured query strings for each pharmacy and extracting the place_id, matched name, matched address, and latitude/longitude from the top result.
Query Construction
Each record’s query string is assembled by concatenating available address components in the order: pharmacy name, street address, city, province, and the fixed suffix “South Africa.” This format was chosen because the Text Search API performs best when the query resembles a natural language search rather than a raw address string, and including the pharmacy name substantially improves match precision in cases where multiple businesses share a building or street number. Records missing one or more components are handled gracefully, with the available fields concatenated and the missing ones omitted.
API Execution
Queries are submitted to the place/textsearch/json endpoint with a type=pharmacy filter to bias results toward pharmaceutical establishments. Each run is checkpoint-resumable: completed query strings are tracked in a checkpoint CSV and skipped on subsequent runs, allowing the extraction to be interrupted and restarted without re-querying completed records or incurring redundant API costs. A 0.2-second sleep between requests respects API rate limits. Progress is saved every 100 records. Each result is flagged with a needs_review boolean if the API returned a non-OK status, returned no place_id, or returned null coordinates.
Hospital Coordinate Fill-In
Pharmacies located within public hospital campuses frequently fail the Places Text Search because the hospital’s primary listing dominates the result. For these records a name-based match is attempted against the government hospital shapefile: pharmacy names and hospital names are both cleaned by lowercasing, removing punctuation, and stripping generic institutional terms (hospital, clinic, medical, centre), and a direct string match on the cleaned key transfers the hospital’s coordinates to any pharmacy with an otherwise-null location. This fills a targeted subset of missing coordinates without requiring additional API calls.
SAPC Integration and Coordinate Transfer
Following geocoding of the insurance-network list, SAPC-registered pharmacies are fuzzy-matched to the geocoded records (described in detail under Registration Validation). A significant portion of SAPC records that failed to match are found to correspond to Places-only rows that were geocoded successfully: the same pharmacy exists in both datasets but under slightly different name spellings. In these cases the Places-derived coordinates are transferred to the SAPC row and the redundant Places row is removed, preserving the richer SAPC regulatory metadata while retaining the geocoded position.
Spatial Validation
All geocoded coordinates undergo a two-tier spatial check. The outer check flags any coordinate falling outside a South Africa bounding box (approximately 22°S to 35°S, 16°E to 33.5°E) as outside_sa. The inner check verifies that each coordinate falls within a buffered bounding box for its expected province: Gauteng (25.1°S–26.8°S, 27.4°E–29.2°E) and KwaZulu-Natal (26.7°S–31.2°S, 28.7°E–33.0°E), each expanded by approximately 5 miles to avoid discarding genuine pharmacies near province boundaries. Records failing the inner check receive a outside_province flag. Records with null coordinates are flagged as no_coords. A 5-mile buffer was chosen as a balance between catching genuine boundary pharmacies and excluding clearly mismatched geocodes.
Re-Geocoding Passes
Records flagged as outside_province, outside_sa, or no_coords undergo up to two re-geocoding attempts. The first pass applies a simplified query using only name, city, and province — omitting the street address, which is the most frequent source of geocoding errors for South African addresses — and applies the spatial checks again to any returned result before accepting it. The second pass targets SAPC rows that emerged from the fuzzy match with no coordinates at all, using SAPC-specific field names. Both passes apply the spatial validation before updating the master record, ensuring that a re-geocoded result outside South Africa does not overwrite a previously null coordinate.
Geocoding Limitations:
- Address Format Sensitivity: South African addresses follow varied conventions including informal locality descriptors, postal codes used as location identifiers, and non-English place names. The Places Text Search interprets these inconsistently, and the top result may match to the correct suburb rather than the specific street address, resulting in centroid-level rather than building-level precision.
- Same-Name Province Mismatch: A common failure mode is a query returning a pharmacy in the correct chain but in the wrong province — for example, a Dis-Chem in Gauteng being returned for a KwaZulu-Natal query because the name match is stronger than the geographic signal. Spatial validation catches these cases, but the underlying record is left with null coordinates after re-geocoding unless the simplified query resolves it.
- API Budget Constraints: Google Places API billing limits the total number of queries feasible within the project budget. The checkpoint system and skip-on-resume pattern minimize redundant calls, but records that exhaust re-geocoding attempts without a valid in-province result remain in the master file with null coordinates and a
no_coordsspatial flag. - Coordinate Precision: The Text Search API returns a single representative point per place, which for pharmacies in shopping centers or medical complexes may represent the complex entrance rather than the pharmacy unit. This introduces sub-block positional uncertainty that is generally acceptable for ward-level accessibility analysis but would affect any fine-grained network routing.
- Hospital Name Matching Sensitivity: The hospital coordinate fill-in relies on cleaned name string equality. Pharmacies with names that diverge substantially from their hospital’s official name — for example, a private dispensary within a public hospital using a branded name — will not be captured by this method.
Spatial Accessibility Calculation
Access metrics quantify the relationship between pharmacy locations and population distribution. The project implements multiple accessibility measures to allow for methodological comparison and dual-metric diagnostics, where each metric serves as a check on the other’s failure modes. All spatial computations use EPSG:32735 (UTM zone 35S, meters). Pharmacy coordinates originate in WGS84 (EPSG:4326) and are projected for distance calculations.
Distance Metrics
Straight-Line Distance (Euclidean): The simplest approach calculates straight-line distance from population centroids to the nearest pharmacy using a scipy cKDTree built from projected pharmacy coordinates, providing a lower-bound baseline that ignores road networks entirely.
Network Distance (Walking): Walking network distance provides accessibility measures relevant for populations without vehicle access. The OSMnx “pedestrian” network type provides calculation of actual walking routes from SAL centroids to pharmacy locations, accounting for street network topology rather than assuming straight-line travel. This form of accessibility is particularly relevant for understanding access barriers for lower-income populations.
Network Distance (Driving): Network distance analysis calculates travel distance along the road network, which better represents actual travel requirements but assumes private vehicle access. Driving network data is extracted via OSMnx using the “drive” network type. Drive graphs are converted to undirected for Dijkstra routing, which slightly underestimates true drive distances in areas with one-way constraints, primarily affecting dense urban cores.
Enhanced Two-Step Floating Catchment Area (E2SFCA)
The project implements the Enhanced Two-Step Floating Catchment Area method with network-based routing and negative exponential distance decay, providing a supply-demand perspective that captures competition between pharmacies and populations rather than purely distance-based measures.
Distance decay function. Negative exponential decay with β = 0.0003:
f(d) = exp(-0.0003 × d)This produces weight 1.000 at 0 m, 0.741 at 1 km, 0.549 at 2 km, 0.407 at 3 km, 0.223 at 5 km, and 0.050 at 10 km. The parameter was selected to produce meaningful distance discrimination within the catchment range. An earlier version used β = 0.0001, which produced almost no decay (37% weight at 10 km) and was replaced.
Step 1: Supply ratio (\(R_{j}\)). For each pharmacy, single-source Dijkstra traversal outward on the road network up to the catchment cutoff identifies all reachable SAL centroids. The decay-weighted population of reachable centroids is summed, and the supply ratio is computed as:
Rj = 1000 / max(weighted_pop, MIN_WEIGHTED_POP)where MIN_WEIGHTED_POP = 50 prevents runaway ratios from pharmacies in sparsely populated catchments. The factor of 1000 is a scaling constant for readability. Without the population floor, a pharmacy with 3 people in its catchment produced \(R_{j}\) = 333, creating scores 5,000x the mean.
Step 2: Accessibility score (\(A_{i}\)). For each SAL, single-source Dijkstra traversal outward on the road network sums the decay-weighted \(R_{j}\) values from all reachable pharmacies:
Ai = Σ (Rj × f(d_ij))where the sum is over all pharmacies j reachable from SAL i within the catchment cutoff.
Province-specific catchment distances:
| Province | Walk cutoff | Drive cutoff | Rationale |
|---|---|---|---|
| Gauteng | 2 km | 5 km | Dense, urban |
| KwaZulu-Natal | 3 km | 10 km | Sparse, mixed rural-urban; ~7x larger than Gauteng |
The walk and drive networks for each province are run separately, producing four province-mode combinations.
Bug fixes applied during implementation: (1) The original code used .set_index("node") to build the population lookup, silently dropping populations when multiple SALs snapped to the same road network node; the fix uses .groupby("node").sum() to aggregate all populations at shared nodes. (2) The MIN_WEIGHTED_POP = 50 floor was added to cap maximum \(R_{j}\) at 20, preventing extreme outlier scores.
Snap Distance Analysis
When a SAL centroid snaps to a distant network node, the Dijkstra routing starts from a displaced origin, introducing systematic measurement error into both the 2SFCA and k-nearest distance results. To quantify this effect, the Euclidean distance between each SAL centroid and its nearest network node is computed for all four province-mode combinations.
| Province | Mode | Median (m) | Mean (m) | P95 (m) | Max (m) | Flagged >500m (%) |
|---|---|---|---|---|---|---|
| Gauteng | Walk | 57 | 83 | 223 | 25,702 | 232 (1.1%) |
| Gauteng | Drive | 67 | 101 | 290 | 26,962 | 398 (1.9%) |
| KZN | Walk | 83 | 226 | 985 | 9,951 | 2,149 (12.3%) |
| KZN | Drive | 102 | 302 | 1,318 | 8,435 | 2,884 (16.5%) |
Farm SALs in KZN have 59.9% flagged at the 500m threshold; Traditional SALs in KZN have 20.4%; Urban SALs in either province have less than 1%. This directly quantifies the differential quality of OpenStreetMap road network coverage across settlement types. SALs with snap distance exceeding 500m receive a binary flag in the final dataset, allowing downstream analysts to filter or weight results by data quality confidence.
K-Nearest Pharmacy Distance
The k-nearest distance metric captures absolute physical reachability rather than supply-demand competition. Distances from each SAL centroid to the 3 nearest pharmacies are computed using Euclidean, pedestrian network, and drive network methods.
k=1 captures basic reachability (can the community reach any pharmacy); k=2 captures redundancy (is there a fallback if the nearest pharmacy closes); k=3 captures choice (does a functioning local market exist rather than a single dependency point).
Network routing: Single-source Dijkstra from each pharmacy node individually, with a 50 km cutoff. For each SAL centroid (snapped to its nearest network node), the distances from all pharmacy sources are collected, sorted, and the 3 shortest are retained as k=1, k=2, k=3.
Circuity ratio: The ratio of network distance to Euclidean distance for k=1 is computed as a diagnostic. Values of 1.2–1.5 are typical for urban grids; values exceeding 3.0 suggest network data gaps or physical barriers (rivers, highways, rail lines). Gauteng walk median circuity is ~1.38, drive ~1.36, both consistent with international benchmarks.
NaN handling: 4,931 SALs (12.8%) have NaN walk distance at k=1 (no path within 50 km on the walk network); 5,001 SALs (13.0%) have NaN drive distance at k=1. These SALs are either genuinely isolated or in areas with sparse OSM coverage. NaN distances are treated as exceeding all thresholds in downstream analysis.
Distance Threshold Exceedance
Binary policy thresholds answer: what proportion of SALs (and people) cannot reach their k-th nearest pharmacy within a given distance? This provides a policy-legible metric that avoids decay functions and provincial quantile comparisons.
For each combination of k level (1, 2, 3), transport mode (walk, drive), and threshold distance, a binary flag indicates whether the SAL exceeds the threshold.
Thresholds applied:
| Mode | Thresholds (km) |
|---|---|
| Walk | 1, 2, 3, 5 |
| Drive | 5, 10, 15, 20 |
| Euclidean | 1, 3, 5, 10 |
Exceedance rates are computed as both SAL-count proportions and population-weighted proportions (using sal2023_est). Cross-tabulations are produced by province, settlement type (EA_GTYPE), and economic status (econ_status).
Policy reference thresholds (k=1):
| Province | Mode | Threshold | % SALs Exceeding | % Population Exceeding |
|---|---|---|---|---|
| Gauteng | Walk | 3 km | 16.5% | 15.8% |
| KZN | Walk | 3 km | 61.3% | 62.7% |
| Gauteng | Drive | 10 km | 1.9% | 1.2% |
| KZN | Drive | 10 km | 35.6% | 35.7% |
Combined Access Table and Access Typology
The three analytical streams (2SFCA scores, k-nearest distances, threshold flags) are merged into a single SAL-level dataset. EA_CODE alignment is verified across all source tables (38,380 intersection, 0 orphans).
Access typology suggestion: A six-category classification combining k=1 distance, k=3 redundancy, Ai score, and snap flag to produce a single policy-legible label per SAL:
| Category | k=1 distance | k=3 / options | Ai score | Snap flag | Meaning |
|---|---|---|---|---|---|
| Well-served | < 3 km | Gap < 5 km | Above median (nonzero) | No | Pharmacy nearby, alternatives exist, not overwhelmed |
| Overcrowded | < 3 km | Gap < 5 km | Bottom tercile (nonzero) or zero despite proximity | No | Pharmacy nearby with options, but demand outstrips supply |
| Fragile | < 3 km | Gap ≥ 5 km or k=3 NaN | Any | No | One pharmacy nearby but no meaningful alternatives |
| Underserved | 3–10 km | Any | Any | No | Requires transport to reach any pharmacy |
| Pharmacy desert | ≥ 10 km or NaN | Any | Any | No | Effectively no access |
| Data-uncertain | Any | Any | Any | Yes | High measurement uncertainty due to OSM gaps |
The classification is evaluated top-to-bottom: the snap flag check for Data-uncertain is applied first to separate measurement artifacts from genuine access conditions. Among the remaining SALs, the k=1 distance determines the primary tier (nearby, reachable, or absent), then k=3 redundancy and Ai score refine the diagnosis within the “nearby” tier. Tercile and median boundaries are computed from nonzero Ai values within each province-mode combination to prevent zero-inflation from distorting the classification. The typology is computed for both walk and drive modes.
Accessibility Calculation Limitations
- Straight-Line Distance: Assumes unobstructed travel in all directions, fails to account for barriers such as highways, rivers, railway lines, or fenced properties that may substantially increase actual travel distance. Urban areas with irregular street grids may have network distances much larger than straight-line distances.
- Network Distance (General): Network quality depends on OpenStreetMap completeness, and informal roads, footpaths, and shortcuts common in township areas may be unmapped, overestimating travel distances for local residents.
- Driving Distance: Assumes vehicle availability and ignores traffic congestion, parking availability, and fuel costs. Also fails to account for public transit options (minibus taxis) that may provide alternative access patterns. The dominant transport mode for the populations most likely to face pharmacy access barriers (low-income, non-vehicle-owning households in townships and traditional areas) is the minibus taxi, which is captured by neither the walk nor drive network.
- Walking Distance: Standard walking speed assumptions (5 km/h) may not hold for elderly, disabled, or mobility-impaired populations, and also does not account for safety concerns, terrain difficulty, or weather conditions.
- 2SFCA Zero-Inflation: Walking \(A_{i}\) is exactly zero for 53.2% of Gauteng SALs and 74.3% of KZN SALs. Not all zeros represent genuine pharmacy deserts: many arise from centroid-to-node snap displacement and catchment boundary edge effects. For Gauteng walk, 53.2% of SALs have zero \(A_{i}\) but only 24.6% exceed 3 km to the nearest pharmacy; the 28-point gap quantifies the artifact fraction. The k-nearest distance metric serves as the diagnostic check on this failure mode.
- Uniform Supply Assumption: All pharmacies are treated as having equal capacity (Sj = 1). A high-volume chain pharmacy with multiple pharmacists is weighted identically to a single-pharmacist dispensary. Pharmacy capacity data (staffing, dispensing volume, hours of operation) are not available.
- Province-Level Catchment Thresholds: Catchment distances are set by province rather than by settlement type, creating a discontinuity at the province boundary: SALs on the KZN side receive larger catchments than adjacent SALs in Gauteng’s East Rand despite comparable urbanization. A variable-catchment approach assigning thresholds by EA_GTYPE (Urban, Traditional, Farms) is the documented upgrade path.
- Decay Parameter: β = 0.0003 was selected for analytical discrimination but is not empirically calibrated to South African travel behavior. Calibration would require observed pharmacy utilization data (patient origin-destination patterns).
- Threshold Arbitrariness: Policy thresholds (3 km walk, 10 km drive) are reference points, not natural discontinuities. A SAL at 3.1 km is classified differently from one at 2.9 km despite near-identical access. The multi-threshold approach (four thresholds per mode at three k levels) mitigates this by showing the full exceedance curve.
- Population Centroid Assumption: All methods calculate distance from geometric (unweighted) SAL centroids. For compact urban SALs, this closely approximates the population center. For large, irregular rural SALs (some exceeding 600 km² in KZN), the geometric centroid may fall in uninhabited terrain, introducing systematic measurement error that propagates through the entire pipeline.
- Province Boundary Truncation: Road networks are downloaded by province and terminate at the provincial edge. SALs near province boundaries have their networks artificially truncated, forcing Dijkstra routing into within-province detours even if a pharmacy across the border is closer. In a national-scale deployment, this boundary effect disappears.
Validation and Quality Assurance
Data quality assurance proceeds throughout the pipeline:
Registration Validation: SAPC registration status is validated through a fuzzy matching process that links each SAPC-registered pharmacy to its corresponding record in the Google Places-derived geocoded list. This matching serves two purposes: confirming that a pharmacy appearing in the insurance network lists is also registered with the SAPC (supporting legitimacy), and enriching geocoded records with regulatory metadata including Y-number, owner, responsible pharmacist, licence number, registration date, and inspection history that the insurance sources do not contain.
Matching is performed at the province level: each SAPC record is compared only against geocoded candidates in the same province to reduce false positive risk. This restriction is enforced through a province normalization step that maps all province field variants (including abbreviations such as KZN and name fragments such as “Natal”) to a canonical lowercase form before grouping.
The match score for each candidate pair is a weighted combination of name similarity and address similarity. Name similarity uses token sort ratio, which reorders tokens alphabetically before comparison to handle word-order variation (e.g. “Clicks Pharmacy Sandton” versus “Sandton Clicks”). Address similarity uses token set ratio, which identifies the common token subset and is more tolerant of partial address matches. These two scores are combined as:
combined_score = 0.6 × name_score + 0.4 × address_score
The 60/40 weighting reflects that name similarity is generally a stronger discriminator than address similarity for South African pharmacy records, where address formatting is inconsistent across sources. Both scores are computed on cleaned text with punctuation removed, all characters lowercased, and generic pharmacy-related terms stripped (pharmacy, pharm, chemist, apteek, dispensary, medical, clinic, and related variants) to prevent these ubiquitous words from inflating similarity scores between unrelated records.
A chain guard short-circuits matching to a score of zero for records belonging to confirmed but different retail chains. A curated list of major pharmacy chains (Dis-Chem, Clicks, Medirite, Alpha Pharm, Mopani, and others) is checked against both records before scoring. If both records are identified as chain pharmacies but under different chain names, the pair cannot be a match regardless of name similarity, preventing cross-chain false positives that would otherwise achieve high scores on generic words like “pharmacy” and “store.”
Three confidence tiers are applied:
| Tier | Score Range | Disposition |
|---|---|---|
auto_match |
≥ 80 | SAPC regulatory fields appended to geocoded record automatically |
review |
70–79 | Record flagged for human inspection; treated as unmatched for master file build |
no_match |
< 70 | SAPC record routed to separate geocoding pass (04_sapc_geocode_unmatched.ipynb) |
The match_type field is preserved in all output files as an audit trail, allowing downstream users to filter to only automatically matched records or to inspect the borderline cases. The lowest-scoring auto-matches are surfaced in a preview table during the notebook run to support threshold calibration before export.
Registration Validation Limitations: The fuzzy match is sensitive to threshold selection. The auto-match threshold of 80 was chosen conservatively to minimize false positives at the cost of routing more records to the re-geocoding pass. Lowering the threshold would increase SAPC coverage but risk linking unrelated pharmacies. The province-constraint assumption that a pharmacy will have consistent province assignment across both the SAPC registry and the insurance network sources may not hold for pharmacies near province boundaries or for records with missing province data that were inferred through the city lookup table. Matching is also performed on a single best candidate per SAPC record: if the correct geocoded match is not in the top-scoring position due to address formatting, the record is routed to geocoding rather than matched.
Spatial Validation: Geocoded coordinates are verified against expected province boundaries where pharmacies geocoding outside Gauteng or KwaZulu-Natal boundaries require manual review.
Duplicate Detection: Post-geocoding spatial analysis identifies potential duplicates based on coordinate proximity (within 50-100 meters) combined with name similarity, addressing records that lack PRACTICE_NUM for identifier-based deduplication.
Outlier Review: Accessibility metrics are examined for outliers that may indicate data errors, like pharmacies in incorrect locations, or genuine access deserts requiring policy attention.
Validation Limitations:
- Registration Currency: SAPC registration status reflects the time of query and pharmacies may have been registered when data was collected but subsequently deregistered, or vice versa.
- Spatial Proximity Thresholds: The 50-100 meter threshold for duplicate detection is heuristic, so pharmacies in the same shopping center may legitimately be within this range, while geocoding errors may place duplicates further apart.
- Name Similarity Matching: Fuzzy name matching for duplicate detection may produce false positives (different pharmacies with similar names) or false negatives (same pharmacy with variant name spellings across sources).
- Outlier Interpretation: Extreme accessibility values may reflect genuine conditions (true access deserts or highly accessible areas) or data errors, and distinguishing these conditions require contextual investigation that may not be feasible at scale.
Summary of Methodological Limitations
The methodology employed in this project involves multiple stages, each introducing potential sources of error that may propagate through the analysis pipeline. Key limitations are summarized below:
| Stage | Primary Limitation | Mitigation Strategy |
|---|---|---|
| Data Collection | Network participation bias, missing independent pharmacies | Cross-validation with Google Places API and SAPC registry |
| PDF Extraction | OCR errors in addresses | Manual review of geocoding failures, iterative address cleaning |
| Province Inference | Lookup table incompleteness | Iterative refinement based on dropped record review |
| Deduplication | Missing PRACTICE_NUM for ~1,500 records |
Post-geocoding spatial deduplication, conservative retention |
| Geocoding | Variable coverage, address format inconsistency | Dual-source geocoding, confidence score filtering, re-geocoding passes |
| Population Step-Down | Frozen 2011 distribution assumption; 1,270 SALs with zero population regardless of post-2011 development | Multi-method comparison, flagging high-divergence areas |
| Centroid Placement | Geometric centroids in large rural SALs may fall in uninhabited terrain | Documented upgrade path: building-footprint-weighted centroids for SALs above 75th percentile in area |
| Network Coverage | OSM completeness varies by settlement type; province boundary truncation forces within-province detours | Snap distance diagnostic flags; national-scale deployment eliminates boundary effects |
| Transport Mode | No minibus taxi mode; walk overestimates difficulty, drive overestimates ease for target population | True access picture lies between walk and drive metrics; transit integration via GTFS is the upgrade path |
| 2SFCA Methodology | Province-level catchment thresholds create boundary discontinuity; uniform pharmacy capacity; uncalibrated decay parameter | Variable-catchment by EA_GTYPE (code exists, commented out); MIN_WEIGHTED_POP = 50 floor applied |
| Pharmacy Data Coverage | Private-sector and government-employee sources only; clinic dispensaries and mobile units excluded | Conservative (pessimistic) view of access; actual medication access points likely higher |
| Threshold Analysis | Binary cut at policy thresholds ignores near-threshold equivalence | Multi-threshold approach (four thresholds per mode at three k levels) shows full exceedance curve |
Users of this analysis should interpret results as indicative rather than definitive, particularly in areas where multiple limitations may compound. In this regard, policy decisions should be informed by but not solely determined by these accessibility measures, with ground-truthing recommended for priority intervention areas.
Exploratory Data Analysis
Pharmacy Distribution
| Province | Count |
|---|---|
| Gauteng | 1,453 |
| KwaZulu-Natal | 699 |
| Total | 2,152 |
