Table 1 Studies comparing different methods of calculating travel impedance to health services
# Ref | Author (year) Country | Data sources | Impedance measures | Outcome | Comparison method | Favoured measure/conclusion |
|---|---|---|---|---|---|---|
Low- and middle-income countries | ||||||
1 [14] | Okwaraji (2012) Ethiopia | 1. Geocoded households | 1. Euclidean distance | Under 5 child mortality | 1. Correlation coefficient | Actual travel distance |
2. Geocoded health center | 2. Raster travel time | 2. Compare measures of effect | ||||
3. Land cover, Ethiopia Mapping Agency | 3. Actual travel distance | |||||
4. Digital elevation model from Shuttle Radar Topography Mission (NASA) | ||||||
2 [15] | Noor (2006) Kenya | 1. Geocoded homesteads | 1. Euclidean distance | Predicted specific facility use by febrile children; Proportion of people within one hour of HF | 1. Kappa statistic (agreement between predicted and observed facility use) | Raster travel time (transport network model) adjusted for competition |
2. Geocoded HFs | 2. Raster travel time (termed transport network model) | 2. Linear regression (R2) | ||||
3. Population density at 100 m resolution (Kenya Census 1999) | 3. Raster travel time (transport network model), adjusted for competition between facilities | 3. Scatter plots | ||||
4. Road network (Africover, plus manual updates) | 4. Spatial mapping | |||||
5. Topography (Africover, plus updates & Livestock research institute, Nairobi & Park & reserve digital map from Kenya Wildlife Service) | ||||||
3 [16] | Costa (2003) Brazil | 1. Admissions data from national public health database | 1. Euclidean distance | None | 1. Maximum difference in distances | “Real” distance |
2. Extracted district of residence from postal codes from national database | 2. “Real” distance, estimated as city bus itinerary from district centroid to hospital, adjusted for residence district area | |||||
3. GIS coordinates for 14 public hospitals | ||||||
4. City transit network map, bus routes | ||||||
High-income countries | ||||||
3 [17] | *Cudnik (2012) USA | 1. Patient location via EMS data | 1. Euclidean distance | None | 1. Wilcoxon signed rank test | Reasonable to use Euclidean distance |
2. HF location via addresses | 2. Network distance | 2. Spearman rank | ||||
3. Road network (ArcGIS StreetMap; commercially available) | 3. Actual transport distance (in EMS vehicle) | 3. Linear regression (R2) | ||||
4 [9] | *Delamater (2012) USA | 1. Population (US Census 2010) | 1. Network travel time | Proportion of state classified as limited access area (LAA) | 1. Percentage change in proportion LAA | Depends on research question |
2. Road network (Michigan Center for Geographic Information 2009) | 2. Network distance | 2. Mapping | ||||
3. Raster travel time | ||||||
4. Raster distance | ||||||
5 [18] | ~*Lian (2012) USA | 1. Incident breast cancer cases (Missouri cancer registry) | 1. Network travel time | Incident odds of late-stage breast cancer | 1. Spearman rank | 2SFCA |
2. Population coordinates (US Census 2000) | 2. Average of 5 shortest network travel times | 2. Kappa coefficient | ||||
3. HF coordinates (FDA) | 3. Service density | 3. Moran I index | ||||
4. Road Network (US Census/ TIGER) | 4. Two-step floating catchment area (2SFCA) | 4. Comparison of effect measures on risk of outcome | ||||
6 [19] | *Jones (2010) USA | 1. Population location (Insurance claims data) | 1. Euclidean distance | None | 1. Wilcoxon’s signed rank sum tests | Network distance |
2. HF location via addresses | 2. Network distance | 2. Scatter plots | ||||
3. Road network (no source listed) | ||||||
7 [20] | *Apparicio (2008) Canada | 1. Population coordinates (Statistics Canada) | 1. Euclidean distance | None | 1. Spearman rank | Network distance |
2. HF coordinates (Quebec Ministry of Health and Social services) | 2. Manhattan distance | 2. Absolute differences in measures | ||||
3. Road network (CanMap street files, commercially available) | 3. Network distance | 3. Spatial mapping | ||||
4. Network travel time | ||||||
8 [21] | Fone (2006) UK | 1. Population via postal survey from Gwent Health Authority | 1. Euclidean distance | Perceived accessibility | 1. Kruskal-Wallis | Minimal advantage in using sophisticated measures |
2. Population location via census | 2. Network travel time | 2. Spearman rank | ||||
3. HF locations from Gwent Health Authority | 3. Network distance | |||||
4. Road network (MapInfo Drivetime software, commercially available) | ||||||
9 [22] | Haynes (2006) UK | 1. Hospital-based patient questionnaire (with post-codes) | 1. Euclidean distance | None | 1. Spearman rank | No evidence that GIS estimates better than Euclidean |
2. Geocoded HF location | 2. Network travel time | 2. Linear regression (R2) | ||||
3. Road network (Ordinance Survey Meridian, digital map) | 3. Actual travel time | |||||
10 [23] | *Fortney (2000) USA | 1. Population location from previous study sample | 1. Euclidean Distance | None (travel time as gold standard) | 1. Correlation coefficients | Marginal gains in accuracy using network measures |
2. HF location from physician desk reference database (State licensing board) | 2. Network distance | 2. Linear regression | ||||
3. Road network (US Census Bureau) | 3. Differences between measures | |||||