Methodology

1. Collect data from the census for population estimate on various geography

2. Polygon preparation

   • Scrape brand store locations from official sources (e.g. official website)

   • Manually go through all brand store locations, cross reference with 3rd party source for store location (e.g. Google Maps), correct brand store locations if required

   • Using (corrected) brand store locations, draw the store boundary

   • Manually vet all drawn polygons in a consistent manner

     • Corporate office should not be included

     • Only polygons drawn with minimal guesswork should be included1

     • Brand stores built under any additional structure (e.g. apartment blocks) are excluded, unless they are overhead car park

3. Selecting a panel

   • This is the entire population with some filters to ensure consistent data collection behaviour in our panel;

4. Quantifying sampling bias on a geographical level

   • Compare the geopgrahical distribution of devices in our dataset against that in the census

   • Compute the up-sampling(/down-sampling) factor needed to apply to each geopgrahical region and devices

5. Detect continuous visit within drawn polygons

6. Extract visit durations and remove outliers

   • Outliers include exceptionally short visits (possibly due to location data inaccuracies)

   • Outliers include exceptionally long visits (possibly due device belong to the store or store worker)

7. Apply sampling bias adjustment factor

   • Down-sampling the oversampled regions

   • Up-sampling the undersampled regions

8. Apply time series smoothing

   • Using a simple rolling median methodology

9. Compute visit time per panel capita metric

   • By divde the sampling biased visit time sum by panel population count in the same month;

---

1

For example, if the store is in a shopping mall and no shop boundary is shown on a map nor is there any external building features we could use to infer it, we will not be include that store.