To-Do: Build Business History Dataset

Purpose

Build a business-change dataset that can support the parking minimum study at two levels:

1. an **aggregated baseline** using NaNDA tract-level annual business counts and employment

2. a later **business-level POI history dataset** using Yelp, Foursquare, OSM, and manual validation

The main idea is to start with a tract-based corridor proxy so we can learn the data, test corridor comparisons, and get an early read on signal before committing to the harder business-level pipeline.

Why This Revision Makes Sense

The original plan jumped quickly into business-level longitudinal reconstruction.

That is still a good long-run goal, but the Nanda dataset give us a (low effort/ high reward) baby step in that direction:

• annual tract-level panels
• counts and employment
• multiple retail and service categories
• long time range
• standardized geography

Examples from the downloaded NaNDA files:

• eating/drinking places
• retail stores
• grocery stores
• liquor/tobacco stores
• personal care / laundromats
• recreation stores
• dollar stores

The tract CSVs appear to contain fields like:

• `tract_fips10`
• `year`
• `totpop`
• `aland10`
• category-specific `count_*`
• category-specific `emps_*`
• category-specific `den_*`
• category-specific `aden_*`

That makes NaNDA a good baseline for:

• annual corridor-proxy business counts
• annual corridor-proxy employment
• pre/post trend inspection
• early cross-city comparison

It does **not** solve true corridor-only attribution by itself, because tracts are larger than the corridor buffers we care about. But it gives us a practical first dataset.

Framing

We should treat this as a two-stage business-history strategy.

Stage A

Build a **corridor-proxy aggregated panel** using:

• corridor geometry
• corridor buffers
• census tract overlap
• NaNDA annual tract-level business counts and employment

Stage B

Build a **business-level corridor panel** using:

• Yelp
• Foursquare
• OSM
• manual checks

Stage A helps us learn:

• whether selected corridors show plausible movement over time
• which category families appear most responsive
• where the strongest cases are
• how to structure later validation

Revised Goal

Construct a staged longitudinal business dataset aligned with the study objective of detecting corridor-level change before and after parking minimum removal.

Immediate goal

Create an **aggregated corridor-year panel** using NaNDA tract-level business and employment measures.

Later goal

Create a **business-location panel** for selected corridors after the aggregated baseline is working.

Phase I: Aggregated Baseline With NaNDA

1. Define The Analysis Frame

• Confirm final study corridors and backups
• Use the reviewed corridor geometry once available
• Define corridor buffers
• likely `100m` and possibly `200m` sensitivity version
• Assign a unique corridor ID
• Assign city, state, city group, and treatment status

Output:

• corridor geometry layer
• corridor buffer layer
• corridor metadata table

2. Define The Geographic Proxy Strategy

Because NaNDA is tract-level, we need a transparent way to map corridor buffers to tracts.

Recommended default:

• intersect corridor buffers with 2010 census tracts
• compute overlap share
• assign tracts to corridors using one or more rules

Recommended rules to support:

• `binary_touch`
• tract counts if it intersects the corridor buffer at all

• `centroid_in_buffer`
• tract counts if tract centroid falls inside the corridor buffer

• `area_weighted`
• tract contributes based on share of tract area inside the corridor buffer

My recommendation:

• use `area_weighted` as the main specification
• keep `binary_touch` as a robustness check

This will not perfectly isolate corridor-only activity, but it is a defensible first-pass corridor proxy.

3. Prepare NaNDA Inputs

Use the tract-level CSVs inside [`data/nanda`](c:/Users/ylaim/OneDrive/6555-urpl-transport-env/parking_research/data/nanda).

Relevant families:

• eating/drinking
• retail
• grocery
• liquor/tobacco
• personal care / laundromats
• recreation
• dollar stores

Tasks:

• unzip to a working data directory
• standardize tract identifiers
• document year coverage per file
• document category field names
• select a clean subset of variables for the first pass

Recommended first-pass fields:

• tract ID
• year
• total population
• land area
• category-specific business counts
• category-specific employment

Output:

• cleaned NaNDA tract panels by theme
• one variable dictionary for selected NaNDA fields

4. Build Corridor-Tract Crosswalk

For each corridor buffer:

• intersect with tract polygons
• compute:
• tract area overlap
• share of tract area in corridor buffer
• optional share of corridor buffer inside tract

Recommended crosswalk fields:

• `corridor_id`
• `city`
• `tract_fips10`
• `buffer_m`
• `tract_area_m2`
• `intersect_area_m2`
• `tract_overlap_share`
• `corridor_buffer_share`
• `crosswalk_rule`

Output:

• corridor-tract crosswalk table

This crosswalk becomes the key bridge between corridor geometry and NaNDA panels.

5. Build Corridor-Year Aggregated Panel

Join the corridor-tract crosswalk to the cleaned NaNDA tract panels.

Then aggregate tract values to the corridor-year level.

Recommended default:

• weighted sums using `tract_overlap_share`

Example outputs by corridor-year:

• estimated eating/drinking count
• estimated eating/drinking employment
• estimated retail count
• estimated retail employment
• estimated grocery count
• estimated personal care / laundromat count

This yields a corridor-year panel such as:

• `corridor_id`
• `city`
• `year`
• `buffer_m`
• `count_totaleatingplaces_est`
• `emps_totaleatingplaces_est`
• `count_totretail_est`
• `emps_totretail_est`
• additional category estimates

6. Derive Aggregated Change Measures

From the corridor-year panel, derive:

• annual change in business counts
• annual change in employment
• pre/post treatment change
• category composition change
• corridor-level trend slopes

Good first measures:

• level change
• percent change
• rolling average change
• category share change

This is not business churn yet in the strict entry/exit sense, but it gives an early measure of corridor commercial change.

7. Validate The Aggregated Proxy

We should explicitly test whether the tract proxy behaves plausibly.

Checks:

• compare corridor maps to tract overlap maps
• flag corridors dominated by one very large tract
• flag corridors with weak tract alignment
• compare aggregated signals to qualitative expectations from corridor notes
• compare selected corridors to OSM/POI context

Important:

Some corridors will be better candidates for tract-based inference than others.

This validation step should help us identify:

• strong aggregated candidates
• weak aggregated candidates
• corridors that likely require business-level reconstruction sooner

8. Produce Phase I Deliverables

Deliverables:

• corridor-tract crosswalk
• cleaned NaNDA tract panels
• corridor-year aggregated panel
• exploratory plots by corridor and city
• notes on which categories are most informative

This phase should be enough to:

• test the city/corridor design
• assess whether treatment corridors move differently from comparison corridors
• decide where the business-level effort is worth the cost

Phase II: Business-Level POI History

Only after the aggregated baseline is working should we move to the harder business-level build.

9. Build Multi-Source POI Inventory

Use a combined approach:

• Yelp for review/activity timing
• Foursquare for structure and categories
• OSM for validation and baseline context

For each corridor:

• query by corridor geometry or corridor-centered search areas
• collect name, address, coordinates, category, and metadata
• store raw source tables separately

10. Standardize And Merge POIs

• normalize names
• normalize categories
• deduplicate by name + proximity + address logic
• create a unified business-location table

Output:

• one row per candidate business-location

11. Attach Time Signals

Examples:

• Yelp review timestamps
• first observed activity
• last observed activity
• optional Foursquare activity/metadata if useful
• OSM as validation or supplementary presence signal

This will allow us to infer:

• entry proxy
• exit proxy
• persistence
• category change

12. Build Business-Level Panel

Unit:

• business-location

Time:

• yearly or coarse interval

Fields:

• active / inactive
• category
• corridor membership
• source confidence

This is the stage that supports true churn metrics.

13. Validate With Manual Checks

Use:

• Google Maps
• Street View
• manual parcel review where necessary

Focus on:

• false closures
• duplicate businesses
• renamed businesses
• major missed establishments

14. Final Analysis Outputs

Eventually we want two linked outputs:

Corridor-level panel

• annual business counts
• annual employment proxies
• category mix
• churn metrics

Business-level panel

• entry timing
• exit timing
• persistence
• category change

Ready for:

• before/after policy comparison
• treated vs comparison corridor comparison
• cross-city comparison
• linkage to repeal timing

Deliverable

A staged business-history workflow that starts with a tract-based aggregated baseline and later extends to business-level longitudinal reconstruction.

That gives us:

• a faster first analytic read
• a cleaner way to learn the corridor design
• a more defensible path into the harder POI history build