Cannibalization Analysis in Site Selection

A site can score 86 out of 100 and still be a bad expansion decision.

The demographics are strong. The traffic looks good. The co-tenancy works. The real estate team likes the visibility. The model says the candidate can hit its sales target.

Then the new unit opens, and the nearby store starts giving back volume. The candidate is performing, but the market barely grew. The company added rent, labor, buildout cost, and operational complexity to serve customers it was already serving.

That is the expensive version of cannibalization: approving a good site in a market you already own.

Cannibalization analysis exists to prevent that mistake. It separates sales that are truly incremental from sales that simply move across the network. For multi-unit operators, the core question is bigger than “Will this site work?”

What happens to the network if we open here?

1. What is cannibalization analysis?

Cannibalization analysis estimates how the demand for a proposed new location would be redistributed across an existing network.

In retail, restaurant, healthcare, financial services, fitness, and other multi-unit businesses, cannibalization usually refers to demand that shifts from an existing same-brand location to a new same-brand location. That demand may appear as sales, visits, transactions, delivery orders, memberships, appointments, patients, deposits, or service volume.

CARTO describes cannibalization analysis as evaluating the impact a new location may have on nearby existing locations, especially whether the new location will draw customers away from them and reduce revenue.

A restaurant chain asks:

How much of this candidate’s projected volume comes from customers we already serve?

A healthcare operator asks:

How much patient volume will shift from existing clinics, and how much will come from better access, competitor capture, or unmet need?

A retailer asks:

Does this site expand market coverage, or does it divide the same demand across more leases, labor, and operating cost?

Cannibalization analysis connects site-level performance to portfolio-level performance.

2. Why cannibalization matters in site selection

Most bad cannibalization decisions begin with a site that looks good by itself.

Strong demographics, traffic, visibility, and co-tenancy can all be real. The mistake is treating those strengths as proof of incremental demand. A candidate can sit in an attractive trade area because the brand already chose the best nearby site years ago.

The business risk shows up in several ways. A sales forecast may count transferred customers as new customers. Same-store sales may weaken at nearby units. Franchisees may object to encroachment. Delivery orders may shift from one kitchen to another without expanding total order volume. A market can appear underpenetrated because the model sees high demand, even though the current network already captures much of the reachable demand.

Sweetgreen’s 2024 10-K uses exactly this kind of operator language. The company says new restaurants in or near existing markets may adversely impact sales at existing restaurants, especially in highly concentrated markets. It also states that the company may selectively open new restaurants in and around areas where existing restaurants are operating at or near capacity, and warns that sales cannibalization may become significant if the delivery radius of an existing restaurant overlaps with that of a new restaurant.

Capacity relief, geographic concentration, delivery-radius overlap, and existing-store impact can all be true at the same time.

3. Gross demand, transferred demand, and net-new demand

A useful cannibalization model separates a candidate’s demand into three buckets.

Gross demand

Gross demand is the total demand a candidate location is expected to capture.

For a restaurant, gross demand may be projected annual sales. For a clinic, it may be annual visits. For a fitness concept, it may be memberships. For a bank, it may be accounts, deposits, or advisory relationships.

Gross demand can come from residents, workers, commuters, visitors, delivery customers, competitor customers, existing same-brand customers, and customers who were previously unserved.

How much volume could this location generate?

It does not answer:

How much value does this location add to the network?

Transferred demand

Transferred demand is the portion of the candidate’s demand that shifts away from existing same-brand locations. This is the core cannibalization measure.

Transferred demand can appear as:

• sales transfer
• visit transfer
• delivery order reassignment
• appointment transfer
• patient transfer
• membership transfer
• transaction transfer
• same-store sales decline

How much of this candidate’s volume were we already serving?

Net-new demand

Net-new demand is the portion of candidate demand that expands the network.

It can come from:

• previously unserved customers
• competitor capture
• reduced leakage
• improved access
• increased purchase frequency
• new delivery coverage
• capacity relief
• convenience-induced demand

Net-new demand = Gross candidate demand - Transferred demand from existing locations

Candidate demand =
  demand transferred from own stores
+ demand captured from competitors
+ demand recaptured from leakage
+ newly induced demand
+ demand of uncertain origin

The final bucket belongs in the model. Attribution is rarely perfect. A credible analysis shows what is known, what is estimated, and where uncertainty remains.

4. Core cannibalization metrics

A serious site brief should report more than a single cannibalization percentage.

Cannibalization rate

Cannibalization rate =
Transferred demand from existing same-brand locations / Candidate demand

Example:

Candidate projected sales = $2.4M
Transferred sales from existing stores = $720K

Cannibalization rate = $720K / $2.4M = 30%

Net-new demand ratio

Net-new demand ratio =
Net-new demand / Candidate demand

Example:

Candidate projected sales = $2.4M
Transferred sales = $720K
Net-new sales = $1.68M

Net-new demand ratio = $1.68M / $2.4M = 70%

Existing-store impact

Existing-store impact =
Transferred demand from affected store / Affected store baseline demand

Example:

Store A baseline sales = $3.2M
Expected transfer from Store A = $410K

Existing-store impact = $410K / $3.2M = 12.8%

This metric can matter more than the candidate-level cannibalization rate. A 30% cannibalization rate may be acceptable overall, while a 15% hit to one strategically important store may fail the approval threshold.

Network increment

Network increment =
Candidate contribution
- lost contribution from existing locations
- added fixed costs
± strategic network effects

The strongest cannibalization analyses end with incremental portfolio value, not just projected candidate sales.

5. Trade area overlap vs cannibalization

Trade area overlap is one of the most useful inputs in cannibalization analysis. It is also one of the most commonly misused.

Trade area overlap measures the shared geography, population, customers, or demand between two locations’ catchments.

Cannibalization measures demand transfer from an existing same-brand location to a new same-brand location.

CARTO’s store cannibalization example makes the distinction concrete. In its worked example, one potential new store shared only 4.7% of its catchment area with an existing store, but that overlap represented 19.3% of the existing store’s population coverage.

That is the lesson operators need to internalize.

Area overlap asks:
How much of the map overlaps?

Demand-weighted overlap asks:
How much of the relevant market overlaps?

Cannibalization analysis asks:
How much demand will transfer?

Those are separate questions. A small polygon overlap can contain dense, high-value demand. A large polygon overlap can contain little relevant demand. A site committee should care about demand exposure and transfer risk, not just the size of overlapping shapes.

CARTO’s retail tooling reflects this distinction by calculating both overlap area and overlap in enriched variables such as population, footfall, or other metrics tied to the business question. Its cannibalization workflow also supports different catchment methods, including buffers, k-rings, and isolines.

6. Why radius-based cannibalization analysis fails

The weakest version of cannibalization analysis is the fixed-radius rule.

Common examples:

Flag every candidate within 3 miles of an existing store.
Reject any site inside a 5-mile protected radius.
Treat two overlapping circles as evidence of cannibalization.

These rules are fast and easy to explain. They are also blunt.

Fixed radii assume customers move evenly in every direction. They ignore road networks, congestion, barriers, bridges, medians, walkability, parking, ingress, commute direction, store attractiveness, daypart, competitors, and channel behavior.

A three-mile radius can include households that would never visit because a highway or river makes the trip inconvenient. It can miss customers six miles away with a direct arterial route and strong purchase intent.

Radius rules still have a place. They work as early pipeline filters, protected-territory screens, and obvious conflict checks. Final approval needs a more realistic model of access, demand, and choice.

7. How cannibalization is measured

Cannibalization analysis usually moves through five levels of sophistication.

Level 1: Proximity screen

This is useful for early review, protected territories, franchise constraints, and obvious encroachment flags.

Level 2: Catchment overlap

Catchments may be fixed radii, drive-time polygons, walk-time polygons, delivery zones, customer-derived trade areas, service-area polygons, isochrones, or custom geographies. For a deeper treatment of catchment design, see Geod’s trade area analysis guide.

Level 3: Demand-weighted overlap

This is where the analysis becomes materially better. The model weights overlapping geography by relevant demand, such as population, households, target customers, footfall, category spend, workers, patients, delivery demand, or observed visits.

The CARTO example is worth repeating because it is so teachable: 4.7% area overlap translated into 19.3% population overlap.

Level 4: Probabilistic demand allocation

At this level, the model estimates how demand is allocated across the candidate, existing same-brand locations, competitors, and outside alternatives.

Methods include Huff models, gravity models, discrete choice models, customer-derived allocation models, and explainable machine-learning models. The same output discipline applies here as in explainable site selection: the model needs to show its work.

Level 5: Post-opening validation

The most mature operators compare pre-opening estimates with actual outcomes. They look at same-store sales, customer origin shifts, loyalty or transaction records, mobile visit patterns, delivery reassignment, matched markets, and competitor response.

This validation step is critical because a sales decline after opening may have many causes: local construction, promotions, staffing, pricing, weather, seasonality, macroeconomic changes, or a competitor move.

8. Huff models, gravity models, and spatial interaction

Cannibalization analysis has deep roots in retail geography, spatial interaction modeling, and competitive facility location.

A Huff-style model estimates the probability that demand from origin i chooses location j.

Pij = (Aj^α × Dij^-β) / Σ(Ak^α × Dik^-β)

Esri describes the Huff model as a spatial interaction model where the probability of a consumer visiting and purchasing at a site depends on distance, site attractiveness, and the distance and attractiveness of competing sites. Esri also notes that distance exponents commonly dampen the probability of choosing distant sites, and that attractiveness can include attributes such as retail floor space, parking, pricing, display area, frontage, advertising, or observed visitation.

For cannibalization, the model compares demand allocation before and after the candidate enters the market.

Demand is allocated among existing same-brand stores, competitors,
outside alternatives, and non-consumption.

Demand is reallocated among existing stores, the candidate, competitors,
outside alternatives, and non-consumption.

The source of candidate demand determines the business interpretation.

Calibration is the hard part. Esri warns that default exponent values may not apply to the specific trade area being modeled and that calibration requires an existing facility layer, customer layer, and potential-sales layer. It also notes that distance decay differs by activity, with grocery shopping generally showing a different travel-distance pattern than furniture shopping.

Academic work supports the practical value of this modeling family. A large-scale transaction-data study found that Huff-style gravity models fit shopping behavior across categories including grocery stores, clothing stores, gas stations, and restaurants.

Another study on multi-facility location models explicitly notes that adding new outlets can cannibalize existing sales, while effective location choices can still increase the chain’s total captured market share.

9. Daypart and channel-specific cannibalization

Cannibalization changes by time of day, trip purpose, and channel.

A coffee shop may cannibalize weekday morning commute visits while adding weekend demand. A QSR may transfer breakfast drive-thru volume from an existing store while capturing competitor dinner traffic. A restaurant may show limited dine-in transfer while heavily reassigning delivery orders.

Esri’s Huff Model tool includes parameters for travel direction and time of day, which matters because accessibility can change depending on whether people are traveling toward stores, away from stores, or moving through the road network at a specific time.

Research on time-aware Huff modeling found that a dynamic Huff model calibrated with mobile-location visit data improved market-share prediction relative to the original Huff model and captured differences by business type and region.

Sweetgreen’s public filing is a strong operator example because it identifies delivery-radius overlap as a specific cannibalization risk. That is a channel-specific risk, not just a store-distance risk.

10. When overlap is strategically valuable

Cannibalization analysis should measure the value created by overlap as well as the volume transferred by it.

Some overlap is rational.

Convenience density

A denser network can increase frequency by making the brand easier to use. This matters for coffee, QSR, convenience, fitness, pharmacies, urgent care, banking, and delivery-heavy restaurants.

Capacity relief

A new location may deliberately pull volume from an overloaded existing location. Sweetgreen says it may selectively open restaurants in and around areas where existing restaurants are operating at or near capacity to serve customers more effectively.

The analysis still needs to quantify the transfer. Capacity relief can justify overlap, but it does not erase the cannibalization question.

Competitive defense

A candidate may capture some demand from existing stores while preventing a competitor from taking the corridor, delivery zone, or retail node. This is common in mature markets where the best sites are scarce.

Density economies

Thomas Holmes’ NBER paper on Wal-Mart’s expansion found that Wal-Mart maintained high store density and a contiguous store network during its rollout. The paper infers the value of density economies partly from the sales cannibalization Wal-Mart was willing to sustain from closely packed stores, and concludes that the benefits of high store density were substantial and likely extended beyond trucking-cost savings.

That is the strategic version of cannibalization: a store can reduce nearby store sales and still improve the network through distribution efficiency, advertising leverage, management efficiency, market control, or customer convenience.

Agglomeration and spillover

Retail locations can benefit from nearby activity. A study of anchor-store effects in Greater London found that anchor stores increased customer traffic to nearby non-anchor stores by 14.2% to 26.5% in the studied areas.

The practical point: overlap is a risk signal. Transfer, contribution, and network value determine the decision.

11. Cannibalization by category

Cannibalization behaves differently across concepts. A useful model should match the category’s trip purpose, frequency, channel mix, and operating constraints.

The matrix is a reminder that “cannibalization analysis” is not one universal overlay. The right model depends on how customers choose, travel, order, and repeat.

12. Market saturation, white space, and cannibalization

Cannibalization analysis sits between white space analysis and saturation analysis.

White space analysis identifies where attractive demand is underserved by the current network. A white space may exist because customers are outside current trade areas, travel times are too long, competitor capture is high, delivery coverage is weak, or existing locations are capacity constrained.

Saturation analysis evaluates whether additional units in a market are producing declining incremental returns.

Cannibalization analysis estimates how demand will transfer after a specific candidate is added.

White space: Where are we under-covered?

Saturation: Where are incremental returns declining?

Cannibalization: How much demand will transfer if we open here?

A market can have strong demographics and still be saturated for a specific brand. High demand does not automatically mean high incremental demand.

13. Data inputs and limitations

Cannibalization analysis depends on the quality of its inputs.

Existing-location data should include sales, visits, transactions, average ticket, contribution margin, daypart performance, delivery mix, pickup mix, appointment utilization, patient volume, membership counts, capacity constraints, local marketing history, and same-store sales trends.

Customer-origin data is especially valuable because it reveals actual catchments. Sources may include loyalty addresses, delivery addresses, patient ZIPs, transaction records, mobile-location panels, account data, survey data, and observed visit origins.

Market demand layers may include population, households, income, age, employment, daytime population, commuter flows, category spending, business density, vehicle ownership, visitor activity, and health need indicators. The Census Bureau describes the American Community Survey as an ongoing source of detailed social, economic, housing, and demographic information from a sample of households across the 50 states, the District of Columbia, and Puerto Rico.

Workplace and commuter data can matter for coffee, convenience, QSR, banking, healthcare, and service concepts. The Census LEHD program provides public-use products including QWI, LODES, OnTheMap, and related workforce data, while also warning that LEHD data are tabulated and modeled administrative data subject to nonsampling errors.

Competitor data should include location, format, size, hours, pricing, reviews, delivery coverage, drive-thru availability, service lines, parking, observed visits, and brand strength.

Accessibility inputs should include drive time, walk time, transit access, congestion, turn restrictions, barriers, parking, ingress and egress, commute direction, delivery time, and travel-time reliability.

The limitations should be stated plainly in the site brief. ZIP codes, tracts, and block groups are approximations. Mobile, loyalty, and transaction datasets all have coverage bias. Existing-store data may contain survivorship bias. Sales can move because of staffing, pricing, hours, service quality, local construction, weather, promotions, macroeconomic shifts, or competitor response. Distance-decay and attractiveness assumptions require calibration. Delivery, pickup, drive-thru, in-store, and appointment channels often transfer differently.

A model earns trust by exposing these limits, not by hiding them.

14. A five-step cannibalization workflow

Step 1: Define the decision and affected network

Start with the business question.

Should we approve this candidate site?
Which existing stores are at risk?
Will this site create enough net-new demand?
Does this opening improve market coverage?
Should we open, relocate, close, or wait?

Define the candidate, existing same-brand locations, franchise territories, delivery zones, competitors, market boundary, affected channels, relevant dayparts, and decision threshold.

Step 2: Build realistic trade areas

Choose the catchment method that fits the category.

Options include:

• drive-time catchments
• walk-time catchments
• delivery zones
• commute corridors
• customer-derived trade areas
• service-area polygons
• daypart-specific catchments
• probabilistic trade areas

A coffee commute model, urgent care access model, and grocery stock-up model should not use the same default radius.

Step 3: Weight overlap by relevant demand

Measure exposure, not just geometry.

Depending on the category, weight overlap by population, households, target customers, footfall, category spend, workers, visitors, patients, delivery demand, customer origins, or observed visits.

This is where the CARTO example remains so useful: 4.7% area overlap can become 19.3% population overlap.

Step 4: Estimate demand transfer

Break the candidate forecast into demand sources.

A calibrated Huff model, gravity model, discrete choice model, or customer-derived allocation model can estimate how demand reallocates when the candidate enters the market. Geod’s methodology page explains how transparent assumptions make these decisions auditable. Esri’s Huff documentation emphasizes the roles of distance, attractiveness, competition, distance decay, and calibration.

Step 5: Produce the decision record and validation plan

The final output should summarize:

• candidate forecast
• transferred demand
• net-new demand
• affected stores
• maximum same-store impact
• daypart and channel risk
• key assumptions
• confidence level
• scenario range
• recommendation
• post-opening validation plan

Every opening should make the next model better.

15. Illustrative scenario: high sales, low net-new demand

The following scenario is illustrative. The numbers are fabricated to show the mechanics of the analysis.

A QSR operator is evaluating a candidate site with strong traffic, good visibility, and a drive-thru.

Candidate site profile

A site-level model projects $2.4M in annual sales.

The network model decomposes that demand.

The candidate is a good sales site and a complicated network decision.

Daypart analysis reveals the transfer pattern.

The real risk is breakfast and lunch route interception.

The operator can now ask better questions:

• Is the breakfast and lunch transfer acceptable?
• Is Store A healthy enough to absorb the impact?
• Does the site block a competitor from the corridor?
• Can marketing or operations shift the candidate toward dinner and late night?
• Does the drive-thru coverage justify the transfer?
• Would relocation create more value than an incremental opening?

A single site score cannot answer those questions.

16. What a defensible cannibalization section should include

A strong site brief should make the cannibalization logic visible.

Recommended structure:

Cannibalization Summary

1. Candidate demand forecast
2. Demand source decomposition
3. Existing-location impact
4. Trade area overlap and demand-weighted overlap
5. Daypart and channel-specific risk
6. Scenario range
7. Key assumptions and confidence level
8. Recommendation
9. Post-opening validation plan

Example language:

The candidate is projected to generate $2.4M in annual sales. The model estimates $1.12M of that volume would transfer from existing same-brand stores, producing a 46.9% cannibalization rate and $1.27M in net-new, conquest, or induced demand. The largest affected-store impact is Store A at 18.0% of baseline sales. Transfer risk is concentrated in weekday breakfast and lunch because the candidate intercepts the same commuter corridor. Approval depends on whether the operator assigns sufficient strategic value to drive-thru coverage and competitor defense in this corridor.

For Geod, this is where cannibalization analysis belongs: inside an explainable site brief that shows the demand source, affected stores, assumptions, confidence level, and post-opening measurement plan.

The useful output is a decision record the real estate team, finance team, operator, franchise partner, and executive committee can interrogate.

17. Common mistakes in cannibalization analysis

Mistake 1: Treating area overlap as demand overlap

CARTO’s 4.7% area overlap vs 19.3% population overlap example shows how misleading pure geometry can be.

Mistake 2: Treating overlap as automatic rejection

Overlap can indicate risk, capacity relief, competitor defense, delivery improvement, or density economics. The decision depends on incremental network value.

Mistake 3: Using one radius for every category

Coffee, grocery, urgent care, fitness, banks, and QSRs have different customer behavior. Their models should differ.

Mistake 4: Ignoring daypart

A site can be low risk overall and high risk at breakfast. Another site may transfer delivery demand while adding dine-in demand.

Mistake 5: Ignoring competitors

A candidate can pull demand from existing units and competitors at the same time. A model that ignores competitors misclassifies demand.

Mistake 6: Counting all candidate sales as incremental

Candidate sales and network growth are different metrics.

Mistake 7: Hiding assumptions

Distance decay, attractiveness, travel mode, competitor set, demand layer, channel mix, and calibration choices should be visible.

Mistake 8: Overstating precision

The best models provide scenarios, confidence levels, and validation plans.

Mistake 9: Ignoring channel overlap

Delivery, pickup, drive-thru, in-store, and appointment channels can transfer differently.

Mistake 10: Approving sites one at a time

Cannibalization is a portfolio effect. Site approval should include the network.

18. FAQ

What is cannibalization analysis in site selection?

Cannibalization analysis estimates how much demand a proposed new location will pull from existing same-brand locations. It separates gross candidate demand from net-new network growth.

What is retail cannibalization?

Retail cannibalization occurs when a new store captures sales, visits, or customers that would otherwise have gone to an existing store in the same retail network.

Is trade area overlap the same as cannibalization?

Trade area overlap shows shared geography or shared demand exposure. Cannibalization measures demand transfer from existing locations to the candidate.

What is a good cannibalization rate?

As a planning heuristic, many operators pressure-test candidate sites in the 15% to 35% cannibalization range, then adjust by category, margin, capacity, market strategy, and affected-store health. There is no universal threshold. A 10% rate can be too high for a weak or low-margin market, while a higher rate may be acceptable for capacity relief, delivery coverage, or competitive defense.

How do you calculate cannibalization rate?

Cannibalization rate =
Transferred demand from existing locations / Candidate location demand

If a candidate is expected to generate $2M and $500K transfers from existing stores, the cannibalization rate is 25%.

What is net-new demand?

Net-new demand is the portion of candidate demand that expands the network instead of shifting from existing same-brand locations.

Why are radius-based cannibalization models weak?

Radius models ignore roads, barriers, congestion, travel time, store attractiveness, competitors, customer intent, daypart, and channel behavior. They are useful for screening, not final approval.

How do Huff models help with cannibalization analysis?

Huff models estimate the probability that demand from an origin chooses each available location based on attractiveness and travel cost. Comparing allocation before and after a candidate enters the market helps estimate demand transfer.

Can cannibalization be good?

Yes. Cannibalization can be acceptable when it improves convenience density, relieves capacity, defends a strategic market, improves delivery coverage, or creates network efficiencies.

What data is needed for cannibalization analysis?

Useful data includes existing store performance, customer origins, demographics, workplace population, competitor locations, mobility or visit data, trade areas, delivery zones, daypart sales, and road-network travel times.

How should cannibalization be presented to executives?

Show candidate demand, transferred demand, net-new demand, affected stores, largest store impact, daypart and channel risk, assumptions, confidence level, and recommendation.

19. Glossary

Cannibalization

Demand captured by a new location that would otherwise have gone to an existing location in the same network.

Cannibalization rate

The share of candidate demand expected to transfer from existing same-brand locations.

Net-new demand

Demand that expands the network rather than shifting from existing units.

Sales transfer

Revenue that moves from existing locations to a new location.

Trade area

The geography from which a location draws customers, patients, members, visits, or orders.

Trade area overlap

The shared portion of two or more locations’ trade areas.

Demand-weighted overlap

Overlap measured by relevant demand, such as population, footfall, spending, visits, patients, or customers.

Drive-time trade area

A catchment based on travel time through a road network.

Isochrone

A polygon showing the area reachable from a location within a specific travel time.

Huff model

A probabilistic spatial interaction model that estimates location choice based on attractiveness, travel cost, and competing alternatives.

Gravity model

A model where interaction between places increases with attractiveness or mass and decreases with distance or travel friction.

Distance decay

The tendency for visitation or interaction probability to decline as travel cost increases.

Market saturation

A condition where additional locations produce declining incremental returns because the market is already well served.

White space

An area with attractive demand that the current network does not serve effectively.

Portfolio optimization

The process of selecting, sequencing, relocating, or closing locations to maximize total network value.

Conclusion

Cannibalization analysis turns site selection into a portfolio decision record.

The final brief should name the worst-affected existing location, quantify its expected baseline impact, separate gross candidate demand from net-new network value, and define the post-opening review window before the lease is signed.

Six months after opening, the operator should compare the model against same-store sales, customer-origin shifts, daypart performance, and channel mix. That review is what turns a one-time approval into a better portfolio model for the next decision.

References

• CARTO: Cannibalization Analysis: what is it and how can I use it?
• Sweetgreen 2024 Form 10-K
• CARTO Academy: Store cannibalization tutorial
• ArcGIS Pro: How Huff Model works
• Validating Gravity-Based Market Share Models Using Large-Scale Transactional Data
• Multi-Facility Location Models Incorporating Multipurpose Shopping Trips
• ArcGIS Pro: Huff Model tool documentation
• Calibrating the dynamic Huff model for business analysis using location big data
• NBER: The Diffusion of Wal-Mart and Economies of Density
• Causal Analysis on the Anchor Store Effect in a Location-based Social Network
• U.S. Census Bureau: American Community Survey
• U.S. Census Bureau: Longitudinal Employer-Household Dynamics data