Methodology — D.C. Robotaxi Policy Sandbox

Tier 1 prototype, v0.1 The Innovation Majority Institute · Draft, not for public release

1. Purpose and limits

This document maps every parameter and lever in the D.C. AV Policy Sandbox to its academic or empirical source. It is the audit trail a skeptical reviewer should be able to follow before treating any output as more than illustrative. We will continue to calibrate this sandbox and welcome any feedback on our methodology. Contact us at mobilitymayor@imajority.institute

Three things this document is not:

1. A forecast. The sandbox expresses the directional logic of regulatory tradeoffs. Output numbers should not be cited as predictions for what would happen in D.C. under any specific rule set.
2. A peer-reviewed model. The closest published analogue is Gao & Li (2023), a Nash-equilibrium model of AV ride-hail regulation calibrated for San Francisco. We borrow structure and parameter ranges from that work but our implementation is materially simpler.
3. A case for or against any policy. The Innovation Majority Institute publishes this tool to inform debate. As a nonprofit, IMI analyzes policy but does not endorse legislation or argue for a predetermined outcome. Where our calibration leans in a direction, we say so in Section 6.

Where the literature is mixed, we have erred toward middle-of-distribution values. Where the literature is thin, we have used industry data or our own judgment and flagged it.

2. Model structure

The sandbox simulates a steady-state equilibrium for a typical weekday in the District. Eight wards. Two regulated operators by default (variable). Demand and supply settle to a market equilibrium each ward, conditional on regulatory levers.

The basic logic per ward:

1. Effective rider fare is the baseline AV fare plus a share of stacked city fees, after partial passthrough.
2. Demand in each ward is a function of fare (with constant elasticity) and an intrinsic ward demand share.
3. Operator margin per trip in each ward is fare minus operating cost (vehicle + fees + parity cost where applicable). Operating cost varies by ward via a density / spread factor.
4. Fleet allocation across wards:
   • Without wait-time parity: vehicles distributed proportional to (demand × max(margin, 0)). Wards with negative margin get zero vehicles.
   • With wait-time parity: vehicles distributed proportional to demand share. Operators absorb a per-trip "parity cost" reflecting the repositioning overhead.
5. Operator viability: deployed fleet scales 0→1 with weighted-average margin. Below a threshold, operators decline to deploy at scale.
6. Per-ward wait time: inverse of local supply/demand ratio, multiplied by the ward's spread factor (capturing the fact that lower-density areas have longer pickup distances per vehicle).
7. Per-ward service level: log-decay function of wait time, floored at 0.05 wherever vehicles are actually deployed.

This is a simplification of the Gao & Li (2023) Nash-equilibrium framework. Two simplifications worth naming up front:

• We do not endogenize the operator's pricing response to regulation. Operators are price-takers on our fee structure; we do not let them re-optimize fares against a vehicle cap, for example. This understates the negative consequences of supply-restricting regulation (see Section 6, item 1).
• We treat the city as eight zones, not a continuous network. This loses the multimodal transit-substitution dynamics Gao & Li capture explicitly. We are downstream of public transit; this matters for equity analysis (Section 6, item 4).

3. Baseline parameters

4. Per-ward parameters

We assign each ward two parameters: a demand share (probability mass of citywide demand originating in that ward) and a cost factor (proxy for density and spread).

Demand shares are calibrated from publicly available D.C. taxi pickup data (DFHV / D.C. Open Data) and Schaller's TNC estimates, adjusted for ward-level population and employment density. (D.C. ride-hail trip data is reported to DFHV but, unlike taxi data, is not published at ward granularity — see Section 6, item 10.) Two known weaknesses: (a) D.C. has not published validated ride-hail origin-destination data at the ward level, so these are inferences from related sources; (b) the AV-addressable share within each ward may diverge from the human ride-hail share — early AV demand likely skews even more toward Wards 2, 6, 1.

Cost factors are the most opinionated single piece of the model. They serve two purposes simultaneously: a multiplier on per-vehicle operating cost (lower-density wards have higher cost-per-trip due to longer pickup distances and lower utilization) and a multiplier on wait time (more spread = longer per-vehicle response). Using one parameter for both is a simplification. In principle these should be separately calibrated; in practice we have not seen ward-level operating-cost data for any U.S. AV operator. See Section 6, item 3.

4.5 Income distribution per ward

For equity decomposition (Section 4.6) we assign each ward a three-class income distribution: Low (below ~60% AMI), Medium (60–120% AMI), High (above 120% AMI). Values are rough estimates from ACS 5-year tabulations; DDOT or MWCOG ward-level breakdowns would replace them in Tier 2.

The shape of the distribution matches D.C.'s known income geography (highest income in Wards 2 and 3, lowest east of the Anacostia) but the exact split is approximate. The model is sensitive to these values primarily through the social equity calculation.

4.6 Equity decomposition

Following Gao & Li (2023), we decompose transport equity into two components:

• Spatial equity — how evenly is access distributed across wards?
• Social equity — how evenly is access distributed across income classes?

For each (ward × income class) pair we compute an access score:

access(ward, class) = service_level(ward) × affordability(class)
affordability(class) = max(0, 1 − estFare / fare_threshold[class])
fare_threshold = { low: $25, med: $45, high: $100 }

Service level is the per-ward value already computed in the operator-equilibrium model (0–1, derived from wait time). Affordability captures the class-specific fare sensitivity following Gao & Li's class-weighted cost parameter — low-income riders are effectively priced out as fares approach $25/trip; medium-income at $45; high-income at $100.

Per-ward and per-class means are weighted by demand_share × class_population_share. We then compute the coefficient of variation across wards (CV_spatial) and across classes (CV_social) and map each to a 0–100 score:

spatial_equity = max(0, min(100, (1 − CV_spatial) × 100))
social_equity  = max(0, min(100, (1 − CV_social)  × 100))

Limitations of this approach: Coefficient of variation is a simpler dispersion measure than the Theil index used by Gao & Li. CV is more interpretable but doesn't decompose cleanly into within-group and between-group components the way Theil does. Tier 2 should adopt the full Theil decomposition with proper utility weighting from the multinomial-logit accessibility measure.

5. Policy lever logic and evidence

5.1 Vehicle cap (per operator)

Lever: Maximum AVs each permitted operator may deploy. Citywide fleet = cap × operators.

Model behavior: Citywide deployed fleet is bounded by cap × operators × deployed_fraction. When the cap binds, per-ward capacity falls, supply/demand ratio in each ward worsens, wait times rise, fewer trips are served. Fleet does not redistribute toward higher-margin wards under the cap binding because allocation logic is independent of total fleet size.

Evidence base:

• Castillo et al. (2019), "Regulating TNCs" finds that for human-driven TNCs, imposing a cap on the number of drivers actually hurts driver earnings because the platform retains the benefits of limiting supply by reducing driver pay. The platform captures the rent from limited supply, not the drivers.
• In an AV-only world the labor-market dynamic disappears (the platform owns the vehicles), but a similar dynamic applies through pricing: limited supply allows the platform to maintain or raise fares without competitive pressure. Our model does not capture this. We treat fare as exogenous to cap level. This understates the consumer-welfare cost of low caps. See Section 6, item 1.
• New York City's 2018 freeze on TNC vehicle counts is the most-cited real-world precedent. Mixed empirical results.

Calibration note: Our default cap of 800/operator × 2 operators = 1,600 citywide is below DC's likely steady-state demand at unregulated equilibrium. The "Tax like rideshare" preset (1,200/op × 4 ops = 4,800) is closer to a market-clearing fleet given our parameters.

5.2 Number of operators

Lever: How many companies are assumed to operate AV ride-hail in the District. This is a market assumption, not a regulatory cap: the Allen Bill limits vehicles per operator but does not cap the number of operators, so the count of companies that actually enter is uncertain. The slider lets you vary that assumption.

Model behavior: Each operator gets the per-operator cap. More operators = larger citywide fleet. At ≥4 operators we apply a 6% utilization haircut to reflect lost scale economies (each operator gets a smaller share of demand, leading to lower vehicle utilization).

Evidence base:

• Castillo et al. (2019) discusses platform competition in their Section 8.2 but does not derive comparative statics. They note that under competition, two platforms may split the heterogeneous passengers, with one offering higher fares and lower waiting times to passengers with higher reservation costs — a product differentiation outcome.
• Empirical evidence on TNC competition effects on fares is mixed. A recent (2026) market analysis of NYC ride-hail found Uber riders show modest price sensitivity while Lyft users are extremely sensitive — a $1 price increase on Lyft reduced demand by 64%, suggesting platform differentiation rather than pure price competition.
• Our 6% scale-economy haircut at high operator counts is judgmental; the actual relationship is not well-pinned-down in the literature.

5.3 Wait-time parity (response-time parity)

Lever: Toggle. When on, operators must allocate vehicles such that Wards 7 and 8 see wait times within ~10% of citywide average.

Model behavior: Allocation switches from (demand × margin) to demand share. Adds $1.20/trip to per-vehicle operating cost (repositioning overhead). Operators recover ≈70% of that cost through fares (≈$0.85/trip passed to riders). Applies an 18% utilization haircut.

Evidence base:

• Gao & Li (2023) model the analogous policy as a "minimum service-level requirement" — a maximum allowable wait time per zone. They find that this improves spatial equity (reduces geographic concentration of vehicles) but exacerbates social inequity, because the added vehicles in low-demand areas primarily benefit higher-income individuals in those areas while low-income individuals face higher trip fares.
• As of v3, our model captures this tradeoff via the equity decomposition (see Section 4.6) and the parity-driven fare passthrough. The "Tax heavily, cap tightly" preset (wait-time parity on, fees stacked) produces a visible asymmetry: spatial equity rises while social equity stays roughly flat or modestly worsens. This is the Gao & Li finding made directly visible to users.
• Real-world precedent: New York City requires that 90% of wheelchair-accessible vehicle requests be fulfilled within 15 minutes. Comparable response-time-parity rules exist for paratransit in many jurisdictions but have not been applied broadly to AV regulation.

5.4 Per-mile fee

Lever: City surcharge on every paid mile.

Model behavior: Added to rider fare (full pass-through assumed); added to operator's per-trip cost. Reduces demand via elasticity.

Gas-tax parity reference. The per-mile fee slider carries a reference marker at ≈ $0.0149/mi — the per-mile road-funding contribution a typical gasoline car makes through D.C.'s motor-fuel tax ($0.357/gallon as of October 2025, ÷ ~24 mpg real-world on-road light-duty fuel economy). Because AVs are electric and pay no fuel tax, a per-mile (VMT) fee is the closest equivalent mechanism for having them contribute to road upkeep; the marker lets users judge any VMT fee against today's effective gas-tax burden. For scale, the Allen Bill's $0.15/mi VMT fee is roughly ten times gas-tax parity. The ~24 mpg figure is a national on-road proxy (D.C.-specific fleet fuel economy is not separately published); using new-car economy (~33 mpg) would lower the marker to ≈ $0.011/mi.

Evidence base:

• Per-mile fees are unusual in current TNC regulation but exist for taxis in many cities. The Chicago ride-hail equity literature finds that variable fees produce better geographic equity outcomes compared to flat fares, based on analysis of 97 million ride-hail trips in Chicago.
• Our 100% rider passthrough assumption is consistent with the Chicago and Castillo et al. modeling but is an upper bound; in competitive markets some absorption by the platform is plausible. U.S. ride-hail demand is fairly inelastic (Cohen et al. 2016 estimate roughly -0.4 to -0.6), so our -0.4 elasticity sits at the low-sensitivity end of that range and, if anything, understates how much a fee suppresses demand.

5.5 Per-trip fee

Lever: Flat fee on every trip.

Model behavior: As per-mile, but applied per trip rather than per mile.

Evidence base:

• D.C. already has a per-trip TNC tax. The District raised it from 1% gross receipts to 6% in 2018 to fund Metro: the D.C. Council proposed raising the gross receipts tax on Uber, Lyft, and Via from 1% to 6%, projected to raise about $23 million annually for the transit agency, citing research that ride-hail contributes to congestion. In the tool, the per-trip fee is expressed as a percentage of the base ride fare (the gross-receipts basis), with a reference marker at the current 6% level. The default Allen Bill scenario sets it to this existing 6% tax (≈$0.90 on the $15 base) on top of the bill's new per-mile VMT fee — reflecting that AVs would pay the standard TNC fees in addition to any new bill-specific fee.
• The Chicago 97-million-trip equity study finds flat per-trip fees are regressive relative to per-mile fees: they hit short trips (more common in low-income neighborhoods) disproportionately hard as a share of fare. Our model does not capture this regressivity explicitly because we treat all trips as 3 miles. See Section 6, item 7.

6. Known weaknesses

These are flagged honestly because the most damaging critique of a policy simulator is one its authors should have anticipated and didn't.

1. Operator pricing is exogenous. Our model treats baseline fare as fixed and only modifies it via fee pass-through. In reality, an operator facing a binding vehicle cap or a congestion fee would re-optimize its base fare and per-mile rate. Castillo et al. (2019) show this matters: under a vehicle cap, the platform raises effective prices because it can. Our model undercounts the welfare cost of supply-restricting regulation by missing this channel.

2. Equity decomposition is CV, not Theil. Partially resolved in v3. Equity is now split into two scores: spatial (dispersion of access across wards) and social (dispersion across income classes). We use coefficient of variation as the dispersion measure rather than the full Theil index used by Gao & Li (2023). CV is simpler and more interpretable but doesn't decompose cleanly into within-group and between-group components. The "Tax heavily, cap tightly" preset shows the Gao & Li pattern — spatial equity rises, social equity flat — but the magnitudes would shift under a proper Theil implementation. Tier 2 fix.

3. Ward cost factor does double duty. A single costFactor parameter modulates both operator cost-per-trip and rider wait time in each ward. These are related (density affects both) but not identical. Empirically separating them would require ward-level operating-cost data we don't have.

4. Transit is absent. Gao & Li (2023) explicitly model the AV-transit interaction and find that the most equity-positive policy is taxing AV trips to subsidize transit in underserved zones for low-income riders, replicating NYC and SF precedents. To provide a simple and legible model, this does not have a transit channel and the implicit assumption is that AV ride-hail is the only mobility option being regulated.

5. Baseline fare is judgmental. Our $15 sits between current D.C. Uber/Lyft pricing for a 3-mile trip and current Waymo SF/LA pricing for the same. Lower than current Waymo, higher than current D.C. ride-hail. DFHV (2024) staff fare estimates corroborate the low end — a 3-mile D.C. TNC trip ran roughly $13.60 (base) to $31.50 (with surge), placing $15 just above the observed floor. (Those TNC figures are DFHV's own point-in-time estimates from a fare-estimator site, not operator-reported data.) Defensible but soft.

6. Maximum utilization (18 trips/day) is conservative — and it is the single most consequential supply-side assumption in the model. This one number does double duty: it sets how many trips each vehicle can serve and, through cost-per-trip, whether operators find it viable to deploy at all. So it moves the supply-constrained scenarios more than any other parameter. Waymo's late-2025 operations imply mature-market utilization closer to 30–40 trips/vehicle/day; we deliberately use a lower early-deployment figure, which also implicitly absorbs the high deadhead and offline time discussed in item 11 — rather than docking for those separately and double-counting the same empty-driving penalty. The tradeoff is worth stating plainly: holding the policy levers fixed, raising utilization to a mature ~40 would roughly double-to-triple the trips served in a cap-binding scenario. The Allen Bill, for instance, moves from serving about 8% of addressable trips to roughly 40%, with average waits more than halving — on identical rules. Our conservative choice therefore biases the model toward predicting tighter service, and possible collapse, under aggressive regulation. We prefer to err in that direction — understating AV capability rather than overstating it, consistent with this being an early-deployment model — but a reader should know that a steady-state, mature-operations calibration would tell a materially more optimistic supply story under the same policies.

7. All trips are 3 miles. Real trip distance is heavily right-skewed and varies by ward. This matters for evaluating per-trip vs per-mile fees, since the equity properties differ across the trip-length distribution. The Chicago 97-million-trip study makes this explicit, and DFHV (2024) shows the same skew locally — 60% of D.C. taxi trips are ~2 miles and 85% are under 5.

8. Fee pass-through ratios are guesses. Our 100% pass-through for per-trip and per-mile fees is not grounded in the TNC pricing literature. There is academic work on this we have not yet integrated.

9. No dynamic effects. This is a steady-state weekday model. We do not simulate peak/off-peak demand variation, weather, special events, or growth/contraction over multi-year horizons.

10. D.C.-specific ride-hail data is collected but not published at the granularity the model needs. This is the question we get most often, so to be explicit about why: taxis and ride-hail (TNCs) in D.C. are regulated by the Department of For-Hire Vehicles (DFHV), not DDOT, and operator trip data is reported to DFHV. But DFHV does not publish ride-hail trip volumes at ward or origin-destination granularity. What it releases publicly is aggregate and rate-structure material — for example the 2024 trip-distance distribution and fare estimates we cite in Section 3 — not ward-level O-D counts. DDOT and MWCOG have likewise not published validated ward-level ride-hail O-D data. Because the granular D.C. numbers are held but not released, most of our D.C. calibration is inference from jurisdictions that do publish trip-level data (NYC's TLC, SFCTA, and Chicago's open ride-hail dataset), adjusted for D.C.'s demographics and geography. If you know of a D.C. source we have missed, we would welcome it at mobilitymayor@imajority.institute.

11. Fleet downtime is not modeled as a separate factor. A meaningful share of any electric robotaxi fleet is out of service at any given moment — charging, cleaning, repositioning, and maintenance. Operators do not disclose the figure, and the CPUC's P0 ("offline, not available for rides") category is withheld under confidentiality in public quarterly filings, which report only the in-service periods (P1–P3). We therefore do not model availability as its own dial; it is absorbed implicitly into our conservative maximum utilization of 18 trips/vehicle/day (item 6), which sits well below mature-market figures partly because of this offline time. Distinct from offline downtime: our deadhead ratio of 0.35 — the empty share of vehicle-miles driven by in-service cars — is low against observed data. CPUC reports through December 2025 show Waymo's empty VMT around 44–46% of total mileage in California (deadheading fell from 51.5% in January 2024 to 44.3% by September 2025). A future revision should raise the deadhead baseline toward ~0.45 and, if availability is ever disclosed, model offline share explicitly.

7. References

Castillo, J.C. (2019). "Who Benefits from Surge Pricing?" Working paper, Stanford University.

California Public Utilities Commission (2023–2025). "Autonomous Vehicle Programs — Quarterly Reporting." Per-period (P0–P3) VMT and passenger data reported by AV operators; the P0 "offline" category is withheld under confidentiality (General Order 66-D). Source for the empty-VMT / deadhead figures in Section 6, item 11.

Cohen, P., Hahn, R., Hall, J., Levitt, S., and Metcalfe, R. (2016). "Using Big Data to Estimate Consumer Surplus: The Case of Uber." NBER Working Paper 22627. Primary U.S. source for ride-hail price elasticity: ~50 million UberX trips across Chicago, Los Angeles, New York, and San Francisco.

District of Columbia Office of Tax and Revenue (2025). "Motor Vehicle Fuel Tax." D.C. motor-fuel tax of $0.357/gallon effective October 1, 2025 (23.5¢ tax + 12.2¢ local transportation surcharge). Basis for the per-mile fee's gas-tax-parity marker (Section 5.4).

Federal Highway Administration / U.S. DOE (2023). "Average Fuel Efficiency of U.S. Light-Duty Vehicles" (Highway Statistics, Table VM-1; BTS). Source for the ~24 mpg real-world on-road fuel economy used in the gas-tax-parity calculation (Section 5.4).

District of Columbia Department of For-Hire Vehicles (2024). "Review of Rate Structure and Decision Pursuant to D.C. Official Code § 50-301.17." March 4, 2024. Source for the D.C. taxi trip-distance distribution (85% under 5 miles, 60% ~2 miles) and DFHV staff estimates of D.C. TNC fares by distance (Section 3).

DC Sustainable Transportation / Nelson\Nygaard (2023). "DC Decongestion Pricing Study, Final Report." Figure 31 (DC-wide daily vehicle entries, 672,747) and Scenario 5 (pass-through share); basis for the ~588,000 midday vehicles-present estimate (Section 3). DC DMV (~288k) and FHWA Highway Statistics 2023 Table MV-1 (~344k) supply the registered-fleet figure.

Gao, J. and Li, S. (2023). "Regulating For-Hire Autonomous Vehicles for an Equitable Multimodal Transportation Network." arXiv:2301.05798. Closest published analogue to our model.

Li, S., Tavafoghi, H., Poolla, K., and Varaiya, P. (2019). "Regulating TNCs: Should Uber and Lyft Set Their Own Rules?" arXiv:1902.01076. Foundational source for our pricing and cap logic.

Mark, B. and Tarduno, M. (2021). "Not all fees are created equal: Equity implications of ride-hail fee structures and revenues." Transport Policy. Chicago 97-million-trip equity study.

NYC TLC (2018). "Driver Income Rules." New York City Taxi and Limousine Commission.

Schaller, B. (2017, 2018, 2021). Various reports on TNC VMT and pickup/dropoff impacts in NYC and U.S. metros. Primary source for deadhead ratios and VMT impacts.

SFCTA (2017, 2018). "TNCs Today" and "TNCs and Congestion." San Francisco County Transportation Authority. Primary source for SF-specific TNC data we extrapolate from.

Tarduno, M. (2021). "The congestion costs of Uber and Lyft." Journal of Urban Economics.

Waymo (2025). Various public materials and analyst estimates from Sacra, Contrary Research, and trade publications, on vehicle costs and operational scale.