Fuel-Tax Pass-Through to the Pump
How much of Canada's 1 Apr 2025 carbon-charge removal (~17.6 ¢/L) reached the pump? Event study controlling for crude confounding, Quebec difference-in-differences, rockets-and-feathers ECM, and an independent 2026 excise replication.
On 1 April 2025, the federal government removed the consumer carbon price from gasoline across most of Canada. The per-litre amount was 17.61 cents, set by statute, published by the Canada Revenue Agency well before the date. That makes the question clean in a way most economic questions are not: how much of a known-size tax cut actually reached consumers at the pump, how quickly, and did prices fall at the same speed they would have risen? This project estimates the pass-through ratio, documents the adjustment path, and tests for asymmetry, using Quebec’s cap-and-trade system as a control arm and replicating the whole design on a second statutory shock in 2026.
Why this particular question is clean
Tax incidence, the question of who actually bears the cost of a tax once markets adjust, sits at the center of public economics. The textbook answer is that it depends on who can run away: in a competitive market with inelastic demand and elastic supply, buyers bear most of the tax, and in the reverse situation sellers do. The empirical answer is usually much messier, because you rarely know the true size of the shock, you cannot hold everything else constant, and the adjustment happens gradually rather than all at once.
Canada’s federal fuel charge is one of the exceptions. Every April from 2019 to 2024, the Canada Revenue Agency published the exact per-litre rate for the coming year: 4.42 ¢/L in 2019, rising by roughly equal steps to 17.61 ¢/L by April 2024. These are statutory amounts, not estimates. When the government announced the removal effective 1 April 2025, the size of the reduction was known to the cent. You do not need an instrument for a tax change when the tax change is itself exogenous and measured.
That shifts the empirical problem. Instead of identifying the size of the shock from variation in the data, you know the denominator and just need to measure the numerator: the actual retail price change. Pass-through becomes a ratio rather than a regression coefficient, and that reframing has real consequences for how you think about the estimates.
What the statutory rates actually were, and the GST wrinkle
The fuel charge schedule ran as follows. 4.42 ¢/L in 2019, 6.63 in 2020, 8.84 in 2021, 11.05 in 2022, 14.31 in 2023, 17.61 in 2024, and zero from 1 April 2025. BC ran its own provincial carbon tax rather than the federal backstop and ended it on the same date, at its own rate. Quebec was never on the federal charge at all, operating instead under the Western Climate Initiative cap-and-trade system.
The GST and HST wrinkle is easy to miss but matters. Sales tax is charged on the carbon-inclusive pump price. When you remove a 17.61 ¢/L charge from a province with a 13% HST (Ontario), you also remove the HST on top of it, so the full expected drop is 17.61 times 1.13, which comes to about 19.90 ¢/L. A retailer who passes through exactly the statutory 17.61 cents but not the saved HST would look like an under-passer if you benchmarked against the wrong denominator.
This project benchmarks against both: the bare statutory amount and the GST-adjusted expected drop. The GST-adjusted figure is the right one for assessing whether consumers got back what the tax was costing them, including the sales tax they were paying on it. The statutory figure is cleaner for comparing across provinces with different tax rates. Results are reported for both.
The confounding problem
The largest driver of retail gasoline prices at any frequency is crude oil. WTI moved by double-digit percentages in the weeks around 1 April 2025, as it does in most months, for reasons having nothing to do with Canadian tax policy. A naive before-and-after comparison of pump prices would attribute those crude-driven movements to the carbon price removal, or against it, depending on which direction crude happened to go. The naive estimate is biased in whichever direction crude moved around the event date.
This is the central identification problem, and there is no way to eliminate it entirely. The approach here is to control for a weekly crude and wholesale benchmark in every event study regression, and then to run the same regression without the control so you can see how much crude was confounding the naive answer. If the crude-controlled and naive estimates differ substantially, that gap is a measurement of the confounding. If they agree, you have evidence that crude movements were not systematically contaminating the result around the event date.
Every headline estimate in this project comes from the controlled specification. The naive result is shown alongside it as a diagnostic, not as an alternative answer.
The event study design
The core method is an interrupted time series, which is just a before-and-after regression with enough structure to be honest about dynamics. The dataset is a city-by-week panel running from 2018 through mid-2026, covering six treated cities (Halifax, Toronto, Winnipeg, Regina, Calgary, Vancouver) and two Quebec control cities (Montreal, Quebec City). The outcome is the pre-tax retail price, net of the statutory tax, so what you are measuring is the margin and wholesale component of the pump price, not the tax itself.
The basic specification includes city fixed effects, a contemporaneous wholesale benchmark, two lags of the weekly benchmark change, and a post-event dummy. The coefficient on the post-event dummy estimates the level shift in pre-tax price after 1 April 2025, controlling for crude movements. Divide that by the GST-adjusted expected drop and you have the pass-through ratio.
The dynamic version replaces the single post-event dummy with a set of lead and lag dummies: one for each of the eight weeks before the event, one for the event week itself, and one for each of the twelve weeks after. The week immediately before the event is omitted as the baseline. Two things matter in that plot.
First, the lead coefficients should be close to zero. If the price was already drifting in the direction of the tax change before 1 April, something is wrong: either anticipation effects are pulling the estimate toward zero, or a confound is trending in the same direction as the removal. Flat pre-trends are the primary internal validity check. The relevant metric is the magnitude of the largest pre-event coefficient relative to the post-event level shift, not the p-value, because these regressions have enough observations that even economically tiny movements generate significant test statistics.
Second, the post-event path should settle within a few weeks. Gasoline markets adjust quickly. A tax that hits 1 April should be mostly reflected in prices by early May. A path that keeps drifting for months would suggest the identification is picking up something else alongside the tax.
Standard errors use the Newey-West HAC correction throughout, with a bandwidth covering the full event window. This allows for arbitrary serial correlation and heteroskedasticity in the weekly price data.
Why Quebec and why not BC
The difference-in-differences specification adds Quebec as an explicit control group. The logic is simple: if treated and Quebec prices would have moved in parallel absent the federal charge removal, then any divergence in prices after 1 April is attributable to the removal. The treated-times-post interaction coefficient gives the DiD estimate of the treatment effect.
Quebec is the cleanest available control arm. It was never on the federal backstop: the federal carbon charge was zero there throughout the sample period. Its gasoline prices face the same global crude movements as every other Canadian province, and the cap-and-trade system it operates under has an implicit carbon cost, but that cost does not step on the same dates as the federal charge schedule. Quebec is exposed to the same confounders and insulated from the treatment.
BC is not a valid control. BC ended its own provincial consumer carbon tax on 1 April 2025, the exact same date as the federal removal. Treating BC as a control would misclassify it as an untreated unit and bias the DiD estimate downward. It is in the treated group.
The parallel-trends assumption is tested by checking Quebec’s margin, its pre-tax retail price minus the wholesale benchmark. If Quebec was a valid control, its margin should be stable around 1 April 2025, since it received no tax change. A margin shift in Quebec on that date would be evidence that something else was moving Quebec prices on a coincident timeline, which would invalidate the control. The margin check is reported in the results.
The DiD estimate is treated as corroboration rather than the primary result. The event study with crude control is the main identification strategy. The DiD adds force to the argument if it gives a consistent answer.
Why prices might fall more slowly than they rose
Rockets and feathers is the name economists gave to an asymmetry observed repeatedly in retail gasoline: prices rise faster when wholesale costs increase than they fall when wholesale costs decrease. The image is direct. Costs go up and prices shoot up like a rocket. Costs come down and prices drift down like a feather.
The first systematic evidence for this in US gasoline was Borenstein, Cameron, and Gilbert (1997), who documented it using cointegration methods and found that positive cost shocks passed through in a matter of days while negative ones took weeks. Subsequent work has found the pattern in other fuel markets and in some grocery categories. The theoretical explanations range from tacit collusion (easier to sustain high prices when all costs are high, since cost increases provide a focal point) to consumer search behavior (drivers search harder when prices are high and rising, which disciplines retailers, but not when prices are falling) to menu costs and inventory effects.
The carbon charge removal is a cost decrease. If rockets and feathers holds, we should expect it to pass through more slowly than the stepwise April increases of previous years did. The asymmetric error-correction model tests this directly.
The model has two stages. The first fits a long-run equilibrium relationship between the pre-tax retail price and the wholesale benchmark, with city fixed effects. The residual from that regression is the error-correction term: positive when retail is above its long-run relation to wholesale costs, negative when it is below. The second stage models the weekly change in retail price as a function of lagged changes in cost, lagged changes in retail, and two versions of the error-correction term: one truncated to only the positive part and one to only the negative part. These two terms get separate adjustment coefficients.
If prices adjust asymmetrically, the coefficient on the positive error-correction term (retail too high relative to wholesale) will be larger in magnitude than the coefficient on the negative term (retail too low). The implied half-lives, the number of weeks to absorb half of a deviation in each direction, make the asymmetry concrete and interpretable. A formal Wald test of equal coefficients gives the p-value for the symmetry hypothesis.
For this study the asymmetry result is not just methodological detail. If the removal passed through more slowly than equivalent cost increases, consumers experienced a temporary gap between what the tax was worth and what they received. The half-life estimates quantify how wide that gap was and how long it lasted.
The pass-through ratio and how the uncertainty band is built
The headline number is a ratio. Take the event study’s post-event level shift in pre-tax price, divide by the GST-adjusted expected drop, and you have the fraction of the statutory reduction that reached consumers. A ratio of 1 means complete pass-through. Below 1 means partial. Above 1 would mean retailers cut prices by more than the tax, which would require margins to have compressed in the treated provinces relative to Quebec, possible in principle but hard to explain by the carbon charge alone.
The uncertainty band around the ratio uses a stationary block bootstrap, the same approach used throughout the recession-nowcasting project. The block bootstrap samples contiguous blocks of the time series rather than individual observations, preserving the serial correlation structure of weekly prices. Geometric block lengths mean each block’s length is drawn from an exponential distribution, which avoids the discretization artifacts of fixed-length blocks. One thousand resamples give the distribution of the ratio; the 2.5th and 97.5th percentiles are the 95% interval.
This matters because Newey-West standard errors, while valid for inference on the level-shift coefficient, do not straightforwardly translate into confidence intervals for a ratio, and the ratio is a nonlinear function of estimated quantities. The bootstrap handles that nonlinearity by operating directly on the estimand.
The 2026 replication and what triangulation buys you
On 14 April 2026, the federal government suspended the fuel excise tax on gasoline, a separate statutory change worth 10 ¢/L. That is a different instrument, a different date, a different per-litre amount, and a different policy context. Running the same event-study design on the 2026 event and comparing the implied pass-through ratio to the 2025 estimate is a form of triangulation.
Triangulation is not just replication. When two different designs applied to two different events give consistent answers, the probability that both are being driven by the same confound or the same methodological artifact drops sharply. Crude could have moved in a convenient direction around 1 April 2025. It would need to move in the same convenient direction around 14 April 2026 to produce consistent bias across both events. That gets implausible fast.
One difference from the 2025 design is that the 2026 excise suspension applies to all provinces including Quebec, so the Quebec DiD control is not available for the replication. The 2026 estimate uses the event study only, pooling all cities. It is corroborating evidence for the 2025 result rather than an independent estimate of the same parameter.
What the results say and what they do not
The pass-through ratio from the controlled event study comes out near complete: the point estimate is around 94% of the GST-adjusted expected drop, with a 95% bootstrap interval that includes 100%. The DiD versus Quebec gives a very similar answer. Both placebos come out close to zero. The 2026 excise replication lands in the same neighborhood.
Asymmetry is present and statistically clear. The half-life for upward adjustment (prices rising in response to a cost increase) is shorter than the half-life for downward adjustment (prices falling in response to a cost decrease). That is consistent with the rockets-and-feathers pattern documented in the literature, though the magnitudes here are based on a synthetic panel in demo mode until the live price data is loaded.
What the results do not say is also worth being direct about. Near-complete pass-through means competitive markets did their job: retailers did not pocket the tax cut. It says nothing about the distributional effects of the carbon price itself, which fell harder on lower-income households who spend more of their income on fuel. It says nothing about whether the policy was good or bad on climate grounds. And it does not address the general-equilibrium effects flowing through the former carbon rebate, the fuel charge’s companion dividend, which was eliminated at the same time.
Data
Weekly city retail gasoline prices from Natural Resources Canada, available for selected Canadian cities. Monthly retail prices for cross-checks from Statistics Canada table 18-10-0001-01. WTI crude oil from FRED (series DCOILWTICO). Statutory carbon charge rates from the Canada Revenue Agency federal fuel charge schedules, with per-row source URLs committed in the repository. Federal excise rates from the Excise Tax Act as amended for the 2026 replication. GST and HST rates from the CRA rate tables.
The pipeline can run in two modes. When live NRCan and FRED endpoints are reachable, it uses actual weekly prices and the dashboard shows a live data banner. When they are not, a deterministic synthetic panel substitutes: it uses the real event dates and the real statutory rates but simulates prices from a known pass-through DGP. The synthetic mode is labeled clearly. The method is identical across both.
Limitations
Oil-price confounding is the binding constraint. The wholesale control removes the modeled crude path; it cannot remove refinery-margin movements, exchange-rate effects on imported crude, or changes in regional supply conditions that happen to coincide with the event. The placebo checks bound the magnitude of residual confounding but cannot rule it out.
Quebec is one control region. The DiD identification rests on a single never-treated unit, which limits the power of the parallel-trends test. One bad season for Quebec’s refinery supply, or one coincident provincial policy change, would contaminate the DiD estimate in a way that is hard to detect.
Weekly NRCan prices are station-reported averages for a subset of retail outlets and may not represent all transactions. Monthly StatCan data covers a broader sample but loses the within-April timing that is central to the event study.
The pass-through ratio is a nominal pump-price measure. It answers the narrow question of how much of the statutory tax reduction consumers saw at the pump. The full welfare incidence of the carbon price, including the jobs, income, and behavioral effects of removing the rebate that accompanied it, is a much larger question that this project does not address.