In this paper, I investigate a central policy question in the electric vehicle market: does a tiered subsidy design that rewards higher battery energy density actually shift manufacturers’ product technology choices? The Chinese electric vehicle (EV) purchase subsidy program offers a natural and important case. Since 2016, the central government has provided purchase subsidies to electric vehicle models that are included in an official catalog. The subsidy amount is first determined by the model’s range under the Chinese test cycle. On top of this range-based schedule, the payment is multiplied by an adjustment coefficient that increases with the gravimetric energy density of the battery pack. This two-layer structure creates strong incentives for automakers to choose batteries of different energy densities. My objective is to evaluate whether this adjustment coefficient design changes the battery technology that automakers select when they launch new EV models, and to compare this design with alternative subsidy schedules.
In contrast to most existing evaluations of demand-side subsidy policies, I focus on supply-side responses. I develop a profit-maximization framework with discrete choices over battery technologies. Using a structural approach, I estimate the contribution of each battery technology to the profit of an EV model. Then, with the estimated parameters, I conduct counterfactual simulations. I compare the technology choices that would have been observed under a uniform subsidy multiplier with those observed under the actual tiered adjustment multiplier. I find that tiered subsidies raise the number of newly launched models using battery packs with energy density above 160 Wh/kg by more than 14 percent compared to a uniform schedule. Yet the average energy density of newly launched models rises by only about 1 percent. The reason is that most automakers choose batteries that just meet the threshold of the highest subsidy tier, rather than going far beyond it. In additional counterfactual exercises, I show that if the multiplier were quasi-linearly related to the energy density range, the average energy density could increase by about 10 percent. This evidence provides a new perspective on how the details of subsidy policy design matter for technological progress in the electric vehicle sector.
1. Introduction
Governments around the world use subsidies to promote products that generate positive externalities. Electric vehicles are a prominent example. Because EV adoption reduces greenhouse gas emissions and local air pollution, many countries have implemented large-scale purchase subsidies. Yet policy makers constantly face a trade-off between fiscal costs and environmental benefits. A popular way to enhance the cost-effectiveness of subsidies is to tie the subsidy level to product attributes that are correlated with social benefits. For example, consumers may receive greater subsidies when they purchase EVs with longer range or more efficient components. In China, the central government has gone further and differentiated the subsidy by the energy density of the battery pack, which is a technological attribute of the product. Such “attribute-based” or “tiered” subsidies are becoming common in other sectors as well. Evaluating whether these designs can push firms to improve technology is an urgent research question.
Most empirical studies on attribute-based subsidies concentrate on consumer behavior. However, firms can respond strategically to such policies through product design and technology choices. If the tiered subsidy is designed well, it should raise the aggregate technology level of new products; if it is poorly designed, firms may choose only the minimum attribute level required to move to the next subsidy tier. China’s EV subsidy scheme provides a good laboratory because the subsidy adjustment coefficient has clear cutoffs at 120, 140, 160 Wh/kg in different years. Thus I can observe whether firms’ choices cluster at these cutoffs and judge whether the design is effective.
There are several reasons why answering this question is difficult. First, firm-level costs of battery technologies are generally treated as trade secrets. Without knowing the production cost of a battery, the observed adoption pattern could reflect either technology cost improvements or policy incentives. Second, because all EV models face the same subsidy policy, there is no natural control group in the usual reduced-form sense. Third, the policy changed over time while battery technologies and consumer preferences also evolved, making it hard to separate the policy effect from a time trend. I therefore adopt a structural discrete-choice model. By assuming that firms maximize expected profit and that the unobservable profit shocks follow a type-I extreme value distribution, I can infer the marginal contribution of each battery technology from observed market shares of different battery choices within each range category.
My data cover all pure electric passenger vehicles included in the Chinese recommended-vehicle catalog between 2017 and 2021. This is essentially the universe of models that are eligible for central subsidy. I merge these data with detailed model information from an online database to verify the main parameters. I define a battery “type” as the combination of an energy-density interval and a chemistry route: ternary lithium, lithium iron phosphate (LFP), or other. I then measure the market share of each battery type among new vehicle models in the same range interval and in the same batch of the catalog. The estimating equation is derived from a random-coefficient logit model, in which the difference between the log share of battery type j and that of a baseline battery is regressed on the subsidy differential, the actual range differential, battery-type-by-year fixed effects, range fixed effects, and batch fixed effects.
The baseline estimate implies that an additional 10,000 yuan (approximately USD 1,500) of subsidy increases the relative profitability of a given battery selection substantially. The magnitude is economically meaningful. More importantly, the battery-type-by-year fixed effects reveal a clear dynamic pattern. In 2017, low-density batteries contributed more to model profits than high-density batteries did. By 2021, however, the relationship had reversed: high-density batteries contributed more to profit, reflecting both falling costs of high-density products and increasing consumer demand for longer-range EVs.
I then use the estimated profit contributions to compute the technology choices that firms would have made under a uniform adjustment coefficient, that is, if the subsidy were identical for every battery type. Given the total subsidy outlay is kept unchanged, the uniform schedule distorts less in favor of high energy density. The counterfactual results show that the actual tiered schedule causes more models to use batteries in the 160–170 Wh/kg interval, which is the interval that receives the highest adjustment coefficient in the 2019–2021 period. The number of such models increases by more than 14 percent relative to the uniform case. On average, however, the actual policy raises the average energy density of newly launched models by about 1 percent only. This modest average effect arises because a substantial fraction of models remains at lower density intervals; those that switch tend to jump only to the minimum of the highest-yielding interval.
My study contributes to the literature in three ways. First, I provide evidence on how tiered subsidy designs affect producers’ adoption of higher technology in the EV market. Many papers have looked at consumer demand for EVs, but fewer have examined firm technology choice with a structural model. Second, I introduce a simulation-based policy evaluation method that is useful when no comparison group is available. Third, my findings have direct implications for the improvement of Chinese industrial policies in emerging energy sectors. The same framework can be applied to evaluate subsidies for solar panels, wind turbines, and other clean technologies that currently use or could potentially use tiered payment schedules.
2. Related Literature
My work is related to two broad strands of literature. The first focuses on attribute-based subsidies and regulations. Several papers document that such policies influence consumer choices. For instance, subsidies for energy-efficient appliances increase purchases, but the response depends on income and awareness. “Rebound effects” may also reduce net energy savings. On the supply side, attribute regulations such as fuel-economy standards can alter vehicle weight, which may be opposite to the intended outcome. Yet there is little direct evidence on whether attribute-based subsidies encourage technological progress in the targeted attribute. An exception is a recent study on China’s EV market that found tiered subsidies induce firms to sacrifice other performance dimensions to extend range; that paper does not show whether the intended attribute, namely battery energy density, actually improves. My analysis fills this gap.
The second strand is the large literature on Chinese EV policies. Demand-side studies find that subsidies, license restrictions, and driving restrictions significantly affect EV adoption. Supply-side studies have focused on R&D investment, patents, and subsidy reliance. The results are mixed: some find positive effects on innovation, while others find no significant effect. More recent work examines dynamic consumer responses to subsidy phase-out. However, only a few studies connect policy detail to product design. My paper specifically looks at the “adjustment coefficient” design, which was introduced in 2017 and updated several times. This detail is crucial because it determines the marginal subsidy reward for different battery technology levels.
My approach also differs methodologically. I do not rely on a difference-in-differences strategy, because all firms face the same national policy. Instead, I invert the profit condition implied by firms’ observed technology choices. From the probabilities of choosing one battery type over another, I identify the dollar-value contribution of each battery type to profit. This strategy does not require cost data. It only requires observing the frequency of choice under different subsidy incentives. Therefore, it is particularly suitable for policy evaluation in contexts with limited data.
3. Institutional Background
China has aggressively promoted electric vehicles since the late 2000s. Beginning in 2016, a national purchase subsidy program covered battery electric and plug-in hybrid vehicles. Based on the vehicle’s tested electric range, the program assigns a base subsidy amount. The first range-based subsidy schedule is shown in Table 1. Because range alone can be manipulated by reducing weight or motor power, the government added a battery energy-density adjustment coefficient. The coefficient varies with the battery system’s gravimetric energy density. A model’s final subsidy equals the base range subsidy multiplied by this energy-density coefficient.
| Range interval | 2016 | 2017 | 2018 | 2019 | 2020 | 2021 |
|---|---|---|---|---|---|---|
| 100–150 km | 2.5 | 2.0 | — | — | — | — |
| 150–200 km | 4.5 | 3.6 | 1.5 | — | — | — |
| 200–250 km | 4.5 | 3.6 | 2.4 | — | — | — |
| 250–300 km | 5.5 | 4.4 | 3.4 | 1.8 | — | — |
| 300–400 km | 5.5 | 4.4 | 4.5 | 1.8 | 1.6 | 1.3 |
| >=400 km | 5.5 | 4.4 | 5.0 | 2.5 | 2.3 | 1.8 |
Notes: The monetary unit is 10,000 yuan. The 2016 schedule applies to all energy densities. The 2017 schedule is for batteries between 90 and 120 Wh/kg. The 2018 schedule is for batteries between 120 and 140 Wh/kg. The 2019 and after schedule is for batteries with density at least 160 Wh/kg.
Table 2 summarizes the evolution of the energy-density adjustment coefficients. The thresholds are not simply “more energy is always better”. They are tiered intervals with specific cutoffs. In 2017, only two intervals existed: below 90 Wh/kg was not subsidized and above 120 Wh/kg had a 1.1 multiplier. In 2018, the intervals were redefined with a minimum of 105 Wh/kg. The highest coefficient in that year was 1.2 for density of at least 160 Wh/kg. In 2019–2021, the minimum density was increased to 125 Wh/kg, and the coefficient for batteries below 140 Wh/kg was less than one. Hence the design gradually raises the bar and potentially induces higher energy density adoption.
| Energy density interval (Wh/kg) | 2017 coefficient | 2018 coefficient | 2019–2021 coefficient |
|---|---|---|---|
| < 90 | 0 | — | — |
| 90–105 | — | — | — |
| 105–120 | — | 0.6 | — |
| 120–140 | 1.1 (for >=120) | 1.0 | — |
| 125–140 | — | — | 0.8 (2019 later lower) |
| 140–160 | — | 1.1 | 0.9 / 1.0 across years |
| >=160 | — | 1.2 | 1.0 (with lower base subsidies) |
Because the adjustment coefficient changes the amount of subsidy available to a given model, it directly affects a producer’s marginal return from choosing a battery technology. The figure below depicts the relationship between the subsidy schedule and the location of newly launched models in the energy-density distribution.
Visual inspection of the distribution of newly launched models relative to each cutoff suggests a strong “threshold effect”: the number of models whose energy density barely exceeds a policy threshold is far greater than that of models that would have reached the same density without the subsidy. Bunching at these cutoffs is consistent with a policy response that is costly in terms of total subsidy spending but possibly weak in stimulating the most advanced technology.
4. Conceptual Framework
I model an automaker’s battery choice as a profit-maximizing discrete selection from the available battery products in a given market. A market is defined as all pure electric passenger vehicles included in the same official batch of the recommended catalog and with a range that falls into a particular range interval. The firm chooses a battery technology $j$ for a new model $i$ at time $t$. The expected profit if battery $j$ is chosen can be expressed as:
$$ \pi_{ijt} = X_{ijt}\gamma + \alpha (S_{ijt} – MC_{ijt}) \tag{1} $$
where $X_{ijt}\gamma$ represents the effect of observed vehicle features on pricing power or cost, $S_{ijt}$ is the subsidy applied to model $i$ when battery $j$ is chosen at time $t$, and $MC_{ijt}$ is the marginal cost associated with battery $j$. I use a reduced-form linear specification because it captures the trade-off between extra subsidy and extra technology cost. The realized profit is the sum of this deterministic component and an idiosyncratic shock $\varepsilon_{ijt}$:
$$ R_{ijt} = \pi_{ijt} + \varepsilon_{ijt} \tag{2} $$
I assume that $\varepsilon_{ijt}$ follows an i.i.d. type-I extreme value distribution and that every vehicle model selects one battery from the feasible set. These assumptions yield the standard conditional logit probability. Let $L_t$ be the set of feasible battery technologies at time $t$. Then the probability that model $i$ chooses battery $j$ is:
$$ P_{ijt} = \frac{\exp(\pi_{ijt})}{\sum_{k\in L_t}\exp(\pi_{ikt})} \tag{3} $$
In the data, I do not observe individual model decisions at the registration level; I observe how many vehicle models in a given range interval $r$ at time $t$ use a particular battery technology. Following Berry (1994), I aggregate individual models by range interval and take differences in logs between the share of battery $j$ and a baseline battery $j=0$. The resulting equation is:
$$ \ln s_{rjt} – \ln s_{r0t} = \gamma_1 (k_{rjt} – k_{r0t}) + \alpha(S_{rjt} – S_{r0t}) – \alpha \beta (k_{rjt} – k_{r0t}) – \alpha(b_{jt}-b_{0t}) + \eta’_r + \eta’_t + u_{rjt} \tag{4} $$
where $k_{rjt}$ is the range associated with battery $j$ in range interval $r$ at time $t$, $b_{jt}$ is the battery price (or any battery-specific cost) that may vary by year, $\eta’_r$ and $\eta’_t$ are fixed effects, and $u_{rjt}$ is the error term. The coefficient on the range differential identifies the net contribution of an additional kilometer of range to profit, after both pricing and cost channels are taken into account. The coefficient on the subsidy differential identifies $\alpha$, the marginal profit retained from an extra yuan of subsidy. This coefficient is of key interest.
The empirical counterpart of Eq. (4) is:
$$ \ln s_{rjt} – \ln s_{r0t} = \alpha(S_{rjt} – S_{r0t}) + \beta^*(k_{rjt} – k_{r0t}) + (b^*_{j,z(t)} – b^*_{0,z(t)}) + \lambda_r + \lambda_t + u_{rjt} \tag{5} $$
where $\beta^* = \gamma_1 – \alpha\beta$ is the reduced-form effect of range, and the battery-type-by-year dummy absorbs battery price, battery chemistry preference, and any remaining battery-specific cost component that changes over time. Range fixed effects and batch fixed effects are denoted by $\lambda_r$ and $\lambda_t$. Because the dependent variable is the log difference in shares relative to a baseline battery, no model-level characteristics are needed. The variation used for identification comes from differences in subsidy and range between battery types within the same range category and same batch.
5. Data and Variables
5.1 Data sources
I draw data from the official catalog of recommended electric vehicle models, which is released by the central government in several batches each year. To be eligible for the purchase subsidy, a model must appear in the catalog. Therefore, this source covers essentially every newly launched or substantially modified pure electric passenger vehicle that is intended for the subsidized market. For each model, the catalog provides the vehicle model code and key parameters: battery type, battery energy density, battery capacity, and the tested electric range under the Chinese test cycle.
I supplement the catalog data with information from an online automobile database. I first map each catalog vehicle code to its recognizable model name using the tax-exemption list. Then I extract additional market information such as manufacturer suggested retail price. By cross-checking the parameters between the two sources, I ensure consistency. Since plug-in hybrids use a different logic and are fewer in number, I restrict the analysis to pure electric passenger cars. My final estimation sample contains 2,571 unique models. However, after I divide the sample into range categories, battery types, and batches, the number of market-level observations used in the regressions is 540. In some robustness checks, I remove extreme observations and re-estimate the model on smaller samples.
5.2 Variable definitions
The most important variable is the gravimetric energy density of the battery system. I group the energy density into nine intervals: $<105$, $[105,120)$, $[120,130)$, $[130,140)$, $[140,150)$, $[150,160)$, $[160,170)$, $[170,180)$, and $\ge 180$, all in Wh/kg. This fine grouping captures the discrete thresholds of the subsidy policy. In addition, I group battery chemistry into three routes: ternary lithium (Nickel Manganese Cobalt / Nickel Cobalt Aluminium), lithium iron phosphate, and others such as lithium manganate and lithium titanate.
The cell-level battery type is therefore the cross of the nine density intervals and three chemistry routes, giving 27 possible battery types. Some of these may not exist or may be too rare in a given year. For each market, I compute the share of models in that range category and batch that use battery type j. The baseline battery is always the most common ternary lithium battery with a density of 120 Wh/kg in 2017, 140 Wh/kg in 2018–2019, and 160 Wh/kg in 2020–2021. The subsidy variable is computed as the difference between the product-specific subsidy and the baseline subsidy.
Table 3 provides descriptive statistics for the main variables. The average energy density is 143.2 Wh/kg. The minimum is 91.1 Wh/kg and the maximum is 206 Wh/kg. There is a wide dispersion in subsidies, from zero to 60,000 yuan. The average subsidy is about 16,900 yuan. The average tested range is 348 kilometers, but ranges in the sample go from 100 km to over 1,000 km. The average battery energy capacity is 46.3 kWh, and the average manufacturer suggested retail price is 158,500 yuan.
| Variable | Mean | Std. Dev. | Min | Max |
|---|---|---|---|---|
| Energy density (Wh/kg) | 143.22 | 20.39 | 91.1 | 206 |
| Subsidy (10,000 yuan) | 1.69 | 1.32 | 0 | 6 |
| Range (km) | 348.43 | 117.58 | 100 | 1008 |
| Battery energy (kWh) | 46.26 | 18.70 | 9 | 144.4 |
| Price (10,000 yuan) | 15.85 | 9.90 | 2.68 | 75.43 |
| Ternary lithium (0–1) | 0.72 | 0.45 | 0 | 1 |
| Lithium iron phosphate (0–1) | 0.18 | 0.39 | 0 | 1 |
Table 4 describes the year-by-year evolution of the average characteristics of newly launched EV models. Between 2017 and 2021, annual average energy density increased from 117.5 to 150.1 Wh/kg, while average range increased from 223 to 415 kilometers. At the same time, average battery capacity increased from 32.6 kWh to 54.0 kWh. Prices remained relatively stable at about 150,000–170,000 yuan. The average subsidy fell from 21,100 yuan in 2018 to 13,700 yuan in 2021, illustrating the general phase-out of subsidies.
| Year | Models | Energy density | Subsidy | Range | Battery capacity | Price |
|---|---|---|---|---|---|---|
| 2017 | 358 | 117.45 | 1.54 | 223.22 | 32.63 | 16.39 |
| 2018 | 471 | 140.55 | 2.11 | 311.05 | 42.32 | 15.06 |
| 2019 | 713 | 148.14 | 1.86 | 358.07 | 47.17 | 14.46 |
| 2020 | 547 | 149.90 | 1.49 | 391.20 | 50.56 | 16.61 |
| 2021 | 482 | 150.10 | 1.37 | 415.13 | 54.00 | 16.85 |
Table 5 shows the distribution of energy density choices over time. Several patterns stand out. In 2017, almost one third of new models had battery density below 105 Wh/kg. In 2018, more than half of all new models chose a density between 140 and 150 Wh/kg. This interval contains the cutoff for the highest 1.1 multiplier that year. In 2019 and 2020, the 140–150 interval remained the most important. By 2021, the modal interval shifted to 160–170 Wh/kg, which is the interval that, together with higher intervals, receives the largest adjustment coefficient under the then-current policy. This descriptive trend is consistent with an economically meaningful policy response, but it is not causal because unobserved technology costs fell over time. I therefore turn to the structural model.
| Energy density interval | 2017 | 2018 | 2019 | 2020 | 2021 |
|---|---|---|---|---|---|
| <105 Wh/kg | 32.51 | 0.21 | 0.42 | 2.90 | 3.05 |
| 105–120 | 15.15 | 2.97 | 0.70 | 4.17 | 7.93 |
| 120–130 | 28.93 | 17.20 | 5.87 | 7.26 | 6.10 |
| 130–140 | 12.40 | 6.79 | 5.17 | 7.26 | 7.72 |
| 140–150 | 9.09 | 53.50 | 49.51 | 31.22 | 22.15 |
| 150–160 | 1.93 | 11.89 | 5.45 | 3.27 | 7.52 |
| 160–170 | 0 | 7.99 | 20.21 | 21.24 | 22.97 |
| 170–180 | 0 | 0.21 | 2.38 | 11.25 | 13.01 |
| >=180 | 0 | 0 | 1.12 | 8.71 | 9.55 |
6. Empirical Results
6.1 Baseline estimates
I estimate Eq. (5) using ordinary least squares with fixed effects. The dependent variable is the log share difference between the specific battery and the baseline battery in the same range-and-batch market. The key explanatory variable is the subsidy difference measured in 10,000 yuan. Table 6 reports the baseline results. Column (1) controls for battery-type-by-year fixed effects only. Column (2) adds batch and range fixed effects. Column (3) reports the same specification with standard errors clustered at the range level. In columns (4)–(6), I add the log of the distance such as the difference in the log of the actual range.
| Dependent variable: $\ln(s)-\ln(s_0)$ | (1) | (2) | (3) | (4) | (5) | (6) |
|---|---|---|---|---|---|---|
| Subsidy (10,000 yuan) | 1.828*** (0.371) |
0.960** (0.382) |
0.960*** (0.143) |
1.683*** (0.359) |
0.888** (0.373) |
0.888*** (0.136) |
| $\ln(\text{range})$ | 2.412*** (0.661) |
1.484** (0.580) |
1.484** (0.413) |
|||
| Observations | 540 | 540 | 540 | 540 | 540 | 540 |
| Adjusted R² | 0.360 | 0.575 | 0.574 | 0.377 | 0.580 | 0.579 |
| Battery x year FE | Yes | Yes | Yes | Yes | Yes | Yes |
| Batch FE | No | Yes | Yes | No | Yes | Yes |
| Range FE | No | Yes | Yes | No | Yes | Yes |
| Cluster | No | No | Range | No | No | Range |
In the preferred specification in column (6), the coefficient on subsidy is 0.888 and statistically significant at the 1% level. This coefficient implies that if a battery type receives an additional 10,000 yuan of subsidy relative to the baseline battery, the relative odds that an EV model chooses that battery increase substantially. More formally, since the model is a linearized conditional logit, the coefficient can be interpreted as the marginal value of subsidy to a firm’s profit. The coefficient on log range is positive and statistically significant. A 1% increase in actual range increases the relative profit from that battery by about 0.0148 units on the log-odds scale.
6.2 Robustness checks
I conduct several robustness checks. Table 7 shows that the main result is stable when I change the baseline battery, exclude extreme battery shares, drop less common chemistry routes, add additional fixed effects, cluster standard errors at the batch and range level, or use the absolute value of range instead of the logarithm. In all cases, the coefficient on subsidy remains positive and statistically significant, and the implied magnitude is close to the baseline estimate.
| Dependent variable: $\ln(s)-\ln(s_0)$ | (1) | (2) | (3) | (4) | (5) | (6) |
|---|---|---|---|---|---|---|
| Subsidy | 0.782** (0.234) |
0.817*** (0.122) |
0.969*** (0.162) |
0.630** (0.141) |
0.888* (0.332) |
0.882*** (0.141) |
| $\ln(\text{range})$ | 1.349* (0.525) |
1.193** (0.399) |
1.778** (0.529) |
1.474** (0.380) |
1.484** (0.485) |
0.004** (0.001) |
| Observations | 468 | 499 | 414 | 539 | 540 | 540 |
| Adj. R² | 0.651 | 0.571 | 0.569 | 0.617 | 0.574 | 0.582 |
I also recognize that unobserved battery prices may vary over time. The battery-type-by-year fixed effects absorb systematic differences in average battery price across years. There remains the possibility that battery price differs across batches within the same year, but the subsidy jump created by the policy is large and transparent. Any measurement error would induce attenuation bias, implying that my estimate of the subsidy coefficient might be conservative.
6.3 Battery contributions to model profits
The battery-type-by-year fixed effects from Eq. (5) provide an estimate of how much each battery type, relative to the baseline battery, contributes to a carmaker’s profit in a given year. Table 8 reports these relative profit contributions for each energy-density interval. I use each year’s most common density interval as the baseline, so the baseline is normalized to zero. A positive number means the battery type is more profitable than the baseline; a negative number means it is less profitable.
| Energy density interval | 2017 | 2018 | 2019 | 2020 | 2021 |
|---|---|---|---|---|---|
| <105 | 0.123 | — | — | — | — |
| 105–120 | — | -1.189 | 0.730 | — | — |
| 120–130 | -0.066 | -1.169 | -0.916 | -0.418 | 0.233 |
| 130–140 | -0.356 | -1.340 | -1.336 | -0.646 | 0.510 |
| 140–150 | -0.153 | 0 (base) | 0 (base) | -0.212 | 0.319 |
| 150–160 | -1.071 | -1.139 | -1.271 | -1.146 | -0.180 |
| 160–170 | -0.428 | -1.220 | -0.820 | 0 (base) | 0 (base) |
| 170–180 | — | -2.217 | -0.767 | 0.699 | — |
| >=180 | — | -1.737 | -0.964 | 0.211 | — |
These fixed-effect estimates show that low-density batteries were relatively profitable in 2017, whereas high-density batteries became increasingly profitable by 2021. The change reflects a rapid reduction in the cost premium for high-density batteries and a strong consumer preference for models with longer range. If I combine these profit estimates with the subsidy differences under the actual policy, I can simulate the choice probabilities and infer the effect of replacing the actual tiered multiplier with a uniform multiplier.
7. Counterfactual Policy Analysis
7.1 Comparing tiered and uniform subsidy multipliers
The first counterfactual exercise answers the central question: What would the technology distribution of newly launched electric vehicle models have looked like if the subsidy were distributed as a uniform multiplier rather than as a tiered multiplier? In the uniform case, every eligible model receives the same multiplier of one. The base subsidy for a range interval remains the same. Because the total subsidy outlay is not held constant in this simple comparison, I compare the model composition rather than levels. In a second exercise, I also hold total subsidy expenditure constant by reallocating the total subsidy amount according to each simulated policy.
To implement the counterfactual choice probabilities, I use the structural estimates. For each range interval, batch, and battery type, I compute the deterministic profit as the estimated subsidy coefficient times the counterfactual subsidy differential plus the estimated range coefficient times the observed range differential plus the battery-type-by-year fixed effects. I then use the logit formula to generate the probability of each battery type within that market. Multiplying by the actual number of models in that range interval gives the simulated number of models choosing each battery technology.
Table 9 compares the number of newly launched models under the uniform multiplier and under the actual tiered multiplier. Overall, the tiered system increases the number of models in the 160–170 Wh/kg interval from 418 to 477, a 14% increase. It also increases the model count in the two highest intervals by about 9%. Meanwhile, models with lower energy density decline by 3–16%. The main winners are not the top-bin batteries but rather batteries that just clear the 160 Wh/kg threshold. This is exactly what one would expect from a tiered schedule with a fixed maximum multiplier.
| Energy density interval | Uniform multiplier | Tiered multiplier | Change (%) |
|---|---|---|---|
| <105 | 58 | 51 | −12 |
| 105–120 | 66 | 56 | −16 |
| 120–130 | 266 | 257 | −3 |
| 130–140 | 195 | 174 | −11 |
| 140–150 | 909 | 880 | −3 |
| 150–160 | 155 | 150 | −3 |
| 160–170 | 418 | 477 | 14 |
| 170–180 | 133 | 145 | 9 |
| >=180 | 97 | 106 | 9 |
I also examine the effect of the tiered system on different battery chemistries. Table 10 shows that the tiered design mainly encourages ternary lithium batteries. Models using low-density ternary lithium decline, while models using high-density ternary lithium increase substantially. The effect on lithium iron phosphate and other chemistries is generally negative or negligible. This is not surprising because ternary lithium chemistry has historically delivered higher energy density, and the subsidy design rewards density irrespective of other attributes.
| Energy density interval | Ternary change (%) | LFP change (%) | Other change (%) |
|---|---|---|---|
| <105 | −12 | −13 | −12 |
| 105–120 | −20 | −9 | 0 |
| 120–130 | 0 | −13 | 0 |
| 130–140 | −11 | −13 | −3 |
| 140–150 | −3 | −4 | −2 |
| 150–160 | −4 | 0 | 0 |
| 160–170 | 14 | 0 | 14 |
| 170–180 | 9 | 0 | 13 |
| >=180 | 9 | — | — |
Although the tiered adjustment coefficient increases the relative share of high-density battery modules, the economic implication for the average technology level is weak. I compute the average energy density of newly launched models under the uniform schedule and under the actual tiered schedule. The actual schedule adds less than 1.5 Wh/kg to the annual average, which is approximately 1% of the average density. In other words, the bumpy distribution at the threshold creates a lot of bunching but does not push the frontier forward in a meaningful way. If the government’s objective is to induce a general upward shift in the density distribution, the current design is under-powered.
7.2 Alternative subsidy designs
I now perform counterfactual exercises to show how alternative designs would perform within the same total subsidy budget. The first alternative is a stronger tiered schedule: I increase the highest multiplier to 1.2 and lower the multipliers for low-density intervals to 0.6. The second alternative is a quasi-linear schedule in which the adjustment coefficient is linearly assigned to each energy-density interval, from a minimum of zero for the lowest interval to a maximum that keeps total expenditure equal to the actual total expenditure. In other words, each interval receives a subsidy of a constant proportion of the base subsidy multiplied by a coefficient that increases smoothly with energy density.
Figure 2 summarizes the simulated changes in the number of models at each density interval relative to the actual policy. The first alternative, with larger subsidy differences, causes more models to move from the 140–150 interval to the 150–160 interval and above. The second alternative, with a nearly linear relationship, has an even larger effect on the highest density bins. The number of models using battery packs with density above 180 Wh/kg becomes substantially higher than what the existing tiered policy generates.
Figure 3 shows the average energy density under the actual policy and under the two alternative policies. In the actual policy, the average density of newly launched models is only marginally higher than in the uniform case. The “strong tiered” alternative produces a modest additional gain. The quasi-linear schedule, however, raises average energy density by roughly 10% relative to the actual policy. This substantial difference occurs because a quasi-linear schedule treats every additional unit of energy density as equally valuable, thus eliminating the incentive to bunch at the threshold.
From a policy perspective, a quasi-linear subsidy is not common in practice. Governments often choose tiered schedules because they are administratively simple and politically easy to communicate. Yet my simulation indicates that such schedules can be inefficient. A tiered schedule effectively creates a “notch” in the firm’s profit function. If the firm’s marginal cost curve is flat, all firms whose optimal density lies just below the cutoff will jump to just above it, resulting in a mass point at the cutoff. The benefit of the subsidy program is therefore captured by inframarginal jumps rather than by larger and more costly technology improvements. Linear subsidies spread the incentive along the entire attribute margin and avoid this bunching problem.
8. Discussion
My estimates indicate that the coefficient on the subsidy difference remains above 0.8 in all robust specifications. This magnitude suggests that subsidy design is not neutral. Firms do respond to the financial incentive when choosing the battery technology for a new model. However, the same response also means that unless the subsidy schedule is closely aligned with the social marginal benefit of technological improvement, the policy can deliver a distorted distribution of attributes. In the Chinese electric vehicle market, the historical tiered schedule with a lump-sum-like cutoff at 160 Wh/kg caused too many models to be moved to 160 rather than above. A more elastic design could achieve the same technological average at a lower fiscal cost or a higher technological average at the same cost.
There are several reasons why the average effect is small. The first is that the subsidy amount itself was decreasing over time. In 2021, the maximum subsidy for a model was only 18,000 yuan, and the coefficient multiplier interval was compressed from 1.2 to 1.0. As subsidies phase out, the role of the adjustment coefficient fades. The second reason is that household demand for energy density is non-linear. Many consumers care about total range but not directly about battery density. Thus automakers may prefer a battery whose density is high enough to maximize subsidy, but not necessarily much higher. Third, the availability of high-density battery cells in the market is limited by upstream suppliers. If suppliers cannot supply batteries beyond a threshold at the same cost, automakers have no feasible way to move further up. My model accounts for this through battery-type fixed effects, which capture differences in cost or availability across years.
9. Conclusion and Policy Implications
This paper asks whether a tiered subsidy adjustment coefficient, under China’s electric vehicle purchase subsidy program, shifts firms’ choices of battery technology toward higher energy density. I construct a structural model of automobile manufacturers’ battery choices and estimate the contribution of each battery technology to model profits using data on all pure electric passenger vehicles in the recommended catalog during 2017–2021. The empirical results confirm that subsidy differentials significantly influence battery choices. Counterfactual simulations show that compared with a uniform subsidy multiplier, the actual tiered multiplier increases the number of newly launched vehicle models with energy densities above 160 Wh/kg by more than 14%. Yet because the schedule provides no further reward for exceeding this ceiling, the average energy density increase is only about 1%.
My findings have direct policy implications for the electric vehicle industry and other renewable energy sectors. First, when governments use subsidies to promote technical progress, the marginal subsidy should not stop abruptly at an arbitrary threshold. A quasi-linear subsidy that increases continuously with the targeted attribute is more effective. Second, if a tiered schedule is preferred for administrative reasons, policy makers should carefully calibrate the differentials between tiers to exceed the incremental cost and demand gaps of high-technology products. Third, subsidy policy should be updated as fast as the technology itself. In an industry where battery energy density rose by a factor of more than two in a decade, a static subsidy schedule quickly becomes irrelevant. The recent phase-out of China’s EV purchase subsidies reflects the completion of the market-building stage, but my analytical framework remains valuable for the design of other attribute-based policies, such as vehicle fuel-economy standards and subsidies for energy-efficient appliances.