Evidence from Transportation - Steps to Clean Smog Check Data

A Steps to clean smog check data A.1 Smog Check Data Our data from the Smog Check Program are essentially the universe of test records from January 1, 1996 to December 31, 2010. We were only able to obtain test records going back to 1996 because this was the year when the Smog Check program introduced its electronic transmission system. Because the system seems to have been phased in during the first half of 1996, and major program changes took effect in 1998 we limit our sample to test records from January 1998 on. For our analyses, we use a 10% sample of VINs, selecting by the second to last digit of the VIN. We exclude tests which have no odometer reading, with a test result of ”Tampered” or ”Aborted” and vehicles which have more than 36 tests in the span of the data. Vehicles often have multiple Smog Check records in a year, whether due to changes of ownership or failed tests, but we argue that more than 36 in what is at most a 12 year-span indicates some problem with the data.22 A few adjustments must be made to accurately estimate VMT and emissions per mile. First, we adjust odometer readings for roll-overs and typos. Many of the vehicles in our analysis were manufacturer withe 5-digit odometers–that is, five places for whole numbers plus a decimal. As such, any time one of these vehicles crosses over 100,000 miles, the odometer “rolls over” back to 0. To complicate matters further, sometimes either the vehicle owner or Smog Check technician notices this problem and records the appropriate number in the 100,000s place, and sometimes they do not. To address this problem, we employ an algorithm that increases the hundred thousands place in the odometer reading whenever a rollover seems to have occurred. The hundred thousands are incremented if the previous test record shows higher mileage, or if the next test record is shows more than 100,000 additional miles on the odometer (indicating that the odometer had already rolled over, but the next check took this into account). The algorithm also attempts to correct for typos and entry errors. An odometer reading is flagged if it does not fit with surrounding readings for the same vehicle–either it is less than the previous reading or greater the next–and cannot be explained by a rollover. The algorithm then tests whether fixing one of several common typos will make the flagged 22 For instance, there is one vehicle in particular, a 1986 Volvo station wagon, which has records for more than 600 Smog Checks between January 1996 and March 1998. The vehicle likely belonged to a Smog Check technician who used it to test the electronic transmission system. 63 readings fit (e.g. moving the decimal over one place). If no correction will fit, the reading is replaced with the average of the surrounding readings. Finally, if after all our corrections any vehicle has an odometer reading above 800,000 or has implied VMT per day greater than 200 or less than zero, we exclude the vehicle from our analysis. All of our VMT analyses use this adjusted mileage. Emissions results from smog checks are given in either parts per million (for HC and NOx ) or percent (O2 , CO, and CO2 ). Without knowing the volume of air involved, there is no straightforward way to convert this to total emissions. Fortunately, as part of an independent evaluation of the Smog Check program conducted in 2002-2003, Sierra Research Inc. and Eastern Research Group estimated a set of conversion equations to convert the proportional measurements of the ASM test to emissions in grams per mile traveled. These equations are reported in Morrow and Runkle (2005) and are reproduced below. The equations are for HCs, NOx and CO, and estimate grams per mile for each pollutant as a non-linear function of all three pollutants, model year and vehicle weight. The equations for vehicles of up to model year 1990 are F T P HC = 1.2648 · exp(−4.67052 +0.46382 · HC ∗ + 0.09452 · CO∗ + 0.03577 · N O∗ +0.57829 · ln(weight) − 0.06326 · M Y ∗ + 0.20932 · T RU CK) F T P CO = 1.2281 · exp(−2.65939 +0.08030 · HC ∗ + 0.32408 · CO∗ + 0.03324 · CO∗2 +0.05589 · N O∗ + 0.61969 · ln(weight) − 0.05339 · M Y ∗ +0.31869 · T RU CK) F T P N OX = 1.0810 · exp(−5.73623 +0.06145 · HC ∗ − 0.02089 · CO∗2 + 0.44703 · N O∗ +0.04710 · N O∗2 + 0.72928 · ln(weight) − 0.02559 · M Y ∗ −0.00109 ∗ M Y ∗2 + 0.10580 · T RU CK) 64 Where HC ∗ = ln((M ode1HC · M ode2HC ).5 ) − 3.72989 CO∗ = ln((M ode1CO · M ode2CO ).5 ) + 2.07246 N O∗ = ln((M ode1N O · M ode2N O ).5 ) − 5.83534 M Y ∗ = modelyear − 1982.71 weight = Vehicle weight in pounds T RU CK = 0 if a passenger car, 1 otherwise And for model years after 1990 they are: F T P HC = 1.1754 · exp(−6.32723 +0.24549 · HC ∗ + 0.09376 · HC ∗2 + 0.06653 · N O∗ +0.01206 · N O∗2 + 0.56581 · ln(weight) − 0.10438 · M Y ∗ −0.00564 · M Y ∗2 + 0.24477 · T RU CK) F T P CO = 1.2055 · exp(−0.90704 +0.04418 · HC ∗2 + 0.17796 · CO∗ + 0.08789 · N O∗ +0.01483 · N O∗2 − 0.12753 · M Y ∗ − 0.00681 · M Y ∗2 +0.37580 · T RU CK) F T P N OX = 1.1056 · exp(−6.51660 + + 0.25586 · N O∗ + 0.04326 · N O∗2 + 0.65599 · ln(weight) −0.09092 · M Y ∗ − 0.00998 ∗ M Y ∗2 + 0.24958 · T RU CK) Where: HC ∗ = ln((M ode1HC · M ode2HC ).5 ) − 2.32393 CO∗ = ln((M ode1CO · M ode2CO ).5 ) + 3.45963 N O∗ = ln((M ode1N O · M ode2N O ).5 ) − 3.71310 M Y ∗ = modelyear − 1993.69 weight = Vehicle weight in pounds T RU CK = 0 if a passenger car, 1 otherwise 65 B Steps to clean DMV data We deal with two issues associated with the DMV data. The main issue with the DMV data is that, often, entries for the same addresses will have slightly different formats. For example, 12 East Hickory Street may show up as 12 East Hickory St, 12 E. Hickory St., etc. To homogenize the entries, we input each of the DMV entries into mapquest.com and then replace the entry with the address that mapquest.com gives. Second, the apartment number is often missing in the DMV data. This has the effect of yielding a large number of vehicles in the same “location”. We omit observations that have over seven vehicles in a given address or more than three last names of registered owners. C Details of the gasoline tax policy simulation For the intensive margin, we estimate a regression as in column 6 of Tables 3 and 4, except that we interact ln(DPM) with quartile of fuel economy, vehicle weight, and emissions of HC, NOx , and CO, and dummies for vehicle age bins, again using bins of 4-9, 10-15, and 16-29 years, and control for the direct effects of quartiles of HC, NOx , and CO emissions. As in Table 5, we use quartiles calculated by year and age bin. The coefficients are difficult to interpret on their own, and too numerous to list. However, most are statistically different from zero, and the exceptions are due to small point estimates, not large standard errors. As in Section 7, we compress our dataset to have at most one observation per vehicle per year. Each vehicle is then assigned an elasticity based on their quartiles and age bin. Vehicle i’s VMT in the counterfactual with an additional $1 tax on gasoline is calculated by: V i M Tcounterf actual =V i M TBAU  ·  Pi + 1 · βi , Pi i where V M TBAU is vehicle i’s actual average VMT per day between their current and previous Smog Check, Pi is the average gasoline price over that time, and βi is the elasticity for the fuel economy/weight/HC/NO/CO/age cell that i belongs to. For the extensive margin, we estimate a Cox regression on the hazard of scrappage for vehicles 10 years and older, stratifying by VIN prefix and interacting DPM with all five type of quartiles and age bins 10-15 and 16-29. Similar to the intensive margin, we assign each vehicle 66 a hazard coefficient based on their quartile-age cell. Cox coefficients can be transformed into hazard ratios, but to simulate the affect of an increase in gasoline prices on the composition of the vehicle fleet, we must convert these into changes in total hazard. To do this, we first calculate the actual empirical hazard rate for prefix k in year t as: OrigHazardkt = Dkt , Rkt where Dkt is the number of vehicles in group k which are scrapped in year t, and Rkt is the number of vehicle at risk (that is, which have not previously been scrapped or censored). We then use the coefficients from our Cox regression to calculate the counterfactual hazard faced by vehicles of prefix k in quartile-age group q during year t as:23  N ewHazardqkt = OrigHazardkt · exp  1 · γq , M P Gk where M P Gk is the average fuel economy of vehicle of prefix k and γq is the Cox coefficient associated with quartile group q. We then use the change in hazard to construct a weight Hqkt indicating the probability that a vehicle of prefix k in quartile group q in year t would be in the fleet if a $1 gasoline tax were imposed. Weights greater than 1 are possible, in which should be interpreted as a Hqkt − 1 probability that another vehicle of the same type would be on the road, but which was scrapped under “Business as Usual.” Since the hazard is the probability of scrappage in year t, conditional on survival to year t, this weight must be calculated interatively, taking into account the weight the previous year. Specifically, we have: Hqkt = t Y (1 − (N ewHazardqkj − OrigHazardkt )). j=1998 We also assign each vehicle in each year a population weight. This is done both to scale our estimates up to the size of the full California fleet of personal vehicles, and to account for the ways in which the age composition of the Smog Check data differs from that of the fleet. We construct these weights using the vehicle population estimates contained in CARB’s EMFAC07 software, which are given by year, vehicle age, and truck status. Our population weight is the number of vehicles of a given age and truck status in a each year given by 23 Note that age group is determined by model-year and year. 67 EMFAC07, divided by the number of such vehicle appearing in our sample. For instance, if EMFAC07 gave the number of 10 year old trucks in 2005 as 500, while our data contained 50, each 10 year old truck in our data would have a population weight of 10. Denote the population weight by Ptac , where t is year, a is age, and c is truck status. There is an additional extensive margin which we have not estimated in this paper, that of new car purchases. To ensure that the total vehicle population is accurate, we apply an ad hoc correction based on Busse, Knittel, and Zettelmeyer (2009), who find that a $1 increase in gasoline prices would decrease new car sales by 650,000 per year. Since California’s vehicle fleet makes up about 13% of the national total, we decrease the population of model years 1998 and later by 84500 when constructing the population weight for the counterfactual. We apply 40% of the decrease to trucks, and 60% to passenger cars. Denote the “new car effect” nc . We estimate the total annual emissions by passenger vehicle in California of NOx , HC, CO, and CO2 as actually occurred, and under a counterfactual where a $1 gasoline tax was imposed in 1998. Let i denote a vehicle, a vehicle age, c truck status, then the annual emissions of pollutant p in year t under “business as usual” are: Emissionpt BAU = X i Ptac · V M TBAU · ri (p) · 365, i and under the counterfactual they are: Emissionpt counterf actual = X i (Ptac −1(model year >= 1998)·nc )·Hqkt ·V M Tcounterf actual ·ri (p)·365, i where ri (p) is the emissions rate per mile of pollutant p for vehicle i. For NOx , HC, and CO, this is the last Smog Check reading in grams per mile, while for CO2 this is the vehicle’s gallons per mile multiplied by 19.2 pounds per gallon. D California versus the rest of the US Given that our empirical setting is California, it is natural to ask whether our results are representative of the country as a whole. At the broadest level, the co-benefits from carbon pricing are a function of the per capita number of miles driven, the emission characteristics of the fleet of vehicles, and the marginal damages of the emissions. We present evidence that 68 the benefits may, in fact, be larger outside of California. The reason for this is that while the marginal damages are indeed larger in California, the vehicle stock in California is much cleaner than the rest of the country because California has traditionally lead the rest of the US in terms of vehicle emission standards. The results in Muller and Mendelsohn (2009) provide a convenient way to test whether California differs in terms of marginal damages. Table A.25 presents points on the distribution of marginal damages for NOx , HCs, and the sum of the two, weighted by each county’s annual VMT.24 Figure A.11 plots the kernel density estimates of the distributions. We present the sum of because counties are typically either “NOx constrained” or “VOC (HC) constrained” and the sum is perhaps more informative. As expected, the marginal damages are higher in California for HCs, but lower for NOx , as California counties tend to be VOC-constrained. The sum of the two marginal damages is 78 percent higher in California. Higher points in the distribution show an even larger disparity. This effect is offset, however, by the cleaner vehicle stock within California—a result of California’s stricter emission standards. To illustrate this, we collected county-level average per-mile emission rates for NOx , HCs, and CO from the EPA Motor Vehicle Emission Simulator (MOVES). This reports total emissions from transportation and annual mileage for each county. Table A.25 also presents points on the per-mile emissions and Figure A.12 plots the distributions.25 Mean county-level NOx , HCs, and CO are 67, 36, and 31 percent lower in California, respectively. Other points in the distributions exhibit similar patterns. Finally, we calculate the county-level average per-mile externality for each pollutant, as well as the sum of the three. Table A.25 and Figure A.13 illustrates these. As expected the HC damages are higher, but the average county-level per-mile externality from the sum the three pollutants is 30 percent lower in California compared to the rest of the country; the 25th percentile, median, and 75th percentile are 35, 30, and 9 percent lower, respectively. These calculations suggest that, provided the average VMT elasticities are not significantly smaller outside of California and/or the heterogeneity across vehicle types is not significantly 24 All of the points on the distribution and densities discussed in this section weight each county by its total VMT. 25 We note that these are higher than the averages in our data. This may reflect the fact that smog checks are not required for vehicles with model years before 1975 and these vehicles likely have very high emissions since this pre-dates many of the emission standards within the US. 69 different (in the reverse way), our estimates are likely to apply to the rest of the country. 70 .0015 Density .0005 .001 0 0 2000 4000 6000 8000 Damage from HCs ($/ton) Rest of the US Density 0 .0005 .001 .0015 .002 California 10000 -6000 -4000 -2000 0 Damage from NOx ($/ton) Rest of the US Density 0 .0002.0004.0006.0008 California 2000 -2000 0 2000 4000 6000 8000 Sum of Damages from NOx and HCs ($/ton) California Rest of the US Figure A.11: Distributions of marginal damages from Muller and Mendelsohn (2009) for California and the rest of the US 71 30 Density 10 20 0 .2 .4 .6 .8 HCs/mile Rest of the US 0 2 Density 4 6 8 10 California 1 1.2 1.4 1.6 NOx/mile 2 2.2 Rest of the US 0 Density .5 1 1.5 California 1.8 5 7 9 11 CO/mile California 13 15 Rest of the US Figure A.12: Distributions of per-mile emissions for California and the rest of the US 72 3000 2000 1000 0 Density .005 0 .001 Damage from HCs ($/mile) Damage from CO ($/mile) California .0075 California Rest of the US .01 Rest of the US .0125 .002 0 .002 .004 .001 .002 Damage from NOx ($/mile) California .003 Rest of the US .01 Rest of the US .006 .008 Damage from HCs, NOx, and CO ($/mile) California .012 .004 Figure A.13: Distributions of per-mile damages for California and the rest of the US Density 1500 1000 500 0 2000 1500 Density 1000 500 0 2000 1500 Density 1000 500 0 73 Table A.23: Robustness Check—Intensive Margin Interacting NOx Quartiles With Other Controls (1) (2) (3) ∗ (4) (5) (6) ln(DPM) * NO Q1 0.0406 (0.0231) 0.0381 (0.0250) 0.0678 (0.0339) 0.0605 (0.0335) 0.0590 (0.0333) 0.0666 (0.121) ln(DPM) * NO Q2 -0.0617∗ (0.0261) -0.0581∗ (0.0269) -0.0453 (0.0309) -0.0478 (0.0310) -0.0484 (0.0308) -0.0410 (0.121) ln(DPM) * NO Q3 -0.158∗∗∗ (0.0271) -0.155∗∗∗ (0.0272) -0.166∗∗∗ (0.0282) -0.165∗∗∗ (0.0291) -0.165∗∗∗ (0.0294) -0.157 (0.120) ln(DPM) * NO Q4 -0.288∗∗∗ (0.0300) -0.298∗∗∗ (0.0302) -0.355∗∗∗ (0.0325) -0.353∗∗∗ (0.0332) -0.351∗∗∗ (0.0331) -0.344∗∗ (0.120) NO Q2 0.378 (0.800) 0.327 (0.735) -2.622 (1.622) -3.925∗ (1.693) -3.954∗ (1.673) -4.916∗∗ (1.732) NO Q3 -1.246 (1.012) -1.447 (0.899) -5.233∗∗∗ (1.447) -6.846∗∗∗ (1.524) -6.793∗∗∗ (1.508) -7.987∗∗∗ (1.566) NO Q4 -2.297∗ (1.116) -2.951∗∗ (1.084) -9.696∗∗∗ (2.257) -11.39∗∗∗ (2.253) -11.26∗∗∗ (2.271) -12.60∗∗∗ (2.301) Quartile-Time Trend Interactions Yes Yes Yes Yes Yes Yes Vintage-Quartile Interactions No Yes Yes Yes Yes Yes Quartile-Year Interactions No No Yes Yes Yes Yes Quartile-Lagged Odometer Interactions No No No Yes Yes Yes Quartile-Demographics Interactions No No No No Yes Yes Calendar Month Fixed-Effects No No No No No Yes 2979289 2979289 2979289 2979289 2979289 2979289 N Note: All regressions include vehicle fixed-effects, year fixed effects, vintage/truck fixed effects, a quadratic time trend, a sixth order polynomial in the odometer reading at previous Smog Check, and ZIP code level demographic characteristics. 74 Table A.24: Robustness Check—Intensive Margin Interacting NOx Quartiles With Other Controls (1) (2) (3) ∗∗∗ (4) ∗∗∗ (5) ∗∗∗ (6) DPM * NO Q1 -2.676 (0.359) -2.807 (0.350) -2.294 (0.301) -2.412 (0.347) -2.421 (0.345) -5.089∗∗∗ (0.696) DPM * NO Q2 -3.337∗∗∗ (0.359) -3.358∗∗∗ (0.357) -3.075∗∗∗ (0.334) -3.128∗∗∗ (0.355) -3.129∗∗∗ (0.354) -5.339∗∗∗ (0.631) DPM * NO Q3 -3.925∗∗∗ (0.389) -3.941∗∗∗ (0.391) -3.858∗∗∗ (0.397) -3.881∗∗∗ (0.395) -3.875∗∗∗ (0.394) -5.728∗∗∗ (0.631) DPM * NO Q4 -4.642∗∗∗ (0.425) -4.720∗∗∗ (0.433) -4.970∗∗∗ (0.444) -4.974∗∗∗ (0.440) -4.957∗∗∗ (0.442) -6.482∗∗∗ (0.653) NO Q2 0.958 (0.674) 0.821 (0.613) -5.997∗∗∗ (1.330) -7.404∗∗∗ (1.391) -7.433∗∗∗ (1.384) -4.917∗∗ (1.567) NO Q3 0.242 (0.889) -0.00702 (0.798) -8.999∗∗∗ (1.563) -10.74∗∗∗ (1.619) -10.69∗∗∗ (1.605) -6.708∗∗∗ (1.527) NO Q4 0.615 (1.015) 0.124 (0.999) -12.63∗∗∗ (2.173) -14.43∗∗∗ (2.181) -14.34∗∗∗ (2.205) -9.222∗∗∗ (2.227) Quartile-Time Trend Interactions Yes Yes Yes Yes Yes Yes Vintage-Quartile Interactions No Yes Yes Yes Yes Yes Quartile-Year Interactions No No Yes Yes Yes Yes Quartile-Lagged Odometer Interactions No No No Yes Yes Yes Quartile-Demographics Interactions No No No No Yes Yes Calendar Month Fixed-Effects No No No No No Yes 2979289 2979289 2979289 2979289 2979289 2979289 N ∗∗∗ ∗∗∗ Note: All regressions include vehicle fixed-effects, year fixed effects, vintage/truck fixed effects, a quadratic time trend, a sixth order polynomial in the odometer reading at previous Smog Check, and ZIP code level demographic characteristics. 75 Table A.25: Percentage Difference Between California and the rest of the US 25th Percentile Median 75th Percentile Mean NOx g/mi -0.230 -0.291 -0.338 -0.282 NOx Damage/ton (MM) -0.439 -0.525 -0.558 -0.685 NOx Damage/mi -0.595 -0.657 -0.712 -0.761 HC g/mi -0.262 -0.321 -0.410 -0.354 HC Damage/ton 1.475 2.558 5.318 1.821 HC Damage/mi 0.602 1.134 3.358 1.035 CO g/mi -0.226 -0.321 -0.366 -0.320 CO Damage/mi -0.226 -0.321 -0.366 -0.320 NOx + HC Damage/ton (MM) 0.0191 0.994 2.337 0.787 NOx + HC + CO Damage/mi -0.353 -0.299 -0.0883 -0.295 Notes: The table reports the coefficient on the California dummy divided by the constant. All differences are statistically significant at the 0.001 level, except for NOx g/mi and HC Damage/mi at the 25th percentile (significant at the 0.05 level), and NOx Damage/mi. 76