Measuring the Impact of Urban Sprawl on Vehicle Usage and Fuel Consumption

1
Measuring the Impact of Urban Sprawl on Vehicle
Usage and Fuel Consumption

• Tom Golob
• University of California Irvine
• tgolob_at_uci.edu
• ITLS - Sydney Seminar
• November 2005

2
Objective

• Accurately estimate the impacts of land use
  density on car usage
• Important for evaluation of policies concerning
• sustainable growth
• greenhouse gas emissions
• Evidence in the debate about car dependency

3
Measuring car usage

• Total distance driven by all household vehicles
• result of many travel demand choices
• car ownership
• trip generation
• mode choice
• including drive vs. car passenger
• destination choice
• Total fuel usage on all vehicles
• vehicle type choice
• implicit choice of fleet fuel efficiency
• vehicle allocation in multi-vehicle households

4
Measuring land use density

• Census data (U.S., 2000, with updates)
• Typical variables
• housing units per sq. mi. (per area unit)
• population per sq. mi.
• jobs per sq. mi.
• Resolution (U.S.)
• Census tract (average size 4,000 persons)
• Census block groups (average 1,000)
• Other GIS functionality available

5
Previous studies aggregate

• Compare averages for cities, zones, neighborhoods
• Impossible to control effectively for differences
  in
• Household characteristics
• Transport infrastructure
• Transport levels of service
• Arrangement of land uses
• Culture

6
Previous studies disaggregate

• Household observations
• Must control for self-selection with respect to
  residential location
• density related to neighborhood attributes
• housing quality
• transport level of service by mode
• transport preferences
• schools, recreation sites,
• cultural and ethnic identity

7
Our approach to the problem

• Make choice of residential density endogenous
• Simultaneous equations with three endogenous
  variables
• residential density
• annual mileage
• fuel consumption
• All endogenous variables explained by household
  characteristics
• The residential density variable affects the two
  travel variables

8
Simultaneous system 3 endogenous variables
9
Data requirements

• Annual mileage for all household vehicles
• derived from odometer readings or imputed
• Fuel usage calculations for all vehicles
• according to vehicle make, model and vintage
• Census data on land use density
• matched to household location

10
Data availability

• The 2001 U.S. National Household Transportation
  Survey (NHTS) data
• annual mileage for all household vehicles
• fuel usage for all household vehicles
• census data on land use density
• 24-hour travel diaries for all members
• 28-day record of long-distance travel (50 mi.)
• demographics and socio-economics

11
2001 U.S. NHTS data

• National sample
• about 26,000 households
• 82 have complete data on fuel usage
• N 21,347
• Residential density in terms of
• housing units per sq. mi. at census block level
• six categories
• scaled in terms of category means

12
Mileage, fuel usage by residential density
13
Vehicle ownership by residential density
14
Demographics by residential density
15
The missing data problem
16
Biases due to missing data

• Probability of being missing related to levels of
  the endogenous variables
• Classical sample selection problem
• Reference
• Tom Golob and Dave Brownstone (2005)
• The Impact of Residential Density on Vehicle
  Usage and Energy Consumption
• Working paper EPE-011, University of California
  Energy Institute
• on the web at
• University of California eScholarship Repository

17
Correcting estimates

• Structural approach
• Heckman selection modeling
• Add equation to construct a new hazard for sample
  inclusion
• Problems
• Results are sensitive to model specification
• Inconsistency when variable sets overlap

18
Correcting estimates

• Weighting
• Weighted Exogenous Sample Maximum Likelihood
  Estimator (WESMLE)
• Problem
• incorrect coefficient (co)variances
• standard errors will be under-estimated

19
Estimation method

• Weighted estimator (WESMLE)
• Estimates using weighted data are robust
• Standard errors seriously downward biased
• Standard errors are accurately estimated using
  Wild Bootstrap method
• Heteroskedasticity consistent covariance matrix
  estimator
• Cannot reject that errors are exogenous using
  Structural (Heckman) approach

20
Model fit on U.S. national data

• Model structure
• 19 exogenous variables
• recursive structure for the 3 endogenous
  variables
• 48 free parameters
• Weighting is important
• estimates different from unweighted estimates
• bootstrap tests reject alternative specifications
• Model fits well
• All overall goodness-of-fit statistics excellent

21
National results
Increase in density of 1,000 households / sq. mi. Increase in density of 1,000 households / sq. mi. Increase in density of 1,000 households / sq. mi. Increase in density of 1,000 households / sq. mi.
Change in annual total mileage on all household vehicles Change in annual fuel consumption (gals./yr.) Change in annual fuel consumption (gals./yr.) Change in annual fuel consumption (gals./yr.)
Change in annual total mileage on all household vehicles Due to mileage Due to fleet fuel economy Total
- 1,630 - 74 - 16 - 90
22
Interpretation

• Comparing two households identical in terms of
• income, retirement status
• numbers of drivers, workers, children
• education of head
• race and ethnicity
• Household A, living in density of 3-5,000 hh./sq.
  mi.
• will drive 3,300 fewer miles on all vehicles
• consuming 180 less gallons of fuel annually
• than
• Household B, living in density of 1-3,000 hh./sq.
  mi.

23
Important exogenous variables

• Income
• Number of drivers
• Number of workers
• Whether household single-person
• Number and age of children
• Education of head(s)
• Whether household retired
• Race/ethnicity

24
Some exogenous effects
25
Tests of alternative models

• Error term correlations
• all can be rejected (no correlation with sig. t )
• ??2 6.35 3 d-o-f (not sig.)
• Feedbacks
• drive more move to higher density
• (t 1.27) ??2 1.52 1 d-o-f (not sig.)
• higher fuel usage move to higher density
• (t 1.03) ??2 1.02 1 d-o-f (not sig.)
• Base model best according to Bayesian criteria
  (CAIC)

26
Applications to individual areas

• Need approximately 225 observations
• rules-of-thumb based on
• number of variables
• number of free parameters
• Translates to 275 at 82 non-missing data
• 2001 U.S. NHTS data will support modeling for
• 30 states
• 17 metropolitan areas

27
Contrasting results for 3 NHTS samples

• National
• N 21,347
• Oregon (including 2 counties in Washington State)
• N 325
• California
• N 2,079

28
Residential densities for 3 NHTS samples
29
Densities for 3 other NHTS samples
30
Results by area
Increase in density of 1,000 households / sq. mi. Increase in density of 1,000 households / sq. mi. Increase in density of 1,000 households / sq. mi. Increase in density of 1,000 households / sq. mi. Increase in density of 1,000 households / sq. mi.
Change in annual total mileage on all household vehicles Change in annual total mileage on all household vehicles Change in annual fuel consumption Change in annual fuel consumption Change in annual fuel consumption
Change in annual total mileage on all household vehicles Change in annual total mileage on all household vehicles Due to mileage Due to fuel economy Total
U.S. - 1,630 - 74 - 16 - 90
OR - 1,340 - 57 - 20 - 77
CA - 1,000 - 43 - 16 - 59
31
Extensions

• Results similar, but less precise when using
  other available NHTS density variables
• Can be extended to estimate effects of
  residential density on specific aspects of travel
• e.g., trips by public transport
• a different estimation method should be used for
  limited dependent variables (those with large
  spikes at the value zero)
• that estimation method requires larger sample
  sizes (perhaps 1,500 minimum)

32
Conclusions methodological

• In measuring the effects of residential density,
  it is important to control for
• selectivity bias in residential location choice
• missing data related to the endogenous vars.
• Survey data needs
• odometer readings
• vehicle specs. (make, model, vintage of all)
• residential location
• Appropriate land use data can easily by added to
  survey data sets using GIS

33
Conclusions Empirical

• Lower residential density does lead to greater
  vehicle usage, controlling for other influences
• Greater fuel consumption is due to both longer
  distances driven and vehicle type choice
• Results show the importance of using disaggregate
  data and controlling for self selection
• many household characteristics
• including race and ethnicity