The Effect of Road Fuel Prices on House Prices
July 5, 2015
Wessel D. Kosterman
Master thesis
Msc Business Economics: Finance and Real Estate Finance University of Amsterdam, Amsterdam Business School
Thesis supervisor: prof. dr. M.K. Francke
Abstract In this thesis the relationship between road fuels and house prices is investigated empirically using residential transaction data from England and Wales spanning two decades. This thesis is unique because the general effect and difference in effects between urbanization classifications are estimated with a dynamic regression model, allowing for autocorrelation and delayed reactions to changes in macroeconomic variables, including road fuels. It is shown that, after controlling for other factors, a 10% increase (decrease) in the price of road fuels leads up to a 2% decrease (increase) in house prices over the next year. In an international sample of 27 European Union member countries this change is 1.1%. No convincing evidence is found for a difference in effects between suburban and urban house price reactions to road fuel price changes.
Keywords Road fuel, gasoline, oil price, commute, transportation, urban, suburban, residential real estate, house price, repeat sales regression, macroeconomic determinants.
Statement of Originality
This document is written by Wessel Kosterman who declares to take full responsibility for the contents of this document.
I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
Table of Contents
1 Introduction ... 2
2 Literature review ... 5
3 Method ... 10
3.1 Repeat sales regression ... 10
3.2 Response to road fuel price changes... 12
4 Data ... 16
4.1 Geography levels ... 16
4.2 Commute measure ... 16
4.3 House prices ... 19
4.4 Road fuel prices ... 21
4.5 Control variables ... 23
4.6 Summary statistics ... 24
4.7 European panel ... 26
5 Results ... 28
5.1 Overall house price regressions ... 28
5.2 The difference in effect between commute distance ... 32
6 Robustness ... 38
6.1 Variations in the estimations ... 38
6.2 Overall house price reaction in an international sample ... 39
7 Discussion and limitations ... 42
8 Conclusion ... 44
1
Introduction
The price of road fuels has fluctuated greatly in the past decades. At the end of 2014, the price of gasoline and diesel dropped by 6% per month in the UK (DECC, 2015). Road fuel prices affect consumer purchase decisions of durable goods in several ways. Barsky and Kilian (2004) showed that there is no evidence for a decrease in car sales resulting from an increase in oil prices. However, Busse et al. (2013) did find an effect in the type of cars that are sold. They estimate that an increase of $1 in the gasoline price per gallon increases the price of fuel-efficient cars by $354 relative to fuel-inefficient cars. This poses an important question to the relationship between road fuels and an even more durable good: the home. The home is perhaps the most important and long-lived asset of a household. House prices have been shown to move with macroeconomic variables (Adams and Füss, 2010) and the location of a household is a factor in the price elasticity of demand for road fuels (Spiller and Stephens, 2012). In this thesis the following research question is posed: How do house prices react to changes in road fuel prices?
A large body of literature is concerned with the impact of heating fuel prices and energy efficiency on real estate prices. For example, Halvorsen and Pollakowski (1981) find that the 1973 oil embargo, which raised oil prices relative to gas prices, had a significantly negative effect on the value of oil heated houses. Others have found a low discount rate used by homeowners in the valuation of energy efficient improvements (Nevi and Watson, 1998), and that certified energy efficient office buildings rent at a premium (Eichholtz et al., 2010).
However, the amount of research on the effect of road fuels on house prices is limited. Beltratti and Morana (2010) estimate the impact of global oil price changes on global fluctuations in house prices. An oil price shock is found to have a significantly opposite effect on house prices. The oil price is estimated to explain between 1 and 7% of the variability in house prices.
Others have studied the relative change in house prices of central and more remote locations within a region or metro area. The hypothesis here is that an increase in transportation costs decreases the prices of remote homes relative to central homes. In this way, transportation is seen as a substitute for a central location. Indeed, Newman and Kenworthy (1989) found that suburban households use more gasoline than urban households as they live further from work and other facilities. Coulson and Engle (1987) found a significant increase in house prices for central locations relative to non-central locations when gasoline and other money costs increased. Molloy and Shan (2010) included more macroeconomic controls and did not find a significantly larger reaction to road fuel price changes for remote locations. Finally, Morris and
Neill (2014) did find heterogeneity in the response to contemporaneous road fuel price changes across locations, although not explicitly decreasing with centrality of the location.
This thesis extends the literature in several ways. Firstly, the regression models allow for autocorrelation in house prices and lagged effects of road fuel prices and control variables. Compared to Beltratti and Morana (2010), variation in road fuel prices across countries is introduced by estimating the effect of road fuel prices including tax instead of global oil prices. This study is the first to consider both general reactions to fuel prices as well as the difference in reaction between suburban and urban areas. Also, the analysis is unaffected by usage of
alternative transportation methods because of the unique suburban-urban classification method. Finally, the sample used has low elasticity of supply for new housing, which allows an eventual effect of road fuel price changes to be capitalized into house prices.
On the basis of the existing literature two hypotheses are formed. The first hypothesis states that a change in the price of road fuels causes a change in house prices in the opposite direction, after controlling for other variables. The second hypothesis states that the opposite reaction to road fuel price changes is more severe for suburban house prices than for urban house prices.
Results indicate that a 10% increase in road fuel prices leads to a 2% decline in house values spread out over a year, even after controlling for other macroeconomic variables. However, no convincing evidence is found for a difference in the reactions between suburban and urban house prices to road fuel price changes. The estimates suggest that suburban areas have an approximately 5% bigger reaction than urban areas, although they were in most specifications statistically indistinguishable from zero.
The results are persistent with different road fuel variables, house price index estimation methods, and when estimating a subsample of only major metro areas. As a robustness check the general response of house prices is estimated using an international sample. These confirm the response to road fuel price changes although the effect is somewhat smaller (1.1% decline of house prices to a 10% increase of road fuel prices).
The study of road fuels is especially relevant in the light of the large swings of road fuel prices in the last two decades, and the desire of governments to internalize the societal costs of consuming fossil fuels in the form of fuel duties. Additionally, it sheds light on the possible implications of technological innovations in personal transportation, such as the electric vehicle and the self-driving car. These innovations have the potential to reduce both the financial and non-financial costs of transportation in general, and commuting in particular.
The remainder of this thesis will proceed as follows: Firstly, an overview of related literature is given in section 2. Secondly, the research method is explained in section 3, the data collection and the properties of the data used are discussed in section 4. The results are
presented in section 5 and robustness checks are made in section 6. Finally, the results are discussed in section 7 and a conclusion is given in section 8.
2
Literature review
In this section, first the theory underlining the hypotheses is outlined. Secondly, the previous empirical literature on the subject is reviewed and finally a comparison is made between this thesis and the existing literature.
Road fuel expenditure is part of household expenditure for every consumer. Directly by the fuel bought for commuting and other car trips, and indirectly by the fuel costs embedded in other goods. A meta-analysis by Brons et al. (2008) of the price elasticity of demand for road fuels has shown the short-run elasticity to be around -0.34. Because the demand is inelastic the households are expected to spend more on road fuels after a price increase. Therefore a rise in road fuel prices is expected to reduce the amount of disposable income left for other goods, such as housing. The reverse could also be true: a decline in road fuel prices frees up current and future disposable income to spend on housing.
As mentioned, the road fuel purchased directly by the consumer is only part of the total road fuel consumed. Reinders et al. (2003) found that about half of the total energy requirements of an average household in the European Union are indirect. Therefore, it can be argued that there are two effects of fuel price changes on households. There is the effect that is mostly invariant to the household’s location within a local region. For example, there is the cost of the fuel used in producing a good and transporting it to the store. But the fuel costs used to
transport the good from the store to the home does depend on the household’s location within the local region: the further from the store, the higher the costs. For the location invariant part, a larger proportion of expenditures on road fuel is expected to decrease demand and thereby decrease the price of residential real estate. This leads to the formation of the first hypothesis: A change in the price of road fuels causes a change in house prices in the opposite direction, after controlling for other factors.
The location-varying effect could best be understood by use of the monocentric-city model. In the monocentric-city model (Alonso, 1964; Muth, 1969) there is one central business district (CBD) where all employment is located. Commute costs are seen as the only relevant factor in the locational value of a property. In equilibrium, location rent equals household income minus transportation costs and the fixed expenditure on other goods. Therefore, the further away a household lies from the CBD, the lower the location rent. This is commonly referred to as the negative rent gradient.
When transport costs decrease, the locational value of living close to the CBD decreases and rents in the city center decline. The development in suburban rents depends on the
preferences of households. If households prefer to use all the savings in transport costs to consume more of goods other than housing, suburban rents will decline, but less so than for urban rents. In this case the city will not expand. If households instead use all of the savings to use more land, the city will expand and suburban rents will increase. The reversed effect is predicted when transportation costs increase. Urban rents (and house prices) will increase and suburban house prices will either decline or increase by a little. Figure 1 shows the case of a decrease in transportation costs graphically. The slope of the negative rent gradient flattens regardless of the spending decisions of households. This effect has an opposite sign to the general (hypothesis 1) effect for urban house values but it has the same sign as the general effect for the suburban house values. This leads to the formation of the second hypothesis: The opposite reaction to road fuel price changes is more severe for suburban house prices than for urban house prices.
Figure 1 Change in rent gradients after a decrease in transportation costs
Figure 1 shows the change in rent gradient for a decrease in transportation costs, indicated by the arrows. On the left the change is shown when all savings are used to expand households’ living space and on the right the change is shown when all savings are used for the consumption of other goods.
The earliest empirical analysis of the response in house prices to road fuel price changes was by Coulson and Engle (1987). They studied the period following the first oil shock in the US in 1973. Yearly house price data from 1974 to 1979 from six metro areas where public transport is not widely available was used (Atlanta, Detroit, Houston, Los Angeles, Minneapolis and San
Diego). The urbanization classification used in this study is whether a residential property lies within the city limits of the central city in the metro area. Using hedonic regression, the
difference in price between a property within city limits and an identical property outside the city limits is estimated. This difference for each city is regressed on estimated average transportation costs in each city. The transportation costs are separated into three costs: time, gasoline and other money costs. Only time costs are found to be significant in explaining the difference between urban and suburban house prices. However, when gasoline and other money costs where combined, the effect was significant. The authors conclude that the gasoline cost savings are overcapitalized in urban house prices, but still plausible. The effect of the time cost savings is considered unrealistically high. These results could be biased due to the exclusion of other macroeconomic factors, such as GDP and CPI changes.
Where Coulson and Engle (1987) were limited in the sample size of 6 US metro areas, the study by Molloy and Shan (2010) uses data on most US metropolitan areas over 27 years. This larger dataset allows the inclusion of macroeconomic control variables. They use repeat sales regressions to create house price indices on a zip code level. The proportional yearly change in house prices is regressed on proportional gasoline price changes interacted with a long
commute dummy. None of the included gasoline lag interactions (1 to 4 years ago) are
significant. However, the effect of gasoline prices on new home construction is also estimated. It is found that an increase in fuel prices of 1% causes a 1% decline in construction after 4 years for remote locations. This result is robust to the inclusion of contemporaneous macroeconomic variables, but models with lagged macroeconomic variables or lagged dependent variables are not tested. The authors explain the absence of a reaction in the house prices as due to the elasticity of new housing supply, which is very high in the US (Sánchez and Johansson, 2011). When the gasoline price rises, new supply in remote locations is low. In reverse, when the gasoline price declines new supply in remote locations is high. Therefore the effects of gasoline price changes are not capitalized into existing house prices in more remote locations.
Alternatively, the yearly lags could be too crude to pick up on an effect in house prices whereas they are suited for an effect in the amount of new home construction.
Morris and Neill (2014) use an area with a somewhat less elastic supply. The metro area of Las Vegas, the focus of their study, is at around the one-third mark of most supply
constrained metro areas in the US. A hedonic regression is employed with, besides property specific variables, the proportional change in real monthly gasoline price included. The results show a positive relationship between contemporaneous gasoline prices and house prices. A 1% change in gasoline prices is associated with a 0.05% change in house prices in the same direction.
The effect is found to differ substantially between neighborhoods. The 10th percentile
neighborhoods response to a 1% gasoline price increase is a 0.08% drop in house prices and the 90th percentile response a 0.18% increase in house prices. Graphically, the cross price elasticity seems to move down when considering areas further from the Las Vegas city center, but no formal tests are performed. Furthermore, no macroeconomic controls (and lags thereof) are added and no gasoline price lags are included.
The heterogeneity of the impact of gasoline price changes is confirmed by Spiller and Stephens (2012). They find a 30% more negative impact of gasoline tax increases for rural households compared to urban households.
Several studies have pointed to the rise in gasoline prices as the trigger for the 2007 US housing collapse. Cortright (2008) showed that the decline in value in 2007 was higher for suburban, more remote homes than for urban homes. This coincides with a doubling of the real price of gasoline from 2004 to 2007. Kaufman et al. (2011) show that expenditure on energy is one of the main contributing factors in the post-2005 increase in mortgage delinquency rates. Sexton et al. (2012) present a theoretical model linking the monocentric-city model (Alonso, 1964; Muth, 1969) with the Poterba (1984) model of housing investment. They show in
simulations that unanticipated increases in gasoline prices increased defaults and lowered house prices far away from the city center in the 2007 US housing collapse. In these simulations, gasoline price increases are found to affect suburban areas in California more than urban areas because of the longer commutes and lower incomes in the suburban areas.
Among other macroeconomic variables, the effect of oil prices on overall house prices has been studied by Beltratti and Morana (2010). Quarterly house price fluctuations from 1980 to 2007 are studied for the US, the Euro area, the UK, Canada and Japan. Using a factor-vector autoregressive model, a significantly negative response of house prices to oil price shocks is found for all countries except for Japan. Oil price shocks are found to explain 0.9 to 7.1% of the variability in house prices. They explain 2.9% of the short-term variability of house prices in the UK and 0.9% in the Euro area. The smaller effect of oil shocks in Europe compared to North America might be due to the higher amount of duties on road fuels in Europe. These duties generally do not directly depend on the oil price but are instead levied by volume or weight. Therefore, an oil price shock might cause a smaller relative change in consumer expenditures on road fuels than for example in North America.
The analysis presented in the following chapters improves on the existing literature on several aspects. Firstly, this study recognizes the dynamic aspects of the relationship between road fuels and house prices. It considers lags of macroeconomic variables and house price
changes which could be related to both (lags of) road fuel prices and current house prices. Secondly, the road fuel pump price per country is used instead of the oil price used by Beltratti and Morana (2010). This introduces variation in the road fuel price across countries and captures idiosyncratic shocks such as road fuel duty increases. This also gives a transportation cost
measure that is closer to the costs that are actually incurred by homeowners. Thirdly, this study is the first to jointly analyze the general effect of road fuels on house prices as well as the
difference in effect between suburban and urban areas. Previous literature each focused on just one of the effects. Also, the commute measure used to divide areas into urban and suburban relates directly to the car use, thereby effectively controlling for the effect of alternative transportation methods. That is: an area with better alternative transportation methods will be more likely to be identified as urban than an otherwise equal area. However, the effect of shifting car usage to alternative transportation methods will still be present, as desired. Finally, this study focuses on England and Wales for the analysis of heterogeneity in effects of road fuels across commute distance. The UK has a much smaller elasticity of new housing supply than the US (Sánchez and Johansson, 2011), the focus of all previous research on this topic. This could allow decreases in road fuel prices to be capitalized in suburban house prices. Also, the large dataset of house transactions allows the use of a shorter time interval of months instead of years to get a more fine-grained estimate of the potential effect. Table 1 shows the features of the empirical literature on the heterogeneity in response of house prices to road fuel price changes in comparison to this thesis.
Table 1 Empirical literature on the effect of gasoline prices on house prices
Study Coulson and
Engle (1987)
Molloy and Shan (2010)
Morris and Neill
(2014) This thesis
Time period 1974-1979 1981-2008 1967-2010 1995-2014
Location 6 US metro areas US Nevada, US England and Wales/EU
Dependent
variable House prices
House prices,
construction House prices House prices Road Fuel
variable
Gasoline fuel price per city
Expenditure on
gasoline US pump price
Country-level road fuel prices including and excluding tax
Time interval Yearly Yearly Yearly Monthly/quarterly
Observations Few (not specified) 145,444 930,720 9,526,512
HPI method Hedonic Repeat sales Hedonic Repeat sales
Housing supply Elastic Elastic Somewhat inelastic Inelastic Commute
classification
Within - outside central city limits
Commute time:
normal - high No classification
Below – above local median commute distance by car Macroeconomic
controls Not included Contemporaneous Not included Contemporaneous and lags Table 1 shows the features of the empirical literature on the heterogeneity in response of house prices to road fuel price changes.
3
Method
In this chapter the methods for testing the hypotheses are covered. First the method for
construction of the house price indices for England and Wales are explained. Next the methods for testing the relationship between road fuel prices and general house prices in England and Wales, and for the international sample of European Union members is dealt with. Lastly, the methods for testing the hypothesis of a difference in the relationship between road fuels and urban house prices versus the relationship between road fuels and suburban house prices are described.
3.1 Repeat sales regression
The scope of this analysis of the impact of house prices on gasoline prices is twofold. Firstly, the impact on house prices overall will be studied in a national (England and Wales) and
international (European) setting. Secondly, the heterogeneity in impact across residents’
commute distance will be studied using the England and Wales dataset. The overall house prices are published in most European countries, including the house prices for England and Wales. However, to test whether gasoline prices have a different impact on suburban/rural house prices versus urban house prices, a house price index has to be constructed for both groups. In an ideal situation one would directly observe the price of each house in each period. In reality, market prices are only observed when a house is sold. One could instead look at the average house price of the houses that are sold in a period, but this would be sensitive to changes in the mix and quality of the houses that are sold. Hedonic regression or the more elementary mix-adjustment method control for these differences across periods. While especially the hedonic regression method does not require a large amount of observations, both methods require a large amount of property characteristics to be reliable. Due to the limited amount of property characteristics available, like the number of rooms and the house or lot size, a repeat sales model will be used to create the price indices (Bailey et al., 1963). This method in combination with this dataset has been shown to produce reliable house price indices in the UK (Lim and Pavlou, 2007), and is the method used for the construction of the official house price index in the UK.
Instead of controlling for differences in houses sold per period, like in hedonic and mix-adjustment methods, one could also track sales of the same house over time. This is what is done in a repeat sales regression. The focus on price changes of the same house aims to ensure that the object being sold is in fact comparable. Therefore only houses are used if they are sold at least twice in the data period.
There are some caveats. Firstly, the physical properties of the house are assumed to remain constant over time. This does not hold in reality, houses depreciate and are being renovated. However, for the purpose of this study we are interested in the house prices of certain areas relative to house prices of other areas. It is reasonable to assume that physical depreciation and investments are not related to road fuel price changes, the subject of this study. Also, by removing transaction pairs with abnormal high annualized returns from the sample, it is relatively straightforward to ignore apparent renovated properties.
Secondly, there is a potential sample selection bias. Homes that are sold more often have a greater influence on the price index than homes that are sold infrequently. Starter homes or lemons could for example be sold more often. This could cause the index to be more
representative to those groups than desired (Clapp and Giaccotto, 1992). To mitigate these issues somewhat, the largest available data period will be used (20 years), and homes that are sold very frequently (more than 9 times in the data period) are removed. Additionally, the panel
regressions for England and Wales are weighted by an approximation of the number of
households per area. This prevents areas with a relatively high housing turnover from being over represented.
To identify transactions as belonging to the same property, location and house characteristics are used. Transactions are matched if the property type, land tenure, primary address, secondary address, street, locality, town/city, district, county and postcode are all the same.
This transaction matching is needed to estimate the Bailey et al. (1963) repeat sales model, which is specified as follows:
,
, = ∗ , , (1)
Where , is the initial sale of the transaction pair of property at time 1, , the final sale of the transaction pair of property at time 2, the house price index at time 1, the index at time 2 and , , the error term.
To estimate the index values, it is useful to transform the model to a linear one by taking logarithms:
ln , − ln , = ln − ln + , , (2)
Time dummies are fitted to predict each sales pair on the left hand side for each month on the right hand side. These dummies are equal to -1 for an initial sale and equal to 1 for a final sale, and zero otherwise. The regression model is thus:
After estimation by ordinary least squares (OLS), the exponent of the month coefficient minus the previous month coefficient, subtracted with 1 is the estimated monthly change in house prices.
The repeat sales occur with varying time intervals, and the standard errors could depend on the time interval. For this reason the standard errors might not be i.i.d., as assumed in the previously described model. Therefore the Weighted Repeat Sales (WRS) method as proposed by Case and Shiller (1987) will be used alongside the original (Bailey et al., 1963) method. This method consists of three steps: First the coefficients are estimated in a ‘standard’ repeat sales method as outlined above. Secondly the residuals from this prediction are regressed on a constant and the corresponding time intervals between sales. In the third step weighted least squares is used with weights obtained in the second step to estimate the coefficients on the time dummies (Case and Shiller, 1989).
3.2 Response to road fuel price changes
First, the response of overall house prices on fuel price changes is studied in the England & Wales and the European dataset. The model for the regressions of the proportional change in the house price index ( ) is:
%ΔHP , = ! + ∑# "#%Δ$%&'(# + ∑+* )*%ΔHP , (*+ , ,- . + , , (4)
where $%&' is the included road fuel variable, , ,- are the control variables, 0 the time period in
months or quarters, 1 the areas for the panel regressions, the amount of road fuel lags and 2 the amount of lags of the dependent variable included. %Δ indicates that the proportional price change is used, for example: the 3th lag of the road fuel, %Δfuel
# =
Δ$%&'1,0−3
$%&'1,0−3−1. In some regression
models quarterly lags are used while the contemporaneous changes, including the dependent variable, are on a monthly time interval. For example, the 3th lag of the road fuel becomes
89:;<, =>?(89:;<, =>(?A )
89:;<, =>(?A ) . This is done to have a fine-grained estimate of the relationships between
changes in current house prices and current macroeconomic variables while simultaneously estimating the broader reactions to lagged values.
Control variables that will be considered are (lags of) CPI change, CPI change excluding fuel and housing costs, change in the (5-year) fixed mortgage rate, change in the floating
mortgage rate, real GDP growth and change in house construction costs. Lags of changes in the HPI are included as well. Lastly, dummies representing the calendar months and quarters, included for the monthly and quarterly models respectively, capture the seasonality in house prices.
To model the overall house price response to road fuel price changes in England and Wales, first a regression is run with the monthly overall estimated house price index (HPI) on the three month relative change in road fuel prices and controls. That is, equation (4) is estimated with monthly time periods 0, lag lengths of three months and without 1’s. Besides exploring the effect of road fuel prices on aggregated house prices, these regressions will serve to determine the macroeconomic control variables and their lags used in the remainder of the study.
The procedure for the model construction is as follows, with all lags included as
subsequent quarterly percentage changes: First a regression is run with the house price indices as dependent variable, 4 lags of the house prices (instrumented by lags of levels), 4 lags of a road fuel variable, dummies indicating the quarter and the contemporaneous macroeconomic controls (GDP, CPI excluding road fuels, and construction). Subsequent lags of macroeconomic controls are added one by one to see if they significantly contribute to the model. Next, insignificant lags of house price indices are removed one by one, moving from the fourth lag to the first.
Insignificant lags of other macroeconomic controls are removed moving from both sides, as there is a priori no specific ordering in the importance of the lags. It is necessary to be
parsimonious with the amount of exogenous variables included because in the two-stage least squares regressions (see below) the instruments might otherwise not be relevant. The closest lag of the road fuels in each specification is set to the previous period, but the effect may in fact start later as house prices are well known to be slow to adjust (Case and Shiller, 1989), especially downwards (Genesove and Mayer, 2001). A priori testing will be bounded by lag periods of 1 month up to a year to limit the accumulation of the probability on a type I error.
The overall house price will be analyzed in a panel setting as well. The HPI of 123 different travel-to-work areas (see paragraph 4.1) will be regressed on the road fuel and control variables determined above. This is a dynamic panel because the lagged HPI is included as a regressor. In this case, taking first differences causes the error term to be correlated with the lagged dependent variable. This produces biased estimates (Nickell, 1981). Therefore, as
proposed by Anderson and Hsiao (1982), the lags of HPI changes are instrumented by levels of HPI from at least two periods ago. This regression model is estimated by two stage least squares, where the lags of house price changes are each estimated by the exogenous variables of equation (4) and lags of HPI levels as instruments. Standard errors are clustered by area to allow for correlation of the errors within the areas.
The England and Wales dataset has no variation in the road fuel variables and control variables (with the exception of lags of the dependent variable) across areas. To introduce variation between areas, a dynamic panel regression will also be run on an international sample.
Because of the small number of countries in this dataset (an N of 27), the Anderson and Hsiao (1982) instrumental variable method will produce biased estimates. Furthermore, a weak identification test is performed to determine whether the instruments would be relevant. The Kleibergen-Paap (2006) rank Wald F-statistic is used. Critical values are obtained from Stock and Yogo (2005). The null hypothesis of weak identification is barely rejected at 5%, whereas the instruments in all models using the England and Wales dataset are highly relevant. A Kiviet (1995) correction to a least squares dummy variable estimation is more appropriate in this setting of small N and T. This estimation adapted for unbalanced panels (Bruno, 2005) is used for the European sample. Bun and Kiviet’s (2001) bootstrap procedure is used with 2000 iterations to get an estimation of the standard errors.
Next, returning to the England & Wales sample, the regression from equation (4) will be run for urban and for suburban house price indices separately. This is done to begin exploring the differences on reactions to road fuel price changes. The lag length for each lag in these regressions is set to one month to get a more detailed view of the timing of the responses. The "#’s for the urban house prices are expected to be higher than the "#’s for the suburban house
prices. Both sets of "# are expected to be negative when macroeconomic controls are included. Compared to the overall house price regressions, significantly higher coefficients in the urban regression contribute to the hypothesis of central locations as a substitute for commute costs. Successively, significantly lower coefficients in the suburban regression contribute to the hypothesis of remote locations as complementary to commute costs.
In addition to the estimation of effects on either the suburban or urban house prices, the difference in effects will also be estimated. This is done by estimating two house price indices for each area in England and Wales: the suburban and urban HPI. The estimation is the same as in equation (4), except that the effect of fuel prices is allowed to differ for the suburban and urban HPI. In effect, it is assumed that the control variables have the same impact across commute distance. The equation becomes:
%HP ,C, = ! + ∑# D"#%Δ$%&'(#+ ∑# DE#%Δ$%&'(# ∗ F + ∑+ )*%ΔHP ,C, (*
* D + ,-. +
,C, , (5)
where S is a dummy variable indicating the HPI is from the suburban part of the area (has a higher than the mean commute distance by car, see paragraph 4.2).
The coefficients of interest in this specification are the E#’s. A coefficient significantly
below zero shows, after controlling for unobserved factors, a negative relationship between gasoline prices and suburban house prices relative to urban house prices.
Lastly, the transaction pairs will also be classified in ten deciles (G) of commute distance. This way the difference in effect of a road fuel price change on house prices when moving up one bin is represented by the E#’s in equation (6). This will also be done for the top and bottom
three quantiles only. In this case G is a dummy variable equal to one when the commute distance is in the top three quantiles (longest commute) and zero when the commute distance is in the bottom three quantiles (shortest commute).
%HPH, = ! + ∑# D"#%Δ$%&'(#+ ∑# DE#%Δ$%&'(# ∗ G + IG + ∑+ )*G
4
Data
In this chapter the collection of the datasets is explained and the properties of the variables in the dataset are shown. Starting with the England and Wales dataset, the geography levels, commute measure, house prices, road fuel prices and control variables are described, in that order. Next the summary statistics for the England and Wales dataset is presented and lastly the international sample is described.
4.1 Geography levels
In the England & Wales dataset two geography levels are used. The lower level is the middle layer super output area (MSOA). A MSOA is a continuous area with between 2000 to 6000 households. There are in total 7,201 MSOAs in England and Wales. The MSOAs are geographically grouped together to form the higher geography level: travel to work areas
(TTWAs). A TTWA is defined as the smallest area where 70% of residents work and where 70% of workers reside. Therefore each property in a single TTWA can reasonably be considered as belonging to the same local housing market. There are 178 TTWAs in England and Wales together. For the purpose of this study small TTWAs are merged together with adjacent TTWAs to ensure a minimum of 20 MSOAs per travel to work area. This is needed to prevent the creation of indices based on very few transactions per period. The random sale price noise can cause a high amount of volatility in small local price indices (Francke, 2010). Besides that there are simply fewer transactions in an area with few households, a larger portion of households will be part of the MSOA that lies on the median of commute distance, and will therefore be
excluded (see paragraph 4.2). The need to merge arises frequently in more remote and less
population dense areas, such as northwest Wales. After merging, 123 travel to work areas remain.
4.2 Commute measure
A measure of the commute distance by car is used as the indicator of urbanization in an area. This measure is used because driving amounts to over half of the total amount of travel in the UK, and commuting amounts to the largest share of the total distance driven by car (Jones and Le Vine, 2012). The commute distance is estimated with use of UK 2011 census data.
Respondents are grouped by the MSOA of residency. The geography level of MSOA ensures a detailed local view on commuting patterns while limiting random noise. For each respondent the straight line distance from the home address postcode to the work address postcode is
calculated, including respondents working mainly from their home.1 This is of course not the actual commute distance, but the factor between the real and straight-line distance is expected to be similar across areas.
This distance is combined with the main method of transportation the respondent uses to commute. The resulting values are weighted to the total workforce in that MSOA and
presented in 9 frequency bins from 0 to over 60km one-distance commutes. The midpoint of the bins (and 70km for the 60km and over bin) are multiplied by the frequency count and summed together. This gives the total commute distance per MSOA. This value is divided over the total number of working residents in the area to get to a measure of an average commute. This is done for both total commutes and car commutes only, although the car commute measure has the main focus in this thesis. The measure of car commute can thus be thought of as the average commute by car of the total working population, including those that don’t commute by car.
The distribution of average total and car commutes per area can be found in figure 2. The distribution looks like a typical lognormal distribution with its positive skew.
For most regression models used in this study, the MSOAs will be divided into quantiles based on commute distance. For example, when using two quantiles: the MSOAs with an average commute above the median MSOA in that travel to work area will be marked as remote and the MSOAs below the median as non-remote. For these two groups separate house price indices will be estimated which will form the basis of the subsequent regression analysis. A spatial representation of the commute classification for part of the travel to work area of
Leicester can be found in figure 3. Leicester is a typical medium to large sized city in the heart of England. As can be seen from figure 3 the urbanization groups are spatially reasonably well clustered. There is only one remote area fully enclosed in non-remote areas. It is interesting to note that the commute distance is not a simple function directly increasing with distance from the city center (the middle of the figure), although that is the general trend.
Figure 2 Commute distances
Figure 2 shows a histogram of the average total- and car commute per MSOA in kms. Figure 3 Commute classifications in and surrounding Leicester
Figure 3 shows the 2-quantile commute classifications for the areas in and surrounding Leicester. Blue (dark) areas are classified as remote and red (light) areas are classified as non-remote. The color-graded area is at the median of commute distances and the black areas are part of another travel to work area.
0 400 800 1200 0 5 10 15 20 25 N u m b e r o f a r e a s
Average commute per method in km
Car Total
4.3 House prices
House transaction data from the UK Land Registry is used to construct the repeat sales indices that form the independent variable in the regression models. Land Registry’s ‘Price Paid’ dataset covers nearly all single-unit residential property transactions for full market value in England and Wales. This dataset contains around one million transactions per year from begin 1995 to the end of 2014. For each transaction the price is recorded, as well as location, transaction date, land tenure, property type and whether it is a new or existing building. Figure 4 shows the overall house price developments over time. While the price trend is in general upwards, there are also some periods of house price declines, most notably following the 2008 financial crisis. What is clearly visible in this figure is the increased random measurement error when estimating house price indices of smaller areas: the HPI for Banbury is a lot more volatile than that for London. The UK Land Registry publishes its own house price index. The overall house price index constructed in this paper tracks the Land Registry’s index reasonably well. The change in the constructed HPI’s have a correlation of 0.942 (OLS) and 0.938 (WRS) with the change in the Land Registry HPI. Figure 5 shows the monthly percentage changes for the constructed OLS index and the official index.
Figure 6 shows the difference in monthly suburban and urban house price changes. Visually, no clear pattern can be discerned in the monthly changes. The standard deviation does not change when one of the two indices is lagged by one period. Therefore there is no evidence that one of the groups lags the other group in price changes.
Figure 4 House price indices for the England and Wales dataset
Figure 4 shows the house price index based on data from the entire England and Wales sample, and the largest, a medium and the smallest area (London, Leicester and Banbury respectively).
Figure 5 Comparison between the official and constructed HPI
Figure 5 shows the monthly percentage changes in the official and constructed house price indices. Both indices are estimated by OLS.
Figure 6 Differences in urban versus suburban house price changes
Figure 6 shows the monthly and cumulative difference between monthly suburban and urban house price changes. A positive value on the monthly difference shows that suburban house prices increased more than urban house prices in that period. A positive value on the cumulative difference shows that suburban house prices increased more since Jan 1995 than urban house prices. An area is defined as urban if that MLSOA has an average car commute below the median car commute for the travel to work area it belongs to.
4.4 Road fuel prices
Road fuel prices are obtained from the UK Department of Energy & Climate Change.2 Monthly price estimates are based on surveys on 10 oil companies and 4 supermarkets, covering 90% of the road fuel market. The fuels used for the purpose of this study are unleaded petrol (gasoline) and diesel. This covers nearly the entire market for consumer road fuels.
In the year 2000, petrol consumption was roughly three times as large as diesel consumption excluding lorries and busses. In the following decade the consumption of diesel increased and the consumption of petrol decreased upon being roughly equal in 2012 (UK Department for Transport, 2014). Tax revenue data confirms this trend pre-2000 and after 2012 (UK HM Revenue & Customs, 2015). Extrapolating yields a petrol share of 85% in January 1995, decreasing by 0.17% per month to 45% in January 2015. This mix of fuel consumption
2 This dataset is available on gov.uk, historical table QEP 4.1.1 as part of the Road fuel and other petroleum product
will be used in the model to weigh the estimated impact of petrol and diesel price changes on house prices.
These fuel prices will be considered both including VAT and duties (pump price) and excluding VAT and duties. The pump price has the benefit of showing exactly what consumers are paying and thus will likely be the most relevant. However, fuel tax changes are predictable. From 1993 to 2000 duty rates were increased yearly with a pre specified rate from 3 to 6% above inflation (Seely, 2011a). This ‘road fuel escalator’ was withdrawn in 2000 in favor of annual duty increases in line with inflation. Also, fuel duties have had a dampening effect on fuel price increases form 2000 onward. Initially the dampening was implicit in the postponement or
cancellation of duty increases when prices were high (Seely, 2011b). In 2011 the UK government indicated that it would explicitly dampen fuel prices by introducing the ‘fair fuel stabilizer’ (Seely, 2014). In this taxation system, tax rates will be increased on oil companies when the oil price is high and on fuel duty when the oil price is low. These patterns make changes in pump prices more predictable and could potentially bias the results. An overview of the road fuel price developments during the sample period can be found in figure 7. Diesel and gasoline prices are quite close together as the UK, unlike any other EU-member country, taxes both fuel types by the same amount. Judging from the graph, prices including and excluding taxes behave similarly, except pre-2000.
The percentage change of all fuel types are positively serially correlated with the previous month, both when including and when excluding tax. Lags of more than one month are not significant. Several studies have pointed out that the serial correlation of pump prices is asymmetric: prices react faster to crude oil increases than decreases (Borenstein et al., 1997; Manning, 1991; Reilly and Witt, 1998). A simple regression of only price increases or decreases on lags confirms the asymmetry: price decreases have significant positive autocorrelation with the previous month but price increases do not.
Figure 7 Road fuel prices in the UK
Figure 7 shows monthly road fuel prices in pence per liter. Pump prices are the average prices for that month including duties and VAT. Excluding tax are the average prices for that month excluding duties and VAT. The crude oil price is indexed with 2014 = 100.
4.5 Control variables
Hamilton (2008) argues that besides supply, speculation and demand in the form of GDP growth are important predictors for the increase in oil prices in the run up to the 2008 financial crisis, and the subsequent decline of oil prices during the following recession (Hamilton, 2009). Historically, increases in oil prices have been regarded as exogenous to macroeconomic
conditions (Hamilton, 1983). Further, it could be argued that the recent drop in oil prices is unrelated to the macro economy (Kilian, 2014). While oil prices might be fairly exogenous to the macro economy, house prices are not. GDP growth has also been shown as a determinant of house prices (Englund and Ioannides, 1997; Adams and Füss, 2010) and is interlinked with oil prices (Hamilton, 1996). House price changes are also related to past house prices and interest rates (Englund and Ioannides, 1997; Adams and Füss, 2010). Therefore, common
macroeconomic variables will be included such as inflation, GDP growth, and interest rates. Data on consumer price indices and real GDP growth rates are obtained from the Office of National Statistics. A construction cost index for new private housing is taken from the UK Department for Business, Education & skills. Interest rate statistics are obtained from the Bank of England. Specifically, the floating and five year fixed mortgage rates are used. Standard controls also include year dummies 1995-2015, month dummies 1-12 and area fixed effects for each of the 123 travel to work areas. The quarterly GDP growth rates and the quarterly change in construction costs are linearly interpolated to monthly time intervals to suit the monthly regression models.
4.6 Summary statistics
Table 2 shows descriptive statistics of monthly house prices, road fuel variables and other macroeconomic variables in England and Wales. The HPI of the whole dataset is displayed, as well as for London, Leicester and Banbury. The mean appreciation is 0.55% per month for England and Wales and 0.71% for London. London is by far the largest TTWA in the dataset, Leicester is a typical medium to large TTWA and Banbury is among the smallest ones. The noise in the price indices for smaller areas is clearly visible by the increasing standard deviation when moving from larger to smaller TTWAs. The standard deviation in the London HPI is 1.14%, in the Leicester HPI 1.66% and in the Banbury HPI 3.11%. This is likely not due to a volatile housing market for the less dense areas but due to imprecise measurement in those areas. Analytical weights based on the number of MSOAs will be used in the regressions to reflect the precision of the estimation of house prices. House prices, road fuel prices and macroeconomic variables except the interest rates have a positive arithmetic mean. Standard deviations are between 2 and 2.5% for the road fuels including tax and around 5% for the road fuels excluding tax. The mean appreciation of road fuels excluding tax per month is also higher with 0.5% versus 0.3%. Standard deviations for the monthly changes in GDP, CPI excluding road fuels and construction costs are much lower with 0.59%, 0.35% and 0.39% respectively.
Table 3 shows the correlations between monthly changes in the independent and dependent variables. Urban house prices have a correlation of 0.9 with suburban house prices. The correlations between other variables and house prices are comparable for suburban and urban house prices. The biggest difference lies in the correlation with fixed interest rate changes (0.21 and 0.26 respectively). In general GDP, fixed interest rates, floating interest rates and petrol prices are clearly positively correlated with house prices.
Table 4 shows a correlation matrix between all road fuel types and mixes. Price changes in diesel and petrol are in general highly correlated (0.92 and 0.87), but not perfectly so.
Including and excluding taxes for road fuel of the same type is roughly correlated with 0.8. Predictably the constructed mixed fuel types are highly correlated with the real road fuels. GDP is more highly correlated with fuels including tax and CPI with fuels excluding tax.
Table 2 Descriptive statistics
Variable Observations Mean Δ Std. Dev. Δ Min Δ Max Δ
HPI England and Wales
(7201 MSOAs) 240 0.55% 0.93% -2.71% 3.12% HPI London (1139 MSOAs) 240 0.71% 1.14% -3.69% 3.27% HPI Leicester (106 MSOAs) 240 0.52% 1.66% -4.33% 5.50% HPI Banbury (20 MSOAs ) 240 0.65% 3.11% -13.40% 10.62% Gasoline 240 0.32% 2.52% -10.73% 7.18%
Gasoline ex. tax 240 0.57% 6.98% -24.28% 33.44%
Diesel 240 0.34% 2.28% -7.61% 8.99%
Diesel ex. tax 240 0.55% 5.91% -15.48% 33.08%
Equal mix 240 0.33% 2.36% -9.17% 7.65%
Equal mix ex. tax 240 0.56% 6.23% -19.80% 31.76%
Usage mix 240 0.29% 2.08% -8.06% 7.23%
Usage mix ex. tax 240 0.51% 5.78% -17.86% 30.90%
GDP 240 0.52% 0.59% -2.19% 1.65%
CPI excluding road fuels 228 0.16% 0.35% -0.96% 0.97%
Construction costs 233 0.38% 0.39% -0.49% 2.10%
Fixed mortgage rate 240 -0.43% 2.77% -11.75% 12.37%
Floating mortgage rate 240 -0.21% 2.34% -15.14% 5.09%
Table 2 shows descriptive statistics for the month-to-month percentage change of the independent variables and selected HPIs.
Table 3 Correlations between independent and dependent variables
Changes in Suburban house price Urban house price House price overall
GDP 0.3277 0.3238 0.3339
CPI 0.0318 0.0611 0.0454
Fixed interest rate 0.2053 0.2569 0.2384
Floating interest rate 0.326 0.3376 0.3394
Petrol (pump price) 0.1717 0.1543 0.1652
Diesel (pump price) 0.116 0.0978 0.1081
Petrol (excluding tax) 0.1906 0.1915 0.196
Diesel (excluding tax) 0.1178 0.119 0.1213
Crude oil 0.0866 0.0823 0.0881
Suburban house price 1 0.9083 0.9761
Urban house price 0.9083 1 0.9765
House price overall 0.9761 0.9765 1
Table 3 shows the correlations between monthly changes in the dependent variables (top) and contemporaneous independents variables (left).
Table 4 Correlations with road fuel prices Changes in P et ro l (p u m p ) D ie se l (p u m p ) C ru d e o il P et ro l (e x. t ax ) D ie se l (e x. t ax ) E q u al m ix ( p u m p ) U se m ix ( p u m p ) E q u al m ix ( ex . t ax ) U se m ix ( ex . t ax ) Petrol (pump) 1 Diesel (pump) 0.9208 1 Crude oil 0.5049 0.4873 1
Petrol (ex. tax) 0.8621 0.7173 0.4950 1
Diesel (ex. tax) 0.7384 0.7407 0.4937 0.8696 1
Equal mix (pump) 0.9819 0.9780 0.5067 0.8094 0.7545 1
Use mix (pump) 0.9920 0.9459 0.4960 0.8401 0.7427 0.9898 1
Equal mix (ex. tax) 0.8328 0.7527 0.5112 0.9723 0.9610 0.8109 0.8225 1
Use mix (ex. tax) 0.8300 0.7084 0.4872 0.9918 0.9070 0.7879 0.8189 0.9854 1 GDP 0.0825 0.1238 0.0499 0.0472 0.0748 0.1043 0.0983 0.0619 0.0541 CPI 0.1501 0.1420 0.0727 0.2111 0.2057 0.1486 0.1501 0.2124 0.2089 Fixed interest rate 0.2914 0.2295 0.1629 0.2732 0.1979 0.2674 0.2772 0.2470 0.2558 Floating interest rate 0.1345 0.1154 0.0698 0.1542 0.1328 0.1280 0.1320 0.1494 0.1510 Table 4 shows the correlations between road fuel types (top and left) and macroeconomic variables (left). The equal mix consists of half petrol and half diesel price changes. The use mix is the Gasoline and Diesel price changes weighed by the estimated household consumption of the fuel type over time. 4.7 European panel
Data on quarterly house prices, GDP and residential construction costs for European Union members are obtained from Eurostat. Data on CPIs excluding road fuels is obtained from the federal reserve bank of St. Louis (FRED). Data on the quarterly road fuel prices for each country are obtained from the UK Department of Energy & Climate Change. Road fuel prices reported are gasoline including tax, diesel including tax, gasoline excluding tax and diesel excluding tax. These prices are also combined to form an equal mix of fuels, both including and excluding tax. The data on road fuels, house prices and other macroeconomic indicators is available from Q1 2005 to Q4 2014 for most countries. Therefore this is the data period considered in the analysis.
Summery statistics of the variables are shown in table 5. The quarter-on-quarter percentage changes, the standard deviation of the changes per country and the average change per country are reported. For each of these three statistics the minimum, median and maximum values are reported. Inter-country correlation coefficients of the variables are shown in table 6. The low average correlation coefficients show that there is a reasonable amount of variation in
the included variables, especially considering that the transformed average correlations are somewhat overestimated (Silver and Dunlap, 1987).
Table 5 Summary statistics for the international sample
Variables Proportional change Standard deviation Average
min med max min med max min med max
HPI -21.51% 0.51% 13.10% 0.85% 2.41% 6.97% -1.45% 0.39% 1.89% Gasoline -20.05% 1.19% 26.41% 3.81% 5.94% 7.74% -2.30% 1.35% 2.37% Gasoline ex. tax -38.28% 1.61% 45.02% 5.53% 11.31% 13.49% -4.22% 2.40% 2.90% Diesel -19.40% 1.30% 27.27% 2.96% 6.09% 7.75% -1.95% 1.51% 2.12% Diesel ex. tax -29.00% 1.93% 31.49% 3.88% 9.40% 10.16% -3.37% 2.00% 2.43% Equal mix -19.22% 1.24% 25.39% 3.24% 5.79% 7.47% -2.14% 1.41% 2.08% Equal mix ex. tax -23.11% 1.37% 24.17% 3.61% 7.37% 8.82% -2.72% 1.74% 2.12% GDP -13.05% 0.43% 6.25% 0.57% 1.05% 2.55% -0.49% 0.31% 0.97% CPI -2.91% 0.47% 5.84% 0.36% 0.79% 2.64% 0.27% 0.54% 1.21% Construction -13.25% 0.46% 21.70% 0.42% 0.84% 6.17% -0.07% 0.54% 1.64% Table 5 shows the minimum, median and maximum quarter-on-quarter percentage changes of all variables and minimum, median and maximum country level standard deviations and averages of those changes.
Table 6 Inter-country correlations of macroeconomic variables
Variables Min Average (transformed) Max
HPI -0.45 0.27 0.88
Road fuel mix including tax 0.05 0.73 0.96
GDP 0.01 0.58 0.89
CPI excluding fuel -0.45 0.29 0.84
Construction costs -0.40 0.35 0.81
Table 6 shows the correlations between countries of the proportional quarter-on-quarter changes in the corresponding variable. Average correlations are obtained by the average of the Fisher z-transformation of the coefficients, transformed back to a Pearson correlation coefficient.
5
Results
In this chapter the results of the regressions are presented. The analysis of the effect of road fuel prices on house prices is presented first and the analysis of the difference in effect across
commute distance second.
5.1 Overall house price regressions
The first regressions are performed with the change in overall house price index as dependent variable as in equation (4). The regression results can be found in table 7. The usage mix
including tax is used as the road fuel variable in the models reported. Coefficients are similar for other road fuel variables. Without controls included (model 1), the three-month relative change in road fuel price has a positive relationship with house prices. This is in line with Morris and Neill (2014). A 1% increase in road fuel prices coincides with an increase in the HPI of 0.03%, whereas the effect is 0.05% in the Morris and Neill (2014) study. Further lags are significantly negative, with coefficients up to -0.09 in the road fuel price of three quarters ago. This could however be due to factors not yet included in the model, other macroeconomic factors could be correlated with both house and road fuel prices.
Contemporaneous macroeconomic variables are included in the second model as well as the lags of the change in house prices. The three-month house price change has a positive and highly significant coefficient. This confirms the existence of autocorrelation in house prices, as for example in Adams and Füss (2010). The CPI coefficient also has a significantly positive coefficient of 0.66. This shows that in this model, after controlling for other factors, a 1% higher price level excluding road fuels coincides with 0.66% higher house prices. The quarterly
percentage change in the 5-year fixed mortgage rate and quarterly GDP growth are also significantly positive. The coefficient on the most recent road fuel price is no longer positive. However, the second and third lags of the proportional change in the road fuel price are negative and significant at 1%.
In the third model, dummies are included for the calendar month. In this model only the third lag remains significant at 1%. The first and second lags are significant at 10% and the fourth lag is not significant at all. The CPI coefficient is now also no longer significant, likely due to the seasonality in the CPI. The month dummies show seasonality in house prices with
somewhat lower prices in January, September and October. The biggest estimated difference in HPI change is 1.3%, between the months of January and April. This means that the proportional
change in house prices is predicted to be 1.3% higher in April than in January, which is quite a small difference.
Throughout these models the coefficients on the mortgage rates are insignificant and positive, contrary to previous findings (Englund and Ioaniddes, 1997). Therefore these variables will be excluded from the control variables used in the remainder of this thesis. In contrast to the study by Adams and Füss (2010), the relationship between house prices and construction costs is insignificant. However, in that study the quarterly construction costs are used. Although the three-month relative change in construction costs does not have a significant relationship, the sign is positive as expected. Therefore this variable will remain in the final model. Only the first lag of HPI changes is significant, therefore the other variables will be excluded. And finally, the first quarterly lag of GDP growth is included as it is found to be significant for some models.
The final set of control variables (model 4) are: the three month changes in house prices from one month ago, the contemporaneous consumer price index and its quarterly lags up to a year ago, GDP growth and its first quarterly lag, the first quarterly lag in relative change in new homes construction costs, and calendar month dummies. As for the road fuel coefficients, a 10% change in road fuel prices causes a 0.28% opposite change per month in house prices three quarters later. The coefficients on the first and second lags are about half of that and the coefficient of the fourth lag is even less pronounced. The combined predicted effect of a 10% increase in road fuel prices in a quarter is a 1.82% decrease in monthly house prices spread out over a year. This is calculated by ∏N# K(1 + "#)LM − 1, where the "#’s are the road fuel lag
coefficients, in this case (1 − 0.0140)L∗ (1 − 0.0149)L∗ (1 − 0.0284)L∗ (1 − 0.0092)L− 1. The third power of
the coefficients is taken because the lags are in quarterly price changes whereas the house prices are in monthly changes. A change in a quarterly fuel price will thus appear three times in
subsequent time periods for each lag.
Model 5 shows the dynamic panel regression with the same variables included as in model 4. It can be seen that this model has greater statistical power than the previous models. Although the magnitude of the road fuel coefficients is the same, the coefficients are highly significant. In this prediction, an increase of road fuel prices of 10% causes a decline of house prices of 2.00% spread out over the following year.
Figure 8 presents the breakdown of the predicted values of the monthly HPI change versus the actual observed HPI change. The value predictions are based on model 4 in table 7 and displayed from January 2007 to January 2015. This period is characterized by large swings in both road fuels and house prices. The relative contribution to the prediction of the road fuel variable can most clearly be seen in periods where all independent variables predict the same
direction of the change in HPI. For example, in 2008 during the biggest decline in house prices, all explanatory variables predicted a decline in house prices. In this period the proportional impact of past road fuel price changes on the total prediction is around 20%. In the late stage of the recovery following this period the proportion is even higher, up to around half of the full prediction. In the following period the increased road fuel prices impact the prediction
negatively, but some other explanatory variables impact the prediction in the opposite direction. Therefore the relative contribution of road fuel prices on house prices in this period is less clear. Overall the impact of road fuel prices on house prices during the housing crisis and the
following financial crisis seems economically significant and in line with findings from Sexton et al. (2012).
Table 7 Overall house price regressions England & Wales
Model (1) (2) (3) (4) (5)
Dependent variable Aggregate
HPIº Aggregate HPIº Aggregate HPIº Aggregate HPIº HPI per TTWAº Road fuel usage mix lag 1^ 0.0262 (1.63) -0.0090 (-0.90) -0.0161* (-1.89) -0.0140* (-1.86) -0.0082** (-2.51) Road fuel usage mix lag 2^ -0.0293* (-1.78) -0.0304*** (-3.18) -0.0165* (-1.69) -0.0149 (-1.50) -0.0211*** (-10.10) Road fuel usage mix lag 3^ -0.0436*** (-3.30) -0.0256*** (-3.01) -0.0276*** (-3.12) -0.0284*** (-3.04) -0.0359*** (-14.82) Road fuel usage mix lag 4^ -0.0374*** (-2.81) 0.0007 (0.08) 0.0007 (0.09) -0.0092 (-1.15) -0.0081*** (-2.85) HPI lag 1^ 0.1905*** (7.72) 0.1914*** (6.36) 0.1634*** (8.16) 0.1832*** (9.31)
HPI lag 2^ -0.0363 (-1.48) 0.0090 (0.23)
HPI lag 3^ 0.0047 -0.0442
(0.18) (-1.30)
HPI lag 4^ 0.0020 (0.11) 0.0024 (0.12)
CPI ex. road fuelsº 0.6646*** (5.03) 0.0515 (0.25) 0.1002 (0.54) 0.2303** (2.43)
CPI ex. road fuels lag 1^ -0.0415 (-0.35) -0.1118*** (-3.24)
CPI ex. road fuels lag 2^ -0.0881 (-0.71) 0.0209 (0.67)
CPI ex. road fuels lag 3^ -0.2045** (-2.06) -0.0526** (-2.03)
CPI ex. road fuels lag 4^ 0.1494 (1.40) 0.0109 (0.48)
Fixed interest rateº 0.0357** (2.12) 0.0199 (1.40)
Floating interest rateº 0.0076 0.0229
(0.35) (1.04)
Gross domestic productº 0.3352*** (3.55) 0.2546*** (3.56) 0.3057** (2.57) 0.2114*** (6.41)
Gross domestic product lag 1^ -0.0168 (-0.42) -0.0044 (-0.25)
Construction costsº 0.1433 (1.10) 0.1302 (1.13)
Construction costs lag 1^ 0.0253 (0.74) 0.0151*** (2.71)
Constant 0.0068*** (9.50) -0.0002 (-0.22) -0.0063*** (-2.77) -0.0037 (-1.28) -0.0043*** (-5.27)
Table 7 Overall house price regressions England & Wales
Model (1) (2) (3) (4) (5)
Area fixed effects No No No No Yes
Observations 239 219 219 208 26,370
Root mean squared error 0.0087 0.0057 0.0051 0.0050 0.0230
Combined effect road fuel -0.229 -0.178 -0.165 -0.182 -0.200
Robust t-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1, ^ Δ / V( , º ΔW/W(
Table 7 shows the regression results for the monthly overall house price index on the road fuel variable and controls (equation 4 with 0 equal to 1 month and lag lengths of 3 months). The road fuel variable in the models presented here is a mix of gasoline and diesel according to the estimated actual use of UK consumers across time. This mix is used including VAT and duty. Lags are the preceding three-month percentage change in a variable, i.e. lag 1 is the percentage change from 4 months ago to 1 month ago, lag 2 from 7 to 4 months. Model 5 is weighted by number of MSOA’s per TTWA, HPI lags are instrumented by previous levels of HPI and standard errors are clustered by TTWA. The variables indicated with ^ are included as quarter-on-quarter proportional changes (Δ / V(), the variables indicated with º are included as month-to-month proportional changes (ΔW/W(). The combined effect is calculated by ∏N# K(1 + "#)LM − 1,
where the "#’s are the road fuel lag coefficients.
Figure 8 Contributions to predicted HPI changes from Jan 2007 to Jan 2015
Figure 8 shows the contribution of the independent variables to the prediction of the dependent variable per month from January 2007 to January 2015. The contributions to the predictions are presented in bars stacked on top of each other. The red line displays the actual observed HPI changes.
-3% -2% -1% 0% 1% 2%
HPI lag variable Other independent variables
CPI variables GDP variables Road Fuel variables Actual observed HPI
5.2 The difference in effect between commute distance
In this section the difference in effect of road fuel price changes on house price changes across commute distance will be tested for the England and Wales sample.
Firstly, the England and Wales overall HPI per commute classification is regressed on monthly road fuels and monthly controls (equation 4 with lag lengths of 1 month). This is done to explore the timing of the effects and to spot any differences between effects on urban and suburban house prices. Table 8 presents the road fuel lag coefficients and t-values. It might seem unintuitive at first that the coefficients on the monthly lags individually have around the same size as the quarterly lags. However, a change in the road fuel price is reflected three times per lag with quarterly lags whereas it is reflected only once in the monthly lags.
In general, the 2nd and 9th lags of road fuels have a significant effect on both urban and suburban house prices. The lag of month six is significant at 10% for the urban house prices but not for suburban house prices. In general, the effect looks similar for urban and suburban house prices with sometimes a bit larger coefficient on the urban lags. The effect seems to be more spread out in the suburban areas than the urban areas. For example, the coefficient of road fuel lag 9 is more negative with respect to urban house prices, but the coefficients of lag 8 and 10 are more negative for the suburban house prices. The combined effect is similar to previous results. A 10% increase in the road fuel price causes a decline of 2.38% and 2.36% for urban and suburban homes respectively. The most interesting takeaway is that most of the coefficients are negative, indicating that, holding all other variables constant, an increase in fuel prices coincides with a decrease in house prices within the next year for both urban and suburban house prices.
