_{1}

When purchasing a car in Switzerland, consumers can consult three quality labels. First, a reliability label provides the probability that a given car will have a breakdown in the next year. Second, an energy efficiency label, based on fuel consumption and carbon dioxide emissions, ranks cars from A to G. Third, a safety label, based on crash tests, ranks cars from five to one stars. The aim of this paper is to establish whether these labels have an impact on the price of new cars in Switzerland. Using a hedonic regression model, we show that the reliability label influences the price. More precisely, a decrease by one per thousand of the breakdown’s probability increases, on average, the car price by 639 Swiss francs. For the energy efficiency and the safety labels, the hedonic model attains its limits due to collinearity issues and the question cannot be answered in a satisfactory way.

The Swiss car market is very interesting as no domestic car manufacturers are present. It provides therefore an appealing benchmark to study the behaviour of consumers without facing a domestic bias. In this small country of about 8 million inhabitants, around 4.5 million cars are registered. Founded in 1896, the Touring Club of Switzerland (TCS) has 1.6 million members and is an inevitable actor in the Swiss automotive sector. Among others, TCS provides useful information for people who want to buy a car. For instance, they have a database containing comprehensive characteristics for the majority of car models available in Switzerland. Furthermore, they also perform independent tests for most of them. These tests are greatly prized by people facing the decision of choosing a new car. They also contribute to disseminating three quality labels that guide consumers in their choice. The first one is a reliability label and is computed by the German automobile club Allgemeiner Deutscher Automobil-Club (ADAC). ADAC offers to more than 18 million members breakdown assistance and, from this, compiles breakdown statistics each year and publishes reliability indices for car models. More precisely, this index is the probability for a car of a given model and a given age to have a breakdown in the next year [

The aim of this paper is to establish, using a hedonic approach, whether these three quality labels have an impact on the price of new cars. Hedonic pricing supposes that the price of a product can be explained by its characteristics and has become a popular method among researchers. It permits, among other things, estimating the value of each single characteristic of the product, or, in other words, each single attribute. Though Haas [

For cars, hedonic price indices have been widely used as they clearly outperform standard indices [

Besides indices, hedonic pricing has also been extensively used to estimate the value of a car’s attributes. For the attributes considered in the present paper, namely, reliability, energy efficiency and safety, we can mention, among others, the following studies.

For the reliability attribute, [

For the attribute of energy efficiency, the results are disparate. Reference [

For the safety attribute, [

In the present paper, we establish that the reliability influences the price of new cars in Switzerland. More precisely, a decrease by one per thousand of the breakdown’s probability increases, on average, the car price by 639 Swiss francs. The research proposed in this article is original in two ways. To our knowledge, it is the first time that the ADAC reliability label has been used in a hedonic pricing model. The second original contribution is the choice of the sample. Indeed, to avoid overwhelming the regression with many options, we only took into consideration car models with homogeneous optional features.

The present paper is organised as follows. In Section 2, we present the methodology and data used for this study. In Section 3, the findings are presented and discussed. Finally, in Section 4, we conclude and outline further research.

Hedonic regression is a powerful tool for estimating the revealed value of the salient attributes characterizing a given product or service. Roughly speaking, it consists of explaining the price, the dependent variable, by different salient attributes, the predictors, using standard regression techniques. For attributes of a qualitative nature, so-called dummy variables are necessary. The dummy variable is set to 1 if the characteristic is present and 0 otherwise.

For our study, the hedonic regression can be written as

P i = β 0 + ∑ j β j x i j + ε i , (1)

where P i represents the price of the car model i, x i j the value of the attribute j for the car model i, β j the regression coefficient for the attribute j, β 0 the intercept coefficient, and ε i an error term for the price prediction of the car model i. As usual, one seeks the coefficients that minimize the sum of the squared error terms. The regression coefficients are a good estimate of the value of the corresponding attribute. Very often, a semi-logarithmic model is used. In such a model, the price is replaced with ln ( P i ) and the regression coefficients give an estimate of the price elasticity of demand for the corresponding attribute.

The principal difficulty in hedonic regression is the choice of the predictors that will be retained in the model. Obviously, predictors with a strong significance must be incorporated in the model and predictors with a low significance must be discarded. Sometimes, this task is not so easy, as the significance of a predictor depends on the predictors that are already incorporated in the model. The classical way to handle this problem is to use stepwise regression.

Although very controversial, stepwise regression remains a widely used method for selecting predictors in a hedonic regression. At each step, variables are included or excluded in the model considering a given criterion. Usually the criterion chosen is a t-test, an F-test, or the Akaike information criterion (AIC) proposed by [

It is beyond the scope of this article to discuss the weakness of stepwise regression (see for example [

It is well known that a strong correlation between two or more predictors causes problems for multiple regression (see for example [

A smarter way to handle the problem of collinearity is to use a principal component regression. For this method, first a principal component analysis (PCA) is performed on all the predictors considered [

For this study, we considered cars’ technical data that are easily available to Swiss consumers. As mentioned earlier, the TCS provides a user-friendly website gathering all the information of interest for people wanting to buy a car [

The predictors considered for the hedonic model are listed in ^{1} the SFOE energy efficiency label and the Euro NCAP safety label. Note that in our study, all cars have a safety label better than or equal to three stars.

For the qualitative variables, the modelling was done with so-called dummy variables. For instance, for the drive wheels characteristic, the variable DRIVE can take three different values (front-wheel, rear-wheel and four-wheel) and two dummy variables are necessary for the model. The first dummy variable, FRONT-WHEEL, is set to 1 in the case of a front-wheel drive and 0 otherwise. The second variable, REAR-WHEEL, is set to 1 in the case of a rear-wheel drive and 0 otherwise. The third dummy variable, namely FOUR-WHEEL, is obviously redundant and should not be included in the regression as it could create collinearity problems. Similarly, for each qualitative variable, one dummy variable is redundant and must be discarded.

^{1}More precisely, as this study considers new cars, we take the statistics for the youngest cars.

In Switzerland, as in many other countries, car manufacturers propose, for each model, a basic version that can be personalised by adding optional features. The list of options is considerably long and, if they are taken into account in the model, will necessitate many variables that will overwhelm the hedonic regression. In this study we can avoid this issue by taking into account, for each car

Quantitative variables | ||
---|---|---|

Name | Description | |

BREAKDOWN | Probability to have a breakdown in the next year. This is the reliability factor. | |

POWER | Power in PS. | |

LENGTH | Length in centimetres. | |

DOORS | Number of doors. | |

PLACES | Passenger places. | |

LOAD | Car load in kilogrammes. | |

WEIGHT | Weight in kilogrammes. | |

FUEL | Fuel consumption in litres per hundred kilometres. | |

CO2 | Carbon dioxide emissions in grammes per kilometre. | |

TORQUE | Engine torque in Newton-meter. | |

Qualitative variables | ||

Name | Description | Dummy variables |

ENERGY | Energy efficiency label | A, B, C, D, E, F and (G) |

SAFETY | Safety label | 5-STARS, 4-STARS and (3-STARS) |

ENGINE | Engine position | FRONT-ENGINE and (REAR-ENGINE) |

DRIVE | Drive wheels | FRONT-WHEEL, REAR-WHEEL and (FOUR-WHEEL) |

CLASS | Class according to TCS | SUPERIOR, MEDIUM, LOWER-MEDIUM, |

CROSS-COUNTRY, MINIVAN, SMALL and (MICRO) | ||

BRAND | Brand | AUDI, BMW, CHEVROLET,... |

The redundant dummy variables are in brackets.

model, only one version: the most basic version with the minimal features given in

In Switzerland, the market is dominated by the petrol engine, that accounts for 76 percent of the market, and we therefore did not consider diesel, liquefied petroleum gas (LPG), or electric engines. Though the share of automatic transmission has drastically increased, from 20 to 45 percent in the last two decades, manual transmission still remains more popular. For this study, we therefore consider only the version with manual transmission. Few brands proposed basic models without air conditioning or a global position system (GPS). For these cases, we chose the most basic version having both options. Finally, the six last features in

To conclude, note that the sample contains 65 distinct models made by 22 different manufacturers and covers the vast majority of cars sold in Switzerland.

To select the predictors for the hedonic regression, we used the three standard methods of stepwise regression. For the criterion, we considered two alternatives:

Petrol engine |
---|

Manual transmission |

Air conditioning |

Global position system (GPS) |

Anti-lock braking system (ABS) |

Electronic stability program (ESP) |

Power steering |

Airbag driver |

Airbag passenger |

Airbags lateral |

the AIC and a two tailed t-test. For both criteria, we additionally imposed that variables which are not strongly significant be discarded from the model. For this purpose, we used a two tailed t-test and fixed the significance level at 0.01. The numerical computations were done with the R software package.

First, we investigated linear regressions considering all predictors. When a problem of collinearity was encountered, the less significant variable was excluded from the model. The different methods lead to the hedonic pricing model

P ≈ 258.81 ⋅ P O W E R − 639.14 ⋅ B R E A K D O W N + 8345.68 ⋅ S U P E R I O R (2)

with an adjusted R-squared of 0.9797. Details of the regression can be found in

The resulting model has an excellent adjusted R-squared (0.9797) with only three predictors. In addition, each of the three predictors has a very good significance (two sided t-test p-values below 3.15∙10^{−4}). Note that although the predictor TORQUE has a good p-value (0.0018), it has been excluded from the regression due to the high correlation (0.94) with the predictor POWER. Finally, note that although two brands (HYUNDAI respectively MERCEDES) have good p-values (0.0079 respectively 0.0040), they have been excluded from the regression due to their dependence on the variable BREAKDOWN. To establish this, we performed a Kruskal-Wallis test and reject the hypothesis that the variables BRAND and BREAKDOWN are independent (p-value = 0.0005). We did not use the more classic analysis of variance (ANOVA) as the normality assumption was not met (Shapiro-Wilk test of normality with p-value 8.41∙10^{−5}). Note that ANOVA, even if not used, confirms the Kruskal-Wallis test (p-value = 3.42∙10^{−9}). The exclusion of the variable TORQUE is obvious but the exclusion of the variable BRAND is less evident. We therefore also investigated models where this variable was not excluded. The different methods lead to four different hedonic models: a first model containing 10 predictors with an adjusted

Predictor | Estimate | Std. Error | t value | p-value (two sided) | |
---|---|---|---|---|---|

POWER | 258.81 | 6.84 | 37.84 | <2∙10^{−16} | *** |

BREAKDOWN | −639.14 | 125.71 | −5.08 | 3.66∙10^{−6} | *** |

SUPERIOR | 8345.68 | 2186.53 | 3.82 | 3.15∙10^{−4} | *** |

***p-value < 0.001 (two sided). Adjusted R^{2}: 0.9797. R^{2}: 0.9806.

^{2}A Shapiro-Wilk test of normality with p-value 1.56∙10^{−6} shows that the normality assumption is not met for the variable POWER and the use of ANOVA is therefore not recommended. However, ANOVA would confirm the result (p-value = 0.0003).

^{3}ANOVA was not used as the normality assumption is not met. However, ANOVA would confirm the result (p-value = 0.0045).

Finally, to avoid collinearity problems in a smarter way, we also conducted Principal Component Regressions (PCR). The results were not as good as the linear regression. This comes from the fact that the relevant variables explaining the price are, with the principal component analysis, diluted into many principal factors. These factors are computed to explain, in an optimal way, the variance of all predictors considered but they are not necessarily, as in our case, optimized to explain the variation of the price. For our case, if we handle collinearity by excluding variables, the standard linear regression outperforms PCR. However, to be fair, we have to mention that PCR outperforms standard regression when collinearity is not treated. We also recognize that in many cases PCR is clearly the best method to deal with collinearity problems.

In the proposed hedonic model, the car price increases, on average, by 259 Swiss francs when the power increases by 1 PS. The price decreases, on average, by 639 Swiss francs when the probability of having a car breakdown during the next year increases by one per thousand. We also see that if the car belongs to the superior category, its price increases, on average, by 8346 Swiss francs. This can be explained by two reasons. First, in this category, the quality of materials and the finishing touches are better than in other categories. Second, as we mentioned earlier, for each model, we only took into account the basic version satisfying some minimal criterion. In the superior category, the basic version includes features that are optional in other categories.

The retained hedonic model contains only three predictors. This was possible since, for each car model, we only took into consideration the basic version with the minimal options described in

In this study, we imposed that variables which are not strongly significant be discarded from the model. As mentioned earlier, we discarded predictors with a p-value above 0.01. This selective choice is justified for this study, as the adjusted R-squared of the model is very high (0.9797) but could be misleading for other cases.

Using the hedonic model presented in Equation (2), we can now answer our initial question: “what are the prices of reliability, energy efficiency, and safety?”. For the reliability label, we can assert the following proposition.

Proposition 1 In Switzerland, the price of a new car depends on the ADAC reliability label.

Indeed, we can confidently reject the hypothesis that the price does not depend on the reliability label (two sided t-test, p-value = 3.66∙10^{−6}). This result is in line with comparable studies [

^{2}A Shapiro-Wilk test of normality with p-value 1.56∙10^{−6} shows that the normality assumption is not met for the variable POWER and the use of ANOVA is therefore not recommended. However, ANOVA would confirm the result (p-value = 0.0003).

For the energy efficiency label, the two sided t-test p-values lie between 0.8783 (A label) and 0.4647 (E label). These values are above all usual significance levels and we cannot reject the hypothesis that the price does not depend on the energy efficiency label. However, this result has to be taken with caution. Indeed, the energy label and the engine power are not independent, as a Kruskal-Wallis test^{2} on the variables POWER and ENERGY shows (p-value = 0.0041). It is therefore not improbable that a possible effect of the variable ENERGY on the price was in fact taken into account in our model by means of the variable POWER. We reach here the limit of our quantitative model and have to admit that no definitive answer can be provided. This result has to be put into perspective with other studies obtaining mixed results [

For the safety label, the two sided t-test p-values lie between 0.6117 (three stars) and 0.2705 (four stars). These values are above all the usual significance levels and we cannot reject the hypothesis that the price does not depend on the safety label. However, like the energy label, the safety label is not independent of the engine power, as a Kruskal-Wallis test^{3} on the variables POWER and SAFETY shows (p-value = 1.17∙10^{−4}). This result has to be taken, therefore, with extreme caution and, as with the energy efficiency label, we must admit that our question cannot be answered. Furthermore, studies using a similar hedonic model seem to contradict this result [

We found that reliability is an important salient attribute for cars and estimate the marginal price of reliability. As a subsidiary question, we would like to estimate if the price of reliability benefits from possible economies of scale. To do so, we will see the problem from the point of view of the manufacturer. The production costs related to reliability can be split into fixed costs and variable costs that depend on the production volume. Reasonably, we assume that variable costs are linear. We can write this as follows.

^{3}ANOVA was not used as the normality assumption is not met. However, ANOVA would confirm the result (p-value = 0.0045).

C ( s ) = C 0 + C 1 ⋅ s , (3)

where C ( s ) represent the total costs for reliability, C 0 the fixed costs, C 1 the marginal costs and s the quantity produced. Hence, the reliability costs per car are given by

C ( s ) s = C 0 s + C 1 . (4)

In the hedonic model presented in Equation (2), the second term (i.e. C 1 ) was considered, whereas the first term (i.e. C 0 / s ), which represents the economies of scale, was not. We therefore extend the regression model to also incorporate this first term as a new predictor. Data for production volumes were obtained from [

In this study, we investigated, using a hedonic approach, whether the price of new cars in Switzerland depends on three popular quality labels and established the following results.

• The price depends on the ADAC reliability label. More precisely, a decrease by one per thousand of the probability of having a breakdown during the next year increases, on average, the car price by 639 Swiss francs.

• Due to collinearity issues, we cannot establish whether the SFOE energy label has an impact on price.

• Due to collinearity issues, we cannot establish whether the Euro NCAP safety label has an impact on price.

These results illustrate superbly the strengths and limitations of hedonic pricing. When used with proper judgement, hedonic regression can be a powerful tool to estimate the salient attributes value of a good or a service. However, before making any conclusion, the hedonic model has to be analysed with a critical point of view. Special attention should be paid to collinearity and to the predictors retained in or discarded from the model. For example, in this study, out of three questions, we were able to answer only one.

The research proposed in this article is original in two ways. To our knowledge, it is the first time that the ADAC reliability label has been used in a hedonic pricing model. We were able to show that this popular label in Switzerland has an impact on the car price. This confirms results from other studies which showed that reliability is an important salient attribute. The second original contribution is the choice of the sample. Indeed, to avoid overwhelming the regression with many options, we only took into consideration car models with homogeneous optional features. Actually, for each car model, only the basic model with some minimal options was considered. We were surprised not to find other studies concerning cars following the same method, as it leads to a model with strong explanation and few predictors.

For further research, it would be interesting to tackle the problem with conjoint analysis, another method that permits estimating the value of salient attributes [

This work was supported by RCSO Economie et Management (HES-SO).

The author declares no conflicts of interest regarding the publication of this paper.

Moresino, F. (2019) A Hedonic Approach to Estimate the Price of Reliability, Energy Efficiency and Safety for New Cars in Switzerland. American Journal of Industrial and Business Management, 9, 468-481. https://doi.org/10.4236/ajibm.2019.93031