E
STIMATING
R
EAL
E
CONOMIC
D
EPRECIATION
OF
R
EAL
E
STATE IN
T
HE
N
ETHERLANDS
B
Y
:
J.H.J.
W
IKKERINK
A
BSTRACT
:
Following recent developments, real estate practitioners are increasingly interested in a more economic approach to the valuation of real estate. In order to serve this interest, this research quantifies real economic depreciation in The Netherlands and attributes the effect to multiple factors. Using a unique, Dutch dataset containing a total of 2016 price observations and 516 rent or gross initial yield observations, three hypotheses are tested with OLS regressions and SURE estimations. Testing for depreciation rates shows that a non-constant specification of age is preferred over constant specification. This is in contrast with practice and most existing literature, in which depreciation is accounted for as a constant rate. The paper is unable to verify whether depreciation is driven by a larger extent by lower rents instead of higher yields; however it shows that depreciation is dependent on the type of real estate. The real depreciation rate that results from the quantitative analysis specifies an average of 1.82% in year 1 and 0.19% in year 38. Following these results, the impact on practice is demonstrated by two exemplary calculations. The calculations show a possible underestimation of total depreciation of 4.8% of initial value in 5 years in real terms.
Masters’ thesis MSc Economics MSc Finance Word count: 13,817 S1889060
I would like to thank DTZ Zadelhoff for providing the dataset, Dr. Brunia for counselling and ir. Van Polanen
1.
I
NTRODUCTION
The last ten years were tumultuous for the Dutch real estate market. As described in „De Vastgoedfraude‟ by Van der Boom & Van der Marel (2014) a large real estate fraud network was brought to light in 2007 by Dutch authorities in the „Klimop‟ case. Following this large case, the 2007/2008 mortgage crisis hit worldwide preluding a period of recession. During this recession, the real estate market faced the aftermath of both the Klimop case and the crisis. As a result, public opinion of real estate worsened. The Dutch authorities monitored real estate practitioners more closely, similar to the way accountants have been monitored for a longer amount of time. This resulted in initiatives by practitioners and authorities, such as the recently founded Dutch real estate appraisers register: Nationaal Register Vastgoed Taxateurs (NRVT). Following this scrutiny of the real estate market, real estate practitioners are increasingly interested in a more economic approach to determining the value of real estate. The goal of these efforts is to consistently estimate parameters in the valuation process. These consistent estimates should in turn lead to the accurate determination of value. The increased accuracy enhances the quality of delivered services and is able to take away concerns that customers or authorities might have regarding estimated value. Additionally, clearly communicating these parameters could improve transparency and reduce information asymmetries in the market. Transacting or owning real estate assets, practitioners hire appraisers to estimate market value. These market values are estimated to adjust book value of assets or to determine transaction prices. In the appraisal process, two dominant methods are used. The direct capitalisation method estimates rent for the property based on references and the appraiser’s experience. The estimation of a competitive yield results in the market value of the property. The
discounting method forecasts income generated by the asset and discounts those incoming cash flows correcting for the
time value of money and riskiness of the investment. Often the discounting method assumes the sale of the property after 5 or 10 year. The proceeds from this sale are based on the direct capitalisation method in the year of the forecasted sale using a cap rate creep to account for increased riskiness due to aging. Practitioners use the direct capitalisation method more often than the discounting method as the latter is more complicated than the former.
Real estate assets tend to lose value over time. This depreciation is caused by structure deterioration or by economic obsolescence. Valuing real estate assets, the effect of depreciation has to be taken into account. Discrepancy between the estimated loss of value and the actual loss of value leads to either capital gains or loss. In case of capital gains, it is likely that the transacted property was transacted under market price, leading to a net loss for the seller. However, when the buyer has to re-evaluate the actual value of her/his property to a lower value, it can lead to constraints on liquidity. Hence, incorrectly estimating depreciation rates can lead to the obstruction of valid investments or even bankruptcy. So in order for the real estate market to function healthily, depreciation has to be estimated as correctly as possible. In practice, depreciation rates are not the result of a comparative analysis as with the determination of market rent or required return on capital. Gaining insight in and further quantifying depreciation rates can be used to improve valuations based on both the income capitalisation method and the discounting method. This research contributes to practice by providing quantitative results that can be used as input in valuation techniques, leading to more accurate valuations.
Academics have estimated depreciation rates for a long time. However, these studies are mostly performed from an accounting point of view and recent studies have primarily focussed on using new databases with old techniques. Where accounting measures are used for taxation policies, it is in the interest of investors to take a more economic point of view. Following the work of Bokhari and Geltner (2016), this research takes an investor’s point of view. Bokhari and Geltner (2016), show a considerable amount of academic relevance in terms of quantifying the drivers of depreciation. However, their work does not extensively show to what extent their findings are significant. This research complements their research by providing an extensive analysis concerning the significance of similar results. By doing so, this research contributes to the academic field of real estate economics by further specifying real economic depreciation of real estate and its drivers, as well as the verification or falsification of previous research findings.
transaction price, corrected for inflation, is estimated and analysed in the time period from 2002 to 2015. The set shows an average depreciation rate of 1.18% per annum in the first forty years of a property, ranging from 1.82% in the first year and 0.19% in the thirty-eight year. The quantitative results are unique and intuitive, serving to consistently and adequately quantify depreciation of Dutch real estate.
Following the findings on real depreciation, this paper presents resulting implications for practice that impact investment decisions to a considerable extent. It is common practice to estimate the depreciation rate as a constant rate, while the cap rate creep is defined as a fixed percentage point increase per year. Using a relatively simple example, the results from this research lead to a recalibration of real depreciation for a holding period of 5 years by 4.8% for an Office property and 0.04% for an Industrial property. The same example shows that cap rate creep should be recalibrated by a total of 0.19% in 5 years.
The research question addressed in this research is formulated as:
“What are the factors driving real economic depreciation on the Dutch real estate market and how can this depreciation be quantified?”
2.
R
ESEARCH
F
RAMEWORK
This section will start with a short review of the leading and relevant existing literature in the field of real estate economics, and sums up to where this research will elaborate in the field of real estate economics. Following the review, the concepts of real economic depreciation, redevelopment and value of real estate are discussed. These concepts define the framework in which the research is conducted and imply some restrictions, which are pointed out along with the corresponding concept.
2.1 L
ITERATURE REVIEW
Research on depreciation of real estate with regression analysis is not innovative in the sense that it has been researched to a great extent. Most literature on depreciation uses a fiscal point of view, leading to recommendations for taxation policy. The paper by Taubman and Rasche (1969) is considered to be one of the fundamental papers on depreciation of real estate. They propose a model in which the building depreciates to total obsolescence and at that point is fit for redevelopment. The authors define the economic depreciation of a real estate asset to be equal to the change in present value of the asset. Assuming this change in present value is equivalent to the change in price, economic depreciation is approximated by the marginal effect of age on price. By estimating a constant depreciation rate, Taubman & Rasche (1969) calculate the economic life span between 65 and 80 years for offices.
Consecutive research on real estate depreciation is mainly focussed on owner-occupied houses. The relative homogeneity of these assets and the availability of data allows for regression models with high explanatory power. However, when considering depreciation of commercial real estate, literature is sparser. Furthermore, these papers mainly consider commercial real estate from a fiscal point of view, like Taubman and Rasche (1969). Recent contributions to the field of real estate depreciation focus on applying existing methods to new datasets.
Fisher et al (2005) is considered to be one of the latest large contributions to the field. They use the National Council of Real Estate Investment Fiduciaries (NCREIF) dataset, consisting of 1516 income properties1 that are traded mostly
by asset managers. The authors analyse the marginal effect of age on transaction price to proxy economic depreciation. Defining real depreciation as the conventional constant rate of initial value, they find that depreciation is approximately 2.7% per annum for the whole set, based on total value and 3.25% per annum for the structure only. Using these rates, they manage to estimate the economic life of real estate assets and show that the estimated economic lifespan is 30.5 years. This is in contrast to the taxable life at that time of 27.5 years. The authors argue that this discrepancy could lead to the misallocation of capital in the U.S.
Elaborating on the work of, among others, Fisher et al. (2005), Bokhari and Geltner (2016) propose an investors’ perspective in their paper. Their interest is in the economic aspects of depreciation instead of the often used fiscal
point of view. Instead of estimating an economic lifespan, Bokhari and Geltner (2016) empirically test drivers of depreciation. Their article is controversial in the sense that the authors test a non-constant depreciation rate of total initial value2, instead of the predominant constant variant. Additionally, the authors have access to a much larger dataset in comparison to earlier research, as they work with 55,913 observations. They find an overall average depreciation rate of 1.3% per annum, ranging from 2.2% per annum for new properties and 0.4% for 50-year old buildings. Their analysis on the drivers of depreciation is thorough and focuses on differences between regions, types and on the extent depreciation is caused by a decrease in Net Operating Income (NOI) and an increase in cap rate. However, their analysis does not go into details about the significance of these differences. In their regression analyses, the age parameters are significant; however there is no thorough analysis on whether the marginal effect of age is significantly different for a specific age.
Testing the model used by Bokhari and Geltner (2016) on the Dutch market with an extensive dataset allows for research into whether their findings describe an economic phenomenon that occurs globally or solely in the U.S. Furthermore, testing for statistical differences of marginal effects might shed light on the validity of their results. In performing such analyses, this research contributes to existing literature by either falsifying or validating their results.
2.2 C
ONCEPTS
In order to gain insight into the underlying assumptions of this research, the two main concepts underlying economic depreciation are keynoted. First the driving forces behind depreciation are explained, using a framework based on Baum (1993). Secondly, the implications of the recurring redevelopment cycle are set out. Finally, the underlying economic assumptions of value are set out.
Baum (1993) describes the full taxonomy of depreciation and shows the relation between depreciation and obsolescence. His framework defines the drivers of depreciation and is visualized in Figure 1. The framework defines depreciation to be a result of both property and tenure factors. Tenure factors are changes in value due to tenant related aspects such as lease contracts and government intervention. Property factors are driven by either value differential of the site or depreciation of the building. Value changes of the site can be attributed to either market movements in comparison to other regions or environmental depreciation. Baum (1993) states that in practice, total depreciation is mostly a result of building depreciation. Building depreciation, in turn, can be due to physical deterioration or building obsolescence. This deterioration refers to the property succumbing to the rigors of time. Building obsolescence refers to the concept that a building loses its usefulness in course of time, as it cannot fulfil aesthetic, functional or social needs.
FIGURE 1: STYLISED VISUALISATION OF BAUM’S (1993) FRAMEWORK
The framework shows that the value of the total asset decreases over time due to multiple factors. These factors are not only physical but intangible as well, such as changing user preferences and riskiness. These factors give a deeper understanding of the possible drivers behind depreciation of real estate, which in turn gives theoretical grounds for differences in depreciation.
The Baum framework regards depreciation from the perspective that a property is built and depreciates until it is completely worthless. However, Baum’s perspective does not align completely with this research. This paper takes a
recurring redevelopment cycle point of view. In this point of view, the asset depreciates due to multiple factors, but in line
with appraisal practice, the building can be redeveloped to „Highest best use‟ (HBU).
Bokhari and Geltner (2016) propose a framework in which four values are regarded: the HBU value, the Land value, the Redevelopment option value and the Property value. In this framework, the HBU represents the value of the property, when it is in the best possible state, given the location. Re-achieving this state can only be achieved by redeveloping the asset. In line with Baum (1993), land does not deteriorate but is prone to market dynamics. This implies that HBU value is contingent on land value as HBU value is the sum of the land and the construction. The redevelopment option value is dependent on the value that can be created by redeveloping, given the location’s restrictions. Right before the redevelopment, the option has the highest value. The redevelopment option will be at its minimum right after redevelopment. As time progresses, the potential increase in value in case of redevelopment increases and with it the right to exercise this redevelopment increases in value. The fourth concept is Property value. This value relates to the value that will be transferred during a transaction, given a willing buyer and seller. This framework, as depicted in Figure 2 assumes that when Property, Land and the Redevelopment option values are equal, the optimal point to redevelop has been reached. The redevelopment increases transactional value to HBU value leading to the start of a new redevelopment cycle. This research assumes that properties are transacted according the corresponding transaction price in Figure 2.
FIGURE 2: STYLISED VISUALISATION OF THE RECURRING REDEVELOPMENT CYCLE, FOLLOWING BOKHARI AND GELTNER (2016)
From the Baum framework and the recurring redevelopment cycle, follow the reasons for depreciation and the cycle of depreciation. However, in order to measure depreciation, a definition for in the recurring redevelopment framework is needed. Considering the first principles of finance, the Gordon Growth Model (GGM) defines the value of a growing perpetuity of cash flows as the free cash flows in the consecutive year, divided by the asset’s required opportunity cost of capital rate minus the long term growth rate of the cash flows:
(1)
, where is value at time , is the free cash flow at time , is defined as the opportunity cost of capital
at time and is defined as the expected growth rate of cash flows at time . The opportunity cost of capital is driven by the asset’s riskiness and the compensation for the time value of money. The long term growth rate reflects to what extent the cash flows grow in the long run.
However, when valuing financial assets, the perpetuity condition is often violated, as investments mostly have finite horizons. To control for this violation, value is defined as the value of a growing annuity of cash flows. This definition corrects the value of a growing perpetuity with a correction term. This term is dependent on the growth rate of cash flows, the opportunity cost of capital and the amount of periods that the asset generates income. The growing annuity of cash flows is defined as:
[ ( ) ]
(2) , where is the amount of periods the asset generates cash flows and the other parameters have the same meaning as in equation (1).
sold asset at time , discounted to present value using the discount rate. Including this value in the equation leads to: [ ( ) ] (3)
Valuing real estate has its roots in these principles of finance, but the real estate application differs on a few aspects. Income property is often valuated using an Income Capitalisation Approach. Following the report by Hungria-Garcia (2004), this approach can be divided into two groups:
- Capitalisation of a perpetuity - Discounted cash flow method
The two real estate valuation approaches are based on a comparison of transactions from both the user market and the investment market. Whereas the user market determines the income of a property, the investment market determines the required return on capital needed to fund the property. Any valuation of income property should therefore reflect both markets in order to estimate an adequate value of property.
The capitalisation of a perpetuity is based on the expectations of income for the first year and reflects the market’s expectations. In practice, the income parameter used in these calculations is not free cash flows, as in equations (1) - (3), but Net Operating Income (NOI) as this parameter is well-used. NOI is calculated detracting costs made when renting out the property, such as the operating and maintenance costs along with taxes, from the received rents. Acquisition costs are not taken into consideration for this income parameter. The use of NOI instead of free cash flows implicates the first discrepancy between the first principles of finance and real estate practice.
The assumption in the capitalisation approach is that the income of the first year will be received for forever, similar to the Gordon growth model as defined in (1). Another difference is that real estate practitioners use the income parameter of the current year, instead of the use of the cash flow from the following year in the first principles of finance.
(4)
, where signifies the transaction price3 at time , is defined as Net Operating Income in and is defined as the Net Initial Yield (NIY) at time . The NIY is used often in real estate and is defined as the opportunity cost of capital for the income parameter minus the perpetual growth of the income parameter:
(5)
The NIY represents a „value weighted average‟ of both and and can be interpreted similarly to a valuation multiple. In the case of business valuations, multiples as EBITDA and EBIT multiples are often used by professionals when valuing an investment in order to increase the comparability between valuations and are in general more comprehensible to practitioners.
The real estate market is facing issues regarding market transparency, leading to information asymmetries. This problem trickles down to practice in, among other ways, the fact that information on costs is often unavailable. That is why these costs are in practice estimated using a fixed percentage of rent, or as a fixed amount per meter. These estimations are based on the experience of the appraiser. The unavailability of information regarding costs leads to the fact that practitioners use a contingent capitalisation approach: direct capitalisation of rent. Practitioners use rent instead of NOI, because information on rent is more available. The definition of value, according to the capitalisation of perpetual rent is as follows:
(6)
Where is defined as the rent generated by the income property and is defined as the Gross Initial Yield (GIY). As the NOI embodies more information than rent does, the NIY is based on more information than the GIY is. In this line of reasoning the valuation outcome of the direct capitalisation of rents can be considered to be an inferior measure with regard to the direct capitalisation of NOI. The risk associated with the property partially determines the yield. As a result the NIY tends to be lower than the GIY, as there is less uncertainty regarding costs.
Due to data limitations, the direct capitalisation of rents is used in this research. Therefore, this research assumes that the costs identifying the difference between rent and NOI are a constant fraction of rent. Whereas this research will define the rent as rent per meter, this is in line with the common assumption of costs being a fixed proportion of the income or per square meter.
In order to model for the depreciation of real estate, is implemented in the framework, representing the real annual depreciation rate:
(7)
, in which is the real depreciation rate at time . The importance of correct estimation of the depreciation rate follows from equation (7). Incorrectly estimating the depreciation rate leads to discrepancy between expected underlying value and actual underlying value. When holding a real estate asset in The Netherlands, the asset has to be on the balance sheet as market value. Correctly specifying the depreciation rate lowers the probability of unexpected amortisations or capital gains due to incorrectly specifying depreciation rate when the property is sold. Furthermore determining important investment parameters such as optimal holding period and internal rate of return are influenced by incorrectly estimating depreciation. A rational investor would therefore attempt to estimate depreciation as correctly as possible.
Regarding the discounting method, depreciation plays the most prominent role regarding the exit value of the asset. This method is similar to the growing annuity of cash flow equation as defined in equation (2). The discounting method differs in that NOI is allowed to differ between periods and therefore the growth rate is not included in the equation. The value of the property is assumed to equal the summed present value for each cash flow in the considered timeframe. For real estate, the discounting method is defined as:
∑ (8)
, where is defined as the price of the property at the end of a holding period of years, or exit value, and the
other parameters have their usual meaning.
In practice is calculated assuming that the NOI will grow with inflation. On the denominator side, NIY will
increase to reflect a higher risk for older property. This increase in yield is referred to as the cap rate creep.
∏ ∏
(9)
Where is the inflation rate at year and is the cap rate creep of the NIY in year and the other parameters have their usual meaning.
2.3 Hypotheses
As stated in the introduction, the research question is defined as:
“What are the factors driving economic depreciation on the Dutch real estate market and how can the depreciation process be quantified?”
Three hypotheses are formulated to answer the research question:
H1: Real economic depreciation of Dutch real estate is best described using a non-constant depreciation rate.
H2: Real economic depreciation of Dutch real estate is to a larger extent dependent on changes in rent than on the cap rate creep effect.
H3: Real economic depreciation of Dutch real estate is dependent on the type of real estate.
Hypothesis 1 challenges the usual constant depreciation rate used in practice. Hypothesis 2 is derived from the concept of value as defined by the direct capitalisation method and reflects market discipline for age by either the user or investor market. Hypothesis 3 follows from Baum’s framework, where differences in market dynamics are expected to reflect in differences in depreciation. Combined, the three hypotheses test the functional form, the impact of the underlying real estate markets and whether there should be diversification in depreciation for different real estate types.
H
YPOTHESIS1
As an alternative to the constant rate of depreciation, as real estate depreciation is often regarded, Bokhari and Geltner (2016) propose the depreciation rate to be non-constant. First of all, this non-constant rationale can be related to the non-linear derivative of in equation (8). Furthermore, this is in line with the reasoning that Akerlof (1970) proposes in his famous „lemon cars‟. Akerlof (1970) argues that market asymmetries due to the technical complexity of assets result in quality uncertainty. This in turn is likely to result in non-constant depreciation of cars in the second hand market. This rationale could be applied to real property as well, as technical complexity can be argued to lead to information asymmetries in the real estate market as well. Additionally, a state-of-the-art premium can be argued to exist for property that includes the most advanced technologies and architecture on the market. State-of-the-art tends to develop over time. This development forces property to forgo such a premium, when it is not new anymore. Furthermore, economic theory originating from behavioural economics states that an agent’s time preference is a diminishingly decreasing function4. As yield is a function, of among others, agent’s time preferences concerning risk and time preference, it can be expected to behave exponentially as well.
Following evidence by Bokhari and Geltner (2016) and the presented theoretic arguments, it is expected to find that depreciation rates of real estate are non-constant, in contrary to current practice.
H
YPOTHESIS2
Following equation (6), the value of a real estate asset has two possible drivers of depreciation: either as the numerator decreases or the denominator increases.
The increasing denominator refers to the cap-rate creep. The risk on investment increases as the property deteriorates or becomes obsolete. Following the cyclical framework, a large redevelopment might be needed earlier in time when the transacted property is older. The investor will take the asset’s increased riskiness into account and assign a higher return on invested capital to compensate for the risk. Due to this increased requirement, the asset’s value decreases leading to depreciation of the property. The cap rate creep is also an input parameter for the discounting method of valuation in (9).
On the numerator’s side of equation (6), rents are expected to decrease as the building ages. Following Baum’s Framework, a property can either deteriorate or become obsolete. Both of those effects lead to lower end-user demand for older property which in turn lowers rent, given supply.
Hence, total depreciation can be attributed to either cap rate creep or decreasing rent. However, the degree in which these effects cause depreciation is not clear. These two movements are a result of market dynamics in either the investment market or user market. Whereas the required cap rate is defined the investors’ market, rents are, in the long run, a result of supply and demand dynamics. If increasing cap rates would determine depreciation to a large extent, this would imply that the decreasing value is driven by investors’ required return on investment. However, when depreciation would seem to be mainly driven by rents, it would suggest that users have market power, leading to rents that reflect the age of the building. Given the supply surplus present in the Dutch real estate market and considering that Bokhari and Geltner (2016) state that most of the depreciation of real estate is due to lower net income, the expectation is that the effect of rent will be larger than the cap rate creep effect.
H
YPOTHESIS3
Following the Baum framework, the course of total depreciation of real estate is partly determined by user’s preferences. The user of a logistics centre has different preferences than the user of a high-end office building in a business district. These preferences influence price, importance of locational factors, aesthetic requirements and preferred time span of usage for example. Differing types of real estate have differing functions and therefore different users, with differing sets of preferences. Additionally, the market for space is likely to be contingent with the market their users face. Different users would in turn lead to different market dynamics. The differences between the real estate types can, in line with Baum (1993) results in differing depreciation. Following these rationales, hypothesis 3 states that depreciation differs for different types of real estate5.
3
R
ESEARCH
D
ESIGN
Keeping in mind the implications of the presented literature and concepts, this section describes the strategy used to test hypothesis 1, 2 and 3. After defining the appropriate models, the DTZ Zadelhoff database is presented with data definitions and descriptive statistics. The data is analysed in order to affirm some of the expectations presented in the preceding sections. This section concludes with validity checks of the models and a preliminary presentation of results.
In order to make a distinction between depreciation and other effects on price, influences on price that are not due to ageing have to be separated from the ageing effect. Wherethe model has to capture the effect of age on price as purely as possible, the hedonic regression approach is followed. This approach identifies and controls for price effects that are not due to ageing. The hedonic regression approach estimates changes in transaction price, rent or capitalisation rates based on characteristics of the property. The underlying theory assumes that agents implicitly price real estate characteristics. Aggregating demand and supply allows for the derivation of marginal contribution to price per property characteristic. Property characteristics that are used in such regressions can include size, age, location, energy label and other type-specific characteristics. The set of control variables in this research is divided in three categories: hedonic, agent and spatiotemporal characteristics.
Following the method of Bokhari and Geltner (2016), the variables are combined into a regression, leading to the estimation of the following equation.
(10)
, where is defined as the transaction price per square meter, corrected for inflation, per square meter and is defined as age.
Variable matrix in equation (10) is defined as hedonic property characteristics. To control for differences in hedonic characteristics between real assets, residential hedonic regression models refer to specific characteristics such as the
presence of a balcony, the house having large windows or a large living room. An example of such a model can be found in Liu (2013). However, for commercial real estate such information is often unavailable. Therefore, research using the hedonic regression method for commercial real estate6 is followed in the selection of hedonic variables. As this research analyses multiple types of real estate, common denominators for all types are formulated. Variable matrix consists of: size, property type, whether the property is unused and a dummy denoting a high probability of excess land.
Variable matrix is defined as agent characteristics. The model controls for agent characteristics in line with behavioural economics theories. Bokhari and Geltner (2016) use seller and buyer characteristics in their hedonic pricing model. These characteristics control for behavioural differences. In one of their preceding articles, Bokhari and Geltner (2011) find that the behavioural concepts of loss aversion and anchoring have effect on real estate prices in the US. Whereas loss aversion represents a higher utility differential when confronted with loss instead of gain, anchoring refers to the phenomenon where price-setting affects ultimate transaction price. Additionally, Clayton et al (2009) find that real estate pricing is affected by investor sentiment. However they conclude that fundamentals are still the key driver of those same prices. Furthermore, Wiley (2012) shows that corporate investors are buying real estate with a premium and selling it with a discount. He attributes this effect to specific actor characteristics such as impatience. The results from these empirical papers ratify the modelling of transacting agent characteristics in the regression model. Similarly to Chegut et al. (2015), the transacting agents are divided in four categories that can be either Dutch or international. Variable matrix contains a dummy variable for each category, whether the transacting party is national and whether the transaction is a sale-leaseback construction.
Variable matrix is defined as location characteristics and variable matrix consists of time effects. The influence of time and place on transaction price, or spatiotemporal effects for short, has been analysed extensively in real estate literature7. The cyclical nature of the economy influences transaction prices over time. Furthermore, regions tend to deviate in level of development, due to a range of factors as described in the field of urban economics. Previous hedonic regression studies controlled for these effects by embedding these spatiotemporal differences in the model. Following these studies, the model accounts for spatiotemporal differences. Variables in matrix are in which region the property is located and the distance to logistic channels such as train-stations, airports and highways. Variable matrix consists of year dummies.
denotes the error term of the corresponding estimation. The corresponding parameter vectors denote the
regression parameter, where indicates the independent variable of the regression and indicates the corresponding dependent variable. Because the dependent variable is stated in logs, the parameters capture the marginal contribution to transaction price of hedonic, agent and spatiotemporal characteristics as a fraction of total value.
Complementary to regression (10), the following two regressions are estimated:
(11)
(12)
, where is defined as the rent per square meter, corrected for inflation, per square meter and is defined as the gross initial yield of the investment.
Following the method Bokhari and Geltner (2016) propose the right-hand variables are identical for all three regressions. Additionally, Fuerst & McAllister (2011) and Shimizu (2015), Saderion et al. (1993) and Sirmans, Sirmans and Benjamin (1989) estimate differences in rent and cap rate using models equal or highly similar to the model for price as well. Price, rent and GIY are by definition interdependent. The interdependence of the three left-hand variables reduces the regression’s efficiency, as the error terms are likely to be correlated across the three regression equations. However, as the right hand side of the regressions are identical, the resulting estimators are valid. Zellner (1962) proposed the seemingly unrelated regression equation (SURE) method to account for the correlated error terms and . Following this method makes the estimation’s standard errors more efficient.
However in this case, the left-hand variables are not only interdependent; they describe a definition. As a result it is not possible to estimate the three regression equations simultaneously, as the covariance matrix of the error terms becomes singular. Therefore, three SURE regressions are estimated containing all three possible combinations of two dependent variables. Comparing these three regressions shows that the results are equal in terms of regression coefficients, standard deviation and R-squared parameter for each dependent variable. Hence, the results from these three estimations can be reduced to one regression outcome table for each left-hand variable.
In the analysis of models, the models estimated with the SURE only report an squared instead of an adjusted R-squared. Because the SURE is based on a General Least Squares (GLS) instead of an Ordinary Least Squared (OLS), the R-squared is a flawed measure for comparison. Another disadvantage of the SURE is the loss of observations as it requires the availability of information on rent and GIY. Loss of observations leads to a reduced ability to infer results. Therefore, analyses that only solely involve price are based on the OLS as defined in equation (10). Analyses that require rent and GIY results are based on the SURE estimation.
In order to check the validity of the model, the underlying assumptions are checked. To test for heteroscedasticity, the Breusch-Pagan test is executed. In case this test is significant, robust error terms are used in the corresponding regression. This will not be possible when the SURE estimation is used, as the error terms are already defined by SURE. Multicollinearity refers to models where independent variables show high correlation with each other. This causes explanatory variables to be dependent of each other leading to the fact that one explanatory variable can be used to estimate the other. In order to test the variables in this model, the Variance Inflation Factors (VIFs) are calculated. As these VIF-coefficients increase, the multicollinearity problem becomes larger. A rule of thumb is that this value is not to exceed 10. However, the inclusion of categorical or exponential variables can force such VIF-coefficients to be relatively high without multicollinearity being a problem. Following Pregibon (1980), the model is tested for correct specification. The residuals and the squared residuals are regressed on the dependent variable. When the squared residuals are significant in this regression, there is reason to believe that the inclusion of the squared term of a variable or a cross-product improves the model. In order to test for functional form, the Ramsey F-test is executed. When this test is not significant there is reason to believe that the functional form of the current model is incorrect. Concluding the regression robustness procedure Cook’s distance test is performed, which checks whether there are influential observations that might disturb the estimated model.
From regressions (10), (11) and (12), the average cumulative depreciation and the depreciation rate per annum as a function of age can be estimated. For price regression (10) the average cumulative depreciation function is defined as:
(13)
, where represents the cumulative depreciation of price at age as a fraction of value at age 0. The parameters for age ( ) and age squared ( ) are deduced from regression (10).
The depreciation rate per annum for price is defined as the first derivative of age for equation (10):
(14) , where denotes the average depreciation rate of transaction price per annum.
The cumulative average change in rents per meter and the cumulative average cap rate creep ( and ), along with the change in rent per annum and cap rate creep per annum ( and ) have the same definition as and :
(15)
(16) , where denotes the idenotes the index for left-hand variables and .
(17)
(18)
, where is defined as the change in price per annum caused by left-hand variable . and are the
regression coefficients for the corresponding left-hand variables from regressions (10), (11) and (12).
As redevelopment data is not available, it is not possible to model the effect of redevelopment on value as discussed in the recurring redevelopment cycle. However, the effects of redevelopment are influencing the results. Redevelopments cause the absolute value of the redeveloped property to increase. As the average property ages, the percentage of properties that underwent a redevelopment increases. In case redevelopment data is available, age can be defined as the difference between the year of last redevelopment and transaction year. In order to minimize the bias due to the unavailability of redevelopment data, this research will exclude all observations older than 75 years. This leads to a smaller set, where one redevelopment curve is likely to be present. By analysing the resulting convex function, the
optimal average redevelopment point can be derived. Whereas the cumulative depreciation function is expected to be
convex, the minimum of this function is likely to represent this optimal average redevelopment point.
Hypothesis 1 states that the depreciation rate per annum of Dutch real estate is non-constant. By estimating the model with and without the quadratic term for age in the regressions (10), (11) and (12), this hypothesis is tested. It is the inclusion of the quadratic term that makes the marginal effect of age non-constant, as the first derivative of age is defined as a function of age. By comparing the linear and quadratic regression model, hypothesis 1 can be tested. Furthermore, the only difference between the quadratic estimation and the linear estimation is the quadratic term for age. This means that the linear model is nested in the quadratic model. The nested model in turn enables the log likelihood test to be executed. Using the log likelihood test, it is possible to test whether the full model is preferred over the model nested within. The null-hypothesis of the log likelihood test states that the full model is not statistically preferred over the nested model. This way, the log likelihood test can provide statistical evidence for non-constant depreciation
The depreciation rate per annum is defined as the marginal effect of age on the logarithm of price. In this regard, the annual decrease in , can be calculated, along with its confidence interval for each value of . Comparing the marginal effect over multiple years and determining whether these marginal effects are significantly different allows testing hypotheses 2 and 3.
Hypothesis 2 states that lower rent drives depreciation to a larger extent than the cap rate creep, as Bokhari and Geltner (2016) show. The Wald-test of equality is performed to test whether the transaction price differential due to rent is different from the transaction price differential due to yield:
(19)
( ) (20)
,for multiple values of .
To test Hypothesis 3 a similar Wald-test is performed, however the method differs slightly. To test whether the marginal effect of age on the logarithm of price differs between real estate types, equation (10) is complemented with the interaction terms of types and age to arrive at the following regression equation:
(21)
, where denotes the dummy variable for types . The addition of the interaction terms allows for the comparison of marginal effect of age on the logarithm of price between groups. Significant values for or indicate
that the quadratic or linear component for age differs between types. However, differing coefficients of age does not have to lead to a significant difference in marginal effect. Generating the marginal effect and its standard error per age and per group, allows for the combined Wald-test:
(22)
(23)
, for multiple values of and .
3.2 D
ATA
This research uses the extensive research database of DTZ Zadelhoff. This firm is the largest real estate agent and appraiser in the Netherlands. The database consists of 3 categories: Availability markets, Occupier markets transactions and
Investment market transactions. The Availability database keeps track of the current spaces available on the market. By
logging the current availability in the market, DTZ Zadelhoff assesses current market characteristics. The Occupier
markets transactions database compiles information DTZ Zadelhoff encounters in fulfilling occupancy needs for its
clients. The Investment market Transactions database is where real estate agents keep track of investment transactions in the market, led by DTZ Zadelhoff or its competitors. This research will use the Investment market transactions database to perform the regression analysis. The remaining two categories are used for statistical reference in hypothesis 3. The raw dataset contains all 7369 observations dating from 2002 to 2015 with a transaction price higher than 500,000 euros. The lower limit of the considered timeframe excludes any confusion whether the transaction was entered in guilders or euros; the upper limit is the limit of the dataset. The lower limit of 500,000 euros follows Bokhari and Geltner (2016) to ensure that it is likely that the property was transacted by a professional or at least based on professional advice. This way it is reasonable to assume that the transaction is rational and the transacted property is income property. The remaining variables are complemented with data from the Dutch Municipality Administration (BAG) system. This governmental system keeps track of real property owner, size and location. Information that was not known in the DTZ Zadelhoff-database has been filled using data from the BAG-database. Especially the year of construction originates from the BAG-database.8 Table 1 shows the specifications of the origin of the raw dataset.
TABLE 1: DEFINITION AND SPECIFICATION OF RAW DATA FROM DTZ ZADELHOFF
SYMBOL VARIABLE DESCRIPTION SOURCE
Nominal Transaction Price less purchasers cost, in € DTZ Zadelhoff Research
Nominal Rent in € per year DTZ Zadelhoff Research
Gross initial yield denoted as a fraction DTZ Zadelhoff Research
Transaction year from 01-01 to 31-12 DTZ Zadelhoff Research
Type of property Office, Industrial, Retail or Other DTZ Zadelhoff Research
Size office space in m2 DTZ Zadelhoff Research
Size industrial space in m2 DTZ Zadelhoff Research
Size retail space in m2 DTZ Zadelhoff Research
Total size in m2 DTZ Zadelhoff Research
Seller / Buyer type Institutional, Real estate, Private or Other DTZ Zadelhoff Research
Seller nationality DTZ Zadelhoff Research
Buyer nationality DTZ Zadelhoff Research
Year of construction DTZ Zadelhoff Research
Region one of 13 DTZ Zadelhoff regions9 DTZ Zadelhoff Research
BAG-ID reference DTZ Zadelhoff/BAG DTZ Zadelhoff Research
Sale leasback dummy DTZ Zadelhoff Research
Building status type used or unused DTZ Zadelhoff Research
Construction year BAG-administration
X-coordinates WSG 84 coordinates BAG-administration
Y-coordinates WSG 84 coordinates BAG-administration
Parcel size in m2 BAG-administration
Dutch inflation in year as a fraction CBS Statline
Missing rent or GIY is calculated if the other of the two is defined, given equation (6). For the spatial variables to be generated, the coordinates of the property have to be known. These coordinates are appended using the BAG-database. No coordinates can be pinpointed for portfolios as there is no singular location. Therefore, portfolios are excluded along with all other properties with missing coordinates. The location of each logistic channel can be found in Appendix Figure 3. Age is defined as the year of transaction minus the year of construction. When the variable
total size is missing, size is defined as the sum of the separate office, industrial or retail space. Following Bokhari and
Geltner (2016), a dummy variable is determined for a high probability of excess land as part of the property. As no information is available on whether the transaction involved the purchase of the underlying land, this research assumes that the underlying land is included in the transaction.
The raw data contains two monetary variables: price and rent. As these variables are stated in the amount of euros at time , the prices are in nominal terms. The monetary variables are adjusted for inflation, following Bokhari and Geltner (2016). The adjustment leads to the construction of real prices at 2002-level. The inflation for each year in the transaction data is extracted from the Dutch Central Bureau of Statistics (CBS) Statline database. Using the inflation, a correction term is constructed that nets out inflation to the benchmark year: 2002. Multiplying all monetary variables by this correction term yields real prices. Table 2 shows the construction of the dataset.
TABLE 2: CONSTRUCTION OF DATA SET USED FOR ANALYSES
Symbol Variable Calculated Condition
Inflationary correction term
Real Transaction Price
Real Rent
Real Rent Missing Rent
Gross Initial Yield Missing GIY
Distance to Trainstation generated based on coordinates Distance to Highway generated based on coordinates Distance to Airport generated based on coordinates
Seller is Dutch dummy variable
Buyer is Dutch dummy variable
DTZ Zadelhoff Region generated based on coordinates
Size in m2 Missing
Real Transaction Price per m2
Real Rent per m2
Building Age
Excess Land Flag dummy variable
New Building dummy variable U = unused
Multiple „sanity checks‟ are defined to exclude observations that are implausible, given the nature of the Dutch real estate market or the characteristics of the variable. Such checks include nominal price and rent per meter over respectively 5,000 and 700 euros per metre and negative age. Properties with reported size exceeding the largest building in the Netherlands (160,000 square metres) are considered to be unrealistic. Transactions with a reported or implicitly calculated gross initial yield lower than 3% or higher than 20% are considered implausible and are dropped from the set. Table 3 shows the impact of the data filters and the inclusion of variables on the amount of full set observations.
TABLE 3: LOSS OF OBSERVATIONS IN COMBINED AND FULL SET PER FILTER AND INCLUDED VARIABLE
Action Loss Result
Full set Combined set Full set Combined set
Unbalanced set 7801 1700
Add Size 2626 474 5175 1226
Add Age 676 214 4499 1012
Add Distance 944 194 3555 818
Add Buyer type 156 8 3399 810
Add Seller type 443 0 2956 810
Add Buyer nationality 2 89 2954 721
Add Seller nationality 0 0 2954 721
Drop for age above 75 759 156 2195 565
Sanity check GIY 25 12 2170 553
Sanity check price 50 9 2120 544
Sanity check size 104 28 2016 516
TABLE 4: DESCRIPTIVE STATISTICS OF THE FULL SET1
Variable Mean Median Standard
Deviation Min Max Skew-ness tosis Kur- JB
2 2008 2007 3.90 2002 2015 0.29 1.95 120 5,328,000 1,851,000 10,580,000 403,684 161,800,000 5.75 52.80 219,420 2,571 1,749 2,858 56 23,582 3.06 15.09 15,414 2,328 1,610 2,535 73 22,663 3.42 19.34 26,362 48,899 41,402 37,488 211 184,583 1.37 4.65 857.21 0.03 0 0.18 0 1 5.21 28.12 62,118 0.93 1 0.26 0 1 -3.27 11.70 9,952 0.90 1 0.31 0 1 -2.59 7.72 4,125 0.09 0 0.29 0 1 2.78 8.72 5,336 6,318 2,787 10,262 500 104,000 4.12 25.16 46,952 946 727 729 14.14 4,982 1.96 8.15 3,519 22.02 18.50 16.35 0 75 0.97 3.64 348 0.24 0 0.43 0 1 1.21 2.47 518
1: the amount of observations is 2016 for each variable
2:Jarque Bera statistic. The critical value for with 2 degrees of freedom is 9.21
TABLE 5: DESCRIPTIVE STATISTICS OF THE COMBINED SET FOR PRICE, RENT AND GIY ANALYSIS1
Variable Mean Median Standard
Deviation Min Max Skew-ness Kur-tosis JB
2 2008 2007 3.88 2002 2015 0.40 2.15 29 5,445,000 1,905,000 15,510,000 500,846 161,800,000 5.69 51.84 54072 636,645 227,514 1,093,000 22,884 10,520,000 4.06 25.75 12545 2,555 1,738 2,834 56 23,582 3.08 15.36 4100 2,324 1,595 2,536 73 22,663 3.43 19.44 6821 48,545 41,209 37,088 211 184,583 1.38 4.73 227 0.03 0 0.181 0 1 5.14 27.42 15093 0.93 1 0.263 0 1 -3.24 11.48 2447 0.90 1 0.309 0 1 -2.55 7.50 993 0.10 0 0.295 0 1 2.74 8.52 1303 6,437 2,850 10,353 500 104,000 4.08 24.67 11524 0.088 0.085 0.021 0.037 0.186 1.25 6.41 383 1,060 829 731.2 37.59 4,706 1.75 6.95 599 87 71.43 52.35 3.124 376.5 1.56 7.00 552 21.91 18 16.36 0 75 0.97 3.64 90 0.243 0 0.429 0 1 1.20 2.43 130
1: the amount of observations is 516 for each variable
2:Jarque Bera statistic. The critical value for with 2 degrees of freedom is 9.21
The variable age is relatively skewed towards younger properties. The amount of observations of buildings older than 60 years is significantly lower than that for the first few years. This might also be a result of the fact that properties are likely to be traded at a younger age. The year of transaction is relatively well-distributed over the set. The distribution of observations over ages can be found in Appendix Figure 1. This observation leads to the redevelopment bias possibly being of a smaller degree than when the dataset was not skewed. Whereas the redevelopment bias is likely to be larger due to ‘older’ properties, the skewedness towards ‘younger’ properties can be expected to relieve some of the redevelopment bias.
Table 6, Table 7, Table 8 and Table 9 display the distribution of observations for the categorical variables transacting agent, region, real estate type and transaction year. In Table 6, it is clear that private investors transact most of the property in the database, ranging from 60 to 77 per cent of the observations. More agents from the category Other are involved as sellers than buyers. Comparing the full and combined sets of observations, the distribution tends to be relatively equal for both sets. Institutional agents seem to be more highly represented in the combined set for buyers than the full set (by 5 percentage points) and on the Seller side, the category Other is more highly represented in the full set (by 5 percentage points).
TABLE 6: DISTRIBUTION OF OBSERVATIONS OVER TRANSACTING AGENT CATEGORIES
Seller Buyer
Full set Combined set Full set Combined set Agent Freq. % Freq. % Freq. % Freq. %
Institutional 247 12.25 72 13.95 190 9.42 76 14.73
Real Estate Investor 160 7.94 49 9.5 169 8.38 53 10.27
Private Investor 1,197 59.38 317 61.43 1,547 76.74 366 70.93
Considering the distribution of observations over regions, as displayed in Table 7, the transactions are relatively well distributed over regions. There are some regions with a small contribution to the sample, such as Hoofddorp and Enschede. The big cities (Rotterdam, Amsterdam and Utrecht) contribute a little more to the set. Those differences are in line with expectations as large regions tend to transact more property. The difference in distribution between the two sets is small. The only exception is Groningen, which is represented better by 5 percentage points in the combined set. The differences between representation in the full and combined set might be attributed to differences in data collection precision or data availability between DTZ Zadelhoff offices in the country.
Table 8 displays the distribution of both sets over real estate types. The full set and the combined set seem to be distributed similarly in terms of real estate types. Furthermore, Office and Industrial real estate are the largest groups and account for 75 per cent of the dataset. About 17% of the dataset consists of Retail real estate. The remaining
Other category is relatively small.
As for Table 9, the distribution of observations over transaction years show that the period 2005-2008 has relatively many observations. This is in accordance with the Dutch real estate market that relatively thrived in that time period (DTZ Zadelhoff Research, 2015). The periods 2002-2003, 2008-2010 and 2013 show an amount of observations that is below average. This can be attributed to respectively the dot-com bubble, the securitized mortgage crisis and the euro-crisis. In these years the real estate market was affected to a large extent by these crises. The crises explain why there are fewer observations in the set as fewer transactions took place during these periods. The years 2005-2007 have a higher contribution in the set, as well as year 2015. These are years that can be indicated as good years for real estate, leading to relatively many transactions.
TABLE 7: DISTRIBUTION OF OBSERVATIONS OVER REGIONS
Full set Combined set
DTZ Region Freq. % Freq. %
Amsterdam 266 13.19 61 11.82 Arnhem 192 9.52 36 6.98 Breda 133 6.6 34 6.59 Eindhoven 188 9.33 52 10.08 Enschede 74 3.67 12 2.33 's-Gravenhage 193 9.57 57 11.05 Groningen 161 7.99 63 12.21 's-Hertogenbosch 112 5.56 21 4.07 Hoofddorp 38 1.88 10 1.94 Limburg 112 5.56 28 5.43 Rotterdam 198 9.82 47 9.11 Utrecht 242 12 67 12.98 Zwolle 107 5.31 28 5.43
TABLE 8: DISTRIBUTION OF OBSERVATIONS OVER REAL ESTATE TYPES
Real estate type
Full set Combined set
Freq. % Freq. %
Office 715 35.47 190 36.82
Industrial 871 43.2 222 43.02
Retail 349 17.31 88 17.05
Table 10 displays the pairwise correlation matrix of all non-categorical variables. Analysing the correlation coefficients, it is possible to detect possible multicollinearity and whether the selected variables control variables are justified. The matrix shows that price is highly correlated with the control variables. The control variables are correlated with rent to a comparable degree. However, yield is not correlated with the control variables as strong as price and rent. Hence, it is expected that the model for yield has lower explanatory power. As price, rent and yield are by definition interdependent, the correlation coefficients between these variables are high. The correlation parameters for the control variables are mainly significant, meaning that they are overall non-zero. However, the degree of correlation is likely not large enough to encounter multicollinearity issues.
The research framework section establishes that the marginal effect of age on price is non-constant, based on theoretic grounds. In order to underline this expectation, Figure 3 displays average transaction price per age, while Figure 4 shows the average rent per age and Figure 5 shows the average GIY per year. These three figures are accompanied by an optimal quadratic trend line. Based on these plots it can be stated that a quadratic line seems to fit the scatterplots quite well and therefore a non-linear effect of age on the dependent variables can be expected. Whereas the average price and GIY have a clear quadratic form, the non-linear effect of age on rent is less obvious, but still present.
TABLE 9: DISTRIBUTION OF OBSERVATIONS OVER TRANSACTION YEAR
Full set Combined set
Year Freq. % Freq. %
2002 87 4.32 46 8.91 2003 93 4.61 46 8.91 2004 120 5.95 27 5.23 2005 204 10.12 54 10.47 2006 221 10.96 55 10.66 2007 304 15.08 74 14.34 2008 101 5.01 25 4.84 2009 132 6.55 25 4.84 2010 102 5.06 21 4.07 2011 141 6.99 52 10.08 2012 85 4.22 22 4.26 2013 79 3.92 13 2.52 2014 113 5.61 13 2.52 2015 234 11.61 43 8.33
TABLE 10: PAIRWISE CORRELATION MATRIX OF NON-CATEGORICAL VARIABLES
FIGURE 3: AVERAGE TRANSACTION PRICE/M2 PER AGE FIGURE 4: AVERAGE RENT/M2 PER AGE
FIGURE 5: AVERAGE GIY PER AGE
FIGURE 6: AVERAGE PRICE/M2 PER AGE, FOR REAL ESTATE TYPES
Estimating the models presented in the research design, allows for detecting possible issues. Furthermore, the first results from the estimation can be analysed and put into context. This way, it can be assessed whether the model fits the data and whether the coefficients make sense.
For the OLS model on price, rent and GIY heteroscedasticity is a problem, leading to the estimation of the OLS model for price with robust standard errors. The models for rent and GIY will be estimated using the SURE; hence the standard errors are already defined. Multicollinearity issues are present in all three estimations as well. Analysing the VIF-factor for the variables reveals that predominantly the dummies for categorical variables and age along with its squared component have multicollinearity issues. Therefore, no actions are required. The normality condition is violated for all three regression equations. Following the Central Limit Theorem, the sample size is assumed to be large enough to guarantee the validity of the results. Both the rent and price regression have no issues with specification or functional form, indicating that these estimations are robust. However, the GIY estimation shows issues with specification and functional form. Generating the scatterplot of age and GIY in Figure 5 reveals that the pattern fits a quadratic function very well. Additionally, no influential observations are present in the estimations. Table 11 shows the regression results from the quadratic and linear OLS estimation of transaction price with robust standard errors and the SURE results for price, rent and GIY. The table reveals that both the linear and quadratic age coefficients are significant on a 0.01 level, except for the quadratic component of the rent regression, which is significant on a 0.05 level. The signs of the linear and quadratic components of the regression are in line with expectations. Where the marginal effect of age was expected to be a convex function on rent and price, it was expected to be a concave function on GIY. Appendix Table 5 contains the linear and quadratic OLS models for rent and GIY, used for hypothesis 1.
0 1000 2000 3000 4000 A v e ra g e P ri c e /m 2 0 20 40 60 80 Age
Average Price/m2 per age Fitted values
Offices 0 1000 2000 3000 4000 A v e ra g e P ri c e /m 2 0 20 40 60 80 Age
Average Price/m2 per age Fitted values
Industrial 0 1000 2000 3000 4000 A v e ra g e P ri c e /m 2 0 20 40 60 80 Age
Average Price/m2 per age Fitted values
Retail 0 1000 2000 3000 4000 A v e ra g e P ri c e /m 2 0 20 40 60 80 Age
Average Price/m2 per age Fitted values
TABLE 11: REGRESSION RESULTS FOR OLS WITH ROBUST STANDARD ERRORS FOR THE QUADRATIC AND LINEAR MODEL OF EQUATION (11) AND FOR THE SURE ESTIMATION FOR EQUATION (10), (11) AND(12)A
SURE FOR OLS FOR
VARIABLE Linear Quadratic
Age -0.0197*** -0.0110*** 0.00872*** -0.00486*** -0.0182***
(0.00405) (0.00383) (0.00160) (0.000872) (0.00242)
Age Squared 0.000283*** 0.000143** -0.000141*** NA 0.000220***
(6.52e-05) (6.17e-05) (2.57e-05) NA (3.80e-05)
log(Size in m2) -0.161*** -0.153*** 0.00729 -0.226*** -0.218*** (0.0237) (0.0224) (0.00934) (0.0178) (0.0177) Industrial -0.608*** -0.576*** 0.0318 -0.619*** -0.618*** (0.0566) (0.0536) (0.0223) (0.0314) (0.0311) Retail -0.397*** -0.547*** -0.149*** -0.182*** -0.194*** (0.0693) (0.0656) (0.0273) (0.0443) (0.0440) Other -0.0801 -0.0597 0.0203 0.145 0.126 (0.129) (0.122) (0.0509) (0.0929) (0.0930) New building 0.0845 0.0539 -0.0306 0.111** 0.0647 (0.0740) (0.0700) (0.0292) (0.0451) (0.0454)
Excess Land Flag 0.156*** 0.155*** -0.000906 0.129*** 0.123***
(0.0513) (0.0486) (0.0202) (0.0291) (0.0289)
Institutional Seller 0.313*** 0.278*** -0.0356 0.256*** 0.287***
(0.0837) (0.0793) (0.0330) (0.0518) (0.0513)
Real Estate Seller 0.118 0.116 -0.00199 0.121** 0.133**
(0.0915) (0.0867) (0.0361) (0.0594) (0.0589)
Private Seller 0.000383 0.0395 0.0391 0.0106 0.0303
(0.0696) (0.0659) (0.0275) (0.0393) (0.0387)
Institutional Buyer 0.192 0.194* 0.00238 0.462*** 0.439***
(0.121) (0.114) (0.0476) (0.0889) (0.0879)
Real estate Buyer 0.248* 0.251** 0.00309 0.336*** 0.314***
(0.127) (0.120) (0.0502) (0.0856) (0.0847) Private Buyer -0.0500 -0.00587 0.0441 0.00458 0.00575 (0.113) (0.107) (0.0447) (0.0756) (0.0750) Buyer is Dutch -0.392*** -0.218*** 0.174*** -0.429*** -0.420*** (0.0763) (0.0722) (0.0301) (0.0537) (0.0530) Seller is Dutch 0.157* 0.0791 -0.0777** 0.0703 0.0681 (0.0856) (0.0811) (0.0338) (0.0574) (0.0574) SAL -0.00598 -0.0201 -0.0141 0.124** 0.140*** (0.0954) (0.0903) (0.0376) (0.0549) (0.0540)
log(Dist. to Train station) -0.0641*** -0.0613*** 0.00283 -0.0579*** -0.0556***
(0.0234) (0.0222) (0.00923) (0.0144) (0.0143) log(Dist. to Highway) -0.0185 -0.0404* -0.0219** -0.0125 -0.0117 (0.0228) (0.0215) (0.00898) (0.0141) (0.0139) log(Dist. to Airport) -0.0274 -0.0471 -0.0197 -0.0710** -0.0672** (0.0471) (0.0445) (0.0186) (0.0281) (0.0280) Constant 9.694*** 7.340*** -2.353*** 10.49*** 10.49*** (0.581) (0.550) (0.229) (0.346) (0.344) Observations 516 516 516 2,016 2,016 R-squared 0.500 0.472 0.359 0.459 0.468 Adjusted R-squared NA NA NA 0.447 0.456
Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
The relevance of the controlling variables in the explanatory variables set is verified by analysing the corresponding parameters. First the OLS model is analysed, after which the SURE estimation is compared with the results of the OLS model. On hedonic variables, the parameter for size estimates that a 1% increase in size, leads to a 0.22% decrease in price per meter, comparing OLS results for price with the SURE results for price shows that the effect is smaller in the SURE regression: 1.6%. Furthermore, it is confirmed that the type of real estate has an influence on the price. Industrial and Retail have respectively 62% and 18% lower prices than category Offices, but the Other category does not significantly differ in price. For the SURE estimation of price, Industrial and Retail are significantly different from Offices (60 and 40 per cent), while Other is not. Additionally, there is no evidence that unused buildings are more expensive compared to used buildings in any of the models. The defined flag for excess land corresponds with 12% higher prices in the OLS results and 16% in the SURE results. The findings from these regression coefficients concur with the intuition that hedonic characteristics have a significant marginal contribution to price, as papers such as Bokhari and Geltner (2016), Fisher et al. (2005), Chegut et al. (2015), Liu (2015) and Fuerst and McAllister (2011) show.
As for agent characteristics, institutional sellers sell real estate for higher prices than sellers from the category ‘Other’ by 29% per cent according to the OLS and 31% according to the SURE. Real estate related agents sell their real estate assets for 13% more than the Other category, according to OLS, but this effect is not significantly different from zero according to the SURE. Institutional buyers transact with prices 44% higher according to the OLS, while the SURE does not indicate a significant difference. Real estate related buyers buy for 32% more according to the OLS and 25% (on a 0.1 level) according to the SURE. According to the SURE, Dutch agents buy for 39% less, while Dutch sellers sell for 16% more (on a 0.1 level), compared to their international counterparts. The OLS shows a discount of 42% (39% according to SURE) for Dutch buyers, while sellers only significantly sell for a premium in the SURE estimation (16%). The findings from these coefficients are in line with previous research on the effect of behavioural differences on price. Dutch agents seem to be better acquainted with the Dutch real estate market on average and realize better prices. The results on institutional parties partly correspond with findings by Wiley (2012). The premium for institutional buyers corresponds with these findings, while the discount in case of institutional seller does not hold in this model.
