1
UNIVERSITEIT VAN AMSTERDAM
“Being green worth more?” –Examining the existence of green premium in the western European property market
Abstract
This paper examines via panel dataset compiled for the time period between 2007-2014 whether obtaining green certification entails higher transaction price on the western European office market (Italy, Spain, Germany, Netherlands, UK, and France). The paper accounts for sample selection bias and a balanced panel dataset is set-up by matching using the propensity score approach. The second step of the analysis investigates the existence of green premium by using hedonic regression model. The paper concludes that based on the examined dataset there is no green premium effect after controlling for selection bias as well as property, investor and market timing characteristics.
Supervisor: Professor Milena Petrova Author: Adam Halasz
Stud. Nr.: 10599398
Faculty: Business and Economics Specialization: Real Estate Finance Year: 2014
2 Acknowledgements
I would like to express my gratitude to my supervisor Professor Milena Petrova for the useful comments and engagement through the learning process of this master thesis. Furthermore I would like to thank the RCA Databank for providing me with the relevant dataset. Without their assistance this thesis would not have been possible.
3 Table of Contents
Acknowledgements ... 2
Table of figures... 4
Introduction ... Error! Bookmark not defined. Literature review ... Error! Bookmark not defined. Methodology... 17
Propensity Score Model ... 18
Key Assumptions... 20
Matching methods ... 21
Hedonic regression model ... 28
Model specification ... 28
Dependent Variable ... 30
Independent variables ... 31
Data and descriptive statistics ... 35
General description of the dataset – before matching ... 35
“Simple t-test” data description – before matching ... 37
Regression outcome – before matching (controlling for limited amount of pretreatment characteristics) ... 38
Propensity score data description ... 38
Output of the different matching methods ... 40
4
Location (country level) ... 44
Time dummies ... 44
Status of buyer company ... 45
Results ... 46
Hedonic model – without fixed effect ... 47
Hedonic model – fixed effect regression (controlling on country level) ... 48
Hedonic model – differentiation between certifications, without fixed effect ... 49
Hedonic model – differentiation between certifications, controlling on country level ... 49
Robustness check... 53
Conclusion ... 54
References ... 57
Table of figures 1. Table: Justification of using log variant of “sqmeter” variable ... 33
2. Table: Description of treatment entities ... 35
3. Table: Descriptive statistics – before matching ... 37
4. Table: Inferior bound ... 39
5. Table: Radius matching sensitivity analysis ... 42
6. Table: Descriptive statistics – after matching ... 43
8. Table: Dispersion of observations within years ... 45
9. Table: Dispersion of buyer companies' aspect... 46
5
6
alyze if “green premium” exists within the examined markets. Should this be the case, then real estate experts are a step closer in coming to a consensus decision about the favorable prospect of green buildings.
The possible retentiveness to the uptake of sustainable practices is the assumed higher construction and development costs (Global Green Building Trends report, 2008) and the so called “vicious circle of blame”. The former phenomenon was justified by a survey conducted by the World Business Council for Sustainable Development. It was revealed that sustainable properties were thought to have a hypothesized 17% construction premium. This assumption seemed to be a misstatement since Barlett and Howard (2000) found that cost consultants misrepresent the trade-off between energy-efficient construction practices and operational cost savings.
In fact, a number of other studies proved that green buildings on average are less expensive than it is perceived. Similar results have been drawn up by Kats (2003) and Berry (2007) about construction cost premium associated with achieving green certification. Their findings suggest merely an estimated 2% higher financial burden on average. The most recent and authoritative studies have come from Davis Langdon (global construction consultancy). Their most recent study compared 83 building projects with a primary goal of LEED designation to 138 similar building projects without the goal of sustainable design (Mathiessen and Morris 2006). Confirming the findings of earlier studies, they found no significant difference in average costs for building projects with a primary goal of LEED certification as compared to non-labeled buildings.
The “vicious circle of blame” is a more subtle anomaly. As the illustration shows below, every stakeholder in the real estate market blames someone else for hindering the spread of
7
sustainable buildings. There would be a demand for green buildings according to the Owner & End Users side, however there is a limited amount of available properties. The response for that criticism from the Designers & Constructors side is lack of developers’ order.
Their business activity is defended by blaming the shortage of solvent demand from the Investors side. Finally, investors blame the Owner & End User segment wherefore there is no solvent demand for green properties.
To break this circle and foster the spread of green properties, all market participants need to act proactively. Policy makers should implement a supportive system via incentives (e.g. governmental aid) thus investors and developers would be more inspired to invest in green properties.
In parallel, financial sector and insurance entities should grant better lending and insurance conditions since sustainable buildings are less risky and offer loss prevention benefits hence making the deal more attractive. Appraisals, advisors and certifiers also need to clearly communicate the better performing capabilities of green properties, hereby improve and foster investment appetite. Finally, empirical researches and educators should emphasize the merit of being sustainable with publicly available presentations, case studies and researches. The latter one would be critical to contradict the misbelief that having a green building would not be financially beneficial.
8 1. Figure: The vicious circle of blame
Plenty of studies have analyzed the distinction between eco-labeled commercial real estate properties and non-green properties. Energy represents a major part of operating expenses in a property for example, the usage of energy in a typical office building represents 30% of total operating costs. The main advantage of sustainable buildings is that they use less energy. A more efficient use of energy leads to a significant decrease in operating expenses (Eichholtz, Kok, and Quigley 2009), (Fuerst and McAllister 2009).
With lower operating costs the productivity (i.e. the net operating income) of the building increases. Consequently, the capital value of sustainable buildings will be higher relative to comparable conventional buildings (Roper and Beard, 2006).
Wasiluk (2007) argued that being sustainable increases the competitive advantage of a commercial real estate via its capability of attracting higher profile tenants, who pay above market rents. Consequently this leads to higher appraised value of green properties.
9
Fuerst and McAllister (2009) examined the eco-labeling effect on commercial office occupancy rate and compared it with non-labeled ones within the United States via Ordinary Least Square (Hereinafter: OLS) and quantile hedonic regression models.
The paper compared the eco-label effect of 667 LEED certified properties and 1,480 ENERGY STAR certified ones with 24,479 non-certified office buildings on occupancy level. The model controlled the differences in age, size, height, building type and submarket. To avoid internal validity threat of having a biased model, the analysis excluded the single tenanted offices where the occupancy level is typically 100%. Their findings revealed a positive significant relationship between being sustainable and having a higher occupancy rate.
The study suggests an approximate 8% “occupancy rent premium” if the property has LEED certification and 3% if it has an ENERGY STAR label. Similar conclusion was reached by Miller, N., J. Spivey, and A. Florance (2008).
With hedonic and two-stage least square (Hereinafter: 2SLS) techniques Wiley, Benefield and Johnson (2010) also investigated a possible existence of “green premium” within the US office market. The study focused on Class A office buildings within 46 metropolitan areas. The focus was not just purely on the occupancy level as was the case with Fuerst and McAllister (2009).
The paper analyzed if rental rates, occupancy levels and selling prices have a premium for properties referred to as being sustainable. The outcome of the research is consistent with former studies, namely green-certified properties outperform in the leasing market, which resulted in significant higher rental rates, occupancy levels and selling prices. ENERGY STAR labeled properties have an estimated 7% rental premium while the occupancy level can be improved by 10% and the selling premium can reach $30/sq. ft.
10
LEED designation generally results in 17% higher rental income, 18% occupancy improvement and $130/sq. ft. selling premium.
Similarly Eichholtz, Kok, and Quigley (2009) established a model to analyze premium on rental rates, effective rental rates and sales prices owned by green buildings compared to similar non-certified ones. Via a hedonic regression model on average 3% higher rental rates were observed per square foot.
The research estimated an even higher effective rental premium of 7% and an impressive 16% in terms of selling prices. All the comparable properties were located within 0.2 miles of the given green properties.
Findings of Dermisi (2009) also underpin the positive financial effect of sustainable designation on office properties. ENERGY STAR and LEED certification indeed increase both the market and assessed value of the properties. Besides the eco-labeling effect it has been concluded that green offices have lower price volatility than conventional properties. This statement is justified with lower environmental and marketability risk and especially important during real estate cycles. During an unforeseen crisis lower volatility would cause reduced risk of large price fluctuation.
Research shows that corporate social responsibility is associated with better financial performance (Orlitzky and Benjamin, 2001). In accordance, Milgrom and Roberts (1989) find that corporations with social responsibilities attract more investors and customers. Through investing in sustainable buildings real estate investors may pursue a corporate social responsibility agenda and hence increase financial performance. Furthermore, a sustainable building also increases the reputation of the occupants (Frombrun and Schanley, 1990), which can be used to attract and retain the most talented employees (Paevere and Brown, 2008), and hence to increase financial performance. Indeed, established firms like Microsoft, IBM,
11
Google, Apple, Tesla etc. use social responsibility not only to attract more customers, but also to attract the most qualified and talented employees. These benefits to the tenant, of course, gives the owner of a green certified building a competitive advantage over the owner of a non-certified building.
Miller, Pogue, Gough, and David (2009) approached the merit of being sustainable in a different way, by way of qualitative research. In their case the productivity of green working environment was explored. The outcome was in accordance with the early study of Greg Kats (2003), namely that a sustainable working environment fosters being financially profitable.
Both studies concluded that green office investments have a positive net present value (Hereinafter: NPV) hence real estate investors should be willing to pay premium (5%-10%) based on bargaining power for green properties.
Another advantage of sustainable commercial real estate is that they provide a healthier working environment and thus increase productivity of employees. Findings of Lucuik, Trusty, Larsson and Charlette (2007) confirm this statement.
Their study ascertains that relocating from a conventional office building to a sustainable office building increases employee productivity by 2%-10%. This increase in productivity may have a significant impact on the performance of a company. Romm and Browning (1994) find that a 1% increase in productivity may provide gains to a company, which outweighs its entire energy bill.
Most theories support the practical evidences about the merit of being sustainable. A green premium seems to exist within the US commercial real estate market hence it would be reasonable to assume a similar phenomenon within the western European market as well. This study scrutinizes the same question from a different market view.
12
The core interest of this study is to examine whether real estate investors remunerate green properties with a higher acquisition price within the western European office market. The paper also investigates if different certification types have a different influential effect on the transaction price, in other words whether a labelling premium exists within the examined area or not.
13
Literature reviews Research field Applied model Main findings
Fuerst and McAllister (2009)
Dermisi (2009)
Analyze the effect of LEED and ENERGY STAR eco-labeling on the occupancy rates of offices compared with non-labeled properties within the United States.
Examine the LEED and ENERGY STAR ratings and the level of
OLS and quantile regression controlling for: age, height, building class, quality, submarket
Robust OLS regression, maximum likelihood spatial error regression and fixed effect regression were used during
1. Positive significant relationship between occupancy rate and the eco-labeling.
2. LEED certified offices have an estimated 8% higher occupancy rates while ENERGY STAR certified properties have an approximate 3% occupancy rent premium.
1. “Green premium” is higher during down market in comparison with up market.
14 Eichholtz, Kok, and Quigley (2009)
Miller, N.G., D. Pogue, Q.D. Gough, and S.M. David.(2009)
certification on both market value (MV) and assessed values (AV).
Investigate whether “green rating” buildings have contract rent, effective rent and selling price premium without any differentiation between eco-labels.
Analyze whether LEED and ENERGY STAR labeled buildings can provide more productive working atmosphere than non-green ones.
the analysis.
Hedonic OLS regression, within the model the control buildings are located within 0.2 miles of the green properties.
Measurement of productivity via sick days and self-assessed productivity after moving into a sustainable working environment.
2. An ENERGY STAR and LEED designation increases both the MV and AV of a property substantially. 1. Rental rates are roughly 3% higher
per square foot controlling for the quality and location characteristics of the properties.
2. Effective rents have an even higher premium, an approximate 7%.
3. Selling prices have the highest premium of an estimated 16%.
Investing in green properties has a positive NPV, hence a green premium does exist. This premium amount depends on the bargaining
15 Wiley, Benefield and Johnson (2010)
Examine the effect of eco-certification on rent, sales price and occupancy rate on Class A office buildings over 46 metropolitan areas within the US
OLS regression and 2SLS modeling – using instrumental variables of rent and occupancy.
power; generally it is between 5% and 10%.
1. There is a significant rental premium both for LEED (15%-18%) and ENERGY STAR (7%-9%).
2. Existence of sales premium both for LEED ($130/sq. ft.) and ENERGY STAR ($30/sq. ft.)
3. Green buildings achieved significantly higher occupancy than non-labeled buildings.
Reichardt, Fuerst, Rottke, and Zietz (2012)
Explore whether sustainable office buildings within the US entail rental premium in the period of 2000 – 2010.
Difference in difference – controlling for the systematic difference between labeled and non-labeled properties and
1. Significant rental premium can be observed for both LEED and ENERGY STAR certified properties.
16
fixed effect model. 2. Significant positive effect of ENERGY STAR designation on occupancy level.
Das, Tidwell, and Ziobrowski (2011) Study the rental rate dynamics of green properties in San Francisco and Washington DC and compare them with non-labeled offices.
Paper used first difference, random-effect panel data and cross-sectional regression models.
1. Sustainable office buildings enjoy green premium over non-certified real estates.
2. Premiums seem to be dynamic instead of being constant hence they may provide a hedge in down markets.
17 Methodology
When attempting to analyze if institutional investors on the western European office markets are willing to pay green premium and to evaluate the effect of certification levels on the acquisition price, the key issue is to monitor a sufficient number of transactions. Property transactions of 5 million euro or greater were provided by RCA company. This dataset consisted of all their registered transaction prices within the German, UK, French, Spanish, Dutch and Italian office market. Overall approximately 8,000 properties were collected.
The entire dataset had to be filtered with the eco-labeled property transactions which resulted in another dataset where all the publicly available green certified properties were listed. Those properties had to be compared with the former dataset. In case of the RCA database properties, these were geo-coded based on their address and matched with dataset of green office buildings.
The final table consists of 81 green certified properties (BREEAM: 22; DGNB: 33; HQE: 2; LEED: 24) and approximately 2,497 comparable ones.
Significant difficulty of causal inference occurs when one attempts to estimate the effect of “treatment” (in this case having green certification) within observational studies, where the selection between being appertained to the “treated” and “control” group is not random. Non-random treatment allocation cannot ensure that the status will not be confounded with either measured or unmeasured attributes. The properties which received green certificates may be different from those which did not, thus a direct comparison of the outcome may not be sufficient enough.
The estimation of causal effect reached by comparing the treated observations with non-experimental control units may lead to a biased model.A possible solution to avoid having a
18
biased model is using propensity score matching because it allows to imitate some of the main characteristics of a randomized treatment in the context of a non-randomized analysis. It corrects the observable differences between treated and untreated units (i.e. handles the sample selection bias).
Propensity Score Model
Propensity score model is capable of correcting treatment effects that control confounding factors. The main idea behind this is to reduce biasness when the treated and control entities are as similar as they can be. The model was elaborated by Rosenbaum and Rubin (1983):
𝑝𝑝(𝑋𝑋) ≡ Pr(𝐷𝐷 = 1|𝑋𝑋) = 𝐸𝐸(𝐷𝐷|𝑋𝑋),
where 𝐷𝐷 = {0,1} indicates whether an entity is exposed to treatment (1) or not (0) and “X” includes all the pretreatment characteristics that may affect the assignment into the treated group. In this paper (1) represents green certified properties and (0) stands for non-labeled ones. The pretreatment characteristics are:
• latitude, • longitude, • size, • age,
• number of floors and • time.
19
After assignment of the observations into two groups (certified vs. non-certified), a probit model (binary outcome model) can help to estimate the propensity scores. This model assists in predicting the probabilities of having a treatment when the pretreatment factors are given. Furthermore, instead of matching every single characteristic (“Xs”) individually, the model sums up these values in a propensity score and this score can be used for the final matching procedure. In other words: a match can be found for the treated observations from the control ones. There is a wide range of matching methods: kernel matching, nearest neighbor, radius, stratification, and mahalanobis metric matching.
This paper used the kernel, nearest neighbor, radius and stratification matching methods. Once the matching has been executed, the treatment effect (comparison of the outcome between the treated and control observations) can be calculated:
𝑌𝑌 = �𝑦𝑦𝑦𝑦1 𝑖𝑖𝑖𝑖 𝐷𝐷 = 1
0 𝑖𝑖𝑖𝑖 𝐷𝐷 = 0
During the analysis a counterfactual problem occurs, namely under the current circumstance it is impossible to see the “pure” effect of eco-labels on properties since the removal of treatment from the treated units cannot be carried out. In other words: the outcome of the treated observations cannot be compared with exactly the same treated observations presuming that the latter had not been treated. This problem would not occur if the dataset included information about the selling price of the properties prior to and after receiving certification (repeated sales). Due to the fact that the set of data does not include repeated sales information, the counterfactual problem remains an issue.
Propensity score method helps to examine in the best possible way. In particular it is a balancing score: conditional on the propensity score and the distribution of observed baseline
20
covariates will be similar between treated and untreated subjects. It searches for the closest match from the control group (based on the given variables) and uses their outcome for further analysis.
Key Assumptions
There are three core assumptions that underlie the use of propensity matching method (Baum, 2013):
Balancing condition
Balancing condition means that the assignment for being treated and untreated is independent from pretreatment characteristics (“Xs”) given the same propensity score. So if propensity scores are similar then it is possible to match the entities based on their characteristics, which would guarantee similarity as well.
𝐷𝐷 ⊥ 𝑥𝑥| 𝑝𝑝(𝑥𝑥)
Conditional independence
(𝑦𝑦1 𝑦𝑦0) ⊥ 𝐷𝐷 | 𝑥𝑥
The conditional independence assumption implies that after controlling for “x” (pretreatment characteristics), the selection of being treated and non-treated follows a random walk.
It requires that all variables which are important to the probability of assigning to treatment should be monitored and captured in “x”. This allows the untreated subjects to be used to build an unbiased counterfactual for the treated entities.
21
Based on this assumption, the assignment to receive treatment should be un-confounded as well as given the propensity score.
(𝑦𝑦1 𝑦𝑦0) ⊥ 𝐷𝐷 | 𝑝𝑝(𝑥𝑥)
Common support
0 < 𝑃𝑃(𝐷𝐷 = 1 | 𝑥𝑥 ) < 1
The common support assumption implies that the probability of being assigned to a treatment for each possible value of the vector “x” is in the unit interval, just as is the probability of not receiving treatment.
This assumption enables the model to find adequate matches between the treated and untreated units due to the sufficient amount of overlaps in “x”.
If the aforementioned core assumptions are satisfied, the assignment procedure is said to be “strongly ignorable” in the terms of Rosenbaum and Rubin (Biometrika, 1983).
Matching methods
Propensity matching method can help to improve the analysis. For each treated observation it is necessary to find matches from the control group observations with similar pretreatment characteristics (size, longitude, latitude, age, size squared, age squared, time).
There are two possible ways of executing the analysis: matching without replacement (1) or matching with replacement (2). The former method makes a restriction for the matching procedure: each control unit can only be used once as a match for treated units. Matching without replacement could improve the accuracy of the estimates but also increases the
22
biasness of the model. An additional complication of using this method is the vulnerability of the results as to the order in which the treated observations are matched. (Rosenbaum, 1995).
On the other hand, by applying the matching with replacement method all the control observations can be matched to several treated observations. This process fosters to minimize the propensity score distance between treated and untreated units. The merit of this method is that it reduces the biasness of the model. Due to the limited amount of treated observations it is preferable to apply the matching with replacement method.
It is also important that the analysis used common support matching. This means that the matching procedure was restricted on the common range of propensity scores. The propensity score model executes the following steps during the analysis:
1. Give an estimation of the probit model: 𝑃𝑃𝑃𝑃{𝐷𝐷𝑖𝑖 = 1|𝑋𝑋𝑖𝑖} = 𝜙𝜙�ℎ(𝑋𝑋𝑖𝑖)�, where 𝜙𝜙 denotes the logistic central distribution function and �ℎ(𝑋𝑋𝑖𝑖)� denotes the starting specification which consists of all the covariates as linear function without having higher orders or any interactions.
2. Divide the observations in “n” equally spaced intervals of the estimated propensity score. (Not that “n” can be determined individually, within STATA the default setting is “n = 5”.)
3. Test if average propensity score differs within the intervals of the treated and control subjects.
4. If the average propensity score differs within an interval, the system halves the interval and re-runs the test. It executes this process until all the average propensity scores within treated and control units are equal.
23
5. Test within every interval that the mean values of every characteristic do not differ among treated and control entities. (This is the key condition for the balancing hypothesis).
6. If the mean characteristics differ that means that the balancing condition failed to be satisfied, thus a less parsimonious specification of �ℎ(𝑋𝑋𝑖𝑖)� is necessary.
7. After successfully running the propensity score the treatment effect can be measured.
Kernel matching
Kernel matching uses the weighted average of all individuals in the control group to deliver the counterfactual outcome (Caliendo, Kopeinig, 2005). The weights are inversely proportional to the radius between the treated and control observations.
The lower the weight the larger the distance. Within the matching methods weights are defined by the following:
𝑤𝑤 (𝑖𝑖, 𝑗𝑗) = 𝐾𝐾 (𝑝𝑝𝑗𝑗−𝑝𝑝𝑖𝑖) ℎ ∑𝑛𝑛0 𝐾𝐾 𝑗𝑗=1 (𝑝𝑝𝑗𝑗−𝑝𝑝𝑖𝑖) ℎ ,
where “K” is the kernel function, (𝑝𝑝𝑗𝑗− 𝑝𝑝𝑖𝑖) is the distance between the propensity score of the treated and control observations and “h” is the bandwidth parameter.
Pagan and Ullah (1999) highlighted the importance of the choice of the bandwidth parameter. High parameter is a smoother density function and results in a better fit. It means it fosters to reduce the variance between the estimated and the truth underlying density score. On the other hand, a high bandwidth parameter can lead to a biased estimate. Thus there is a trade-off between an unbiased and minimized variance estimate of the real density function.
24
Nearest neighbor matching
The nearest neighbor method is a one by one matching mechanism. It matches the treated entity to the untreated one where the absolute difference between their estimated propensity score is minimized (Dehejia & Wahba, 2002). If there is a multiple selection possibility (the outcomes of the propensity score are equally close to that of the treated) the property from the control group would be selected randomly. It is also important to note that there is no limitation of the maximum acceptable difference between the propensity scores of two matched subjects.
𝐶𝐶(𝑃𝑃𝑖𝑖) = min| 𝑝𝑝𝑖𝑖 − 𝑝𝑝𝑗𝑗|
where 𝐶𝐶(𝑃𝑃𝑖𝑖) represents the matches on the estimated propensity score, while 𝑝𝑝𝑖𝑖 and 𝑝𝑝𝑗𝑗 are the estimated propensity score for the treated entities “i” and for the control subjects “j”.
Radius matching
𝛿𝛿 > �𝑝𝑝𝑖𝑖 − 𝑝𝑝𝑗𝑗� = min𝑘𝑘∈(𝐷𝐷=0)�|𝑝𝑝𝑖𝑖− 𝑝𝑝𝑗𝑗|�
Nearest neighbor matching can face the risk of having wrong matches (e.g. the closest match is still located far away). With the caliper matching method this phenomenon can be removed. By imposing a tolerance level (𝛿𝛿) on the preferred propensity score level the method guarantees that the closest match cannot be located outside the commanded distance (Cochran & Rubin, 1973). If none of the untreated units are within “𝛿𝛿” from a given treated entity then it would remain unmatched. This leads to a better matching quality.
25
On the other hand, omitting the bad matches would result in a higher estimated variance. A common drawback of caliper matching derives from the uncertainty of the reasonably used tolerance level (Smith & Todd, 2005).
The paper uses a variant of the caliper matching method, the radius one. The idea derives from Dehejia and Wahba (2002) who suggest not just using the nearest neighbors within the calipers but all of the reachable comparison units. The advantage of this process is the favorable feature of oversampling (use more than one nearest neighbor) and decreases the potential of having bad matches.
Stratification matching �∑ 𝛿𝛿𝑘𝑘𝑁𝑁𝛿𝛿𝑖𝑖𝑘𝑘 𝑘𝑘 𝐾𝐾 𝑘𝑘=1 𝑁𝑁𝑗𝑗𝑘𝑘(𝑌𝑌���� − 𝑌𝑌𝚤𝚤𝑘𝑘 ����)𝚥𝚥𝑘𝑘 𝐾𝐾 𝑘𝑘=1
Stratification matching / interval matching method involves stratifying entities into mutually exclusive subsets based on their calculated propensity score. Entities are qualified based on their estimated propensity score then stratified into subgroups based on an advanced defined threshold of the propensity score.
A common approach is to divide subjects into five equal-size groups using the quintiles of the estimated propensity score. The right part of the equation represents the difference between the mean of the treatment group and the control one within each stratum of “k”, within the amount of “K” strata. The left part of the equation represents the weighted value of the mean difference from each strata according to the number of the existing units in the strata.
26
Cochran (1968) proved that this procedure (stratifying on the quintiles of a continuous confounding variable) abolishes approximately 90% of the biasness of the given variables. Within all generated stratum treated and control entities will have approximately the same propensity score value.
After successfully matching the treated entities with the control ones the “average treatment effect” (Hereinafter: ATE) can be used.
∆ = 𝑦𝑦1− 𝑦𝑦0
𝐴𝐴𝐴𝐴𝐸𝐸 ≡ 𝐸𝐸 (∆) ≡ 𝐸𝐸 (𝑦𝑦1|𝑥𝑥, 𝐷𝐷 = 1) − 𝐸𝐸 (𝑦𝑦0|𝑥𝑥, 𝐷𝐷 = 0)
ATE measures the line comparing the outcome results of treated and control observations.
The first part of the equation simply measures the expected outcome of the treated group given it has been selected as treated and so does the second term with the control observations. It analyzes the outcome with a simple “t-test”.
It is easy to justify that this method is suitable for random experiments rather than non-randomized ones. The reason behind this is that if the treated and non-treated subjects do not have similar characteristics then the model would be biased.
In non-random experiments the common approach is using the “average treatment effect” on the treated (Hereinafter: ATET).
𝐴𝐴𝐴𝐴𝐸𝐸𝐴𝐴 ≡ 𝐸𝐸 (∆| 𝐷𝐷 = 1) ≡ 𝐸𝐸 (𝑦𝑦1|𝑥𝑥, 𝐷𝐷 = 1) − 𝐸𝐸 (𝑦𝑦0|𝑥𝑥, 𝐷𝐷 = 1)
This equation has been elaborated in the previous section. Namely, this process measures the difference between the outcomes of the treated and the outcomes of the same entities had they not been given treatment (Hirano, Imbens & Ridder, 2003). In this paper it examines the transaction price of the properties which have received green certification and compares the
27
outcome of “what would have happened” with those properties’ transaction price if they had not been given the certification. As mentioned before, the second term (𝐸𝐸 (𝑦𝑦0|𝑥𝑥, 𝐷𝐷 = 1)) is called a counterfactual problem because it cannot be measured but needs to be estimated. This is the point where the propensity score model plays an important role via its matching capabilities.
𝐴𝐴𝐴𝐴𝐸𝐸𝐴𝐴 ≡ 𝐸𝐸 (∆| 𝑝𝑝(𝑥𝑥)𝐷𝐷 = 1) ≡ 𝐸𝐸 (𝑦𝑦1|𝑝𝑝(𝑥𝑥), 𝐷𝐷 = 1) − 𝐸𝐸 (𝑦𝑦0|𝑝𝑝(𝑥𝑥), 𝐷𝐷 = 1)
The comparison of the outcomes of the treated and control observations can be elaborated based on their matched propensity score. In practice, the estimation is based on the following equation;
𝐴𝐴𝐴𝐴𝐸𝐸𝐴𝐴 = 𝑛𝑛1
1 � �𝑦𝑦𝑖𝑖 ∈{𝐷𝐷=1} 1,𝑖𝑖− � 𝑤𝑤(𝑖𝑖, 𝑗𝑗)𝑦𝑦𝑗𝑗 0,𝑗𝑗�
where "𝑦𝑦1,𝑖𝑖” symbolizes the outcome on the certified properties while in the second “i” term refers to the green properties and “j” represents the non-certified ones. In the formula “w” refers to the weights and “𝑦𝑦0” to the outcomes.
During the estimation, weights up altogether depending on “i’ and “j” which results in the average weighted outcome. Afterwards that is added up based on all the “i” where “D = 1” and is multiplied with the number of observations to get the ATET.
Within this analysis propensity score model plays a significant role. It supports handling sample selection bias, hence helps to compare the eco-certified properties with their possible best matches based on pretreatment characteristics. Once the matching process is done the existence of green premium can be investigated by using hedonic regression model.
28 Hedonic regression model
The formerly introduced propensity score model enables the comparison of sustainable office buildings with non-green properties and concludes that transaction price differences must be attributed to green premium. The next step is examining the price deviation between green and their matching non-green properties after controlling for pre-treatment characteristics. To this aim hedonic regression model is applied. Hedonic price theory assumes there is an existing relationship between transaction prices and property characteristics thus the differences between given transaction prices can be justified by the differences within characteristics. In other words: the model measures the contribution of each observable characteristic to the overall value of the given property.
Model specification
The two most popular hedonic regression models are the fully linear model (1) and the logarithmic-linear model (2). (1) 𝑃𝑃𝑛𝑛𝑡𝑡 = ß0𝑡𝑡 + � ß𝑘𝑘𝑡𝑡 𝐾𝐾 𝑘𝑘=1 𝑍𝑍𝑛𝑛𝑘𝑘𝑡𝑡 + 𝜀𝜀𝑛𝑛𝑡𝑡 (2) 𝑙𝑙𝑛𝑛𝑃𝑃𝑛𝑛𝑡𝑡 = ß0𝑡𝑡 + � ß𝑘𝑘𝑡𝑡 𝐾𝐾 𝑘𝑘=1 𝑧𝑧𝑛𝑛𝑘𝑘𝑡𝑡 + 𝜀𝜀𝑛𝑛𝑡𝑡
In this paper the second equation is preferred due to the log-linear model’s capability of reducing problems caused by heteroskedasticity and to the absence of lot size observation. When lot size observation is missing (like in this case) empirical studies’ preferred approach uses the log-linear model.
29
In the equation "ß0𝑡𝑡" and "ß𝑘𝑘𝑡𝑡" represent the characteristics parameters to be assessed. The characteristics can be continuous variables with / without interaction terms, transformations (like natural logarithm of the transaction price) and categorical ones. The latter is frequent and essential within hedonic regression models.
These variables in practice are represented by dummy variables where the variable takes value of (1) if the property meets a certain condition and takes (0) otherwise.
Notice that characteristic parameters in both equations are allowed to change continuously. This assumption derives from Pakes (2003) who stated the importance of changing environment. In his interpretation market conditions change over time in terms of demand and supply, hence the contributions should not be constant. To simplify the model this paper supports the contra argument of Pakes, which believes that in the short run a simplified assumption can be justified. Due to the fact that market conditions change gradually these characteristic parameters can be handled as constant parameters over a short time period. It is important to note that this statement does not affect the constant parameter. With this assumption the simplified model would be the following:
𝑙𝑙𝑛𝑛𝑃𝑃𝑛𝑛𝑡𝑡 = ß0𝑡𝑡 + � ß𝑘𝑘 𝐾𝐾 𝑘𝑘=1
𝑧𝑧𝑛𝑛𝑘𝑘𝑡𝑡 + 𝜀𝜀𝑛𝑛𝑡𝑡
In this constrained model the parameters can be estimated to all time periods by adding dummy variables. To avoid having perfect co-linearity in the model a dummy variable must be left out. When having time and certification dummies (𝐷𝐷𝑛𝑛𝑡𝑡) the model would become the following: 𝑙𝑙𝑛𝑛𝑃𝑃𝑛𝑛𝑡𝑡 = ß0+ � 𝛿𝛿𝑡𝑡 𝑇𝑇 𝑡𝑡=1 𝐷𝐷𝑛𝑛𝑡𝑡 + � ß𝑘𝑘 𝐾𝐾 𝑘𝑘=1 𝑧𝑧𝑛𝑛𝑘𝑘𝑡𝑡 + 𝜀𝜀𝑛𝑛𝑡𝑡
30 Dependent Variable
The ideal model is built in a way to have the possible smallest standard deviation of the residuals. Compared to the propensity score model hedonic regression model uses the natural logarithm of the dependent variable (in this case: transaction price). The explanation is elaborated precisely by Francke (2013).
The four main reasons are the following:
1. Law of diminishing returns
In reality the value of a property is not proportional to its size, it is less. To create this environment one does not need to take just the natural logarithm of the dependent variable (transaction price) but also has to consider the independent variable (size). The result would indicate how a one percent increase in the size of the house would increase the property’s value.
If the “ß” coefficient is equal to one then the value of the building is proportional to its size. Usually, however, the coefficient varies between 0 < ß 1, which represents the law of diminishing returns.
2. A sound model specification is in factors (multiplicative)
According to Francke’s explanation it is more natural to describe the difference between properties in percentage terms rather than in absolute value. Using the natural logarithm of the transaction price offers a more realistic solution.
31
Using the natural logarithm would result in minimized squared relative residuals, where “Y” refers to the transaction price, “M” stands for the model value and “n” represents the number of observations. This case standard deviation of the regression can be interpreted in relative percentage term.
((𝑌𝑌1− 𝑀𝑀1)/ (𝑌𝑌1)2+ . . . +((𝑌𝑌𝑛𝑛− 𝑀𝑀𝑛𝑛)/(𝑌𝑌𝑛𝑛)2
4. The error term is closer to normality
The merit of using natural logarithm is that resulting residuals follow normal distribution more closely than the acquisition price itself because the outliers impact is less influential when one is taking the log values.
Independent variables
Independent variables which were used within the propensity score model remained nearly the same with only some minor changes occurring: transforming, adding new and dummy variables.
Setting up the explanatory variables within the hedonic regression model is outstandingly important. After scatter plotting the relationship between the natural logarithm of the price and size, one can conclude that “size” variable needs to be converted into natural logarithm form as well. All the collected factors that could influence the transaction price should be included in the regression, otherwise the model would suffer from omitted variable bias. Here it should be noted that it is nearly unfeasible to obtain all the information related to given properties hence in every model some omitted variable bias will be present. Based on the current observations this paper included the following explanatory variables to the hedonic regression model:
32 • Longitude, Latitude
• Natural logarithm of Size
• Time, certification, country, portfolio, renovation dummy variables
Age , Age squared and Renovation dummy
Real estate properties suffer from physical amortization as well as external and functional obsolescence. Due to the continuously changing environment and improving technology older buildings would “suffer” lower price valuations assuming other “external” circumstances constant. Francke (2013) argues that amortization does not follow a linear trend hence in his model he specified the age variable as a combination of linear part and dummy variables. To simplify the model this paper follows the conventional process and assumes a linear amortization trend. However, a turning point can occur. One can assume that the decreasing price effect wears off after a certain point.
A good example for this is the “car market”. After buying a new car its price starts falling, however the price of an “old timer” car would increase. Adding the AgeSqr variable would control this effect. Renovation dummy variable is also added to the hedonic model.
Once a property is renovated the age effect would not include the change hence a comparison with non-renovated buildings would lead to a biased estimation.
Longitude/ Latitude, Location and City dummies
Long and Lat variables refer to the geographical coordinates of a given property. It determines the absolute location of the green buildings. Controlling for relative location is necessary to have a more precise estimation. Later in the analysis it will be clearly observable that controlling on a country basis versus a city basis would change the outcome of the analysis significantly.
33
Natural logarithm of size
Transforming size variable into its natural logarithm base is a vital procedure. As table 4 shows below in order to make a linear regression the logarithm of the variable offers a better solution.
1. Table: Justification of using log variant of “sqmeter” variable
Number of floors
Number of floors can influence significantly the transaction price of a property as well, thus adding this variable would presumably improve the model.
Portfolio dummy
Using portfolio as a control factor variable is also interesting and can have an influential power if the sale of a property was a portfolio sale or not.
Treatment dummy
The core interest of the analysis is investing if green premium exits on the western European office market. Treatment dummy indicates whether a property is built / retrofitted as a sustainable building (1) or not (0). In the regression, while keeping other variables constant, this “green effect” can be examined.
34
Buyer company dummies
These dummy variables refer to the type of investor. During this analysis the paper differentiates between five types of acquisitions, variables.: Private, Public, Institutional, Other and “Unknown” (created where buyer company observations were missing).
Time dummies
Market conditions differ continuously over time. These conditions (e.g. GDP growth, unemployment rate, interest rate, country default spread etc.) usually are not explicitly included in hedonic regression models. A common approach is using time dummy variable which could be a deterministic linear trend. A drawback of this linear assumption is assuming price change as being equal in each period.
Applying all of the aforementioned assumptions the final hedonic regression model would consist of the following characters:
(1)𝑙𝑙𝑛𝑛𝐴𝐴𝑃𝑃𝑙𝑙𝑛𝑛𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑖𝑖𝑙𝑙𝑛𝑛𝑃𝑃𝑃𝑃𝑖𝑖𝑙𝑙𝑙𝑙
= 𝛽𝛽0+ 𝛽𝛽1𝐴𝐴𝐴𝐴𝑙𝑙 + 𝛽𝛽2𝐴𝐴𝐴𝐴𝑙𝑙𝐴𝐴𝐴𝐴𝑃𝑃 + 𝛽𝛽3𝐿𝐿𝑙𝑙𝑛𝑛𝐴𝐴 + 𝛽𝛽4𝐿𝐿𝑙𝑙𝑙𝑙 + 𝛽𝛽5𝐿𝐿𝑛𝑛𝐴𝐴𝑖𝑖𝑧𝑧𝑙𝑙
+ 𝛽𝛽7𝐹𝐹𝑙𝑙𝑙𝑙𝑙𝑙𝑃𝑃𝑁𝑁𝐹𝐹𝑃𝑃+𝛿𝛿1𝐷𝐷𝑃𝑃𝑃𝑃𝑃𝑃𝑡𝑡𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖𝑃𝑃𝛿𝛿2𝐷𝐷𝑅𝑅𝑅𝑅𝑛𝑛𝑃𝑃𝑅𝑅𝑅𝑅𝑡𝑡𝑖𝑖𝑃𝑃𝑛𝑛+ 𝛿𝛿3𝐷𝐷𝐶𝐶𝑃𝑃𝐶𝐶𝑛𝑛𝑡𝑡𝑃𝑃𝐶𝐶 + 𝛿𝛿4𝐷𝐷𝑇𝑇𝑃𝑃𝑅𝑅𝑅𝑅𝑡𝑡𝑇𝑇𝑅𝑅𝑛𝑛𝑡𝑡𝛿𝛿5𝐷𝐷𝐵𝐵𝐶𝐶𝐶𝐶𝑅𝑅𝑃𝑃𝐶𝐶𝑃𝑃𝑇𝑇𝑝𝑝𝑅𝑅𝑛𝑛𝐶𝐶
+ 𝛿𝛿6𝐷𝐷𝑇𝑇𝑖𝑖𝑇𝑇𝑅𝑅 + 𝜀𝜀
The second model investigates how and in what way different eco certifications influence acquisition price. In other words does a labelling premium exist within the western European office market.
35 (2)𝑙𝑙𝑛𝑛 𝐴𝐴𝑃𝑃𝑙𝑙𝑛𝑛𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑖𝑖𝑙𝑙𝑛𝑛𝑃𝑃𝑃𝑃𝑖𝑖𝑙𝑙𝑙𝑙
= 𝛽𝛽0+ 𝛽𝛽1𝐴𝐴𝐴𝐴𝑙𝑙 + 𝛽𝛽2𝐴𝐴𝐴𝐴𝑙𝑙𝐴𝐴𝐴𝐴𝑃𝑃 + 𝛽𝛽3𝐿𝐿𝑙𝑙𝑛𝑛𝐴𝐴 + 𝛽𝛽4𝐿𝐿𝑙𝑙𝑙𝑙 + 𝛽𝛽5𝐿𝐿𝑛𝑛𝐴𝐴𝑖𝑖𝑧𝑧𝑙𝑙
+ 𝛽𝛽7𝐹𝐹𝑙𝑙𝑙𝑙𝑙𝑙𝑃𝑃𝑁𝑁𝐹𝐹𝑃𝑃+𝛿𝛿1𝐷𝐷𝑃𝑃𝑃𝑃𝑃𝑃𝑡𝑡𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖𝑃𝑃𝛿𝛿2𝐷𝐷𝑅𝑅𝑅𝑅𝑛𝑛𝑃𝑃𝑅𝑅𝑅𝑅𝑡𝑡𝑖𝑖𝑃𝑃𝑛𝑛+ 𝛿𝛿4𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐵𝐵+ 𝛿𝛿5𝐷𝐷𝐻𝐻𝐻𝐻𝐻𝐻+ 𝛿𝛿6𝐷𝐷𝐿𝐿𝐻𝐻𝐻𝐻𝐷𝐷
+ 𝛿𝛿7𝐷𝐷𝐵𝐵𝑅𝑅𝐻𝐻𝐻𝐻𝐵𝐵𝐵𝐵𝛿𝛿8𝐷𝐷𝐵𝐵𝐶𝐶𝐶𝐶𝑅𝑅𝑃𝑃𝐶𝐶𝑃𝑃𝑇𝑇𝑝𝑝𝑅𝑅𝑛𝑛𝐶𝐶 + 𝛿𝛿9𝐷𝐷𝑇𝑇𝑖𝑖𝑇𝑇𝑅𝑅 + 𝛿𝛿10𝐷𝐷𝑐𝑐𝑃𝑃𝐶𝐶𝑛𝑛𝑡𝑡𝑃𝑃𝐶𝐶+ 𝜀𝜀
Data and descriptive statistics
General description of the dataset – before matching
The observations were geo-coded and matched with the database which included the green certified properties. At the beginning 81 eco-labeled and approximately 8,000 potential comparable properties remained in the system.
After controlling for all the critical factors (e.g. longitude, age, size etc.) the final dataset consisted of 81 green properties (referred as: treated) and precisely 2,497 non-certified ones (referred as: control group). To elaborate a more realistic analysis the paper sets limits related to property age, number of floors and area size. Office buildings smaller than 100 square meters were handled as outliers and were omitted from the dataset. The same procedure was executed with properties whose age exceeded 200 years. Properties where the number of floor data was not available were handled as having one floor.
After handling for outliers the bedrock of the analysis consists of 2,497 observations where the proportion of the green certified properties accounts for 3.10% while the remaining 96.9% represents the non-labeled ones. Before analyzing the outputs of the various econometrics model the paper briefly summarizes the main statistical facts (reported in Table 2).
36 Treatment Freq. Percent
0 2497.00 96.86%
1 81.00 3.14%
Total 2578.00 100.00%
Table 3 gives a comparison of the basic statistical facts concerning certified (1) and untreated (0) units. From the summary it can be seen that the average transaction price of sustainable properties is 89.30 million euros which varies on average between 6.56 – 573.00 million euros with a standard deviation of 93.10 million euros.
As a comparison the average selling price of the untreated properties is 49.10 million euros which varies from 0.51 to 2,110.00 million euros with a standard deviation of 90.00 million euros.
There is a significant difference between the age of green and non-labeled properties as well. Not surprisingly, sustainable buildings are “younger” on average. Their average age is 7.24 (18.54) years compared to an average 27.61 (37.77) years within the untreated ones (standard deviations are quoted in brackets). The youngest green and non-green properties are brand new (constructed in 2014) but the first eco-certified one was built 122 years ago while the oldest comparable building is 198 years old.
Another interesting outcome relates to the size of the buildings. On average certified buildings have a size of 22,593 square meters (19,167) while the size of non-certified ones is 13,818 (19,092) square meters. The number of floors usually correlated with the size of the buildings hence it is no surprise that green properties have more floors. On average they consisted of 10 floors (11) compared to an average of 6 (6) floors in case of non-certified ones.
37
It is easy to see that executing an analysis with these significant differences would result in biased consequences (as presented later). This phenomenon is called selection bias and the formerly introduced propensity score model is elaborated to make the dataset balanced.
3. Table: Descriptive statistics – before matching
price (million euro) 2497.00 49.51 90.00 0.51 2110.00
Age (year) 2497.00 27.61 37.77 0.00 198.00
sqmeter 2497.00 13818.31 19092.48 165.00 399987.00
FloorNbr 2497.00 41818.00 41802.00 1.00 59.00
Treatment = 1
Variable Obs Mean Std. Dev. Min. Max.
price (million euro) 81.00 89.31 34 274,00 6.56 573.00
Age (year) 81.00 41 844,00 18.54 0.00 122.00
sqmeter 81.00 22593.21 19167.42 2616.00 100000.00
FloorNbr 81.00 41914.00 41954.00 1.00 50.00
“Simple t-test” data description – before matching
As the first step of the analysis a simple “t-test” is executed. The model estimates the effect of being treated on the transaction price without controlling for any other characteristics. The outcome represents that having green certification is significantly different from being untreated related to their transaction price. On average there exists a green premium of approximately 39.8 (10.2) million euros. This analysis has a really low R-squared result (0.0059) which indicates that purely the regressor of being treated is not good enough to predict the values of transaction prices. This model hence suffers from omitted variable bias.
38
Regression outcome – before matching (controlling for limited amount of pretreatment characteristics)
A possible solution to improve the model is adding control variables (Age, Sqmeter, Lat, Long, Trandate, FloorNbr). After adding a dummy variable for being treated (Di = 1) and also
controlling for pretreatment characteristics, the model would provide more realistic results although some variables do not seem to be significantly different from zero.
This is the case with the core component of the model, having a green certification would increase the transaction price on average by 9.06 (6.80) million euros.
However, this result does not significantly differ from zero. The most significant determinant factors of the transaction price are the size and number of floors. Both variables are significant at 1% significance level. A one unit increase in the size (in square meters) – while leaving the other variables constant – would cause on average an approximate 0.0032 (0.000064) million euro increase in the price. In parallel, adding one more floor to the property would cause a value increase of 2.05 (0.19) million euros.
The result shows the expected outcome: a unit change in age would result in price reduction. The magnitude of the decrease is approximately 0.47 million euro (leaving the other variables constant). The R-square of this model has improved a lot (0.5720) but still the model does not function well. As mentioned in the previous section, within a non-randomized model the ATE would not give a credible result.
Propensity score data description
With the algorithm described in the methodology section, blocks where the average propensity scores of the treated and control units differ are to be split in half. This method enables that the propensity score of the entities do not differ anymore. In this analysis the
39
final number of blocks is six. This ensures that there is no difference between the treated and control groups’ mean propensity score. Due to the fact that the examination required a common support condition, block identifiers are missing for the untreated observations outside the common range. After selecting the common support option propensity score varies from 0.00009388 to 0.47679054.
Afterwards the model proceeds to the balancing test for each covariate. If the balancing condition is satisfied then the statistical program tabulates the final dispersion of treated and control units across blocks with their inferior. Hence the original amount of 2,578 observations was reduced to 2,428. The final model includes 81 observations of green properties and the remainder (2,347) accounts for the conventional / comparable ones. Balancing condition is satisfied, Table 4 represents the dispersion of the entities based on their propensity scores.
40 Total 0 1 .0000939 1 387 9 1 396 .025 515 24 539 .05 317 25 342 .1 114 18 132 .2 14 4 18 .4 - 1 1 Total 2 347 81 2 428 Treatment Inferior of block of pscore
Output of the different matching methods
In the upcoming part the paper highlights the most important outcome of the different matching methods. A summary table provided in the result phase highlights the main differences of the outcome between the four examined techniques.
Nearest neighbor matching
The nearest neighbor analysis showed that a green premium does exist on the western European property market. According to the calculation, being a sustainable property would lead to on average a 19.70 (23.90) million euro higher transaction price. However, the statistic also showed that this premium is not significantly different from zero (“t” value is equal to 0.823). After the bootstrapping of the standard errors the outcome does not change significantly. It is important to remember that the nearest neighbor method with common support used only those observations from the control group where estimated propensity scores stand closest to the examined treated unit. Within the common support requirement 81 sustainable office buildings were compared with 72 control observations.
One can argue that nearest neighbor should give the most precise estimation. Note however that the nearest neighbor match is the property with the closest predicted probability based on the matching process. Next to it, due to the limited amount of treated observations, the model
41
does not execute a one to one match, thus the nearest neighbor can differ a lot from the treatment property.
Kernel matching
Kernel method applies a comparison of treated entities to all the control units within the same range of propensity scores. The outcome provides significant result this time. On average the transaction price of green office buildings is higher than that of conventional office buildings by an estimated 28.6 million euros (9.50).
Stratification matching
Stratification matching method also compares the treated units to all the control ones. The result this time, just as was the case with nearest neighbor matching, does not lead to conclude significant consequences. Based on the calculation an estimated 1.88 (14.80) million euro green premium does exist on the market, nevertheless it is statistically not different from zero.
Radius matching
As elaborated above the analysis uses a variant of caliper matching method, namely the radius one. The result shows a significant transaction price premium of green properties. The smaller the tolerance level the smaller the amount of comparable units within the control observation and the smaller the magnitude of the premium. As Smith & Todd (2005) quoted in their paper the appropriate radius tolerance level is not obvious.
Due to the fact that the common support range varies between [.00009388 to .47679054], a tolerance level between 0.00001 and 0.075 would be reasonable. Table (5) represents the sensitivity of the treatment effect based on the change in tolerance level. It can be concluded based on the self-determined range of tolerance level that the green premium varies from 8.30 to 37.90 million euros.
42 5. Table: Radius matching sensitivity analysis
Number of treated units Number of control units Treatment effect Std. Err. "t"
r = 0.075 80 2347 37.61 8.99 4.185 r = 0.05 80 2347 37.91 5.32 7.264 r = 0.01 79 2335 35.21 8.37 4.208 r = 0.0075 79 2333 34.61 11.32 3.102 r = 0.005 78 2330 33.91 8.46 4.012 r = 0.001 77 1417 27.32 6.46 4.185 r = 0.00075 75 1231 24.71 9.44 2.556 r = 0.0005 70 930 21.32 10.40 2.052 r = 0.0001 53 274 8.35 7.32 1.184 r = 0.000075 47 224 12.32 18.62 0.644 r = 0.00005 40 153 8.32 16.32 0.511 r = 0.00001 15 27 26.61 13.71 1.934
Results of radius matching (after bootstrepping)
Note however that the main purpose of executing propensity score model is controlling for selection bias, hence the interpretation of the outcome can be misleading. Having a probability of being treated based on pretreatment characteristics enables the model to compare the most suitable properties based on a limited amount of characteristics. However, there are other characteristics that the model does not control for. The main purpose of executing the propensity model is that it allows the hedonic regression model to use a balanced panel data where the difference between the treated and untreated units is minimized. As it is clearly observable in Table 6, after executing propensity match the difference of pretreatment characteristics has decreased radically, thus a comparison via hedonic regression model would give a more accurate outcome.
43 6. Table: Descriptive statistics – after matching
Treatment = 0
Variable Obs Mean Std. Dev. Min. Max.
price (million euro) 72.00 84.50 99.50 4.46 430.00
Age (year) 72.00 8.44 23.50 0.00 195.00
sqmeter 72.00 20592.67 17490.81 1600.00 100261.00
FloorNbr 72.00 6.80 5.70 1.00 27.00
Treatment = 1
Variable Obs Mean Std. Dev. Min. Max.
price (million euro) 81.00 89.31 34 274,00 6.56 573.00
Age (year) 81.00 41 844,00 18.54 0.00 122.00
sqmeter 81.00 22593.21 19167.42 2616.00 100000.00
FloorNbr 81.00 41 914,00 41 954,00 1.00 50.00
Hedonic regression data description
The main concept of hedonic regression model has been elaborated in the methodology phase, as was the model specification. When one attempts to analyze the existence of green properties’ transaction price premium the percentage price effect is more relevant than the absolute amount. Due to this and other theoretically supported reasons the dependent variable has been transformed to natural logarithm base. The right side of the equation consists of factor and continuous variables. A short summary of the most essential explanatory variables is provided below.
44 Location (country level)
Table 7 shows the dispersion of treated and untreated observations within the western European office markets. As it can be monitored most of the observations were conducted within France and Germany.
Altogether 89% of treatment observations were accounted for within these two countries. Proportion of the properties in the Netherlands also represents a larger part (7.40%) compared to Italy and Spain. This suggests that the analysis could be more credible within these countries. Furthermore, the analysis examines treated observations on city level as well.
7. Table: Dispersion of treated and non-treated observations within countries
Treatment = 0 Treatment = 1
Country Frequency Percentage Country Frequecy Percentage
France 864 34.60% France 16 19.75% Germany 975 39.05% Germany 56 69.14% Italy 143 5.73% Italy 2 2.47% Netherlands 417 16.70% Netherlands 6 7.41% Spain 98 3.92% Spain 1 1.23% Total 2497 100.00% Total 81 100.00% Time dummies
The following table (Table 8) provides information about how many acquisitions occurred within the period from 2007 till 2014. Regarding the sales of untreated, “conventional” office buildings it can be concluded that during the crisis the investment appetite of real estate investors was not buoyant. A decreasing trend in the amount of sales represents well that during the financial crisis the real estate market was hit severely. On the other hand, even if the current dataset does not have a sufficient amount of observations for green property transactions and the consequence may be biased, nonetheless it seems that real estate
45
investors tend to allocate their purchase to green properties. When the financial crisis shocked the market the amount of green investment dropped by an approximate 87%, however after the first year the number of green acquisitions followed an increasing trend in contrast to conventional office sales.
8. Table: Dispersion of observations within years
Treatment = 0 Treatment = 1
Year Frequency Percentage Change ( % ) Year Frequency Percentage Change ( % )
2007 715 28.63% 2007 15 18.52% 2008 462 18.50% -35,38% 2008 2 2.47% 86.67% 2009 192 7.69% -58,44% 2009 6 7.41% 200.00% 2010 242 9.69% 26,04% 2010 7 8.64% 16.67% 2011 308 12.33% 27,27% 2011 16 19.75% 128.57% 2012 252 10.09% -18,18% 2012 18 22.22% 12.50% 2013 281 11.25% 11,51% 2013 17 20.99% -5.56% 2014 45 1.80% Total 2497 100.00% 81 100.00%
Status of buyer company
An interesting topic would arise when one would like to examine if buyer companies’ status would affect the transaction price. In the given analysis the paper differentiates between equity fund, institutional, private, public, other and unknown (when data was not available) types of investments. As the summary table shows below (Table 9) institutional investments are dominant in both conventional and green office sales. This is followed by public and private acquisitions. When a certain type of investor is willing to pay green premium compared to another investment type this reflects that these investors appreciate and see bigger opportunities in green buildings.
46 9. Table: Dispersion of buyer companies' aspect
Treatment = 0 Treatment = 1
Buyer type Frequency Percentage Buyer type FrequencyPercentage
Equity fund 266 10.65% Equity fund 6 7.41%
Institutional 1100 44.05% Institutional 48 59.26% Private 470 18.82% Private 10 12.35% Public 447 17.90% Public 12 14.83% Unknown 110 4.41% Unknown 1 1.23% User/Other 104 4.16% User/Other 4 4.94% Total 2497 100.00% 81 100.00% Results
The first step of this research presented a propensity score matching model that is able to handle selection bias within non-randomized experiments where the treated units differ substantially from their control entities. In practice the model has highlighted that there is a huge difference between treated and untreated units. Due to limited data availability the matching with replacement method is used. It allows to match one control observation more than once. As the analysis below shows the propensity score model concludes different outcomes based on the executed matching methods.
10. Table: Outcome of matching methods
Nearest Neighbor Kernel Stratification Radius (r = 0.001)
Number of treated units 81 81 80 77
Number of control units 72 2347 2348 1417
Treatment effect (million eur) 19.71 28.61 1.88 27.32
Standard error (million eur) 23.91 9.49 14.81 6.46
47
It is important to note that even though the propensity score model is capable of handling selection bias it uses a limited amount of pretreated characteristics, hence a relevant conclusion on a possible existence of green premium can only be drawn up after setting up a hedonic regression model with balanced panel data.
The base concept behind hedonic regression model is the assumption that the characteristics of a given property can influence its transaction price. During the analysis the following hedonic regression models have been run:
Hedonic model – without fixed effect
This first takes a look at the general facts of the first model. The R-squared indicator is a reference point to check validity. It measures statistically how close the data are to the fitted regression line, in other words: it is the percentage of the response variable variation that is explained by the model.
The current outcome of 0.7924 (see Table 11-12) means that 79.24% indicates that the model explains all the variability of the response data around its mean (Stock & Watson, 2012).
Without controlling for the absolute location a significant result cannot be monitored related to the green premium effect. Based on the first model application green properties’ transactional price does not seem to be significantly higher than that of conventional office buildings.
It also seems that institutional investors tend to make bigger investments. They spend larger amounts of capital than equity funds. The magnitude of their premium payment is 16% on average, however the result is not statistically different from zero. The paper cannot draw up further conclusions based on the type of investors.
48
Not surprisingly the bigger the office the higher the sales price. Increasing the size of the property by 1% would lead to a 0.99% increase in the acquisition price. The number of floors cannot influence significantly the selling price. Counterintuitively age also cannot influence significantly the sales price. Causal effect of other variables should not be interpreted since their influential power is not significantly different from zero (time variables, portfolio sales, renovation year). In the second model specification a market fixed effect is included. It controls for absolute location and imposes strictly a linear relationship with the dependent variable. An accurate specification would be using city dummy variables as control ones. However, the problem of including city dummies in the model is that when using a 1-to-1 match this results in an approximate 160 observations and there are about 55 different cities in the matched data, so it loses its flexibility. The country controls model seems to work better.
Hedonic model – fixed effect regression (controlling on country level)
Adding market fixed effect to the hedonic regression would increase the accuracy of the model. 78.39% is the fraction of the sample variance of “LogPrice“ explained by the regressors (Adjusted R-squared = 0.7839). As mentioned above, including market fixed effect would impose strictly a linear relationship with the dependent variable, hence the accuracy and validity improves.
The outcome of the analysis highlights a relevant conclusion: controlling for the relative location by adding market (country) dummies strengthens the hypothesis of the first model that green premium effect does not exist.
Size indicators of properties seem to remain statistically significant. Similarly to the outcome of the first regression, increasing the size of a property by one percent would lead to a 1.01 percent increase in the selling price. All the other variables seem to be insignificant thus further conclusions cannot be drawn up.
