1# RECORD 2bbd25e60ca8ea2625da98d69a5096e00200229f2# TITLE: Does energy efficiency matter for prices of tenant-owned apartments?3# YEAR: 2022.0 VENUE: Environmental science and pollution research international DOI: 10.1007/s11356-022-20482-w4# AUTHORS: David Stenvall; P. Cerin; Bo Sjö; Gazi Salah Uddin5# PAGES: 15 CANDIDATE TABLE PAGES (1-based): [8, 9, 10, 11, 12]6# ABSTRACT: In this study, we use Energy Performance Certificate (EPC) information to investigate the effect of energy efficiency on the selling price of Swedish tenant-owned apartments. While there is a large body of literature on how energy efficiency affects the sales price of single-family houses, none has exclusively focused on tenant-owned apartments. For owners of tenant-owned apartments in Sweden, heating is for a large share included in the monthly fee paid to the tenant association, which usually does not change on a short-term basis. This raises the question whether homebuyers’ incentives for acquiring energy-efficient tenant-owned apartments are large enough to be capitalized into the prices. By hedonic models and matching methods, we found mixed results. In our most optimistic scenarios, tenant-owned apartments enclosed in energy-efficient buildings are sold with a premium of approximately 0.8 to 1.2% compared to apartments in non-efficient buildings. The results in this study are not robust to all model specifications and vary across regions. In comparison with recent studies using data for single-family houses in Sweden, our detected capitalization is smaller. Our results highlight a need for targeted measures if EPCs are to be fully capitalized in prices for all dwelling types in which buyers have different economic incentives for reducing their energy consumption.789===== PAGE 1 (score 9, numeric-paren 0, table-words 0, var-words 9) =====1011Vol.:(0123456789)121 313https://doi.org/10.1007/s11356-022-20482-w14RESEARCH ARTICLE15Does energy efficiency matter for prices of tenant‑owned apartments?16David Stenvall1 · Pontus Cerin1 · Bo Sjö1 · Gazi Salah Uddin117Received: 15 July 2021 / Accepted: 23 April 2022 18© The Author(s) 202219Abstract20In this study, we use Energy Performance Certificate (EPC) information to investigate the effect of energy efficiency on 21the selling price of Swedish tenant-owned apartments. While there is a large body of literature on how energy efficiency 22affects the sales price of single-family houses, none has exclusively focused on tenant-owned apartments. For owners of 23tenant-owned apartments in Sweden, heating is for a large share included in the monthly fee paid to the tenant association, 24which usually does not change on a short-term basis. This raises the question whether homebuyers’ incentives for acquiring 25energy-efficient tenant-owned apartments are large enough to be capitalized into the prices. By hedonic models and matching 26methods, we found mixed results. In our most optimistic scenarios, tenant-owned apartments enclosed in energy-efficient 27buildings are sold with a premium of approximately 0.8 to 1.2% compared to apartments in non-efficient buildings. The 28results in this study are not robust to all model specifications and vary across regions. In comparison with recent studies 29using data for single-family houses in Sweden, our detected capitalization is smaller. Our results highlight a need for targeted 30measures if EPCs are to be fully capitalized in prices for all dwelling types in which buyers have different economic incen-31tives for reducing their energy consumption.32Keywords Energy performance certificates · Housing markets · Tenant-owned apartments · Energy efficiency · Green 33premium34JEL Classification D10 · Q51 · R2035Introduction36Reducing the speed of climate change has in the recent dec-37ade become as a top priority for numerous governments, 38scientists, and institutions. The EU has set up ambitious 39GHG emission targets in a vision for 2050, which could 40pose one important step-stone for the world to reach below 41the Paris 2-degree target and, as the union puts it, keep it 42down to 1.5 °C. This will, consequently, put a lot of pressure 43on the housing sector to abate its considerable large share of 44CO2 emissions compared to other sectors. In this context, a 45common European approach to reduce the energy usage by 46dwellings was introduced through the Energy Performance 47of Buildings Directive (EPBD 2002/91/EC). The directive 48includes a certification scheme aimed towards dwellings, 49namely, the Energy Performance Certificates (EPCs).1 50One of the main purposes of the introduction of EPCs was 51to decrease potential information asymmetries between 52owners and prospective buyers with regard to the energy 53performance of dwellings. Hence, these certifications are 54supposed to make potential buyers aware of the energy per-55formance of their future homes or buildings (e.g., Brounen 56and Kok 2011; Bio Intelligence Service 2013; Cerin et al. 572014; Fuerst et al. 2015; Fuerst et al. 2016a, b; EC 2020). 58Responsible Editor: Philippe Garrigues59 * Gazi Salah Uddin 60 61gazi.salah.uddin@liu.se62 63David Stenvall 64 65david.stenvall@liu.se66 67Pontus Cerin 68 69pontus.cerin@liu.se70 71Bo Sjö 72 73bo.sjo@liu.se741 75Department of Management and Engineering, Linköping 76University, Linkoping, Sweden771 For more information regarding the EPC in EU, see e.g., EC 78(2020). In the Swedish case, more information can be found on the 79website of the National Board of Housing, Building and Planning 80(Boverket 2019).81/ Published online: 4 May 202282Environmental Science and Pollution Research (2022) 29:66793–66807838485===== PAGE 8 (score 103, numeric-paren 73, table-words 5, var-words 11) =====86871 388the number of rooms according to the VIF values (variance 89inflation factors). As both are control variables, they are kept 90in the estimations. The VIF value for our treatment vari-91able for energy-efficient tenant-owned apartments, HIGH, 92is very low and does not suggest any problems with multi-93collinearity. Our R2 over 81% is notably high, but common 94for hedonic price studies on the housing market, especially 95given that our sample consists of apartments which are more 96homogeneous compared to houses. As HIGH is insignifi-97cant, we do not find any effect of energy efficiency on the 98sales price of the tenant-owned apartments.99In column (2), we instead use the 4-digit postcode to 100control for more detailed locational effects. Our R2 now 101increases to 93%, which indicates that a more detailed con-102trol for location is important in explaining the sales price. 103Our treatment, HIGH, is now significant and indicates a 104price premium of approximately 1.1% for energy-efficient 105tenant-owned apartments. This can be interpreted as approx-106imately 11,000 SEK more for every 1 million in sales price 107for the energy-efficient homes, ceteris paribus, compared to 108the reference category (homes enclosed in buildings with 109higher energy performance than 75 kWh/m2). The size of 110the premium is smaller compared with the findings of Cerin 111et al. (2014) and Wilhelmsson (2019) that found larger pre-112miums when using data for houses rather than apartments. 113As emphasized in “Introduction,” a large share of the home-114owners of tenant-owned apartments has heating included in 115the monthly fee paid to the association each month, which 116usually does not change on a short-term basis. The fee is not 117based on the usages per apartment but rather the joint costs 118for the association. That the premium is quite low might 119simply reflect low economic incentives for many buyers of 120tenant-owned apartments. In column (3), we include our 121dummy, LOW, for tenant-owned apartments enclosed in 122tenant-owned buildings with low energy efficiency (hereafter 123denoted as low-efficient apartments). By including dummies 124for apartments with both high and low energy efficiency, the 125reference category is now apartments in between these two 126categories (medium energy efficient). The insignificance of 127LOW indicates that low-efficiency apartments are not sold 128Table 2 Full sample — hedonic price estimations129The dependent variable is the logarithmic sales price. Standard errors in parenthesis. Columns (1), (2), and (3) are estimated with robust stand-130ard errors. Postcode fixed effects is based on 4 digits if nothing else is stated. HIGH is a dummy variable for tenant-owned apartments enclosed 131in tenant-owned buildings with high energy efficiency; LOW is a dummy for tenant-owned apartments enclosed in tenant-owned buildings with 132low energy efficiency; Area is the size of the dwelling in square meters (log); Room is the number of rooms; Building year (age) is a dummy 133equal to 1 if the apartment was constructed during the given range of years (omitted category is pre-1901). Rent is the fee paid to the tenant asso-134ciation each month in SEK (log). We also include a dummy variable equal to 1 for apartments with double construction years (see the explana-135tion in the data section), for which the result is not shown136Statistical significance is denoted by ***p < 0.01, **p < 0.05, *p < 0.1137(1)138(2)139(3)140(4)141(5)142Energy efficiency (EPC)143 HIGH1440.001 (0.007)1450.011** (0.005)1460.011** (0.005)1470.007* (0.004)1480.007* (0.004)149 LOW150 − 0.001 (0.002)151 − 0.001 (0.002)152Dwelling-specific controls153 Area1540.862*** (0.019)1550.746*** (0.012)1560.746*** (0.012)1570.751*** (0.006)1580.751*** (0.006)159 Rent160 − 0.277*** (0.023) 161 − 0.144*** (0.014)162 − 0.143*** (0.014)163 − 0.156*** (0.004)164 − 0.156*** (0.004)165 Room1660.049*** (0.004)1670.060*** (0.002)1680.060*** (0.002)1690.058*** (0.002)1700.058*** (0.002)171 Building year172 1901–1921173 − 0.041***174 − 0.002175 − 0.002176 − 0.006177 − 0.006178 1922–1941179 − 0.143***180 − 0.066***181 − 0.066***182 − 0.070***183 − 0.070***184 1942–1961185 − 0.280***186 − 0.116***187 − 0.116***188 − 0.120***189 − 0.120***190 1962–1976191 − 0.467***192 − 0.165***193 − 0.165***194 − 0.163***195 − 0.163***196 1977–1991197 − 0.304***198 − 0.143***199 − 0.144***200 − 0.139***201 − 0.139***202 1992–2005203 − 0.218***204 − 0.074***205 − 0.074***206 − 0.063***207 − 0.063***208 2006–2019209 − 0.123***210 − 0.001211 − 0.002212 − 0.004213 − 0.004214Constant21513.48***21613.48***21713.48***21813.54***21913.54***220Time fixed effects221Yes222Yes223Yes224Yes225Yes226Robust regression227No228No229No230Yes231Yes232Postal code fixed effects233(2 digits)234Yes235Yes236Yes237Yes238R22390.8122400.9332410.9332420.9432430.943244Obs24521,69624621,69624721,69624821,69524921,69625066800251Environmental Science and Pollution Research (2022) 29:66793–66807252253254===== PAGE 9 (score 51, numeric-paren 3, table-words 18, var-words 11) =====2552561 3257with a discount compared to medium-efficiency homes all 258else equal. In columns (4) and (5), we try to mitigate issues 259related to potential outliers by performing the estimations 260using the robust regression. The effect of high energy effi-261ciency on the sales price is now reduced to a premium of 262approximately 0.7% that now is weakly significant. The 263effect of low efficiency on prices is still insignificant.264In Table 3, we present the result from regional subsam-265ples. The category South consists of the observations from 266the municipality of Malmö while the Mid category is based 267on the sales from Gothenburg and Linköping. As only a few 268observations of the sample from the northern cities Luleå 269and Umeå do qualify as energy efficient according to our 270boundaries, we do not run estimation solely based on these 271observations. As observed, the coefficient for HIGH is larg-272est in the southern sample followed by the Stockholm. The 273effect is, however, not statistically significant for any region.274Table 4 shows the hedonic price estimations using the 275sample of tenant-owned apartments enclosed in a tenant-276owned building with an EPC grade. In (1), we classify 277energy-efficient apartments as those enclosed in a build-278ing with the grades A or B and the other labels (C–G) as 279the reference category. Our results indicate that the effect 280of energy efficiency is positively related to the sales price 281but only weakly significant. The estimated premium is 282approximately 1.5%, all else equal, which is higher com-283pared with the effect of energy efficiency in the full sample 284in Table 2. This 1.5% premium can be compared with the 285ten-basis point reduction in mortgage rate that buyers can 286get by certain banks if they acquire a home with an energy 287grade of A or B. Assuming an annual mortgage rate of 2% 288on a 4 million SEK loan, a 0.10-percentage point reduction 289is approximately 4000 SEK less per year. If not considering 290the time value of money or any other potential cost savings 291from owning an energy-efficient home, a 1.5% premium is 292paid off in 15 years based solely on the mortgage rate reduc-293tion. Therefore, a 1.5% premium is not unreasonable based 294on these economic incentives. In Table 4, column (2), we 295instead define the energy-efficient labels as A–C and com-296pare it with the less efficient grades (D–G). Even though 297the coefficient for the energy efficient grades is positive, the 298effect is not significant. Hence, our results from (1) and (2) 299suggest that the price premium is only present in the most 300energy-efficient labels, A and B. That these two labels also 301provide the largest interest rate reduction for homebuyers 302with green loans might perhaps be a part of the explana-303tion. In column (3), we add dummies for the least energy 304efficient labels (E–G). Our result from this estimation shows 305no significant discount compared with the reference category 306(label D). In (4) and (5), we run the estimations using robust 307regression. Our effect found in column (4) indicates a 1% 308premium for labels A and B. However, this is not enough 309to reach statistical significance. While the robust regression 310barely drops any observations, it still weights the observa-311tions that are not dropped. Hence, the reduced effect from 3121.5% premium in column (1) to 1% in column (4) is likely 313due to the weighting of potential outliers rather than the 314actual dropping of observations.315Robustness check: neighborhood controls 316and matching317As one of the few Swedish studies, we also control for some 318neighborhood characteristics in our hedonic price models. 319In this section, we use the full sample from Table 2. Because 320the neighborhood data is based on DeSO regions, observa-321tions belonging to the same region will get the same neigh-322borhood values. In Table 5, columns (1) and (2) we start the 323investigation by dividing the data into income categories 324based on the income data. We calculate the income quar-325tiles in the sample based on the incomes attached to each 326observation. Apartments with incomes below quartile 1 is 327denoted by IC1, apartments with incomes equal to or larger 328than quartile 1 but smaller than quartile 2 is denoted by IC2, 329and, lastly, apartments with incomes equal to or larger than 330quartile 2 but smaller than quartile 3 is denoted by IC3.331As observed in Table 5, column (1), the income dum-332mies (IC 1–3) are all significant with negative coefficients. 333This means that there is a significant difference in price, all 334else equal, between observations within our included income 335categories and our omitted reference group (high incomes 336above the third quartile). Our energy efficiency variable, 337Table 3 Regression with regional subsamples from the full sample338The dependent variable is the logarithmic sales price. Standard errors 339are in parenthesis. Postcode fixed effects are based on 4 digits. Robust 340standard errors are used in the estimations. The mid region contains 341dwellings from the municipalities Linköping and Gothenburg, and the 342southern region consists of apartments from the Malmö municipal-343ity. HIGH is a dummy for tenant-owned apartments enclosed in ten-344ant-owned buildings with high energy efficiency. Dwelling-specific 345variables include the following: area (log), number of rooms in each 346apartment, rent or fee paid to the tenant association each month (log), 347and dummy for construction year (same as Tables 2 and 3). We also 348include a dummy variable equal to 1 for apartments with double con-349struction years (see the explanation in the data section)350Statistical significance is denoted by ***p < 0.01, ** p < 0.05, * 351p < 0.1352Stockholm353South354Mid355HIGH3560.004 (0.006) 0.015 (0.017) 0.002 (0.008)357Dwelling-specific 358controls359Yes360Yes361Yes362Time fixed effects363Yes364Yes365Yes366Postal code fixed 367effects368Yes369Yes370Yes371R23720.9133730.8853740.912375Obs37611 4683774142378528537966801380Environmental Science and Pollution Research (2022) 29:66793–66807381382383===== PAGE 10 (score 100, numeric-paren 68, table-words 7, var-words 9) =====3843851 3386HIGH, indicates a price premium of approximately 0.8%, 387but it is only of a weakly statistically significant size. How-388ever, when using 3-digit postcodes (a wider geographical 389area) in column (3), controlling for outliers using robust 390regression in (2) and (5), including a full set of neighbor-391hood controls (3)–(5), our energy efficiency variable HIGH 392is no longer significant.393In Table 6, we test if our estimations vary depending 394on the boundaries for which we classify the apartments as 395energy efficient. As observed, the coefficient when narrow-396ing the boundaries, columns in (1) estimate of HIGH, is 397larger compared to the results in Table 5, but still not sig-398nificant. This might be a sample size problem with relatively 399few observations reaching energy efficiency status. If instead 400increasing the boundaries in (2), we observe a non-signifi-401cant premium for energy-efficient tenant-owned apartments 402of approximately 0.5%. In (3), we include an interaction 403between our energy efficiency measure and the net income 404in the DeSO in which each observation belongs to. By this, 405we want to analyze if the effect of energy efficiency on prices 406is different if the energy-efficient apartments are sold in a 407low- versus high-income area. As a comparison, an energy-408efficient apartment in a neighborhood with a net income of 409250,000 SEK (which is slightly below median in the sample) 410will according to our estimates be sold with a premium of 4111.77% compared to a non-efficient apartment. However, in 412a neighborhood with a net income of 320,000 SEK (above 413median), an energy-efficient apartment is sold with a dis-414count of − 0.2%. The interaction between net income and 415HIGH is significant even if we run the estimation with our 416standard boundary of 75 kWh/sqm. One possible explana-417tion is that there might be a larger share of apartments in 418which heating is not included in the monthly fee in low- 419versus high-income areas. In such case, the economic incen-420tives might be lower for high-income individuals. It could 421also be that high-income individuals simply do afford to not 422Table 4 Grade sample — hedonic price estimations423The dependent variable is the logarithmic tenant-owned apartment sales price. Standard errors in parenthesis. Columns (1), (2), and (3) are 424estimated with robust standard errors. Postcode fixed effects is based on 4 digits if nothing else is stated. Area is the of the size of the dwelling 425in sqm (log); Room is the number of rooms; Building year (age) is a dummy equal 1 if the apartment was constructed during the given range of 426years (omitted category is pre 1901). Rent is the fee paid to the tenant association each month in SEK (log); Energy grades are dummy variables 427equal to 1 if the tenant-owned apartment belongs to a tenant-owned building with the specified EPC-grade/grades. In columns (3) and (5), the 428energy grade D is the omitted category, while grades C–G is omitted in columns (1) and (4) and D–G in (2). We also include a dummy variable 429equal to 1 for apartments with double construction years (see explanation in the data section), for which the result is not shown. Columns (1), 430(2), and (3) are ordinary hedonic price estimations while (4) and (5) are hedonic price model estimated by robust regressions431We denote statistical significance by ***p < 0.01, **p < 0.05, *p < 0.1432(1)433(2)434(3)435(4)436(5)437Energy grade (EPC)438 AB4390.015* (0.009)4400.010 (0.008)441 ABC4420.001 (0.005)4430.002 (0.006)444 − 0.001 (0.006)445 EFG4460.002 (0.005)447 − 0.002 (0.005)448Dwelling-specific controls449 Area4500.749*** (0.023)4510.750*** (0.023)4520.750*** (0.023)4530.750*** (0.012)4540.750*** (0.012)455 Rent456 − 0.164*** (0.025)457 − 0.165*** (0.025)458 − 0.165*** (0.025)459 − 0.170*** (0.009)460 − 0.170*** (0.009)461 Room4620.065*** (0.004)4630.065*** (0.004)4640.065*** (0.004)4650.059*** (0.003)4660.059*** (0.003)467 Building year468 1901–19214690.0004700.0004710.0004720.000473 − 0.001474 1922–1941475 − 0.059***476 − 0.059***477 − 0.060***478 − 0.068***479 − 0.068***480 1942–1961481 − 0.129***482 − 0.129***483 − 0.129***484 − 0.136***485 − 0.136***486 1962–1976487 − 0.172***488 − 0.172***489 − 0.172***490 − 0.165***491 − 0.165***492 1977–1991493 − 0.166***494 − 0.166***495 − 0.166***496 − 0.158***497 − 0.158***498 1992–2005499 − 0.091***500 − 0.091***501 − 0.091***502 − 0.079***503 − 0.079***504 2006–20195050.0165060.0195070.0195080.0135090.014510Constant51113.80***51213.80***51313.80***51413.97***51513.97***516Time fixed effects517Yes518Yes519Yes520Yes521Yes522Robust regression523No524No525No526Yes527Yes528Postal code fixed effects529Yes530Yes531Yes532Yes533Yes534R25350.9365360.9365370.9365380.9455390.945540Obs5416555542655554365555446555545655354666802547Environmental Science and Pollution Research (2022) 29:66793–66807548549550===== PAGE 11 (score 49, numeric-paren 18, table-words 9, var-words 10) =====5515521 3553care about energy savings. As we do not have information 554on heating included for the full sample, we are not able to 555control for this.556As described in the data section, ordinary hedonic price 557estimations might, however, suffer from methodical prob-558lems. Using a propensity score method, we can address the 559issue of differences between the group with high energy 560efficiency (treatment) versus the non-energy efficient group 561(control). In Table 8, we use two different methods of pro-562pensity score. First, we estimate the propensity score using 563a probit model and include the score in the hedonic model 564as a covariate. The propensity score in the first step is esti-565mated using the treatment variables (HIGH and AB) as the 566dependent variable and the dwelling- and location-specific 567Table 5 Full sample — hedonic 568prices with neighborhood 569controls570The dependent variable is the logarithmic tenant-owned apartment sales price. Standard errors in paren-571thesis. HIGH is a dummy variable for tenant-owned apartments enclosed in tenant-owned buildings with 572high energy efficiency. IC1 is a dummy for observations that belong to a postcode with low incomes. IC2 573is a dummy for low to median income postcodes while IC3 is a dummy for median to upper incomes. 574Neighborhood controls consists of median net income (log), share of persons with foreign origin, share 575of persons in rental apartments, and share of persons receiving financial aid and number of households 576(log). Dwelling-specific variables are area (log), number of rooms in each apartment, rent or fee paid to the 577tenant-association each month (log), and dummy for construction year (same as Tables 2 and 4). We also 578include a dummy variable equal to 1 for apartments with double construction years (see explanation in the 579data section)580We denote statistical significance by ***p < 0.01, **p < 0.05, *p < 0.1581(1)582(2)583(3)584(4)585(5)586HIGH5870.008* (0.005)5880.006 (0.004)589 − 0.003(0.005)5900.003 (0.005)5910.000 (0.004)592IC1593 − 0.113***594 − 0.092***595IC2596 − 0.053***597 − 0.038***598IC3599 − 0.022***600 − 0.015***601Dwelling-specific controls602Yes603Yes604Yes605Yes606Yes607Time fixed effects608Yes609Yes610Yes611Yes612Yes613Robust regression614No615Yes616No617No618Yes619Postal code fixed effects620(4 digits)621(4 digits)622(3 digits)623(4 digits)624(4 digits)625Neighborhood controls626No627No628Yes629Yes630Yes631R26320.9356330.9446340.9246350.9386360.946637Obs63821,69663921,69164021,69664121,69664221,696643Table 6 Robustness and 644alternative boundaries for 645energy efficiency646The dependent variable is the logarithmic tenant-owned apartment sales price. Standard errors are in 647parenthesis. HIGH is a dummy for tenant-owned apartments enclosed in tenant-owned buildings with high 648energy efficiency. Dwelling-specific variables consist of the following: area (log), number of rooms in each 649apartment, rent or fee paid to the tenant-association each month (log), and dummy for construction year 650(same as Tables 2 and 4). Neighborhood controls include median net income (log), share of persons with 651foreign origin, share of persons in rental apartments, share of persons receiving financial aid, and number 652of households (log). We also include a dummy variable equal to 1 for apartments with double construction 653years (see explanation in the data section)654Statistical significance is denoted by ***p < 0.01, **p < 0.05, *p < 0.1655(1)656(2)657(3)658HIGH (≤ 50 kWh/m2)6590.010 (0.012)660HIGH (≤ 100 kWh/m2)6610.005 (0.003)6620.468*** (0.098)663HIGH (≤ 100 kWh/m2) × net income664 − 0.082*** (0.017)665Neighborhood controls666Yes667Yes668Yes669Dwelling-specific controls670Yes671Yes672Yes673Robust regression674No675No676No677Time fixed effects678Yes679Yes680Yes681Postal code fixed effects682(4 digits)683(4 digits)684(4 digits)685R26860.9386870.9386880.938689Obs69021,69669121,69669221,69669366803694Environmental Science and Pollution Research (2022) 29:66793–66807695696697===== PAGE 12 (score 36, numeric-paren 6, table-words 11, var-words 7) =====6986991 3700variables from the hedonic estimations as independent vari-701ables. Here, we chose to only include dummies with high 702energy efficiency and not for low efficiency. In the second 703method, we match the treatment and control groups based 704on the estimated propensity score using a nearest neighbor 705approach. We thereafter run the hedonic regression on this 706matched sample using weights from the matching process. 707For a discussion of these methods, see, e.g., Stuart (2010).708Table 7 shows the means in the sales price and covari-709ates before and after propensity score matching. As sus-710pected, we can observe that the treatment group in the 711unmatched samples largely differs regarding the age of 712the tenant-owned apartments. The control groups include 713older apartments that are both smaller and cheaper on 714average. However, after the matching procedure these dif-715ferences have been largely reduced. The balance in the 716matched samples is tested by t-tests and mean standardized 717differences, which can be found in appendix Table A3. 718These tests show that there is no significant difference 719between the control and treatment groups after match-720ing. Table 8 shows the results from the propensity score 721estimations.722Table 7 Mean in covariates 723before and after propensity 724score matching725This table shows the treatment and control groups before and after propensity score matching. The treat-726ment in the full sample is the variable HIGH, which is a dummy for tenant-owned apartments enclosed in 727tenant-owned buildings with high energy efficiency. The treatment for the graded sample is dummy vari-728ables equal to 1 if the tenant-owned apartment belongs to a tenant-owned building with the energy grade A 729or B. The matched full sample is equivalent to the sample used in estimations of column (2) in Table 8. For 730the graded matched sample, this is equivalent to the data used in estimations of column (4) in Table 8731Full sample732Graded sample733Unmatched734Matched735Unmatched736Matched737Treatment738Control739Treatment740Control741Treatment742Control743Treatment744Control745Price7463,742,0867473,338,2077483,761,0897493,754,7877503,605,3267513,129,4897523,697,6487533,624,645754Area75570.975663.675771.175872.575970.276065.976170.176269.0763Age76426.576566.676626.876728.876819.2776953.677020.077123.0772Rent7733800.47743495.47753805.17763975.47773841.67783637.17793839.47803814.1781Room7822.77832.47842.77852.87862.77872.57882.77892.6790Obs791139779220,299793137779413777954247966131797394798394799Table 8 Result after propensity 800score estimations801This table shows the hedonic price model with estimated propensity score as a covariate columns (1) and 802(3) as well as on propensity score matched samples (columns (2) and (4)). The dependent variable is the 803logarithmic tenant-owned apartment sales price. Standard errors are in parenthesis. HIGH is a dummy for 804tenant-owned apartments enclosed in tenant-owned buildings with high energy efficiency. Postcode fixed 805effects are 4 digits if nothing else is stated. The treatment for the graded sample is dummy variables equal 806to 1 if the tenant-owned apartment belongs to a tenant-owned building with the energy grade A or B. 807Dwelling-specific variables include area (log), number of rooms in each apartment, rent or fee paid to the 808tenant-association each month (log), and dummy for construction year (same as Tables 2 and 3). We also 809include a dummy variable equal to 1 for apartments with double construction years (see explanation in the 810data section)811Statistical significance is denoted by ***p < 0.01, **p < 0.05, *p < 0.1812Full sample813Graded sample814(1)815(2)816(3)817(4)818HIGH8190.009* (0.005)8200.012** (0.005)821AB8220.011 (0.009)8230.014 (0.009)824Dwelling-specific controls825Yes826Yes827Yes828Yes829Time fixed effects830Yes831Yes832Yes833Yes834Postal code fixed effects835Yes836Yes837Yes838Yes839Regression on matched sample840No841Yes842No843Yes844Propensity score as control845Yes846No847Yes848No849R28500.9278510.9118520.9298530.939854Obs85516,5118562754857363885878885966804860Environmental Science and Pollution Research (2022) 29:66793–66807861862863===== PAGE 13 (score 42, numeric-paren 2, table-words 16, var-words 7) =====8648651 3866As we can see from columns (1) and (3), including the 867propensity score as a covariate largely reduces the effect of 868high energy efficiency on the sales price of tenant-owned 869apartments. The coefficient is positive and ranges from a 0.9 870to 1.1% premium, ceteris paribus, for tenant-owned apart-871ments enclosed in energy-efficient tenant-owned buildings 872compared to non-efficient. While the effect in the full sample 873is weakly significant, the estimated premium in the graded 874sample is insignificant. Looking at the matched samples 875(columns (2) and (4)), the effect is similar with the coef-876ficient ranging from 1.2 to 1.4% price premium. As before, 877the estimations based on the full sample is weakly significant 878but the effect in the graded sample is insignificant. Regard-879ing the insignificance in column (4), it is possible that the 880lack of significance could be a power issue considering the 881few observations. The coefficient for the treatment variable 882(AB) does not change much compared to the hedonic price 883model estimates in Table 4, column (1). We also observe 884that the coefficient for energy efficiency in Table 8, column 885(2), i.e., the non-graded sample, decreases compared to its 886corresponding hedonic estimates in Table 2. In this sample, 887we do not have the same issue with power, which indicates 888that other problems (such as non-randomness or sample 889selection) potentially may explain the differences between 890Tables 2 and 8.891Lastly, we use an alternative matching method, the 892coarsened exact matching (CEM) by Iacus et al. (2012). 893The results of the CEM are dependent on the variables cho-894sen for the matching procedure and the cut points through 895which these variables are coarsened. We chose to match on 896the most important control variables age, log rent, log area, 897and the number of rooms in each tenant-owned apartment. 898While running the CEM in Stata, it provides an algorithm 899to automatically choose the cut points. However, we chose 900the cut points by our own to ensure that this is done based 901on theoretical knowledge.9 We test different specifications 902before we end up with the final cut points. To quantify how 903much our cut points reduce the imbalances between the 904control and treatment observations, we use the multivariate 905L1 distance. The goal is to produce a model with as low a 906distance as possible. The distance in our chosen models can 907be found in Table A4, which are lower compared with the 908multivariate L1 distance produced by the Stata algorithm. 909In accordance with the procedure of Blackwell et al. (2009), 910we add the matched control variables to the hedonic regres-911sion to control for the remaining imbalances. We also add 912the control variables that were not used in the matching. Our 913results are shown in Table 9.914As observed, the results from the CEM are lower com-915pared to the estimates produced from the propensity score or 916hedonic estimation in Tables 2 and 4. The estimated effect 917for the graded or full sample is not statistically significant. 918The overall result from the matching procedures gives 919some conflicting evidence. Although the propensity score 920approach in Table 7 indicates price premiums of energy effi-921ciency close to the hedonic price model in Tables 2 and 4, 922the premiums indicated by the CEM are much lower. While 923results from the CEM are sensitive to the chosen cut points, 924the propensity score estimates are sensitive to which control 925observations the treatments are matched to. It is therefore 926very difficult to conclude which of these results are the most 927credible. In the previous literature, Wilhelmsson (2019) 928found that the price premium for single-family houses was 929substantially reduced when controlling for selection bias 930using propensity score matching. Against this background, 931we are not able to rule out that the results from the hedonic 932price models in Tables 2 and 4, at least to some extent, could 933be affected by sample selection bias. If such bias exists, our 934results from this section show that the price premium for 935energy-efficient tenant-owned apartments is lower compared 936with the results presented in Tables 2 and 4.937Table 9 Hedonic estimations after coarsened exact matching (CEM)938This table shows the results of hedonic price estimations weighted on 939the matched dataset after CEM. Standard errors are in parenthesis. 940The dependent variable is the logarithmic tenant-owned apartment 941sales price. HIGH is a dummy for tenant-owned apartments enclosed 942in tenant-owned buildings with high energy efficiency. Postcode fixed 943effects are based on 4 digits if nothing else is stated. The treatment 944for the graded sample is dummy variables equal to 1 if the tenant-945owned apartment belongs to a tenant-owned building with the energy 946grade A or B. Dwelling-specific variables include area (log), number 947of rooms in each apartment, rent or fee paid to the tenant-association 948each month (log), and dummy for construction year (same as Tables 2 949and 3). We also include a dummy variable equal to 1 for apartments 950with double construction years (see explanation in the data section)951Full sample952Graded sample953HIGH9540.005 (0.005)955AB9560.007 (0.009)957Dwelling-specific controls958Yes959Yes960Time fixed effects961Yes962Yes963Postal code fixed effects964Yes965Yes966R29670.9319680.933969Obs97017,70297143739729 Our cut points are as follows: age (0 14.5 29.5 44.5 59.5 79.5 99.5 973120.5), lrent (0 7.7 8 8.3 8.45 8.52 8.7 10), larea (0 3.1 3.45 3.8 4 9744.2 4.5 4.75 5 6), and rooms (0 2.25 3.75 5.25 10). For the cut points 975of the variables with a logarithmic transformation (larea and lrent), 976these are chosen with respect to corresponding values for the non-977logged variables (rent and area). For instance, the range 0–3.1 for 978larea corresponds to the range 0–22 m2 without any logarithmic 979transformation.98066805981Environmental Science and Pollution Research (2022) 29:66793–66807982