1# RECORD 38bde6e6f3fb947d8b6029d64eef6268c93759cc2# TITLE: Marginal prices of improvements made to blocks of flats: empirical evidence from Romania3# YEAR: 2016.0 VENUE: nan DOI: 10.3846/1648715x.2015.11214154# AUTHORS: Adela Deaconu; D. Lazar; Anuța Buiga; Gheorghe Fatacean5# PAGES: 16 CANDIDATE TABLE PAGES (1-based): [8, 10, 12]6# ABSTRACT: This paper examines a specific segment of the housing markets from Central and Eastern European countries, namely the stock of flats in the urban area. With focus on Romania, we emphasize the improvements made by owners and authorities, aimed to enhance the quality of life and the maintenance of old blocks of flats; are highlighted some aspects regarding the cost of these improvements. In addition, the empirical section investigates whether and to what extent the improvements made to the flats and buildings have a significant effect on sale prices and rental prices. The marginal prices of these attributes are provided by hedonic pricing models, estimated in a generalized linear model framework; gamma and log-linear regression models perform in a similar way in terms of fit and prediction accuracy. The physical condition of the building and finishing works inside the flats, as variables of interest in our models, have a significant effect on sale and rental prices, but with different intensity. A common set of significant attributes, in the hedonic models for both sale and rental price, is suggested.789===== PAGE 1 (score 8, numeric-paren 0, table-words 0, var-words 8) =====1011Copyright © 2016 Vilnius Gediminas Technical University (VGTU) Press12http://www.tandfonline.com/tspm13* Corresponding author. E-mail: dorina.lazar@econ.ubbcluj.ro14MARGINAL PRICES OF IMPROVEMENTS MADE TO BLOCKS OF FLATS: 15EMPIRICAL EVIDENCE FROM ROMANIA16Adela DEACONU a, Dorina LAZAR b,*, Anuta BUIGA b, Gheorghe FATACEAN a17a Department of Accounting and Audit, Faculty of Economics and Business Administration, Babes-18Bolyai University, Cluj-Napoca, Romania19b Department of Statistics, Forecasting, Mathematics, Faculty of Economics and Business Administra-20tion, Babes-Bolyai University, Teodor Mihali 58-60, 400591 Cluj-Napoca, Romania21Received 7 April 2014; accepted 3 July 201522Abstract. This paper examines a specific segment of the housing markets from Central and Eastern 23European countries, namely the stock of flats in the urban area. With focus on Romania, we empha-24size the improvements made by owners and authorities, aimed to enhance the quality of life and the 25maintenance of old blocks of flats; are highlighted some aspects regarding the cost of these improve-26ments. In addition, the empirical section investigates whether and to what extent the improvements 27made to the flats and buildings have a significant effect on sale prices and rental prices. The marginal 28prices of these attributes are provided by hedonic pricing models, estimated in a generalized linear 29model framework; gamma and log-linear regression models perform in a similar way in terms of fit 30and prediction accuracy. The physical condition of the building and finishing works inside the flats, as 31variables of interest in our models, have a significant effect on sale and rental prices, but with differ-32ent intensity. A common set of significant attributes, in the hedonic models for both sale and rental 33price, is suggested.34Keywords: Hedonic model; Flat; Sale price; Rental price; Improvements attributes; Generalized 35linear model361. Introduction37Residential property accounts for a significant 38amount of homeowners’ wealth, representing an 39important asset for most of them. In Europe, the 40value of properties ranges from 0.4 times the GDP 41in Poland to 1.78 times in France, as reported by 42Colleta (2013). On the other hand, there is a strong 43correlation between the housing market, in par-44ticular the real estate prices index, and macroeco-45nomics (Hepsen and Vatansever 2012). Goodhart 46and Hofmann (2008) highlight the interdepend-47ence between housing cost dynamics and economic 48activity, bank credit, inflation, and the implica-49tions of house price fluctuations for monetary and 50public policies.51Following the pattern of other European econo-52mies in the last decade, the rapid development of 53real estate markets in Central and Eastern Euro-54pean (CEE) countries was accompanied by an up-55ward trend in the housing prices, and the expan-56sion of bank loans. Besides fundamentals, these 57movements in housing prices are related to fac-58tors specific to the transition towards a functioning 59market economy. Huynh-Olesen et al. (2013) found 60that the movements in residential property prices 61between 1999 and 2011 were driven by economic 62fundamentals such as disposable income and inter-63est rates, transition-specific factors like fast credit 64growth and funding through remittances, and by 65construction costs.66In the context of real estate market growth, 67an accurate valuation of residential properties be-68came more critical, taking account of some mar-69ket specific features, such as: heterogeneity of 70properties, highly segmented market, infrequent 71trading of properties, depreciation and renovation 72of housing over time, durability, and high trans-73action costs. An accurate estimate of the value of 74a given property avoids arbitrary measurements 75and is relevant for valuation companies, valua-76tors and public authorities. This is critical even 77International Journal of Strategic Property Management78ISSN 1648-715X / eISSN 1648-9179792016 Volume 20(2): 156–17180doi:10.3846/1648715X.2015.1121415818283===== PAGE 8 (score 60, numeric-paren 33, table-words 3, var-words 18) =====8485Marginal prices of improvements made to blocks of flats: empirical evidence from Romania8616387Table 1. Variables for sales price respectively for rental price models: descriptive statistics and ANOVA88Name89Variable90Sales price91Rental price92ANOVA93F (p-value)94%95Mean96price97ANOVA98F (p-value)99%100Mean101price102Price103Attributes104Thermal105Heating106Finishing107Finishing2108PhCond109PhCond2110Sales price of the flat (euro)111Rental price of the flat (euro)112Thermal insulation of the block:113Yes114No (reference category)115Heating system:116Individual boilers117Central heating (reference category)118Finishing work into flat:119Super finishing120Medium finishing121Demi-finishing (reference category)122Finishing work into flat:123Super finishing or medium finishing124Demi-finishing (reference category)125Overall physical condition:126Very good127Good128Unsatisfactory (reference category)129Overall physical condition:130Very good131Good and unsatisfactory (ref. cat.)1328.4 (0.01)13328.5 (0.00)13417.6 (0.00)13535.2 (0.00)13616.9 (0.00)13759.1%13850.9%13962.7%14037.3%14123.4%14256.3%14320.3%14479.7%14520.3%14640.2%14746.4%14813.5%1494059015042295151389481524292715336655154424541554220815633948157422811583394815944117160393191613444116231.6 (0.00)16322.7 (0.00)1648.7 (0.00)16518.7 (0.00)16636.1 (0.00)16731.2%16868.8%16945.5%17054.5%17126.5%17256.8%17316.7%17426.0%17562.0%17612.0%17726.1%17873.9%179220180256181204182245183214184241185220186187187261188204189218190260191206192City193Location194Rooms195Floor196Floor2197Area198Year199Garage200NoFloors201Type202Elevator203Balcony204Year_tr205The city:206Arad207Iasi208Timisoara209Cluj-Napoca (reference category)210Location in the city:211Central and intermediate district212Urban periphery district areas213Suburban areas (reference category)214Number of rooms:215Two rooms216One room (reference category)217Floor level:218Ground floor219Intermediate220Last floor (reference category)221Floor level:222Intermediate223Ground or last floor (ref. cat.)224Floor area (m2)225Year of construction:226fc227a2281990229<1990 (reference category)230Existence of a garage:231Yes232No (reference category)233Number of floors of the block:234>5 floors235cc236b2375 (reference category)238Type of flat:239Detached240Not-detached (reference category)241Existence of an elevator:242Yes243No (reference category)244Existence of balcony:245Yes246No (reference category)247Year of transactions24820112492012 (reference category)2508.9 (0.00)25111.7 (0.00)252137.5 (0.0)2532.2 (0.11)2543.4 (0.07)2550.9 (0.34)25620.6 (0.00)2572.5 (0.10)25820.1 (0.00)2592.6 (0.11)26012.2 (0.00)2610.9 (0.33)26215.5%26325.3%26431.7%26527.5%26644.5%26732.5%26823%26971.2%27028.8%27115.5%27267.3%27317.2%27467.3%27532.7%27636%27764%27810%27990%28031.9%28168.1%28270.6%28329.4%28436.9%28563.1%28665.4%28734.6%28853.6%28946.4%2903438329139596292425222934277829442904295407982963581329744433298310782994016230041337301380523024133730339054304413723054014430648515307397363084193430939961310422513113660231241822313398713144205631537851316411153173998331830.1 (0.00)31930.3 (0.00)32049.0 (0.00)3211.63 (0.20)32227.0 (0.00)323113.7 (0.00)32419.5 (0.00)3251.8 (0.18)32619.2 (0.00)32720.1 (0.00)3283.5 (0.07)3298.0%33025.6%33139.3%33221.1%33321.6%33433.1%33545.3%33667.5%33732.5%33816.5%33967.9%34015.6%34116.7%34283.3%3439.0%34491.0%34527.1%34672.9%34766.8%34833.2%34927.4%35076.2%35163.7%35236.3%35350.9%35449.1%355152356242357191358262359275360217361196362239363180364204365223366228367271368211369346370209371245372208373230374218375249376211377234378196379227380213381Note: The sample size N = 483 for sale transactions, respectively N = 468 for rent transactions.382383384===== PAGE 9 (score 29, numeric-paren 0, table-words 9, var-words 11) =====385386A. Deaconu et al.387164388category and also the results from the analysis of 389variance (ANOVA) method. For the entire sam-390ple, the average sale price is of 40590 Euros, with 391a standard deviation of 12834 Euros. The distri-392bution of prices has a positive skew and is plat-393ykurtic. Two-room flats (71.2%), with an average 394floor area of 52 m2, prevail; for one-room flats the 395average floor area is about 36 m2. Most transac-396tions seemed to occur in the three largest cities, 397respectively Cluj (27.5%), Iasi (25.3%) and Tim-398isoara (31.7%). A large part of the flats are located 399in blocks with less than 5 floors, built before 1990, 400and without elevators.401The attributes measuring the improvements 402made to the flat or building are encoded as Ther-403mal, Heating, Finishing, and physical conditions 404as PhCond. A large part of the blocks were built 405before 1990, but their average price does not seem 406to be significantly lower than that of newer flats, 407revealing potential improvements to the old blocks. 408About 80% of the flats have undergone interior fin-409ishing works, 23.4% being superior-finished and 41056.3% medium-finished. Examining the moderni-411sation features of the blocks, namely physical con-412dition and thermal insulation, it can be noted that 413a large number of the flats are in good or very good 414state in terms of overall physical condition; 59% 415of the blocks were included in thermal rehabilita-416tion programs. The values for the variable heating 417system confirm our preliminary descriptions con-418cerning the prevalence of flat owners who decided 419to quit the district heating in favour of individual 420boilers. We mention that these statistics are drawn 421from our sample, without being generalised to the 422entire market. The data indicates, however, that 423some efforts were made during the last two dec-424ades on the housing market to improve the physi-425cal state and comfort of blocks of flats.426As a preliminary analysis, we examined each 427attribute separately and computed the F-test 428from ANOVA, presenting the results in Table 4291. The null hypothesis that the average prices 430are equal across the groups is rejected for all 431improvement attributes, except for thermal in-432sulation, which is borderline rejected (p-value = 4330.011). Therefore the average price seems to vary 434across the categories, for the physical condition 435of the building, degree of finishing, and heating 436system. As regards the control attributes, the null 437hypothesis is rejected for the age of the building 438(before/after 1990) and year of transaction, and 439borderline rejected for the number of floors in the 440building, flat floor, existence of an elevator. The 441transaction dates cover the period from late 2011 442to 2012, during which the market prices remained 443relatively stable, therefore we expected this con-444clusion for the year of transaction. Upon exami-445nation of the F statistic, the city does not seem 446to hold a very high explanatory power for the 447price; as the descriptive statistics suggest, except 448for Arad where the average price is significantly 449lower, the average prices for the other three large 450cities are comparable to each other.451The entire sample (sample size is 483 transac-452tions) will be used to estimate the hedonic pricing 453models. Table 2 contains the variables with sig-454nificant effect on the sale price, for each of the five 455models, which were estimated using the maximum 456likelihood method. A backward selection approach 457has been applied. All attributes were included in 458the model, then the attribute with the highest p-459value was removed, and the new specification was 460estimated, and these steps were repeated until all 461coefficients had a p-value lower than 0.111. For 462the Finishing variable, the coefficient of dummy 463associated to medium finishing is not significant, 464therefore this variable has been redefined as a 465dummy with two categories Finishing2, as shown 466in Table 1. In a similar manner, the variable Floor 467was redefined as Floor2; there is no significant dif-468ference between the prices of flats located on the 469top floor compared to the ground floor. Also, due to 470few flats being sold in the central area, for some 471cities, the central district and intermediate district 472locations are compressed into a single category.473An overall goodness-of-fit is indicated by 474the likelihood ratio test. The Wald chi-square 475test is applied to test the statistical significance of 476each coefficient. The empirical results seem to be 477consistent with different specifications for GLM; 478generally the same explanatory variables were 479retained as having significant coefficients. Esti-480mates obtained for marginal effects in Gamma-log 481M2 and inverse Gaussian-log M3 are close to those 482provided by the traditional hedonic log-linear M1 483model. In the framework of maximum likelihood 484method, the coefficient4852486148704881489/490R491D492D493=494−495, where4961497D 498is the residual deviance for the estimated model 499and 5000501D the residual deviance for the model includ-502ing only the intercept, measures the proportion 503of the null deviance accounted for by the model 504(Fox 2008). These coefficients suggest a medium 505goodness-of-fit for our models; a similar value of 506this coefficient is obtained for traditional log-linear 50711 The empirical results for models including all attrib-508utes can be provided upon request.509510511===== PAGE 10 (score 181, numeric-paren 130, table-words 15, var-words 14) =====512513Marginal prices of improvements made to blocks of flats: empirical evidence from Romania514165515model12. To compare the (in-sample) performance 516of the three models, we also considered the differ-517ences between observed and predicted prices, and 518computed the mean absolute error MAE13. This 519measure suggests that Gamma-log model M2 pro-520vides a slightly better agreement between observed 521and predicted prices than M1 and M3. The Akaike 522information criterion (AIC)14, R-square and MAE 523usually suggest that the Gamma-log model performs 524better than the inverse Gaussian-log model, in terms 525of goodness-of-fit and prediction accuracy.526Examining the estimated coefficients from 527Table 2, where are indicated exp( )528β , the attrib-529utes have the expected signs. The percentage 530effect of an attribute over the price is given by 531[(exp( )5321] 100533β −534⋅535. As regards our primary goal, to 536measure the effect of the improvement attributes 53712 We mention that the traditional R-square coefficient, 538computed for the log-linear model estimated by OLS 539has the same values, respectively R2 = 0,54 for M1 540and R2 = 0.61 for M4, when the same regressors, as 541in Table 3, are included.54213 Willmott and Matsuura (2005) suggest that MAE is 543better suited than the Root Mean Squared Error to 544measure the model performance.54514 The AIC cannot be used to compare M2 and M3 mod-546els with the traditional log-linear model M1, as long 547as in the log-linear model the prices were transformed, 548being thus on different scales.549on pricing, the physical condition of the building 550and the finishing variables have a significant effect 551on the price in all specifications. As the log-linear 552model seems to perform slightly better, in terms of 553R-square coefficient, we look at the implicit mar-554ginal prices provided by M1. Thus, the prices of 555flats located in buildings that are classified as good 556and very good are higher by 7.8% and by 15.4% 557respectively than the prices of flats from buildings 558with unsatisfactory technical conditions. The fin-559ishing works inside the flat determines price in-560creases of 10%. These findings are in line with the 561empirical results from other studies. The market 562prices of flats and houses in which changes and 563renovations were undertaken appear to be higher 564(Portnov et al. 2005; Kryvobokov, Wilhelmsson 5652007; Sipos, Crivii 2008; Herath et al. 2015).566The marginal effect of the heating system is 567borderline significant. There is a willingness to 568pay a price approximately 6% higher for the exist-569ence of an individual boiler heating system. The 570thermal insulation dummy does not have a sig-571nificant coefficient, but this characteristic may be 572viewed as embedded into the physical condition of 573the building. If the physical condition of the build-574ing is not included in the model, the Thermal vari-575able has a significant effect on pricing, at 5% sig-576nificance level (Wald test = 4.15; p-value = 0.04). 577Table 2. Summary of hedonic models for sales price578Model579M1580M2581M3582M4583M5584Variable585Log-linear586Gamma-log587In.Gauss-log588Log-linear589Gamma-log590Finishing2591PhCond592Very good593Good594Heating595City596Arad597Iasi598Timisoara599Location600Central and interm.601Urban periphery district602Rooms603Area604Garage605NrFloors606Type607Balcony608Elevator609Floor2610Intercept6111.099*(9.4)6121.154* (12.9)6131.078* (4.4)6141.062** (4.6)6150.750* (58.0)6160.941 (3.4)6170.963 (1.3)6181.270* (56.9)6191.191* (27.8)6201.44* (218.4)621–6221.118* (8.1)6231.061** (4.9)6241.121* (20.1)6251.090* (12.2)626–627–62818688.2*629LR = 367.0*630AIC = 13.6631MAE = 6329.3632R2 = 0.546331.084* (7.4)6341.149* (13.3)6351.057 (2.9)6361.062 ** (4.9)6370.752* (58.9)6380.962 (1.4)6390.951 (2.2)6401.250* (52.4)6411.177* (24.8)6421.419* (213.5)643–6441.115* (8.1)6451.046 (3.06)6461.120* (21.1)6471.080* (10.3)648–649–65020216.7*651LR = 345.89*652AIC = 10193.9653MAE = 6319.5654R2 = 0.506551.096* (8.3)6561.183* (16.3)6571.081** (4.5)6581.067** (4.6)6590.736* (54.3)6600.959 (1.2)6610.941 (2.4)6621.260*(44.8)6631.199*(23.6)6641.424* (184.9)665–6661.114* (5.1)6671.058 (3.5)6681.120* (16.8)6691.081* (8.4)670–671–67219382.8*673LR = 285.8*674AIC = 10322.5675MAE = 6431.2676R2 = 0.456771.095* (10.5)6781.150* (15.7)6791.091* (7.0)680–6810.729* (74.4)6820.972 (0.9)6830.955 (2.4)6841.217* (45.7)6851.138* (17.6)686–6871.017* (375.1)688–689–6901.041** (2.8)6911.067* (8.1)6921.062** (6.1)6931.066* (8.1)69412381.2*695LR = 448.4*696AIC =–69.1697MAE = 5472.1698R2 = 0.616991.077* (8.4)7001.127* (41.7)7011.056* (15.4)702–7030.738* (93.6)7040.994 (0.3)7050.944 (3.0)7061.192* (41.9)7071.117* (14.5)708–7091.018* (438.6)710–711–712–7131.061* (7.3)7141.057**(6.2)7151.061*(8.1)71613017.0*717LR = 466.53718AIC = 10061.4719MAE = 5463.2720R2 = 0.62721Notes: For each model are reported Exp(β ) where β are the coefficients of explanatory variables, and the Wald test 722(in parentheses). *Significant at 1%, **significant at 5%. For inverse Gaussian-log model including the Area are obtained: 723LR = 398.4, AIC = 10205.9, MAE = 5936.2, R2 = 0.56.724725726===== PAGE 11 (score 35, numeric-paren 0, table-words 9, var-words 16) =====727728A. Deaconu et al.729166730The chi-square test supports this remark, a signifi-731cant degree of association being observed between 732the PhCond and Thermal variables (chi-square 733statistic = 168.0 and p-value = 0.00; the contin-734gency coefficient = 0.51).735On the other hand, as expected, the number 736of rooms has the highest explanatory power. The 737price of two-room flats is 44 % higher than the 738price of one-room flats. Location inside the city 739also has a significant effect on pricing. The flats 740from blocks located in the central district, inter-741mediate district and urban periphery district areas 742have a significantly higher price than those from 743suburban areas, by 19% and 27% respectively. If 744we compare the prices in the four cities, we find, 745as expected, significantly lower prices in Arad 746(by about 25%) than in Cluj-Napoca. For Iasi and 747Timisoara, the conclusion is not so clear, as there 748seems to be no significant differences in prices 749compared to Cluj-Napoca; in these two cities the 750price does not appear to be significantly different, 751as also suggested by the Student t-test for equality 752of two means (t = 0.18; p-value = 0.85). The coef-753ficients of corresponding dummy variables seem to 754be insignificant in some specifications; therefore, 755they must be interpreted with caution. Generally 756the coefficients for dummies for cities should be 757only interpreted as indicative, due to the small 758sample of transactions from each city. The other 759structural characteristics, namely the existence 760of a balcony and a garage, detached flats, have a 761positive effect on the price. Flats located in high 762blocks (more than 5 floors) seem to be more ex-763pensive than those from blocks with fewer floors. 764The four-floor blocks are generally old buildings; 765therefore, this attribute may incorporate partial 766information about the age of the building.767We conducted a similar empirical analysis for 768hedonic regressions, considering the floor area 769instead of the number of rooms. A backward se-770lection approach was applied for the same speci-771fications as above (S1-normal distribution with 772identity link for logarithm of prices, S2-Gamma 773distribution with log link, and S3-inverse Gauss-774ian distribution with log link). Table 2 reports the 775results for the Log-linear model denoted by M4 776and for the Gamma-log denoted by M5. For inverse 777Gaussian specifications, not reported here, we ob-778tained a lower value for R-square and a higher 779value for MAE, than for M4 and M5. In terms of 780fit and prediction accuracy the Gamma-log model 781(M5) performs slightly better than the log-linear 782model (M4). We can see that the floor area has 783a positive and high significant effect on pricing; 784one additional square meter raises the flat’s price 785by 1.8%, as long as the other attributes are the 786same. The physical condition of the building and 787finishing variable have a significant impact on the 788sale price, and the estimated coefficients are close 789to those obtained from models which include the 790number of rooms as an explanatory variable. If 791the physical condition of the building is removed 792from the models, the thermal insulation becomes 793a significant attribute for the sale price (Wald test 794= 9.07; p-value = 0.00), therefore the overall physi-795cal condition variable seems to capture the effect 796of thermal insulation. The other structural charac-797teristics, with positive effect on pricing are: exist-798ence of a balcony, floor number, detached flats, and 799existence of an elevator.800Therefore, the physical condition of building and 801the finishing works into the flat have a significant 802effect on sale prices for all distributions considered 803for sale price (log-normal, Gamma, inverse Gaus-804sian); our results are robust to the distribution of 805sale price. On the other hand, the multicollinearity 806is often an issue in regression models. The Variance 807Inflation Scores (VIF), a measure usually used for 808detecting the presence of multicollinearity, not ex-809ceeds 3.28 for any explanatory variables, suggesting 810that the multicollinearity is not a major problem. 811Although no definitive rules exist for the threshold 812of VIF, the general approach is that multicollinea-813rity is not a serious problem if the VIF not exceed 8145 (Troy, Grove 2008). A likelihood ratio (LR) test 815for omitted variables is also conducted. According to 816LR statistic, the physical state of the building (LR = 8179.36, p-value = 0.00 for model M1, and LR = 13.2, p-818value = 0.00 for model M4) and the finishing works 819(LR = 10.5, p-value = 0.01 for model M1, and LR = 82015.6, p-value = 0.00 for model M4) have a signifi-821cant contribution to explaining the variation in the 822sale price. Regarding the heteroscedasticity issue, 823in order to avoid a potential bias in the variance of 824estimated parameters, has been used the White’s 825heteroskedasticity consistent covariance matrix es-826timator (White 1980). The provided results were si-827milar to the reported ones; our variables of interest 828have significant coefficients.829Empirical results for rental price830For our sample of 468 rental transactions, the av-831erage monthly rental price is of 220 Euros, with a 832standard deviation of 90 Euros. Table 1 indicates the 833percentage of transactions and average rental price 834determined for each attribute and category, and also 835the results from the analysis of variance method. 836The distribution of rental prices has a positive skew 837838839===== PAGE 12 (score 148, numeric-paren 107, table-words 11, var-words 13) =====840841Marginal prices of improvements made to blocks of flats: empirical evidence from Romania842167843and is strongly platykurtic, therefore into the GLM 844framework we look at the distributions that take ac-845count of this behaviour. The rented flats from the 846three large cities, located mainly in the intermedi-847ate district and urban periphery district areas, and 848in blocks with less than 5 floors, are predominant.849The descriptive statistics computed for improve-850ment attributes, Thermal, Heating, Finishing, and 851overall physical conditions PhCond, indicate low 852percentages for unfinished flats, and for flats lo-853cated in buildings with inadequate technical state; 854the majority of flats are equipped with individual 855boilers. Examining each attribute separately, the 856F-test from ANOVA rejects the equality of average 857rental prices among the categories, for all improve-858ment attributes. The rental prices seem not to vary 859across categories for the following variables: year 860of transaction, type of the flat (detached/ non-de-861tached) and floor of the flat. The descriptive statis-862tics suggest, as expected, that the rental prices are 863significantly lower in Arad than in the other three 864large cities, and that they are significantly higher 865in the central area.866Following a careful analysis of the significance 867of coefficients for dummy variables, in the models 868including all attributes, the variable PhCond was 869redefined, as indicated in Table 1. Also, the urban 870periphery district and suburban areas are consid-871ered a single category. The backward selection ap-872proach provides the empirical results reported in 873Table 3; almost the same variables were selected 874regardless of specification, as the attributes with 875significant effect on rental prices. The set of signifi-876cant attributes, in addition to the physical state of 877the building and finishing, includes: the city, loca-878tion inside the city, number of rooms/ area, exist-879ence of a garage, existence of a balcony, generally 880included also in the hedonic models for sale prices.881We first examine the M1, M2, and M3 models, 882considering the number of rooms. According to the 883AIC information criteria the Gamma-log model 884performs better than the inverse Gaussian-log 885model. The in-sample performance of these models 886is almost the same, as suggested by the MAE, and 887by the R-square coefficient measuring the propor-888tion of the null deviance accounted for by the mod-889el. The physical state of the building and finishing 890variables are significant in all specifications, but 891have lower effect than on sale prices.892According to the log-linear model, for a very 893good physical condition of the building there is a 894willingness to pay a 10.5% higher price, as long as 895the other attributes are the same. As in the case 896of the sale price, this attribute seems to incorpo-897rate the effect of thermal insulation; when it is re-898moved from the model, the thermal insulation has 899a significant effect on rental prices (Wald test = 9006.55; p-value = 0.01). The finishing level of the flat 901has a significant but not strong effect on rental 902prices, as long as this variable remains borderline 903significant. A superior level of finishing in the flat 904has a marginal effect on rental prices, triggering 905Table 3. Summary of hedonic models for rental price906Model907M1908M2909M3910M4911M5912Variable913Log-linear914Gamma-log915Inv.Gauss.-log916Log-linear917Gamma-log918Finishing919Superior920Medium921PhCond2922City923Arad924Iasi925Timisoara926Location927Central928Intermediate929Rooms930Area931Garage932Balcony933Intercept9341.106** (5.7)9351.0669361.103* (9.4)9370.628* (83.7)9381.052 (1.7)9390.879* (13.7)9401.263* (45.6)9411.064** (4.6)9421.298* (94.9)943–9441.290* (30.5)9451.080* (8.8)946147.2*947LR = 320.7948AIC = 95.3949MAE = 42.85950R2 = 0.509511.096**(4.8)9521.065 (3.3)9531.131* (15.7)9540.660* (69.3)9551.082** (4.8)9560.893* (10.2)9571.266* (45.1)9581.051* (3.7)9591.302*(100.5)960–9611.300* (33.1)9621.077* (8.1)963147.2*964LR = 349.1965AIC = 5072.9966MAE = 43.25967R2 = 0.519681.100* (5.5)9691.059 (2.8)9701.140* (16.7)9710.634* (98.7)9721.080 (3.6)9730.905* (7.3)9741.245* (36.8)9751.041 (1.9)9761.299* (100.3)977–9781.274* (18.2)9791.074* (7.4)980147.6*981LR = 324.1982AIC = 5086.1983MAE = 43.38984R2 = 0.499851.118* (7.9)9861.066** (3.9)9871.084** (5.6)9880.575* (129.5)9891.065 (3.0)9900.904* (9.6)9911.199* (27.8)9921.061** (4.7)993–9941.010* (159.6)9951.182* (14.1)9961.067* (6.9)997109.7*998LR = 371.7999AIC = 44.61000MAE = 38.561001R2 = 0.5510021.114* (7.6)10031.060* (3.2)10041.090* (8.2)10050.612* (104.5)10061.100* (7.2)10070.909* (8.7)10081.192* (29.8)10091.061* (4.8)1010–10111.010* (149.8)10121.183* (14.8)10131.072* (8.2)1014110.9*1015LR = 398.61016AIC = 5021.51017MAE = 38.741018R2 = 0.561019Notes: For each model are reported Exp(β) where β are the coefficients of explanatory variables, and the Wald test (in 1020parentheses). *Significant at 1%, **significant at 5%. For inverse Gaussian-log model including the Area, are obtained: 1021LR = 351.0, AIC = 5057.3, MAE = 39.6.2, R2 = 0.53.102210231024===== PAGE 13 (score 28, numeric-paren 0, table-words 6, var-words 15) =====10251026A. Deaconu et al.10271681028a price surge of about 10%. Bala et al. (2014) also 1029conclude that renovated dwellings carry higher 1030monthly rents. Owners are motivated to improve 1031the physical characteristics of properties, in order 1032to rent at higher prices. Tenants are looking for 1033finished flats, to improve their living standards, 1034but are also interested in some improvements that 1035could reduce the cost of utilities.1036The number of rooms, existence of a balcony, 1037garage, and location have a significant effect on 1038rental prices. The price of two-room flats is higher 1039roughly by 30% than the price of one-room flats. 1040Flats from blocks located in the central area carry 1041a significantly higher rent than those from the ur-1042ban periphery district and suburban areas, rough-1043ly by 26%. The lowest rental rates are in Arad, 1044and the highest in Iasi and Cluj-Napoca; in these 1045two cities the rent does not appear to be signifi-1046cantly different, as also suggested by the Student 1047t-test for equality of two means (t = 1.37; p-value 1048= 0.17). These two cities are important university 1049centres, thus a large part of the demand for rental 1050flats comes from students. Due to the small num-1051ber of rent transactions in each city these statistics 1052should be viewed as indicative.1053When the floor area is considered instead of 1054the number of rooms, usually the MAE and R-1055square indicators show better fit; the continuous 1056variables are richer in information than categori-1057cal variables. The empirical results from the log-1058linear model and Gamma-log model are indicated 1059in Table 3 (model M4 and M5). As the estimated 1060hedonic regressions suggest, one additional square 1061meter of floor area triggers a rental price increase 1062of 1%. The overall physical state of the building 1063and finishing variables still remain significant, 1064at 5% significance level, but the effect of the first 1065variable is slightly lower than in previous specifi-1066cations.1067Overall, both variables, the physical state of the 1068building and finishing works, remain significant 1069regardless the distribution of rental price. The 1070multicollinearity is not an issue as long as no VIF 1071exceeded 3.03. The LR test for omitted variables 1072indicates that the physical state of building (LR = 10739.30, p-value = 0.00 for M1, LR = 5.56, p-value = 10740.02 for M4), respectively the finishing works (LR 1075= 5.73, p-value = 0.00 for model M1, and LR = 7.85, 1076p-value = 0.00 for model M4) are variables that 1077significantly influence the rental price. Using the 1078White’ heteroskedasticity robust standard errors, 1079similar results are provided; the variables of inter-1080est are still significant.10815. Conclusions1082This paper examines a specific segment of the Ro-1083manian housing market, namely blocks of flats in 1084the urban area. Similarly to other markets from 1085Central and Eastern Europe, the Romanian real 1086estate market inherits some stock of low-quality 1087blocks of flats, with a small living surface and low 1088comfort, undergoing a process of deterioration. We 1089highlight the steps taken by owners and authori-1090ties to improve the housing maintenance, and the 1091living standards in these dwellings. This analysis 1092may contribute to a better understanding of the 1093housing market from Romania and from the CEE 1094countries.1095The empirical section investigates whether and 1096to what extent the improvements made to the stock 1097of flats have a significant effect on sale and rental 1098prices. As described in Section 2, the total cost of 1099building modernization and finishing works inside 1100the flats, is about 179 Euro/m2, without VAT, as 1101an average for flats with one or two rooms. On the 1102other hand, in our sample, the average sale price 1103is 906 Euro/m2 for flats with very good physical 1104conditions of the building and super-finishing level 1105inside the flats, respectively 717 Euro/m2 for flats 1106with unsatisfactory physical state of the building 1107and demi-finishing level inside the flats; the dif-1108ference between these two average prices is 189 1109Euro/m2. Overall, the marginal price and the cost 1110of modernization can be considered as being com-1111parable, taking account of the difficulty of accu-1112rately estimating the cost and that average values 1113are involved. 1114Two hedonic analyses were conducted, one for 1115sale transactions and another one for rental trans-1116actions, based on samples of flats from blocks lo-1117cated in four large towns from Romania. Hedonic 1118models are estimated in a GLM framework, which 1119considers the following specifications: the log-1120linear model, Gamma distribution with log link, 1121and inverse Gaussian distribution with log link. 1122Generally, for our data, the empirical results re-1123garding the marginal prices of the improvements 1124attributes are consistent with these specifications. 1125The traditional log-linear and Gamma-log models 1126provided close results in terms of fit and prediction 1127accuracy. The inverse Gaussian-log models lead to 1128a weaker agreement between the actual and pre-1129dicted prices. The robustness of empirical results 1130was also investigated from other points of view, 1131namely multicollinearity, omitted variables and 1132heteroscedasticity.1133