Antonello Eugenio Scorcu
University of Bologna

 and The Rimini Centre for Economic Analysis, Italy

Roberto Zanola
University of Eastern Piedmont

 and The Rimini Centre for Economic Analysis, Italy

THE ‘RIGHT’ PRICE FOR ART COLLECTIBLES.
A QUANTILE HEDONIC REGRESSION 

INVESTIGATION OF PICASSO PAINTINGS

Copyright belongs to the author. Small sections of the text, not exceeding three paragraphs, can be used 
provided proper acknowledgement is given. 

The  Rimini  Centre  for  Economic  Analysis  (RCEA)  was  established  in  March  2007.  RCEA is  a  private, 
nonprofit organization dedicated to independent research in Applied and Theoretical Economics and related 
fields.  RCEA  organizes  seminars  and  workshops,  sponsors  a  general  interest  journal  The  Review  of 
Economic Analysis, and organizes a biennial conference: The Rimini Conference in Economics and Finance  
(RCEF) . The RCEA has a Canadian branch: The Rimini Centre for Economic Analysis in Canada (RCEA-
Canada). Scientific work contributed by the RCEA Scholars is published in the RCEA Working Papers and 
Professional Report series.

The views expressed in this paper are those of the authors. No responsibility for them should be attributed to 
the Rimini Centre for Economic Analysis.

The Rimini Centre for Economic Analysis 
Legal address: Via Angherà, 22 – Head office: Via Patara, 3 - 47900 Rimini (RN) – Italy

www.rcfea.org -  secretary@rcfea.org

WP 01_10

mailto:secretary@rcfea.org


 1 

 
 
 

The ‘right’ price for art collectibles.  
A quantile hedonic regression investigation of Picasso 

paintings* 
 

A.E. Scorcu+ and R. Zanola° 
 
 
 
 
 
 
 
 

ABSTRACT 
 

Different art objects are likely to be priced by means of different systems of hedonic 
characteristics; more precisely, different evaluation procedures for high and low price items are 
often postulated. However, the empirical evidence on this point is scant. The main purpose of this 
paper is to fill this gap by using the quantile hedonic regression approach. The empirical evidence, 
based on a data set of 716 Picasso paintings sold at auction worldwide, highlights the critical role 
of the price classes in determining the evaluation criteria of art items. 
 
 
 
 
 
 
 
 
 
 
JEL Classification: D49 
Key Words: hedonic price; auction; quantile regression; painting; Picasso. 

                                                 
* A first version of this paper was completed while the second author was visiting at University of York. Thanks 
are due to Andrew Jones, Guido Candela, Roberto Cellini and Simone Giannerini. The usual disclaimers apply. 
+ University of Bologna, Department of Economics, Italy and Rimini Center for Economic Analysis (RCEA), Italy, 
antonello.scorcu@unibo.it 
° University of Eastern Piedmont, Department of Public Policy and Public Choice, Italy and Rimini Center for 
Economic Analysis (RCEA), Italy, roberto.zanola@unipmn.it 

 



 2 

1. Introduction 

The price determinants of collectibles has often raised the interest of scholars of 

the field. Surveying the relevant empirical literature, Ginsburgh et al. (2006) 

noted that analyses often use the hedonic regression approach, with the price of an 

art item explained by a number of hedonic characteristics (author, genre, 

technique, dimension, etc.), market variables (auction house, city of sale, 

provenance of the object, etc.) and time dummies. The sign and the size of these 

price determinants emerge from an hedonic regression, and the corresponding 

market price index can be obtained from the estimated time dummy coefficients.  

Empirical investigations often analyze the specification of the regression, or the 

stability of the estimated coefficients over time, as the collectors’ preferences and 

the evaluation of the hedonic characteristics might evolve, exogenously or in 

relation to market booms and slumps. In any case, an implicit (albeit restrictive) 

assumption is that in given market and period, a single evaluation system is 

shared by low- and high-price art items. 

Although there is an extensive literature on the hedonic approach [e.g. Candela 

and Scorcu, 1997; Locatelli-Biey and Zanola, 2005; Zanola, 2007; Collins et al., 

2007, 2009], it seems that one basic point has been neglected, namely, the 

possible existence of a segmentation in the art market with respect to the market 

value of the items. One could hardly be surprised to learn that a different criterion 

is used in the appraisal of a masterpiece valued several million dollars and another 

one in the evaluation of a painting whose price might be a few hundred dollars. In 

fact art item characteristics evaluations can change across the sectional 

distribution of art prices, because those who bid for expensive items are likely to 

differ from those who buy relatively inexpensive items [Malpezzi, 2003]. 

Moreover, even the same collector might appreciate differently the characteristics 

in low- and high price items. Finally, paintings that reach an extraordinary 

evaluation within a group, the top lots, seem to behave differently from other 

items [Pesando, 1993; Mei and Moses, 2002]. 

The purpose of this paper is to analyze the existence of different hedonic models 

for cheap and expensive paintings. To this aim, an hedonic quantile regression 

framework is used, which allows the impact of art item characteristics to differ 

across price distribution. More precisely, by using a dataset of 716 Picasso 



 3 

paintings sold at auctions worldwide during the period 1988-2005, we address 

two questions: 

• Is the assumption of homogeneous effects of covariates on prices, implicitly 

determined by the estimation of average effects, justified; or do effects differ 

at different quantiles of the price distribution? 

• How do time-invariant and time-variant characteristics affect the returns from 

paintings? 

 

The rest of the paper is organized as follows. Section 2 defines the model to be 

used. Data and functional form are presented in Section 3. The empirical evidence 

is discussed in Section 4. Section 5 concludes. 

 

2. Theoretical framework 

The hedonic OLS regression is commonly used in the analysis of the art market to 

determine the relationship between a set of characteristics of collectibles and their 

corresponding (hammer) prices. Such an approach relies upon the mean of 

conditional distribution of the dependent variable. However, to the extent that 

characteristics are expected to be valued differently across a given distribution of 

selling prices, the exogenous variables influence the parameters of the conditional 

distribution of the dependent variable differently. Neglecting this possibility 

might undermine the reliability of the results [Koenker and Bassett, 1978; Zietz et 

al., 2007].  

Unlike OLS, quantile regression models allow for a full characterization of the 

conditional distribution of the dependent variable. The standard OLS hedonic 

regression minimizes the sum of the squared residuals: 

 

{ }
,min

2

0
,

0

∑ ∑ 








−

== i

k

j
ijji xyk

jj

β
β

 (1) 

 

where yi is the dependent variable at observation i; xj,i is the j regressor variable at 

observation i; and βj the parameter of the implicit price of the j characteristic. By 

contrast, quantile model involves instead the minimization of a weighted sum of 

the absolute deviations in a median-regression context: 



 4 

 

{ }
,min

0
,

0

∑ ∑
=

−
= i

i

k

j
ijji hxyk

jj

β
β

 (2) 

 

where the weight hi is defined as hi = 2q if the residual for the ith observation is 

strictly positive or as hi = 2-2q if the residual for the ith observation is negative or 

zero. The variable q (0<q<1) is the quantile to be estimated. 

This flexible estimation method might be particularly useful in our framework,  in 

which the price distribution of the art items shows a neat departure from 

normality. Figure 1 displays the empirical distribution function (EDF) for the 

hammer prices. The inverted ‘L’ shape of the EDF shows that many items are 

concentrated in the left-hand tail and that there is a long right-hand tail made up 

of relatively few items that have high selling prices. In these cases, the quantile 

regression procedure is less sensitive to outliers and provides a more robust 

estimator [Koenker and Hallock, 2001; Bassett et al. 2002]. Quantile regression 

models may also have better properties than OLS models in the presence of 

heteroschedasticity [Deaton, 1997]. 

 

 [Figure 1 about here] 

 

3. Data and functional form 

The dataset considers 716 Picasso paintings sold at auction worldwide during the 

period 1988-2005. The data set is collected from the 2006 edition of the Art Price 

Index on CD-Rom. The use of a relatively homogeneous data set based on a 

single, well known painter, should reduce the effect of the several sources of 

artistic and market variability, except for the dimension we are interested in, the 

auction price distribution. 

The publication contains records of paintings sold at the world’s major auctions, 

and provides information on a number of variables: artist’s name, nationality, title 

of the work, year of production, materials used, date and city of sale, auction 

price, pre-sale estimate (when available), dimensions, signature, and a number of 

further information that might vary from case to case. In the analysis we 

complement the data set with a series of indicators about the artistic styles of the 



 5 

painting. No information is provided on the provenance and the previous 

exhibitions of the items. Prices are gross of the buyers and sellers’ transaction fees 

paid to auction houses and recorded in both local currencies and USD; all prices 

are however expressed in USD. In order to provide a sensible distribution of the 

“real” prices of Picasso paintings to be used in the regressions, nominal USD 

prices are deflated using US CPI prices (2000=100). All prices are then logged, in 

line with most of the relevant literature, in order to deal with heteroschedasticity. 

 

[Table 1 about here] 

 

The explanatory variables included in the study are dimension, style, media, 

salesroom and year of sale. The effects of the price distribution on the variables 

are shown in Table 1. We compute the average values for the (log of) prices as 

well for the independent variables, both for the full sample and the quantiles .20, 

.40. .60, .80, .95. In column (1) we show the means of the variables for the whole 

sample; in columns (2)-(6) we show the corresponding averages for the price 

classes considered. Some ‘stylized facts’ emerges neatly. The dimension is, on 

average, lower for the low price items (neglecting style and techniques). The Old 

Picasso Style is often associated with low prices, whereas the opposite holds for 

the Young Picasso Style. More precisely, the variables included in the regression 

are: 

 

• Dimension: the surface of the painting, area, and the squared surface, area2, 

in squared meters.  

• Style: different style periods are identified [Czujack, 1997]; Childhood and 

Youth (1881-1901), style1; Blue and Rose Period (1902-1906), style2; 

Analytical and Synthetic Cubism (1907-1915), style3; Camera and Classicism 

(1916-1924), style4; Juggler of the Form (1925-1936), style5; Guernica and 

the ‘Style Picasso’ (1937-1943), style6; Politics and Art (1944-1953), style7; 

and The Old Picasso (1954-1973), style8 (excluded variable). As shown in 

Table 1, some stylistic period are much more frequent than others. The Blue 

and Rose period represents less than 2 per cent of the observations, whereas 

the most common, the Old Picasso Style (35 per cent of the total sales), is less 

frequent in the high price quantiles. Different styles have different weights, 



 6 

and are unevenly distributed across price classes. In fact, quantile distributions 

are significantly different from the average full sample distribution1.  

• Media: a set of dummy variables, reflecting the technique adopted, is used: 

canvas, oil on canvas, the 67 per cent of the items auctioned; panel, oil on 

panel;  mixed, mixed technique; and medother, all other media (excluded 

variable).  

• Salerooms:  Sotheby’s and Christie’s are known to be the leading auction 

houses in this kind of transaction while the most important art auction markets 

are in New York and London. We consider therefore some interaction 

dummies between salerooms and cities: chriny, for Christie’s New York; 

chrilon, for Christie’s London; sothny, for Sotheby’s New York; sothlon, for 

Sotheby’s London; othauc for all other locations (excluded variable). The 

most important market for the Picasso paintings is New York, followed by 

London, with few differences between Sotheby’s and Christie’s. Overall, the 

items auctioned by Sotheby’s and Christie’s represents the majority of the 

sales collected in the data set, but no obvious pattern in terms of prices classes 

emerges. 

• Year of sale: a set of yearly dummy variables, dt, with t = 1988,.., 2005, are 

introduced for each year between 1988 and 2005 (1988 baseline variable). 

Sales are spread quite evenly across the years under scrutiny, with notable 

exceptions, the years 1988 and 1991, with less than 2 percent of the total 

sales, and the year 1998, in which occurred the largest share of sales, 12.71 

percent, mostly concentrated in the .80 quantile. The quintile distributions, 

however, are not significantly different with respect to the full sample 

distribution.  

 

Formally, the following regression function form is set: 

 

it
i

ii

it

dsothlon

sothnychrilonchrinymixedpanelcanvasstyle

stylestylestylestylestylestyleareaareap

εββ

βββββββ
βββββββββ

++

++++++++
++++++++=

∑
=

33

17
16

1514131211109

876543210

7

6543212

        (3) 
                                                 

1 A Shapiro-Wilks test is used to test for the normality of the differences between the average and the quantile 
yearly sales distributions. The results, not shown for the sake of brevity, are available upon request. 



 7 

where pit is the log-price of painting i sold at time t; and ɛit is the error term.  

 

4.  Results 

In the following, firstly the whole sample standard hedonic price model is 

estimated to set a benchmark. Secondly, the quantile regressions are estimated in 

order to analyze the hedonic characteristics across the price distribution. The 

results are displayed in Table 2. 

 

[Table 2 about here] 

 

Column 1 shows the results for the OLS whole sample specification whereas the 

quantile regression results are shown in columns 2-6. Following Berndt et al. 

(1995), standard errors and variance-covariance matrices of the coefficients have 

been computed by using the White heteroschedasticity-robust procedure.  

In column 1, dimension has a positive effect on prices, but at decreasing rate, as 

the squared area shows a negative sign, in line with several previous studies. All 

styles perform better than The Old Picasso style (the baseline variable): the Blue 

and Rose Period is the most appreciated by the collectors, followed by the Cubist 

period and the young Picasso. Oil on canvas displays a positive coefficient, 

whereas oil on panel is not statistically significant and mixed techniques shows a 

negative coefficient sign, compared to the other media class (the excluded 

variable). Finally, in line with previous studies, even controlling for other hedonic 

characteristics, Christies’ and Sotheby’s auction houses, located both in London 

and New York, sell at a premium with respect to the other auction houses (the 

excluded variable).  

The evidence based on the whole sample shows in other words the expected 

results of a typical hedonic regression. As shown in columns 2-6 of  Table 2, most 

of the coefficients are significantly different from zero at 1 per cent probability 

across quantiles; however, the R-squared of the overall sample regression is larger 

than the corresponding quantile regressions, because of the greater variability. 

As for the quantile regression coefficients, a casual inspection might suggest a 

pattern roughly in line with the whole sample result, with the possible emergence 

of a stable ranking for artistic values. Some differences in the absolute size of 



 8 

coefficients across quintiles emerges for the style variable, whereas in relative 

terms The Blue and Rose Period is always the most appreciated, followed by 

Cubism, Young Picasso and the Juggler of the Form periods. Camera and 

Classicism, Guernica and Politics and Art (even if have a positive estimates and 

therefore perform better than the baseline Old Picasso style) are less appreciated. 

Like in the whole sample regression, the (log of) the price increases with the size 

of the painting, but at decreasing rate, with the size of the linear and (more neatly) 

squared term decreasing for the larger quantiles.  

Differences among quintiles are more evident for the salesrooms, and the media 

coefficients: some variable, significant in the lowest quantile regressions, lose 

their significance in the higher quantile regressions. Also the yearly dummies 

follow idiosyncratic patterns.  

However, a more formal evaluation requires to test for the equality of the 

coefficients across quantile regressions. In order to avoid problems arising in the 

presence of certain departures when solving minimization by linear programming 

[Deaton, 1997], we compute robust standard error estimates by bootstrap, with 

1,000 replications. F-tests for each groups of variables on the lowest quintile 

versus the other quintile models is run and reported in Table 3. The null 

hypothesis of equal coefficients across quintiles cannot be rejected for the 

dimension, style and salerooms variables, whereas is rejected at the usual 

significance level for the media and the trend period. For the media variables this 

result might be in part due to the correlation between the media and the quintile: 

in the low quintiles mixed techniques or other techniques are relatively important, 

whereas in the high quintiles canvases prevail. The most striking outcome 

concerns the yearly dummies: low and high quintiles trends are clearly different, a 

result already known as the masterpiece effect [Mei and Moses, 2002]. 

 

[Table 3 about here] 

 

With different values associated with the mean (OLS) regression and the quantile 

regression estimates, it is natural to evaluate the emergence of significant 

differences in the price dynamics of Picasso paintings. To this aim, the yearly 

dummy coefficients are used to calculate quantile-specific adjacent year 

regressions price indexes. These indexes measure the per cent change in price 



 9 

associated with the change from 0 to 1 of the dummy variables and they are given 

by [100∗(eβj-1)], where βj is the quintile-estimated coefficient. Table 4 and Figure 

2 show both the whole sample and the quintile price index values.  

 

[Table 4 about here] 

 

[Figure 2 about here] 

 

Figure 2 shows how the different specific price classes submarkets share several 

common short run adjustments, making difficult the recognition of the differences 

in the quantile trends. The boom of the late 80s is apparent for every series, as 

well the subsequent burst and the weak recovery that occurred over the period 

1993-2003. Again the boom period of the 2003-05 emerges neatly from the time 

indices.  

The imperfect synchronization of (price) cycles among price classes might affect 

the previous conclusions, as the use of a common base year influences the 

subsequent dynamics of the indexes. In order to develop a more balanced 

approach, in Table 5 we show the various 5-year returns computed over the 

periods 1988-1992, 1989-1993,…, computed using the Table 4 time dummies. 

 
[Table 5 about here] 

[Figure 3 about here]  

 

Within a 5-year horizon, the burst of the bubble of the late Eighties lead to a 

period of significant losses for quantiles in the first years. However, with a careful 

(or lucky) selection of the investment period, remarkable rates of return can be 

obtained, as in the period 1993-1997. The same outcome emerges for specific 

quintiles, as in the period 1991-1995 for the highest quintiles. 

On average, the return is higher for the high end of the market, but also more 

volatile, as shown in Figure 3.  

The evidence in support of specific price characteristics for different quantiles is 

therefore strengthened. 

 



 10 

 

5. Conclusions 

Although the existence of different values for painting characteristics across 

different hammer price distributions may have been considered intuitive 

beforehand, to our knowledge, such a issue has never been empirically 

investigated. By contrast, the homogeneity of the systems of evaluation with 

respect to price levels has been the standard approach in the applied literature. 

Quantile regression provides an intuitive way to evaluate the relevance of the 

price effect in the appreciation of the paintings. In this paper we use the quantile 

regression approach for analyzing the structure of the hedonic characteristics of 

716 Picasso paintings sold at auctions worldwide during the period 1987-2005. 

We estimate a log-linear hedonic model both for the full sample and the .20, .40, 

.60, .80, and .95 quantiles of hammer price distribution. The empirical evidence 

suggests the existence of significant differences both in the way prices respond to 

characteristics (particularly in the media variables) and in the rates of return from 

an investment in Picasso paintings across different price ranges (the so called 

masterpiece effect). 

     



 11 

References 

 
Bassett, G.W.Jr., Tam, M.S., Knight, K. (2002), Quantile models and 
estimators for data analysis, Metrika, 17-26. 
 
Berndt E. R., Rappaport N. J. (2001), Price and quality of desktop and 
mobile personal computers: A quarter-century historical overview, 
American Economic Review, 91(2), 268-73. 
 
Candela G., Scorcu, A.E. (1997), A Price Index for Art Market Auctions. 
An application to the Italian Market of Modern and Contemporary Oil 
Paintings, Journal of Cultural Economics, 21, 175-196. 
 
Collins, A., Scorcu, A.E., Zanola, R. (2007), Sample Selection Bias and 
Time Instability of Hedonic Art Price Indexes, Working Papers 610, 
Dipartimento Scienze Economiche, Università di Bologna. 
 
Collins, A., Scorcu, A.E., Zanola, R. (2009), Reconsidering hedonic art 
price indexes, Economic Letters, 104, 57-60. 
 
Czujack, C. (1997), Picasso paintings at auction, 1963-1994, Journal of 
Cultural Economics, 21, 229-247. 
 
Deaton, A. (1997), The Analysis of Household Survey, Baltimore, MD, 
Johns Hopkins University Press. 
 
Ginsburgh V.A., Mei, J., Moses, M. (2006), The computation of price 
indices, in Ginsburgh, V.A., Throsby, D. (eds.), Handbook of the Art and 
Culture, North-Holland. 
 
Locatelli Biey, M., Zanola, R. (2005), The Market for Picasso Prints: A 
Hybrid Model Approach, Journal of Cultural Economics, 29(2), 127-136. 
 
Koenker. R.W., Bassett, G.W. (1978), Regression Quantiles, Econometrica, 
46(1), 33-50. 
 
Koenker, R.W., Hallock, K. (2001), Quantile Regression: An Introduction, 
Journal of Economic Perspectives, 15(4), 143-156. 
 
Malpezzi, S. (2003), Hedonic pricing models: A selective and applied 
review, In Sullivan, T.O., Gibbs, K. (eds.), Housing economics and public 
policy: Essay in honour of Duncan Maclennan, Oxford, Blackwell. 
 
Mei, M,, Moses, M. (2002), Art as an Investment and the Underperformance 
of Masterpieces, American Economic Review, 92, 1656-1655 
 
Pesando, J.E. (1993), Art as Investment: The Market for Modern Prints, 
American Economic Review, 83, 1075-89. 
 



 12 

Zanola, R. (2007), The Dynamics of Art Prices: The Selection Corrected 
Repeat-Sales Index, Working Papers 76, Department of Public Policy and 
Public Choice, University of Eastern Piedmont. 
 
Zietz, J., Zietz, E.N., Sirmans, G.S. (2007), Determinants of House Prices: 
A Quantile Regression Approach, Department of Economics and Finance 
Working Paper Series, Middle Tennessee State University. 



 13 

 
 
 
      Figure 1. Empirical distribution function (EDF) of selling prices 



 15 

Table 1. Summary statistics 
 Percentile 
 

Full sample 
 

(1) 
.20 
(2) 

.40 
(3) 

.60 
(4) 

.80 
(5) 

.95 
(6) 

 Mean Std. Dev. Mean Std. Dev. Mean Std. Dev. Mean Std. Dev. Mean Std. Dev. Mean Std. Dev. 
price 2,122,799 5,984,123 118,987 83,834.55 350,470 85,597.45 655,989.4 162,247.2 1,512,689 519,238.3 3,866,488 1,362,944 
area .5171 .5275 .1586 .1757 .3285 .3117 .5032 .4270 .6568 .5577 .8371 .6087 
area2 .5452 1.1117 .0558 .1269 .2044 .4772 .4342 .8275 .7402 1.2360 1.0678 1.4972 
style1 .0564 .2309 .0611 .2404 .0282 .1660 .0652 .2478 .0634 .2445 .0588 .2364 
style2 .0184 .1344 0 0 .0211 .1443 .0072 .0851 .0070 .0839 .0196 .1393 
style3 .0589 .2356 .0534 .2258 .0563 .2314 .0580 .2345 .0493 .2172 .0980 .2989 
style4 .1055 .3074 .1221 .3287 .1268 .3339 .0725 .2602 .1056 .3084 .0392 .1951 
style5 .1043 .3058 .0382 .1923 .0775 .2683 .0870 .2830 .1056 .3084 .2157 .4133 
style6 .1595 .3664 .2061 .4061 .0775 .2683 .1014 .3030 .1690 .3761 .2255 .4200 
style7 .1472 .3546 .2137 .4115 .1479 .3562 .1159 .3213 .1479 .3562 .1372 .3458 
style8 .3497 .4772 .3053 .4623 .4648 .5005 .4927 .5018 .3521 .4793 .2059 .4063 
canvas .7251 .4467 .4685 .5008 .7343 .4433 .7552 .4314 .8042 .3982 .8518 .3569 
panel .0580 .2340 .0699 .2559 .0559 .2306 .0909 .2885 .0629 .2437 .0370 .1897 
mixed .0557 .2294 .1608 .3687 .0350 .1843 .0070 .0836 .0210 .1438 .0370 .1897 
other_tech .1896 .3922 .4056 .4927 .1748 .3811 .1468 .3552 .1119 .3163 .0833 .2777 
chrilon .1505 .3578 .1608 .3687 .1259 .3329 .1888 .3927 .1888 .3927 .1204 .3269 
chriny .2666 .4424 .2517 .4355 .2657 .4433 .2587 .4395 .2797 .4504 .2222 .4177 
sothlon .1458 .3530 .1259 .3329 .1888 .3927 .1538 .3621 .1468 .3552 .1481 .3569 
sothny .2868 .4525 .1888 .3927 .2937 .4571 .2517 .4355 .2867 .4538 .4259 .4968 
othauc .1256 .3316 .2587 .4395 .1259 .3329 .1468 .3552 .0979 .2982 .0833 .2777 

 



 16 

Table 1. Summary statistics (contd.) 
 Percentile 
 

Full sample 
 

(1) 
.20 
(2) 

.40 
(3) 

.60 
(4) 

.80 
(5) 

.95 
(6) 

 Mean Std. Dev. Mean Std. Dev. Mean Std. Dev. Mean Std. Dev. Mean Std. Dev. Mean Std. Dev. 
d88 .0140 .1174 0 0 .0140 .1178 .0070 .0836 .0140 .1178 .0278 .1651 
d89 .0475 .2128 .0070 .0836 .0140 .1178 .0350 .0836 .0769 .2674 .0741 .2631 
d90 .0670 .2503 .0140 .1178 .0350 .1843 .0559 .2306 .0909 .2885 .1759 .3825 
d91 .0182 .1336 .0140 .1178 .0140 .1178 .0420 .2012 .0210 .1438 0 0 
d92 .0335 .1801 .0489 .2165 .0559 .2306 .0210 .1438 .0280 .1655 .0185 .1354 
d93 .0517 .2215 .0559 .2306 .1119 .3163 .0489 .2165 .0350 .1843 .0092 .0962 
d94 .0461 .2098 .0699 .2559 .0559 .2306 .0699 .2559 .0070 .0836 .0370 .1897 
d95 .0642 .2454 .0769 .2674 .0420 .2012 .0839 .2782 .0559 .2306 .0556 .2301 
d96 .0517 .2215 .0769 .2674 .0839 .2782 .0420 .2012 .0350 .1843 .0185 .1354 
d97 .0768 .2665 .0839 .2782 .0629 .2437 .0559 .2306 .0839 .2782 .0926 .2912 
d98 .1271 .3333 .0489 .2165 .0839 .2782 .0909 .2885 .1119 .3163 .0741 .2631 
d99 .0824 .2752 .2867 .4538 .1049 .3075 .1049 .3075 .0699 .2559 .0833 .2777 
d00 .0475 .2128 .0280 .1655 .0420 .2012 .0559 .2306 .0280 .1655 .0833 .2777 
d01 .0503 .2187 .0350 .1843 .0699 .2559 .0489 .2165 .0699 .2559 .0370 .1897 
d02 .0531 .2243 .0629 .2437 .0559 .2306 .0559 .2306 .0559 .2306 .0370 .1897 
d03 .0349 .1837 .0280 .1655 .0350 .1843 .0210 .1438 .0559 .2306 .0463 .2111 
d04 .0601 .2378 .0280 .1655 .0420 .2012 .0559 .2306 .0699 .2559 .1018 .3039 
d05 .0377 .1906 .0140 .1178 .0140 .1178 .0559 .2306 .0699 .2559 .0278 .1651 



 15 

Table 2. Full sample and quantile regression (QR) results 
 QR 

 

OLS 
 

(1) 
.20 
(2) 

.40 
(3) 

.60 
(4) 

.80 
(5) 

.95 
(6) 

 Coef. Robust 
Std. Err. 

Coef. Robust 
Std. Err. 

Coef. Robust 
Std. Err. 

Coef. Robust 
Std. Err. 

Coef. Robust 
Std. Err. 

Coef. Robust 
Std. Err. 

area 3.7359*** .2078 3.5903*** .2713 3.5632*** .2691 3.3192*** .2761 3.3111*** .3216 3.1853*** .4439 

area2 -1.1313*** .1023 -1.1613*** .1421 -1144*** .1458 -.9691*** .1490 -.9304*** .1386 -.8047*** .2094 

style1 1.6202*** .2042 .9711** .4494 1.5927*** .3154 1.6891*** .2152 1.7641*** .2497 2.0390*** .4826 

style2 2.2070*** .3544 1.7832*** .5042 1.9731*** .4815 2.0475*** .5686 3.1895*** .7636 3.2695*** .5030 

style3 1.8120*** .1816 1.2825*** .3381 1.6133*** .2443 1.5710*** .2426 1.8667*** .3136 2.2191*** .2804 

style4 .9518*** .1180 .5566*** .1546 .6337*** .1508 .7094*** .1693 .9234*** .2149 1.5589*** .2755 

style5 1.2603*** .1207 1.0189*** .1712 1.0480*** .1389 1.0390*** .1573 1.5893*** .2524 1.8825*** .2148 

style6 .9309*** .1232 .4705*** .1478 .6479*** .2007 1.0169*** .1866 1.2914*** .1704 1.6429*** .2443 

style7 .2710*** .1090 .3308** .1433 .3040*** .1054 .3259*** .1066 .3188*** .1215 .3495*** .1366 

canvas .3880*** .1160 .6498*** .1824 .3617** .1524 .2503** .1207 .1874 .1618 .2322 .2168 

panel -.2826 .2209 -.2214 .5064 -.2554 .2766 -.2133 .2289 .1447 .2800 -.1093 .2830 

mixed -1.0104*** .2855 -1.2402*** .3683 -1.6714*** .5032 -.6297 .6021 -.5600 .3717 -.1662 .3710 

chrilon .3724*** .1406 .4241** .2035 .3357* .1826 .1481 .1761 .0892 .1913 -.1282 .3085 

chriny .3964*** .1401 .4562** .2150 .4087** .1802 .2156 .1761 .2122 .1921 -.0936 .2693 

sothlon .3104** .1455 .2416 .2257 .3723* .1736 .1671 .1890 .1608 .1902 -.2504 .2957 

sothny .5919*** .1371 .5598*** .2140 .5862*** .1736 .3122* .1690 .3658** .1673 .0532 .2900 

cons 10.1871*** .2154 9.6212*** .3073 10.1133 .3060 10.7488*** .2661 11.0038*** .2459 11.6212*** .3491 

 Yearly dummies Yearly dummies Yearly dummies Yearly dummies Yearly dummies Yearly dummies 

R-squared 0.6388 0.4299 0.4014 0.4134 0.4284 0.5022 

 
***,**,*. significance at .01, .05, .10 



 16 

         Table 3. Test for the equality of the quantile hedonic regressions coefficients  
 QR 

 .20 vs .40 .20 vs .60 .20 vs .80 .20 vs .95 

Variable F Prob > F F Prob > F F Prob > F F Prob > F 

Dimension 0.97 0.3257 0.84 0.3606 0.69 0.4074 0.57 0.4505 

Style 1.74 0.1091 1.73 0.1106 1.73 0.1124 1.73 0.1107 

Media 3.81 0.0100 3.82 0.0100 3.82 0.0099 3.86 0.0094 

Salerooms 1.48 0.2074 1.45 0.2166 1.41 0.2280 1.40 0.2324 

Period 2.89 0.0001 2.80 0.0002 2.68 0.0004 2.65 0.0005 

 



 15 

 
 
          Table 4. Adjacent year regression indexes 

Percentile 
 Full sample 

.20 .40 .60 .80 .95 

d88 100.00 100.00 100.00 100.00 100.00 100.00 

d89 179.10 172.25 246.25 170.70 166.60 162.26 

d90 120.58 123.62 143.39 95.05 81.17 156.05 

d91 67.11 72.64 82.79 49.91 46.92 48.82 

d92 39.07 41.68 45.38 34.39 39.36 50.26 

d93 31.85 20.74 58.61 35.54 32.10 37.42 

d94 53.73 38.02 56.57 36.18 58.25 88.43 

d95 52.63 44.55 55.43 37.53 55.30 82.20 

d96 45.34 43.98 58.70 34.44 45.88 75.73 

d97 77.51 59.36 81.50 65.93 85.69 117.23 

d98 47.24 42.80 63.93 43.30 53.69 56.44 

d99 70.29 51.82 81.82 61.66 89.33 102.33 

d00 83.30 67.51 90.74 79.22 81.35 125.86 

d01 67.78 63.83 72.38 50.95 47.59 120.32 

d02 54.52 43.12 79.81 46.71 46.84 134.45 

d03 71.56 75.85 88.33 55.94 53.14 76.95 

d04 104.82 89.92 123.04 91.08 103.19 116.53 

d05 118.79 117.94 149.87 91.66 106.83 198.18 

 
 



 16 

             Table 5. 5-year return 

 
return 
(full) return (0.2) return (0.4) return (0.6) return (0.8) 

return 
(0.95) 

1988-92 39.07 41.68 45.38 34.39 39.36 50.26 

1989-93 17.78 12.04 23.80 20.82 19.27 23.06 

1990-94 44.56 30.76 39.45 38.06 71.76 56.67 

1991-95 78.42 61.33 66.95 75.20 117.86 168.37 

1992-96 116.05 105.52 129.35 100.15 116.57 150.68 

1993-97 243.36 286.21 139.05 185.51 266.95 313.28 

1994-98 87.92 112.57 113.01 119.68 92.17 63.82 

1995-99 133.56 116.32 147.61 164.30 161.54 124.49 

1996-00 183.72 153.50 154.58 230.02 177.31 166.20 

1997-01 87.45 107.53 88.81 77.28 55.54 102.64 

1998-02 115.41 100.75 124.84 107.88 87.24 238.22 

1999-03 101.81 146.37 107.96 90.72 59.49 75.20 

2000-04 125.83 133.20 135.60 114.97 126.85 92.59 

2001-05 175.26 184.77 207.06 179.90 224.48 164.71 



 15 

      Figure 2. Adjacent year regression indexes 

0.00

50.00

100.00

150.00

200.00

250.00

d88 d89 d90 d91 d92 d93 d94 d95 d96 d97 d98 d99 d00 d01 d02 d03 d04 d05

Index (full)

Index (.20)

Index (.40)

Index (.60)

Index (.80)

Index (.95)



 1 

       Figure 3. Average return and standard deviation of 5-year returns 

full samp le

0.6 quant
0.2 quant

0.8 quant

.95 quant

0.4 quant

0

10

20

30

40

50

60

70

80

90

105 110 115 120 125 130

average 5-year return

st
d

 d
ev

ia
tio

n 
on

 5
-y

ea
r 

re
tu

rn