...

Quantile Regression - Econometrics at Illinois

by user

on
Category:

canada

0

views

Report

Comments

Transcript

Quantile Regression - Econometrics at Illinois
Journal of Economic Perspectives—Volume 15, Number 4 —Fall 2001—Pages 143–156
Quantile Regression
Roger Koenker and Kevin F. Hallock
e say that a student scores at the ␶th quantile of a standardized exam if
he performs better than the proportion ␶ of the reference group of
students and worse than the proportion (1–␶). Thus, half of students
perform better than the median student and half perform worse. Similarly, the
quartiles divide the population into four segments with equal proportions of the
reference population in each segment. The quintiles divide the population into five
parts; the deciles into ten parts. The quantiles, or percentiles, or occasionally
fractiles, refer to the general case. Quantile regression as introduced by Koenker
and Bassett (1978) seeks to extend these ideas to the estimation of conditional
quantile functions—models in which quantiles of the conditional distribution of the
response variable are expressed as functions of observed covariates.
In Figure 1, we illustrate one approach to this task based on Tukey’s boxplot
(as in McGill, Tukey and Larsen, 1978). Annual compensation for the chief
executive officer (CEO) is plotted as a function of firm’s market value of equity. A
sample of 1,660 firms was split into ten groups of equal size according to their
market capitalization. For each group of 166 firms, we compute the three quartiles
of CEO compensation: salary, bonus and other compensation, including stock
options (as valued by the Black-Scholes formula at the time of the grant). For each
group, the bow-tie-like box represents the middle half of the salary distribution
lying between the first and third quartiles. The horizontal line near the middle of
each box represents the median compensation for each group of CEOs, and the
W
y Roger Koenker is William B. McKinley Professor of Economics and Professor of Statistics,
and Kevin F. Hallock is Associate Professor of Economics and of Labor and Industrial
Relations, University of Illinois at Urbana-Champaign, Champaign, Illinois. Their e-mail
addresses are [email protected] and [email protected], respectively.
144
Journal of Economic Perspectives
Figure 1
Pay of Chief Executive Officers by Firm Size
Notes: The boxplots provide a summary of the distribution of CEO annual compensation for ten
groupings of firms ranked by market capitalization. The light gray vertical lines demarcate the
deciles of the firm size groupings. The upper and lower limits of the boxes represent the first and
third quartiles of pay. The median for each group is represented by the horizontal bar in the middle
of each box.
Source: Data on CEO annual compensation from EXECUCOMP in 1999.
notches represent an estimated confidence interval for each median estimate. The
full range of the observed salaries in each group is represented by the horizontal bars at the end of the dashed “whiskers.” In cases where the whiskers would
extend more than three times the interquartile range, they are truncated and the
remaining outlying points are indicated by open circles. The mean compensation
for each group is also plotted: the geometric mean as a ⫹ and the arithmetic mean
as a *.
There is a clear tendency for compensation to rise with firm size, but one can
also discern several other features from the plot. Even on the log scale, there is a
tendency for dispersion, as measured by the interquartile range of log compensation, to increase with firm size. This effect is accentuated if we consider the upper
and lower tails of the salary distribution. By characterizing the entire distribution of
annual compensation for each group, the plot provides a much more complete
picture than would be offered by simply plotting the group means or medians.
Here we have the luxury of a moderately large sample size in each group. Had we
had several covariates, grouping observations into homogeneous cells, each with a
sufficiently large number of observations, would become increasingly difficult.
In classical linear regression, we also abandon the idea of estimating separate
means for grouped data as in Figure 1, and we assume that these means fall on a line
or some linear surface, and we estimate instead the parameters of this linear model.
Least squares estimation provides a convenient method of estimating such condi-
Roger Koenker and Kevin F. Hallock
145
tional mean models. Quantile regression provides an equally convenient method
for estimating models for conditional quantile functions.
Quantiles via Optimization
Quantiles seem inseparably linked to the operations of ordering and sorting
the sample observations that are usually used to define them. So it comes as a mild
surprise to observe that we can define the quantiles through a simple alternative
expedient as an optimization problem. Just as we can define the sample mean as the
solution to the problem of minimizing a sum of squared residuals, we can define
the median as the solution to the problem of minimizing a sum of absolute
residuals. The symmetry of the piecewise linear absolute value function implies that
the minimization of the sum of absolute residuals must equate the number of
positive and negative residuals, thus assuring that there are the same number of
observations above and below the median.
What about the other quantiles? Since the symmetry of the absolute value
yields the median, perhaps minimizing a sum of asymmetrically weighted absolute
residuals—simply giving differing weights to positive and negative residuals—would
yield the quantiles. This is indeed the case. Solving
min
␰ 僆ℜ
冘 ␳ 共y ⫺ ␰ 兲,
␶
i
where the function ␳␶⵺ is the tilted absolute value function appearing in Figure 2
that yields the ␶th sample quantile as its solution.1
Having succeeded in defining the unconditional quantiles as an optimization
problem, it is easy to define conditional quantiles in an analogous fashion. Least
squares regression offers a model for how to proceed. If, presented with a random
sample { y 1 , y 2 , . . . , y n }, we solve
冘 共y ⫺ ␮兲 ,
n
min
2
␮ 僆ℜ i⫽1
i
we obtain the sample mean, an estimate of the unconditional population mean, EY.
If we now replace the scalar ␮ by a parametric function ␮ ( x, ␤ ) and solve
冘 共y ⫺ ␮共x , ␤兲兲 ,
n
min
␤ 僆ℜ p i⫽1
2
i
i
we obtain an estimate of the conditional expectation function E(Y兩x).
1
To see this more formally, one need only compute directional derivatives with respect to ␰.
146
Journal of Economic Perspectives
Figure 2
Quantile Regression ␳ Function
In quantile regression, we proceed in exactly the same way. To obtain an
estimate of the conditional median function, we simply replace the scalar ␰ in the
first equation by the parametric function ␰ ( x i , ␤ ) and set ␶ to 21 . Variants of this
idea were proposed in the mid-eighteenth century by Boscovich and subsequently
investigated by Laplace and Edgeworth, among others. To obtain estimates of the
other conditional quantile functions, we replace absolute values by ␳␶⵺ and solve
min
␤ 僆ℜ p
冘 ␳ 共y ⫺ ␰共x , ␤兲兲.
␶
i
i
The resulting minimization problem, when ␰ ( x, ␤ ) is formulated as a linear
function of parameters, can be solved very efficiently by linear programming
methods.
Quantile Engel Curves
To illustrate the basic ideas, we briefly reconsider a classical empirical application in economics, Engel’s (1857) analysis of the relationship between household
food expenditure and household income. In Figure 3, we plot Engel’s data taken
from 235 European working-class households. Superimposed on the plot are seven
estimated quantile regression lines corresponding to the quantiles {0.05, 0.1, 0.25,
0.5, 0.75, 0.9, 0.95}. The median ␶ ⫽ 0.5 fit is indicated by the darker solid line; the
least squares estimate of the conditional mean function is plotted as the dashed
line.
The plot clearly reveals the tendency of the dispersion of food expenditure to
increase along with its level as household income increases. The spacing of the
quantile regression lines also reveals that the conditional distribution of food
expenditure is skewed to the left: the narrower spacing of the upper quantiles
indicating high density and a short upper tail and the wider spacing of the lower
quantiles indicating a lower density and longer lower tail.
Quantile Regression
147
Figure 3
Engel Curves for Food
Notes: This figure plots data taken from Engel’s (1857) study of the dependence of households’ food
expenditure on household income. Seven estimated quantile regression lines for different values of ␶
{0.05, 0.1, 0.25, 0.5, 0.75, 0.9, 0.95} are superimposed on the scatterplot. The median ␶ ⫽ 0.5 is
indicated by the darker solid line; the least squares estimate of the conditional mean function is
indicated by the dashed line.
The conditional median and mean fits are quite different in this example, a
fact that is partially explained by the asymmetry of the conditional density and
partially by the strong effect exerted on the least squares fit by the two unusual
points with high income and low food expenditure. Note that one consequence of
this nonrobustness is that the least squares fit provides a rather poor estimate of the
conditional mean for the poorest households in the sample. Note that the dashed
least squares line passes above all of the very low income observations.
We have occasionally encountered the faulty notion that something like quantile regression could be achieved by segmenting the response variable into subsets
according to its unconditional distribution and then doing least squares fitting on
these subsets. Clearly, this form of “truncation on the dependent variable” would
yield disastrous results in the present example. In general, such strategies are
doomed to failure for all the reasons so carefully laid out in Heckman’s (1979)
work on sample selection.
In contrast, segmenting the sample into subsets defined according to the
conditioning covariates is always a valid option. Indeed, such local fitting underlies
all nonparametric quantile regression approaches. In the most extreme cases, we
have p distinct cells corresponding to different settings of the covariate vector, x,
and quantile regression reduces simply to computing ordinary univariate quantiles
for each of these cells. In intermediate cases, we may wish to project these cell
estimates onto a more parsimonious (linear) model; see Chamberlain (1994) and
Knight, Bassett and Tam (2000) for examples of this approach. Another variant is
148
Journal of Economic Perspectives
the suggestion that instead of estimating linear conditional quantile models, we
could instead estimate a family of binary response models for the probability that
the response variable exceeded some prespecified cutoff values.2
Quantile Regression and Determinants of Infant Birthweight
In this section, we reconsider a recent investigation by Abrevaya (2001) of the
impact of various demographic characteristics and maternal behavior on the birthweight of infants born in the United States. Low birthweight is known to be
associated with a wide range of subsequent health problems and has even been
linked to educational attainment and eventual labor market outcomes. Consequently, there has been considerable interest in factors influencing birthweights
and public policy initiatives that might prove effective in reducing the incidence of
low-birthweight infants, which is conventionally defined by whether infants weigh
less than 2500 grams at birth, about 5 pounds, 9 ounces.
Most of the analysis of birthweights has employed conventional least squares
regression methods. However, it has been recognized that the resulting estimates of
various effects on the conditional mean of birthweights were not necessarily indicative of the size and nature of these effects on the lower tail of the birthweight
distribution. A more complete picture of covariate effects can be provided by
estimating a family of conditional quantile functions.3
Our analysis is based on the June 1997 Detailed Natality Data published by the
National Center for Health Statistics. Like Abrevaya (2001), we limit the sample to
live, singleton births, with mothers recorded as either black or white, between the
ages of 18 and 45, residing in the United States. Observations with missing data for
any of the variables described below were dropped from the analysis. This process
yielded a sample of 198,377 babies. Birthweight, the response variable, is recorded
in grams. Education of the mother is divided into four categories: less than high
school, high school, some college and college graduate. The omitted category is
less than high school so coefficients may be interpreted relative to this category.
The prenatal medical care of the mother is also divided into four categories: those
with no prenatal visit, those whose first prenatal visit was in the first trimester of the
pregnancy, those with first visit in the second trimester and those with first visit in
the last trimester. The omitted category is the group with a first visit in the first
trimester; they constitute almost 85 percent of the sample. An indicator of whether
2
This approach replaces the hypothesis of conditional quantile functions that are linear in parameters
with the hypothesis that some transformation of the various probabilities of exceeding the chosen
cutoffs, say the logistic, could instead be expressed as linear functions in the observed covariates. In our
view, the conditional quantile assumption is more natural, if only because it nests within it the
independent and identically distributed error location shift model of classical linear regression.
3
In an effort to focus attention more directly on the lower tail, several studies have recently explored
binary response (probit) models for the occurrence of low birthweights.
Roger Koenker and Kevin F. Hallock
149
the mother smoked during pregnancy is included in the model, as well as mother’s
reported average number of cigarettes smoked per day. The mother’s reported
weight gain during pregnancy (in pounds) is included (as a quadratic effect).
Figure 4 presents a summary of quantile regression results for this example.
Unlike the Engel curve example, where we had only one covariate and the entire
empirical analysis could be easily superimposed on the bivariate scatterplot of the
observations, we now have 15 covariates, plus an intercept. For each of the 16
coefficients, we plot the 19 distinct quantile regression estimates for ␶ ranging from
0.05 to 0.95 as the solid curve with filled dots. For each covariate, these point
estimates may be interpreted as the impact of a one-unit change of the covariate on
birthweight holding other covariates fixed. Thus, each of the plots has a horizontal
quantile, or ␶, scale, and the vertical scale in grams indicates the covariate effect.
The dashed line in each figure shows the ordinary least squares estimate of the
conditional mean effect. The two dotted lines represent conventional 90 percent
confidence intervals for the least squares estimate. The shaded gray area depicts a
90 percent pointwise confidence band for the quantile regression estimates.
In the first panel of the figure, the intercept of the model may be interpreted
as the estimated conditional quantile function of the birthweight distribution of a
girl born to an unmarried, white mother with less than a high school education,
who is 27 years old and had a weight gain of 30 pounds, didn’t smoke, and had her
first prenatal visit in the first trimester of the pregnancy. The mother’s age and
weight gain are chosen to reflect the means of these variables in the sample.4 Note
that the ␶ ⫽ 0.05 quantile of the distribution for this group is just at the margin of
the conventional definition of a low birthweight baby.
We will confine our discussion to only a few of the covariates. At any chosen
quantile we can ask, for example, how different are the corresponding weights of
boys and girls, given a specification of the other conditioning variables. The second
panel answers this question. Boys are obviously larger than girls, by about 100 grams
according to the ordinary least squares estimates of the mean effect, but as is clear
from the quantile regression results, the disparity is much smaller in the lower
quantiles of the distribution and considerably larger than 100 grams in the upper
tail of the distribution. For example, boys are about 45 grams larger at the 0.05
quantile but are about 130 grams larger at the 0.95 quantile. The conventional least
squares confidence interval does a poor job of representing this range of
disparities.
The disparity between birthweights of infants born to black and white mothers
is substantial, particularly at the left tail of the distribution. At the 5th percentile of
the conditional distribution, the difference is roughly one-third of a kilogram.
4
It is conducive for interpretation to center covariates so that the intercept can be interpreted as the
conditional quantile function for some representative case—rather than as an extrapolation of the
model beyond the convex hull of the data. This may be viewed as adhering to John Tukey’s dictum:
“Never estimate intercepts, always estimate centercepts!”
150
Journal of Economic Perspectives
Figure 4
Ordinary Least Squares and Quantile Regression Estimates for Birthweight Model
Mother’s age enters the model as a quadratic effect, shown in the first two
figures of the second row. At the lower quantiles, the mother’s age tends to be more
concave, increasing birthweight from age 18 to about age 30, but tending to
decrease birthweight when the mother’s age is beyond 30. At higher quantiles, this
Quantile Regression
151
“optimal age” becomes gradually older. At the third quantile, it is about 36, and at
␶ ⫽ 0.9, it is almost 40.
Education beyond high school is associated with a modest increase in birthweights. High school graduation has a quite uniform effect over the whole range of
the distribution of about 15 grams. This is a rare example of an effect that really
does appear to exert a pure location shift effect on the conditional distribution. For
this effect, the quantile regression results are quite consistent with the least squares
results; but this is the exceptional case, not the rule.
Several of the remaining covariates are of substantial public policy interest.
These include the effects of prenatal care, marital status and smoking. However, as
in the corresponding least squares analysis, the interpretation of their causal effects
may be somewhat controversial. For example, although we find (as expected) that
babies born to mothers with no prenatal care are smaller, we also find that babies
born to mothers who delayed prenatal visits until the second or third trimester have
substantially higher birthweights in the lower tail than mothers who had a prenatal
visit in the first trimester. This might be interpreted as the self-selection effect of
mothers confident about favorable outcomes.
In almost all of the panels of Figure 4, with the exception of education
coefficients, the quantile regression estimates lie at some point outside the confidence intervals for the ordinary least squares regression, suggesting that the effects
of these covariates may not be constant across the conditional distribution of the
independent variable. Formal testing of this hypothesis is discussed in Koenker and
Machado (1999).
Selected Empirical Examples of Quantile Regression
There is a rapidly expanding empirical quantile regression literature in economics that, taken as a whole, makes a persuasive case for the value of “going
beyond models for the conditional mean” in empirical economics. Catalyzed by
Gary Chamberlain’s (1994) invited address to the 1990 World Congress of the
Econometric Society, there has been considerable work in labor economics: on
union wage effects, returns to education and labor market discrimination. Chamberlain finds, for example, that for manufacturing workers, the union wage premium at the first decile is 28 percent and declines monotonically to a negligible
0.3 percent at the upper decile. The least squares estimate of the mean union
premium of 15.8 percent is thus captured mainly by the lower tail of the conditional
distribution. The conventional location shift model thus delivers a rather misleading impression of the union effect. Other contributions exploring these issues in
the U.S. labor market include the influential work of Buchinsky (1994, 1997). Arias,
Hallock and Sosa-Escudero (2001), using data on identical twins, interpret observed heterogeneity in the estimated returns to education over quantiles as
152
Journal of Economic Perspectives
indicative of an interaction between observed educational attainment and unobserved ability.
There is also a large literature dealing with related issues in labor markets
outside the United States, including Fitzenberger (1999) on Germany; Machado
and Mata (1999) on Portugal; Garcia, Hernandez and Lopez-Nicolas (2001) on
Spain; Schultz and Mwabu (1998) on South Africa; and Kahn (1998) on international comparisons. The work of Machado and Mata (1999) is particularly notable
since it introduces a useful way to extend the counterfactual wage decomposition
approach of Oaxaca (1973) to quantile regression and provides a general strategy
for simulating marginal distributions from the quantile regression process. Tannuri
(2000) has employed this approach in a recent study of assimilation of U.S.
immigrants.
In other applied micro areas, Eide and Showalter (1998), Knight, Bassett and
Tam (2000) and Levin (2001) have addressed school quality issues. Poterba and
Rueben (1995) and Mueller (2000) study public-private wage differentials in the
United States and Canada. Abadie, Angrist and Imbens (2001) consider estimation
of endogenous treatment effects in program evaluation, and Koenker and Billias
(2001) explore quantile regression models for unemployment duration data. Work
by Viscusi and Hamilton (1999) considers public decision making regarding hazardous waste cleanup.
Deaton (1997) offers a nice introduction to quantile regression for demand
analysis. In a study of Engel curves for food expenditure in Pakistan, he finds that
although the median Engel elasticity of 0.906 is similar to the ordinary least squares
estimate of 0.909, the coefficient at the tenth quantile is 0.879 and the estimate at
the 90th percentile is 0.946, indicating a pattern of heteroskedasticity like that of
our Figure 3.
In another demand application, Manning, Blumberg and Moulton (1995)
study demand for alcohol using survey data from the National Health Interview
Study and find considerable heterogeneity in the price and income elasticities over
the full range of the conditional distribution. Hendricks and Koenker (1992)
investigate demand for electricity by time of day using a hierarchical model.
Earnings inequality and mobility is a natural arena of applications for quantile
regression. Conley and Galenson (1998) explore wealth accumulation in several
U.S. cities in the mid-nineteenth century. Gosling, Machin and Meghir (2000)
study the income and wealth distribution in the United Kingdom. Trede (1998)
and Morillo (2000) compare earnings mobility in the United States and Germany.
There is also a growing literature in empirical finance employing quantile
regression methods. One strand of this literature is the rapidly expanding literature
on value at risk: this connection is developed in Taylor (1999), Chernozhukov and
Umantsev (2001) and Engle and Manganelli (1999). Bassett and Chen (2001)
consider quantile regression index models to characterize mutual fund investment
styles.
Roger Koenker and Kevin F. Hallock
153
Software and Standard Errors
The diffusion of technological change throughout statistics is closely tied to its
embodiment in statistical software. This is particularly true of quantile regression
methods, since the linear programming algorithms that underlie reliable implementations of the methods appear somewhat esoteric to some users. Since the early
1950s, it has been recognized that median regression methods based on minimizing sums of absolute residuals can be formulated as linear programming problems
and efficiently solved with some form of the simplex algorithm.5
Among commercial programs in common use in econometrics, Stata and TSP
offer some basic functionality for quantile regression within the central core of the
package distributed by the vendor. Since the mid-1980s, one of us has maintained
a public domain package of quantile regression software designed for the S
language of Becker, Chambers and Wilks (1988) and the related commercial
package Splus. Recently, this package has been extended to provide a version for R,
the impressive public domain dialect of S (Koenker, 1995). This website also
provides quantile regression software for Ox and Matlab languages.
It is a basic principle of sound econometrics that every serious estimate deserves a
reliable assessment of precision. There is now an extensive literature on the asymptotic
behavior of quantile regression estimators and considerable experience with inference methods based on this theory, as well as a variety of resampling schemes. We
have recently undertaken a comparison of several approaches to the construction
of confidence intervals for a problem typical of current applications in labor
economics. We find that the discrepancies among competing methods are slight,
and inference for quantile regression is, if anything, more robust than for most
other forms of inference commonly encountered in econometrics. Koenker and
Hallock (2000) describe this exercise in more detail and provide a brief survey of
recent work on quantile regression for discrete data models, time series, nonparametric models and a variety of other areas. Some of these developments have been
slow to percolate into standard econometric software. With the notable exceptions
of Stata and Xplore (Cizek, 2000) and the packages available from the web for
Splus and R, none of the implementations of quantile regression in econometric
software packages include any functionality for inference.
5
The median regression algorithm of Barrodale and Roberts (1974) has proven particularly influential
and can be easily adapted for general quantile regression. Koenker and d’Orey (1987) describe one
implementation. For large-scale quantile regression problems, Portnoy and Koenker (1997) have shown
that a combination of interior point methods and effective preprocessing renders quantile regression
computation competitive with least squares computations for problems of comparable size.
154
Journal of Economic Perspectives
Conclusion
Much of applied statistics may be viewed as an elaboration of the linear
regression model and associated estimation methods of least squares. In beginning
to describe these techniques, Mosteller and Tukey (1977, p. 266) remark in their
influential text:
What the regression curve does is give a grand summary for the averages of
the distributions corresponding to the set of x’s. We could go further and
compute several different regression curves corresponding to the various
percentage points of the distributions and thus get a more complete picture
of the set. Ordinarily this is not done, and so regression often gives a rather
incomplete picture. Just as the mean gives an incomplete picture of a single
distribution, so the regression curve gives a corresponding incomplete picture
for a set of distributions.
We would like to think that quantile regression is gradually developing into a
comprehensive strategy for completing the regression picture.
y We would like to thank the editors, as well as Gib Bassett, John DiNardo, Olga Geling,
Bernd Fitzenberger and Steve Portnoy for helpful comments on earlier drafts. The research was
partially supported by NSF grant SES-99-11184.
References
Abadie, Alberto, Joshua Angrist and Guido
Imbens. 2001. “Instrumental Variables Estimation of Quantile Treatment Effects.” Econometrica. Forthcoming.
Abrevaya, Jason. 2001. “The Effects of Demographics and Maternal Behavior on the Distribution of Birth Outcomes.” Empirical Economics.
March, 26:1, pp. 247–57.
Arias, Omar, Kevin Hallock and Walter SosaEscudero. 2001. “Individual Heterogeneity in
the Returns to Schooling: Instrumental Variables Quantile Regression Using Twins Data.”
Empirical Economics. March, 26:1, pp. 7– 40.
Barrodale, I. and F. Roberts. 1974. “Solution
of an Overdetermined System of Equations in
the L1 Norm [F4] (Algorithm 478).” Communications of the ACM. 17:6, pp. 319 –20.
Bassett, Gilbert and Hsiu-Lang Chen. 2001.
“Quantile Style: Return-Based Attribution Using
Regression Quantiles.” Empirical Economics.
March, 26:1, pp. 293–305.
Becker, Richard, John Chambers and Allan
Wilks. 1988. The New S Language: A Programming
Environment for Data Analysis and Graphics. Pacific
Grove: Wadsworth.
Buchinsky, Moshe. 1994. “Changes in U.S.
Wage Structure 1963–1987: An Application of
Quantile Regression.” Econometrica. March, 62:2,
pp. 405–58.
Buchinsky, Moshe. 1997. “The Dynamics of
Changes in the Female Wage Distribution in the
USA: A Quantile Regression Approach.” Journal
of Applied Econometrics. January/February, 13:1,
pp. 1–30.
Quantile Regression
Chamberlain, Gary. 1994. “Quantile Regression, Censoring and the Structure of Wages,” in
Advances in Econometrics. Christopher Sims, ed.
New York: Elsevier, pp. 171–209.
Chernozhukov, V. and L. Umantsev. 2001.
“Conditional Value-at-Risk: Aspects of Modeling
and Estimation.” Empirical Economics. March,
26:1, pp. 271–92.
Cizek, Pavel. 2000. “Quantile Regression,” in
XploRe Application Guide. Wolfgang Hardle,
Zdenk Hlavka and Sigbert Klinke, eds. Heidelberg: Springer-Verlag, pp. 21– 48.
Conley, Timothy and David Galenson. 1998.
“Nativity and Wealth in Mid-Nineteenth-Century
Cities.” Journal of Economic History. June, 58:2, pp.
468 –93.
Deaton, Angus. 1997. The Analysis of Household
Surveys. Baltimore: Johns Hopkins.
Eide, Eric and Mark Showalter. 1998. “The
Effect of School Quality on Student Performance: A Quantile Regression Approach.” Economics Letters. March, 58:3, pp. 345–50.
Engel, Ernst. 1857. “Die Produktions- und
Konsumptionverhaltnisse des Konigreichs Sachsen.” Reprinted in “Die Lebenkosten Belgischer
Arbeiter-Familien Fruher und Jetzt.” International Statistical Institute Bulletin. 9, pp. 1–125.
Engle, Robert and Simone Manganelli. 1999.
“CaViaR: Conditional Autoregressive Value at
Risk by Regression Quantiles.” University of California, San Diego, Department of Economics
Working Paper 99/20. October.
Fitzenberger, Bernd. 1999. Wages and Employment Across Skill Groups. Heidelberg: PhysicaVerlag.
Garcia, Jaume, Pedro Hernandez and Angel
Lopez-Nicolas. 2001. “How Wide is the Gap? An
Investigation of Gender Wage Differences Using
Quantile Regression.” Empirical Economics. March,
26:1, pp. 149 – 67.
Gosling, Amanda, Stephen Machin and Costas
Meghir. 2000. “The Changing Distribution of
Male Wages in the U.K.” Review of Economic Studies. October, 67:4, pp. 635– 66.
Heckman, James J. 1979. “Sample Selection
Bias as a Specification Error.” Econometrica. January, 47:1, pp. 153– 61.
Hendricks, Wallace and Roger Koenker. 1992.
“Hierarchical Spline Models for Conditional
Quantiles and the Demand for Electricity.” Journal of the American Statistical Association. March,
87:417, pp. 58 – 68.
Kahn, Lawrence. 1998. “Collective Bargaining
and the Interindustry Wage Structure: Interna-
155
tional Evidence.” Economica. November, 65:260,
pp. 507–34.
Knight, Keith, Gilbert Bassett and Mo-Yin S.
Tam. 2000. “Comparing Quantile Estimators for
the Linear Model.” Preprint.
Koenker, Roger. 1995. “Quantile Regression Software.” Available at 具http://www.
econ.uiuc.edu/⬃roger/research/rq/rq.html典.
Koenker, Roger and Gilbert Bassett. 1978.
“Regression Quantiles.” Econometrica. January,
46:1, pp. 33–50.
Koenker, Roger and Yannis Bilias. 2001.
“Quantile Regression for Duration Data: A Reappraisal of the Pennsylvania Reemployment Bonus Experiments.” Empirical Economics. March,
26:1, pp. 199 –220.
Koenker, Roger and Vasco d’Orey. 1987.
“Computing Regression Quantiles.” Applied Statistics. 36, pp. 383–93.
Koenker, Roger and Kevin Hallock. 2000.
“Quantile Regression: An Introduction.” Available at 具http://www.econ.uiuc.edu/⬃roger/
research/intro/intro.html典.
Koenker, Roger and José Machado. 1999.
“Goodness of Fit and Related Inference Processes for Quantile Regression.” Journal of the
American Statistical Association. December,
94:448, pp. 1296 –310.
Levin, Jesse. 2001. “For Whom the Reductions
Count: A Quantile Regression Analysis of Class
Size on Scholastic Achievement.” Empirical Economics. March, 26:1, pp. 221– 46.
Machado, José and José Mata. 1999. “Counterfactual Decomposition of Changes in Wage Distributions Using Quantile Regression.” Preprint.
Manning, Willard, Linda Blumberg and Lawrence Moulton. 1995. “The Demand for Alcohol:
The Differential Response to Price.” Journal of
Health Economics. June, 14:2, pp. 123– 48.
McGill, Robert, John W. Tukey and Wayne A.
Larsen. 1978. “Variations of Box Plots.” American
Statistician. February, 32, pp. 12–16.
Morillo, Daniel. 2000. “Income Mobility with
Nonparametric Quantiles: A Comparison of the
U.S. and Germany.” Preprint.
Mosteller, Frederick and John Tukey. 1977.
Data Analysis and Regression: A Second Course in
Statistics. Reading, Mass.: Addison-Wesley.
Mueller, R. 2000. “Public- and Private-Sector
Wage Differentials in Canada Revisited.” Industrial Relations. July, 39:3, pp. 375– 400.
Oaxaca, Ronald. 1973. “Male-Female Wage Differentials in Urban Labor Markets.” International
Economic Review. October, 14:3, pp. 693–709.
Portnoy, Stephen and Roger Koenker. 1997.
156
Journal of Economic Perspectives
“The Gaussian Hare and the Laplacian Tortoise:
Computability of Squared-Error Versus Absolute-Error Estimators, with Discussion.” Statistical
Science. 12:4, pp. 279 –300.
Poterba, James and Kim Rueben. 1995. “The
Distribution of Public Sector Wage Premia: New
Evidence Using Quantile Regression Methods.”
NBER Working Paper No. 4734.
Schultz, T. Paul and Germano Mwabu. 1998.
“Labor Unions and the Distribution of Wages
and Employment in South Africa.” Industrial and
Labor Relations Review. July, 51:4, pp. 680 –703.
Tannuri, Maria. 2000. “Are the Assimilation
Processes of Low and High Income Immigrants
Distinct? A Quantile Regression Approach.” Preprint.
Taylor, James. 1999. “A Quantile Regression
Approach to Estimating the Distribution of Multiperiod Returns.” Journal of Derivatives. 7:1, pp.
64 –78.
Trede, Mark. 1998. “Making Mobility Visible:
A Graphical Device.” Economics Letters. April,
59:1, pp. 77– 82.
Viscusi, W. Kip and James Hamilton. 1999.
“Are Risk Regulators Rational? Evidence from
Hazardous Waste Cleanup Decisions.” American Economic Review. September, 89:4, pp.
1010 –27.
Fly UP