WC find reliable evidence I~JI both book-to-market (9 Nt and dividend yield track
time-series variation in expect& real stock returns over the priod IX!6 91 (in which
B M is stronger) and the subperti 1941-91 (In which dividend yieid is stronger).
A Bay&an bootsmap procedure implies hat an investor with prior belief 0.5 that
expected re:ums on the equal-weigh’ed index are never negative comes away from the
full-period B. hi evidencxz with posterior pro&hi@ O.oII for the hypothesistO.I4 with the
impact of the I933 outher temperedI. Altnough this raises doubts about market c%cicnq.
the post-t940 evidence is consr$tent with expected returns always heins positive.
There is considerable evidentle that stock returns are predictable. The aca-
demic, literature on return predictability goes back at Ieast to the late 1970s.
*Corresponding author.
We thank Ray Bail. Peter Bossaerts. Stwe Caulej-. Gene Fama. K~I French (the r&met. Burton
HoMieN John Long &iii Schwert (the editorl Rob Stambaugh. Jerr: Warner. Robert Whitelaw.
and espectill~ Johnathan Lewellen. Aiso. thanks to participants in semtnars al NBER. the Southern
Finance Association. Virginia T&, London Business School. Dartmouth. Washington University.
Yale. the University of California at Irvine and Los Angelex Camhn~. and Clniversily of
Rochester. We gratefully acknmvledgfinancial wpport from the John M. Ol*n Foundation and the
Bradley Policy Research Center at the Simon School. University of Rochester. Assistant: from K Em
Goetz and Jane Muellner on manuscript preparation is appreciated.
Several studies (e.g., Fama and Schwert, 1977; Keim and Stambaugh, 19886;
and Fama and French, 1988) regress returns on predetermined variabks, includ-
ing past returns, to infer the existence of statistically significant time-variation
in expected returns. Among the predetermined variables, many studies have
found dividend yield to have predictive power in both cross-section (L&en-
berger and Ramaswamy, 1979) and time series (Rozeg, 19@ Shifter, 19w,
Fama and French, 1988a, 1989; and Campbell and Shilfer, 1988). Since the
publication of Fama and French (1992, 1993), the book-to-market ratio
(B/M) has emerged as a strong contender as a determinant of expected returns.
Fama and French, Davis (MM), and Chan et al. (1995). among others+
provide evidence that B/M significantly explains cross-sectional variation in
average returns, although Kothari et al. (1995) show that the effect is weaker
in large firms (also see fought-an, 1996) and argue that the magnitude and
signiGcance of the effect may be overstated due to data mining and a variety
of selection biases in the COMPUSTAT data base (also see Breen and
Korz,, czyk. 1994
The literature emphasizes two views of the predictive power of financial ratios
that are relevant in both time series and cross-section. The first (e@cient
markets) view has been referred to by Fama and French (1988b) as the
‘discount-rate effect (also see Keim and Stambaugh, 1986). Holding expected
cash flows roughly constant, anything that incmases the rate at which cash Sows
are discounted (e.g., changes in risk or liquidity) lowers market value and
irtcmss both eqxcted return and financial ratios. This assumes+ ofcoum that
the most rarmt value in the numerator (dividend or book value of equity) does
an adequate job of controlling for expected future cash f?ow. An alternative
(ineffi&nt markets) view is that financial ratios refkct the extent to which the
market is overvalued (low ratios) or undervalued (high ratios) at a given point in
time. fn the &se of underpricing. for example, future returns (and thus ?rue’
expected mums) will be high insofar as the undervaluation is likely to be
corrected over the given horizon.*
In this research we evaluate the ability of an aggregate B/M ratio to track
time-series variation in expected market index retunrs, and compare its forecast-
ing ability to that of divrdend yield. We employ a vector-autoregressive (VAR)
framework to evaluate the statistical signiRcance of the regression evidem~ of
predictability of expected returns. The bootstrap simulation procedure is used to
test the null hypothesis of no predictability as well as to assess the magnitude
and economic significance of the relation between expected returns and its
detetminants, such as B/M and dividend yield. Finally. we extend the traditional
bootstrap procedure by developing a Rayesian-bootstrap simulation. With the
‘A third motivation for the cro6s-JeFtjo nal investigation of dbidcnd yield is taxes. This study.
howvcr. does not focus on the tax motivation.
S.P. Ko~hwi, J. Sbanken/Joumai ojFinanciai Economics 44 (1997) 169--X83’ 171
-____ -_--~
zMacð and Emanuel(1993)prcscntsevidcmz that price-t&rook hassomsfomcastingabilityfor
anodii @-year mums on the standard and poor’s IRdwrial index from I!246 to 1982 tt is not
dear (to us) from the discussion in tFie paper. however, wh&cr TiK p&2 and book valuer simulated
under the null hypothesii were t&ted in a manner similar tv tha4obsmvd in the haiotil data. If
ov(, there is some question as to the reliability of the rest&a We hare similar conarns about the
hvvtstmp pra.tculu~r vf Gvetzmann and Jorion (1993), in which them ix no relation betwcm
raodvti Etums and tbe (actual) dividends used in computing divihttd yield. It is natural, uttdcr
the null hyphcsb, to have returns iodepcnknt d past divider&, but returns sbotdd anticpate
futorr: dividends in an e&ient market. even if expected returns are constant (e.& see the evidence
in Kothari and !Shankeu, 1992L The VAR framework discussed below allurn this titure ofthe data
to be relleucd in the simulated modd.
172 S.P. Korhari. J. Shanken/Joumal of Financial Economics 44 (1997) 169-203
yield results are comparable for the equal-weighted index, but the dividend yield
evidence is stronger for the value-weighted index in the subperiod.
We report descriptive statistics in Section 2 and initial regression results in
8ection 3. Section 4 describes our bootstrap simulation methodology, while
Section 5 describes the computation of bootstrap p-values and likelihoods based
on the simulations. Section 6 d&cusses the robe of prior information about
expected returns in the context of testing the sensitivity of expected returns to
determinants like B/M and dividend yield. Bootstrap simulation results are
presented in Sections 7 and 8. Section 7 focuses on the magnitude of the slope
coe@cients on the finaucial ratios in the expected return refation, while Section
8 provides tests of I?K hypothesis that expected returns are always nonnegative.
section 9 summarizes our main findings
This section begins with a brief description of the data ~sI, fohowed by
‘descriptive statistics for the time series of B/M. dividend yield, and market index
returns To ensure that the B/M ratio is known to investors at the beginning of
the return year, we measure annual returns from April of year t to March of year
r~1aad~thesefetumsontheratioofbookvaluefor~eart-ltopricc
at the end of Match of year t. The corresponding dividend yiefd equals dividends
paid from April of year t - 1 tohlarchofyearrdividedbypriceattheendof
March of year r. These tlttios and returns an t&erred to as the ‘year r’
observations
The hook value ‘per share’ that corresponds to the Dow Jones industrial
Average is published in The Valve Line fnmamenf +xy. Time series data are
available since 1920. Book value data for 1920-35 are based on estimates by
The V&e Line. The book value for year t - 1 is deflated by the Dow Jones
industrial Average at the end of March of year t + 1 to obt&n the B/M ratio,
which is used to forecast returns for year t.
We begin by presenting some descriptive statistics in Table I for returns, B,IM,
and dividend yield over the sample Period 1926-9! and subperiod 3941-91.
Retcm and dividend yield information is for both equal- and value-weighted
CJSP indexes. All three series are positively skew& with the equal-weighted
returns being the most highly skewed. For the whole period, the average real
return on the equal-weighted (value-weightad) index is 15.3% (9.4%) with
standard deviation 38.8% (23.9%). The minimum is, - 61% ( - 48.4%) in 1937
and the maximum is 214% (85.7%) in 1933. The returns on both indexes are
lower and less vohtile in the frost-1940 subperiod.
The mean of the B/M series for 1926,-91 is 70% with standard deviation 23%.
The minimum is 27% for 1929 and the maximum is 148% for 1933, the same
year in which returns achieve their maximum for both indexes. The mean of the
Wpfive statistics for the return and financial ratio variae
.-._ _“” ...-_ -^ - .-. . ..- - . .._ ---.----..I .-...__.._. ~.~ I _,.._---_ g
Variable Mean Std. Skewness
Mm VI Median 03 Max 2
dvn.
ll~l,_ I-.. --_ . . ._ I ..- ._- ..__ ___--_-“- 8
- -2,
Pld A: Perind I92t%91 s
Equal-weighted return (%) IS.3 38.8 2..30 - 60.9 - 6.3 10.2 27.7 213.9 2
Value-weighted rctum (%) 9.4 23.9 0.36 -
- 48.4 4.3 9.0 21.4 85.7 i
3ook-to-mtuket (o/6) 69.5 22.7 0.87 27.2 52.8 66.5 80.5 147.7
Equal-weighted dividend yield t %t 3.65 1.44 1.09 I.41 2.71 3.22 4.31 8.21 2
2.66 3.33 4.11 5.06 9.27 B
V&c-weighted dividend yicJd t%t 4.40 1.42 1.41
7
:.
Panel B: Patlod IWl-91 9
Equal-weightcd return (%I 12.5 23.2 0.53 - 31.0 - 3.0 9.4 25.0 3
75.”
8
Value-we&)hcd rctum (%) 8.6 17.2 0.01 - 27.1 3.7 8.4 19.7 42.2 3.
Book-to=market (%) 70.7 I9.Q 0.59 42.8 54.7 69.0 81.4 118.6 0
Equal-w&&cd dividend yield (%I 3.67 1.35 0.94 I.85 2.71 3.18 4.51 7.27 h‘
V&c-wcightcd dividend yield W) 4.24 t.23 I.04 2.66 3.21 4.05 5.06 8.39 B
----- ___--. -~____ _--- ___l-----
5.
Equ&wei$htcd return is the CRSP equal-weighted snnual real mum inclusive of dividends from April of year t to March of year I t 1. a
Vaksl-weighted return is the CRSP vaiuc-weighted rnnual real Wurn inclusive of dividends from April of year I to Murch of year I + I. 2
Book-to-market is the ratio of the book value per share at the cud of year I - I of the stocks in the Dow Jones lndustriai Average to the Dow Jones $
industrial Average at the end of March of year I. 2
Equal-weighted dividend yicid is the ratio o~dividendr paid from April of year t - I IO M&r& ofye..~r I IO the accumulated value of one dollar invested in the 2
CRSP cquabwcightcd in&x exclusive of dividends for one year ending on March 3 I of year 1. P
Value-weighted dividend yield is the ratio ofdividends paid from April of year I - t ta March of year r to the accumulated value olonc dollar invested in the s
CRSP value-weighted index cxch~dve of dividends for one year ending on March 31 of yair 1.
S.P. Kotltari, J. ShankenlJonmal of Financial Economics 44 (1999) I@-201 175
3OLSTt?pSfhrrafts
average of 70% by one standard deviation (23%) the expected real return, which
averages 15.3% is forecast to rise by 20.2 percentage points As we shall see
later, adjustments for small-sample bias do not materially alter these con-
clusions. Deleting the 1933 outlier cuts the B/M coefhcient in half, but the
estimate is still very large and highly statistically significant. Truncation also has
a substantud, but smaller. efkct. Turning to the regression results for the
value-weighted index reveafs that the Dow B/M ratio explains a much smaller
fraction of the time-series variation in these returns than for the CBSP equal-
weighted index.
Dividenf? yieM, in contrast, is somewhat better at explaining variation in
returns on the value-weighted index than on the equal-weighted index. For the
full sample, dividend yield explains less of the variation in value-weighted in&x
returns than does B/M, but explanatory power is similar with deletion or
truncation. Although equal-weighted dividend yield steadily incrzases from
1929 to 1932, it drops nearly by half, to 3.44%. at the beginning of 1933. fience,
deleting the 1933 observation actually strengthens the telation between equal-
weighted returns and dividend yiefd. Although value-weighted dividend yield
also begins to drop, at 694% it is still well above its mean 4.40% at the
beginning of 1933 and hence oeietion weakens the value-weighted rdation.
The slightly better abiJity of dividend yield to explain value-weighted index
rctumslnayberrlatcdtothttactt:latalarl4efractiaaoitbcfirmsintbeindexdo
not pay dividends Thedone, these firms’ dividend yields cannot refkct
in expected market return. Since firms that do not pay dividends are, on average,
smaller in market capilaiization, they have a relatively small efkt on the
dividend yield of the value-weighted index.
Multiple regressions of returns on B/M and dividend yiekl over the 192691
period, shown in Table 3, teU essentiahy the same story as the univariate
regre&ons B/M dominates dividend yield in explaining equal-weighted index
returns B/M also outperforms dividend yield for value-weighted ir&x returns
when the 1933 outlier is inch&d, but otherwise the explanatory power of the
variables is similar.
Given the concerns about small-sample biases mentioned earher, we can
hardly have much confidena in the White (19f8tJt asymptotic adjustment for
heteroskedasticity shown in Table 2. We use it as a heuristic for (we hope)
identifying situations in which results may be particnfarly sensitive to he&m-
skedasticity problems or other misspecifkations. This appears to be the case for
the full-sample equal-weighted index regressions, in which the White r-statistic is
about 45% smaller (larger) than the usual OLS t-statistic for the B/M (dividend
yield) slope. Exploratory regzssions (not reported) of squared or absolute
return residuals on B/M suggest a signifkantly positive conditional variance
relation, whereas the refation for dividend yield is negative and statistically
insignificant. We examine weighted least squares and nonlinear regression
specikations for B/M, w*ith little impact on our basic conclusions. There is no
Table 3
OLS muitipic regression msuits: 1926 91
;h:
\4
EL
8
Tabkr 4
OLS regression results: 1941 91
_ _ “. .__..._.-_- _._.,..__..- --- _.-... *
crin ‘$ s(a) in % t-stat h in % n(h) in % I-mtat White Adjusted 3
Index
t-stat R’ in % 6
_.______ -._--. ^ “. ‘. - -~ .-. _.- --------- “r
9. .
Panel A: R,, t =x+/IR:M,+u,~,
Equal weigh: - 19.9 1 I.9 - I.68 g.7 3.1 z.03 2.79 12.3
Value weight - 7.8 9.2 - 0.85 4.4 2.4 1.86 2.01 4.7
PanPI B: R, I , = I c- gDYU, + u, , ,
Equal weight - 2.7 8.R - 1.45 9.3
Value weight - IS.0 R.I -- l.g6 6.9
w - 8) = &+/u,!)E(p - 4, (3)
SP. Korhari. J. Shankm/Jownal of Financial Economics 44 (IQ971 MY--203
where p is the OLS estimate of 4 in Eq. (2). As expected, the residuals in the
return regression equation are strongly negatively contemporaneously corre-
lated with the residuals in the financial ratio autoregression (2). For the 1926-9 I
period, the correlation is - 0.80 in the case of B/M and - O-65 ( - 0.79) for
equal-weighted (value-weighted) dividend yield. Over tbe post-1940 period,
residuals in the two dividend yield VAR’s an more strongly contemporaneously
correlate4about -O.%Sineach~The~ti~correlationisiMtucedbythe
fact that price increases tend to reduce the financial ratios wbik resulting in
positive returns. Thus, by (3), tbe OLS slope estimates are biased upwards.
Combining (3) with the fact that the bias in p is approximately - (1 + 34)
divided by tbe sampk size T (Kendall 1954) suggests the following bias-
adjusted estimator of 8:
b A = b + &r/&l + 3pA)/7-. (4
where s, and s,” are sample moments of the OLS residuals from Eqs. ( I ) and (2)
and PA = (rp + l)/(r - 3) is a bias-adjusted estimator for 9. After computing
the bias-adjusted estimates hA and PA, the intercepts are specitkd by setting tbe
avm fitted vahres in (1) and (2) equal to the sampk means. Corresponding
adjust& (~rf residual pairs are tben obtained from (It and (2), with each
rtsiciual taxies averaging to zero over time. Tbeae residuals are used in the
bootstrap shdatiom which we describe below.
The system is started up with the historical value for x at tbe heginning of the
sampk lreriod. Given a hypothetical value of /3 and substituting tbe estimator
pAforAtimetieriesofreturnsaadeewvalucsoCxanthengatcratedfnmr(l)
and (2) by randomly selecting adjusted (a. r) residual pairs with mt.
Finally, tbe simulated time series of returns, which has the same length as the
original data used to estimate Eq. (1 L is regressed on tbe simulated time series of
X*S.The estimated slope from this regmssion is saved. This procedure is repeati
2StMltimes with the hypothetical @value fixed and each time, starting tbe system
with the historical vale: for .Yat the beginning of tbe sampk period- This yields
an empirical distribution of slope estimates ihat is employ& in a variety of’
inkrence procedures3
The entire process (2500 iterations) is repeated for each of a discrete set of beta
values starting from zero. Thus, we have a differcat empirical distribution of
slope estimates for each hypothetical alternative. Wc focus on ‘nonnegative slope
‘Since the distribution of the r&da& is fiti i~foss simulatioas. as the Iruc slope c&kkttt vat&%
cbestrnittatalvaiianceofmunts inarilscr wil tht prvdkfahk var%atim k~ mturns Ifanything, this
wordd appear to make it more diikuh to lad eAknce of strottg ptcdrctability aad thus might
impart a eonsl?lvatiye hias to our lesub Likewise, with rqwd to the mx! of bias-adjusted rcsidualr
It would be interesting to exphzmz the sensitivity of inferences to altemak pmaxlurcs which
attempt to maintain a fixed total variability of mums.
I82 S.P. &&an’, J. Shankn/Joumal of Financial Economics 44 (1997) 169-203
distribution of the sample statistic under the alternative plays no role in this
context.
In order to obtain probabiMes for the diRerent hypotheses, conditional on
tbe historical evidence, the Bayesian approach ,focuses on the %kelihood func-
tion’, in conjunction with prior (subjective) probabilities for the hypotheses.
With a continuous density function for the sample evidemz (here, the historical
slope), the relative likelihood of one hypothesis as compared to aDOther can
Ioosey be described as follows: it is the ratio of the respective probabilities
(under the different h.qotheses or models) of generating sample evidence within
an infinitesimal aeigh5orhood of the historical slope.
If the VAR disturbances in Eqs (I) and (2) are assumed to be jointly normally
distributed, a dosed-form expression for the iikelihood function can be obtained
(see Kandedand Stambaugh, 19!&). Here, we take a d&rent approach, intro-
ducing what we call B bootstrap likelihood, is, the proportion of simulated
samples (under a given hypothesis) for which the observed slope is in some fixed
neighborhood of the historical slope Although lacking the mathematical el-
egance of more formal Bavesian approack our strategy, like the ‘da&al
bodstrap, is simple and pragmatic, and does not requi= much in the way of
assumptions.* For example, the b&trap procedure aI- for the positive
skewness that we sometimes observe in the VAR residuala A similar approach
c&d be adapted to a variety of situations that an commonly viewed as
analyticaNy intractable an& the&~ not BmenaMe to Bayesian analysis. While
there is a literature on nonparametric density estimation, we are not aware of
any pqxm that have used bootstrapping techniq~ in the manner dthis paper.
Bootstrap likelihoods arr: computed for the same set offi values cons&r4 in
the class&d p-value analysis. These proportions are scakd so that the sum of the
likelihoods over all values of lp conside& is one. As in any Bayesian application,
scaling does not a&ct the relative tikdihoods or the associated posterior
probabilities derived from the likelihoods and a given prior. Moreover, the
scaled proportions can be interpreted directly as posterior probabilities for
the diffireat slope values when prior belie& arc uniformly distributed over the
various slopes. Although we do not claim that this is the ‘right’ prior, it provides
a convenient &&ace point for inteqxeting the data. Addirionally, if the
likelihood is (approximately) zero for the largest slope consider4 posterior
probabiirties would be unaffected by expanding the uniform prior to larger
values of beta. This is the case in mauy of our simuiatious.
*One sboukJOO(push this point loo iu. ils any boo~rap tcdmique is limitd by the inf’ormativc-
mfss oftbe obseme4l sample fmu which the simulations IVC dcvhpcd Also. in princ+ple we wouid
waattocoaditionona~~esetof~tslabisriFslurthe~~~asouroimpMied
procedure may entail wme loss of information, although it is easily exterded to acwotmodate
amditioning on several statistics
184 S.P. Kothari. J. Shanken/Journal of Finuncial Economirl 44 (1997) 169 203
We have explored the impact of varying the width of the neighborhood of the
historical slope used in computing the likelihoods. The sensitivity‘ to this
alteration is minimal (usually identical to two decimals). This robustness
is reassuring for our approach and presumably reflects the fact that the
true density function is approximately linear within the small neighborhoods
considered.
‘See Ma*Owitiz and Usmen (1996) fur additional discussion of thee issues in another financiaf
context.
S.P. Kothari. J. Shank IJownal of Financial Economics 44 (1997) 169~ 203 185
7. Tea!aforexpeeWlremlretnmnriatfen
This section presents bootstrap evidence for the regressions of real equai-
weighted and value-weighted index returns on B;M and dividend yield for the
entire period, 1926-91. and the 1941-91 and 1963-91 subperiods.
bi: is possible to devise models in which the expected real market pwWwn is negative IBoudoukh
et al., 1996). Since real T-bill returns excccdcd 4.5% in the deflationary period of the late ISZOS and
early 1930s. however. the negative premium required to make total exptxted real return acgatiw
durirtg that period wouid have to bc quite large.
186 S.P. Kothari. J. Shanken/Jounzal of Financiul Economics 44 (1997) 169-203
Table 5
Regmsion of one-year CRSP equal-weighted index real returns on book-to-market or dividend
yield: 1926-1991
__ ____ _c_I__-- -~_..-.--- -.-- .-- .-“.-_..----_---.----_._--______
f3 % Mean h(%) p-value LKLHD Mean h(%) pvalue LKLHD
---- ~-
Book-to-market tesults Dividend yield results
~---
0 1.88 0.001 0.001 1.23 0.127 0.041
I 2.77 0.002 o.m3 2.20 a167 0.047
2 3.83 0.004 0.004 3.32 0.213 0.067
3 4.91 0.002 0.003 4.37 0.279 0.075
4 5.91 0.007 O.&-i8 5.39 0.352 0.084
5 6.83 0.004 0.010 6.27 0.426 0.095
6 7.85 O.OlO 0.012 7.26 0.518 0.100
7 8.79 0.014 0.017 8.28 0.617 0.101
8 9.77 0.022 0.03 1 9.31 a692 0.092
9 10.81 O.OF p-036 10.24 0.77 I OB80
10 11.80 0.040 (?NG 11.10 0.8-36 0.069
11 \- -‘lr \l.< - O.W+ : 2.43 0.902 0.051
12 13.84 0.075 0.089 13.29 0.937 0.036
13 14.64 0.103 0.W 14.24 0.955 0.025
14 15.81 0.149 0.148 15.23 0.974 0.016
15 16.80 0.198 16.49 0.9R8 a009
16 17.77 0257 0.227 17.26 0.990 a005
17 _ 18.25 0.996 o.aI3
18 -. 19.30 0.998 0.002
19 - - 20.36 0.999 a001
S.P. Korhari, J. Shanken/Jmmal of Financial Ecpunics 44 (1997) 169-203 187
the minimum expected return is zero. This ensures that the lowest implied
expected return over 1926-91 is the negative of the sample mean return or
- 1X3%, as discussed above.
‘The bootstrap standard error is 4.3% under the null hypothesis that fl = 0,
and it varies very little as the true value of /? is altered in the simulations. It is
higher than, but close to, tbe OLS standard error, 4.1%. Tbe OLS estimate is
biasedupward, however, as discus& earlier. The averap simulated vabre of the
OLS estimator given in the second cohmn tends to exoeed the true vahre by
M-1.9%, a bit higher than the 1.75% predi&d using Stambaugh’s formula.
The bootstrap revalue in the third column of Table 5, which is actually
a simulation estimate of the p-vaJue, is 0.001. This is the proportion of tbe 2500
simulated (under the null) estimates that exceed the historical slope of 20.2%.
The standard error of this estimate, based 00 a binomial analys& is the square
root of (0.001~0999~/2500, which is less than 0.001. The standard error for
a p-value do.01 (/3 = 6) is 0.002. Having strongly mjected the nuU hypothesis of
a zero slope, we would like additional information tqarding the true value offi.
Examining the p-values in column 3, we see tbat values of /3 up to 6% can be
rejectd at the 0.01 levei, whik vahres up to 10% can be r&cted at the 0.05 levd.
At this slope value, a move done-standard deviation in B/M implies an &tease
in expected return of a whopping 1000 basii points.
WenowshiCtthefocusfmm~~ngwhicbvaluePoCBcanbe~~at
conventional significanot leveks to which values have highest likelihood, amdi-
tional on the one-year histori& slope eatima&. The likdi reported in
cohunn4ofTablc5starta.0.001 foraslopeofzeroand - (almost)
monotonically to 0.23 at fl = 16% the maximum co- These likelihoods
(or posterior probabilities under the uniform prior) an computed using a neigh-
borhood consisting of points within 1% of the historical skope, and are plotted
in Fii 1 (see triangles). The values of beta with high likelihood cormspond to
huge levels of expected return variation. Under the (naive) uniform prior. most
of the posterior probability exceeds /I = 8%. which is the largest value of beta
that ensures that conditional expected returns (fitted values) are always non-
negative over the 1926-91 period.
In genera& the ratio of posterior probability for a nul! hypothesis to posterior
probability for an alternative (the ‘poste&r sjds ratio’) equals the prior odds
ratio, p/(1 - p)t times another ratio often r&erred to as the l ayes factor’ (see
Je&eys, 1957). This f&ctor is tbe weighted average likelihood under the null
bypotbesis divided by the weighted average tikelilrood under the alternative,
with weights determined by the prior. In a test of the null hypothesis tbat there is
no variation in expected returns r&ted to B/‘&4, the Bayes faclor is just tbe
likelihood at /3 = 0 divided by a weighted average of the likelihoods for positive
betas. From Fig. 1 it is clear that the Bayes factor wi’tl always be less than one
(over the range of slopes considered) and it will be close to zero for any prior that
puts substantial weight on the larger values of beta. Thus we have a sharp shift
188 S.P. Kolhari. J. Shonken iJournu1 of Finonciol Economics 44 (1997) 11% 203
in beliefs away from the hypothesis /3 = 0, particularly for those individuals who
believe that expected returns are sometimes negative.
The bootstrap evidence from the regression of equal-weighted index returns
on dividend yield is reported in columns 5-7 of Table 5. The maximum /I values
considered correspond to a minimum expected return of - 15.3% in each case.
The range of slopes for B/M is smaller than that for dividend yield since the
minimum value of B/M is more extreme whea measured in units of standard
deviations from the mean. The bootstrap standard error is 4.9% under the nufl
hypothesis fi = 0. It is higher than, but again close to, the OLS standard error of
4.8% in Table 2. The OLS bias (difference between the first two columns) ranges
S.P. Krdwi. J. Shot&en, Journal oj Financial Economics 44 (I 997) 169 203 189
from 1.2% to ! 4%. comparable to the 1.3% predicted using Stambaugh’s bias
formula. The bias is roughly 20% of the value of the historical slope, 6.7.
The bootstrap pvaluil for @ = 0 is 0.13, close to the pvalue of 0.12 based on
the t-statistic of 1.39 in Table 2. Although we cannot reject the null hypothesis of
a zero slope at conventional levels. the likelihood function in Fig. I (see squares)
rises ini:ially, peaks at /3 = 7%. and declines steadily after that. Thus, the Bayes
factor will be less than one, i.e, will favor the conclusion that expected returns
do vary, except for priors that place most of the weight on very large values of
j!I(greater than 11 o/6) Therefore, anyone who believes that expected returns were
positive over the pziod 1926-91 (B 5 9%) will become more confident that
expected returns vary with dividend yield. The maximum possible shift in odds
is by a factor of 2.5 = O.lOl~WMl. the ratio of the maximum hkelihood to the
likelihood at zero. On the other hand, an investor with prior beUs heavily
weighted on #I values greater than 11% will become less confident that expected
returns vary with yield. This is a relative statement, though; such an investor will
still be confident (high posterior probability) that /3 # 0, but this will be due to
the prior dominating the sample evidence.
This section discusses bootstrap simulation results for the 1941-91 and
1963-91 subperiods. The Ml-91 subperiod excludes the Great Depression
period of high return volatility; the more recent ml-l%2 subperiod is of
interest because it coincides with the period over which much crossaectional
examination of book-to-market ‘ratios has been undertaken.
‘Not surprisingly, given the smaller time-series sample and greater persistence
in the independent variables over the subperiods. the estimated slope biases
190 S.P. Kolhari, J. ShankeniJournal of Financial Economics 44 (1997) 169-203
Table 6
Regression of one-year CRSP value-weighted +dex real returns on book-to-market or dividend
yield 19261991
--I__-.-__--_- .-.- ._--.- -- ..- ~_- .--.- - _-- -.-.-- -.-_, ..-- __._.-.--._-._-l__
p % Mean b( %) p-value LKLHD Mean h(%l pvalue LKLHD
--------- -----II__
Book-to-market results Dividend yield results
from Eq. (3) are larger for the subperiod estimators. For example, the simulated
bias for the B/M slope (not rqorted in tables) exceeds 4.0 for both indexes over
196391. &s-adjusted t-statistics are computed by subtracting the simulated
mean value (under the null fi = 0) from the estimator b and dividing by the
standard deviation of the bootstrap slopes.
Table 7 presents bootstrap p-values for /? = 0, along with pvalues based on
the OiS r-statistics and bias-adjusted t-statistics. ksults for 1941-91 are given
S.P. Kothori. J. Shanken/Joumai of Financial Eronomics 44 (1997) 169 -203
in panel A, while results for 1963-91 are in panei B. As expected, the bias-
adjusted pvalues are often much larger than the conventional CXS pvalues.
However, when the values are relatively low, the bias-adjusted pvaks are
smaller than the bootstrap p-values. This apparently reflects nonnormality of
the bootstrap distribution for the slope estimator, with the bootstrap having
a fatter upper tail. For example. the adjusted p-value is OJlNi in the equal-
weighted dividend yield case (Ml-91), while the bootstrap’pvalue is 0.048. The
binomial standard errors are 0.002 for the 0.00s pvafue aad 0.004 for the 0,048
p-VillUe.
We focus on the more conservative bootstrap pvalua. In the post-1940
period, the hypothesis of no related expected return variation is rejected at the
132 S.P. Kothari. J. Shanken /Journal 01 Financial Economics 44 (19%‘) 169. 203
Table 7
P-values for the null hypothesis fi = 0: subperiods 1941 91 and 1963-31
-- _____---. _-___--__----
index Financial ratio OIS Bias-adjusted Bootstrap
Panel A: 1941 91
Panel B: 1963-91
0.05 I+, except far the combination of value-weighted index and ?M. The
smallest pvalue is 0.024 for the regression of value-weighted returns on dividend
yield. In the past-1962 period, rejection at the 0.05 level is only observed for the
eqtutCwe&hted index and dividend yield.
The bootstrap likelihood functions provide additional insights into these
results as well as the potential contrasts between classical and Bayesian
infmnce in all but one case, an investor who 6rmly believes that expected
returns are never negative will come away from the regression evidem~ with
greater conti that expected returns vary with B/M or dividend yield. This
is true .~ven for the post-l%2 scenario with B/M and the equal-weighted
index (with bootstrap pvalue 0.2Sk although the maximum shift in odds
(Bayes factor) is only 1.7 in this case. On the other hand, in most q
the direction of the shift in beliefs for an investor who firmly believes that
expected returns are sometimes negative depends on the specifics of their prior
distribution.
S.P. Korhari. J. Shanken ‘Journal of l;inonciai Economics 44 (1997) 169-203 193
The post-1962 likelihood function for the equal-weighted index and dividend
yield rises steeply and stays well above the likelihood at j3 = 0 for the entire range
of j? value& This will be viewed by ail investors (for the class of priors considered)
as evidence in support of time-varying expected returns. At the other end of the
spectrum, the value-weighted, B/M, post-1862 iikelihood function for the value-
weighted index and B/M steadily deciines as 6 increases, shifting all priors toward
the con&&n that B/M does not track variation in expected mturns for the
v&re+veighted index. One final case wortb highlighting is the post-1962 scenario
for the value-weighted index and dividend yield. Here, the likelihood function
drops below the likelihood at fl= 0 only at the largest value co&&red (j? = 9%).
Thus vittuaUy all investors considered wiii view the evidence as supporting
a relation between dividend yield and expected return, even though the c&s&al
test fails to mject :he null hypothesis of no relation at the 0.10 level.
The Dow Jones B/M has been about 0.4 during this period, fairly low, by
historical standards. However, expectations of a low market return have been
foiled by the stellar market performance in the 1990s to date. The Dow Jones
B/M stood at its historical lowest, 0.27, on April 1, 1995, but the market
performance since then has been one of the best in the post-Work! War II
period. pe Dow Jones B/M is currently (August 1996) only 0.23.
In the remainder of this section, we consider formal tests of the hypothesis
that expected real returns are always non-negative over the given period of
interest. By the law of iterated expectations, if expected returns are non-negative
conditional on all in’ormation availabIe at each point in time, then expected
returns must also be non-negative conditional on any subset, in particular when
conditioning on B,‘M or dividend yield.
From Table 5, we see that the hypothesis that expected returns are non-
negative over the 1926-91 period is rejected at the 0.02 Ieve! (p-value for j? 5 8)
for B/M and equal-weighted index returns. The corrqumdiag pvaiue for B/M
and value-weighted index miums in Table 6 is about 020 (for jf I 5). Thus, we
cannot rejeck at conventional signiticance &eve4 the hypothesis that the nega-
the fitted vahres for the value -weighted index, in Fig. 4, am due to error in
estimating /I. There is no evidence of negative expected returns in the dividend
yidd results.
As explained in Seaion 6 these non-negativity tests implicitly assume that the
sample mean of the return disturbances+ i& is zero for the given period. We now
compare this test to the conventional regmssion procedure for making infer-
ences about a conditional mean. The eslimated expected return @ted v&e) at
the minimum sample value of B/M (27.2% for 1929) is - 22.3% for the
fzqual-weighted index. The corresponding standard error, based on the usual
regression formula, is 17.3%. This yields a t-statistii of - 1.3 and a pvaiue of
0.10 for the hypothesis that the minimum expected return is positive over the
1926-91 period.
The hirly large increase in p-value (0.02 to 0.10) re&cts the fact that the
conventional procedure incorporates uucertaint~ about the sample mean distur-
bance term, whereas our ~-based test con&ions on a t&&about tbe magnitude
of i&l Due to the substantial variability in return surprises, the conventional test
for non-negative expected returns is not powerful. Conditioning on the mean
disturbance permits us to concentrate on the slope cxxfkieut and the impha-
tions of time variation for nonnegative expected returns, holding the average
level of expected return constant at the sample mean. since we focus on
a narrower issue, the power of our test is, not surprisingiy. enhanced.
Now, suppose that (relative to one’s priors) the data indicate that the mean
return surprise, ri, is positive.8 For example, the literature on the equity premium
puzzle(Mehra and Prescott, 1985) suggests a lower prior expected real return on
the equal-weighted index than the sample mean of 15.3%. Alternatively, one
might believe that the observed sample of US. returns is subject to a selection
bias related to survival and grdwth of equity markets at the country level
(Brown et al., 1995).
Suppose, far exarnk~ that half of the mean return is viewed as an ex post
surp* i.e, H = 7.65%. Given the analysis in -ion 6, this has the eflkct of cutting
in half the maximum value offi copsistent with non-negative expested tiurns The
associated p-value (for /I 5 4) is 0.002 as compared to 0.02 (for /I I 8) earlier.
Naturally. incorporating a belief that the return surprises are positive, on avm
leads to a stronger rejection of non-negativity. since our belief about the general
level of expected returns is lowered relative to that of the previous scenario.
1.m
0.90
0.80
0.70
OM
oxl
OAO
O.Xl
0.20
QIO
om
Fig 5. Posterior probability that cxpeded real rcturxs arc ~-negative 1926 1991
Poskrior prob&ilitics arc derived from tk like&xxi timctionr plotted IB Pigs. I and L with prior
distributiom spedid over the ranges of beta values in those Crgurcs. Beta r&es in the first half of
each range aIri cxuwistcnl with nonlltgrtivc expected ITtumi. v-h& betas in tbc z3cxmod half yield
negative expeucd mtums at some point in the 1926-1991 period The prior distributions arc
uniformly distributed within each MF. prior probability is the wet&t assigned to the first ItaIfofthe
b-raagp.
irt which the Bayes factor is one. The low va!ue of the actual ibyes factor in the
equal-weighted B/M case ensures that the posterior probability curve iies beiow
the no-change tine for p between zero and one. This means that. regardless of the
prior (p # O,l), one ends up less confident that expected returns are always
non-negative (i.e., 0 I /I I 8%). In particular, an individual who is initially
198 S.P. Kofhari, J. Shanken/Joumal of Finaaciai Economics 44 (1997) 169-203
uncommitted @ = 0.5) with regard to the hypothesis that expected returns are
non-negative has posterior probability 0.08, while an individual with prior
confidence in the non-negativity condition, say p = 0.8, has posterior probabil-
ity 0.26.
Now, as we did for the classical test of non-negativity in the previous
subsection,we considerthe alternative assumptionthat iJ, the mean surprise
componentof return, is 7.65%. As before,the maximum /I consistentwith
non-negativeexpectedreturnsis reducedto 4%. In addition, the rangeof frs
entertainedas plausiblea priori is likewise cut in half,with a maximum fl of 8%.
At this value,the most negativeexpectedreturn overthe 192691 period is now
- (F-‘ii)=- 7.65%.Given the likelihoods reportedin Tabie 5, the resulting
Bayesfactor(ratio of averagelikelihoodsfrom 0 to 4 and 5 to 8) is 0.226,almost
threetimes the Bayes factor obtained under the si = 0 assumption. The corres-
poadingposteriorprobability for non-negativeexpectedreturnsof the l corn-
mitted’ (I, = 0.5) investor is now 0.18. as comparedto 0.08 above.Truncation
:td\. a mLl;n?al Pffr?! ia *&i; rec;LIx.
The increasein posteriorprobabihtymay be surprising,sincetheclassicaltest
rt$cctd non-negativitymore stronglywhen U = 7.65*/s.The reasonis that the
prior associatedwith I?= 7.65% eliminates the weight previously placed on
the very largevaluesof fl(> 8) that haveextremelyhuge likehhood,given the
historical slopeof 20.2%(seeFig. I). Thus, the increaseis a consequence of the
changein prior implied by the given beliefabout Ti (maximum 8 reducedfrom
16% to 8%), in conjunction with the particuhtr behavior of’ the hkelihood
bmction in this example.
in the value-weightedB/M case,the posteriorprobability curve again lies
belowthe no-changetine,despitethe relativelylargepvalue of020 (for /?s 5%
in TaMe 6) fof the null hypothesisof non-negativeexpectedreturns. The
posteriorprobabilitiesarekss extremethan thosefor the equal-weightedindex,
however.For exam& an observerwith p = 0.5 comesaway from the B/M
evidemzwith posteriorprobability 0.25that expectedreturnsarenon-negative.
Posteriorprobability cutvesfor dividend yield in Fig. 5 he abovethe no-change
line, mflectittg&teased confidencein the non-negativityhypothesis.Curvesfor
the 1941-Wsubperiod,given in Fig. 6, all heabovethe no-changeline as we!1
l.CQ
0.90
om
am
0.60
o.50
04
OS
o.20
a10
a00
F*tX Post&or pmbability that crpadcd teal tcWntb ate nonacyUrcc: IW-1991
PostetCprobabiiities on derived ftuttt the IikeIibood functions aad prior &tributiot~ spedal
overthecomspondiagraagsofbctaval~Becav~afintfwhr;t~oreachoreach~arrcorYsirteot
witb non-negative expdcd returns. while betas in the second ltd+d ibegMive expatcd teturtt5 at
sotne point in the WI-1991 period. The prior distribrttions ate uttironnly distribulcd with crh
Ita& prior probabiiity is the weight assigned to the lirst half of the beta range.
With the equal-weighted stock in&z, the bootstrap pvalue for the null
hypothesis that expected returns are unrelated to B/M @ = 0) rises to 0.005 with
tmncatioo, from G.001 earlier, the pvalue for the hypothesis of non-negative
expected rptums (6 I 8) increases to 0.084, from 0.022 when 1933 is included
(Table 5). Corresponding p-values for the value-weighted index are 0.045 (0.016
earlier) for no relation. and 0.27 (0.20 earlier) for non-negative expected returns.
Perhaps more informative, given the limitations of p-values. are the posterior
probabilities. Now, an uncommitted, i.e., p = 0.5, observer comes away from the
equal-weighted B/M evidence with posterior probability 0.14 (0.08 earlier) that
expected returns were non-negative over the 192691 period. For the vafue-
weighted index. the posterior probability is 0.34 (0.25 earlier). Thus there is still
a shift in beliefs, albeit a weaker one. away from the non-negative expected
returns hypothesis.
9. Summaryad coaehasioas
We find reliable evidence that both dividend yield and B,,M track time-series
variation in eawted real one-year qtock returns over the period 1926-91 and
Lb< sub~“-~iJd ??‘! 141 The P X r&lion is stronger over the full period. while
the dividend yield relation is strongerin the subperiod.
Jhe estimated 1926-91equal-weightedreturn slope on B/M implies chat
expectedreturnmovesby an astounding20percentagepoints(18bias-adjusted)
for a one-standard-deviation drarrgein B/M. The coefRcientis so large that
implied expectedreturnsare sometimesnegative.particularly in the late 1920s
whenthe B/M ratio is relativelylow. The resultsarequite sensitiveto the 1933
data point, however,when both B/M and market index returns attain their
maximum values.Using a +truncation’method that reducesthe impact of this
outher. the pvahte for non-negativeexpectedreturns is 0.08 (0.02 without
truncation).ASthoughthe pvalue of O.Ogmight be interpreted by some as
‘insignificant’,the posteriorprobability for an investorwho assignsprior pro&
ability 0.5 to the hypothesisof non-negativeexpectedreturnsis only 0.14(O.Orr
without truncationt Resultsfor the vahte-weightedindex and B,/M arequalitat-
ively simihtr, though the &i&s in beliefsare far lessdramatic.
Thus,theB/M resultssuggestthat expectedreturn variationover the 192691
periodwasnot drivenentireiy by equilibrium changesin compensationfor risk.
Rather,It appearsthat the market may havebeeniocfficieat,particularly in the
late 1920sand early 193oaa period of greateconomicvolatility. Ihe stronger
resultsfor the equal-weightedindex, which is infiueneedmore by small-firm
returns,is consistentwith this conclusion,although greatervariation in risk-
rehcd predictabiity is also piausiblefor smallerfirms.
The market ineffciency conclusionshould be temperedby severalobserva-
tions. First, them is the familiar elementof data mining which implies that the
sign&ance ofextremerest&sis overstated,althoughby how much it is difhcult
to say.Whiie the6nancialratiosconsideredherecan bemotivatedtheuretically,
dividend yield is examinedin part becauseof its prominenceas comparedto,
say,earningsyield in pas:time-seriesstudies(elg.Fama and French,1988).B/M
.s.i? Kothari. J. ShankenlJomnal of Fincmcial Economics 44 (1997) 169-203 201
Reqer. J.. 1985. !htisticd Da&ion Theory and Bay&an Analysis. Springer. New York.
tBadoukC J, Rtcbhoo. M, Whit&w. R, 1997. Nodincaritie in tbc re&ti*n bctwccn the quity
risk premium 6ad the teml structore. Munag6wkcnt scirnac. forthcan~
&at, WJ, Komjcsyk. IA., MM. On sckdon birsa in book-to-ma&t based tests of awet
pricing mudds. Workimg pap no. 167. Nortbwcstcm University. Evans~oa. IL
Brown, S., t3oamma W, Rosa s. t995. !bvival. Journal of Fiinaz 50,853 -873.
Campbe@. J, Saber, It, 1988 Ibe divided price ratio ad exprcatieas of future divicknds ad
ttismmt factom Review of Fintmcid Stdics 1,195-22R.
Glum LK.C., Je@ceh N, &kohbok, J, 1995. Evduatiyy the prlomru*r ofvaluevcrxus
~uodrsthe~~oCsekaionbL+Jounul~FinsncLlEconanics38269-296.
Davk J.L. 1991. The nosrcaeaIon of realid stock returnr: the pre+COMPlJST.AT evidmcc
Joumd of Filmna 4x1579- 1593.
fxhot.M.. 1982.cominml on Lindky’spsrpbox.JoumafoftheAmcrican std.dcdAssaiation
77.336.339.
Dixon. W, htaswzy. F, 1969. Introduction to !Gta~isticaIAnalysis. McGraw-Hill. Nen York.
Fame, El); Fnmch, K-R, t988a. Dividend yields and expected stock mums. Journal of Financial
E.commii 22 3-25.
Fpllla, EF. Fnwtcb. K.R.. !988b. Permanent and transitory components of stock pricft Joumd of
Po&tiad Ecotmtny 96.24&273.
Fama, EF, French, K.R 1989. Bpriaas cotiditionzz and expected retims on stocks and bonds.
Journal d Fitmtwhl Eaxmomb 2% 23-49.
Fame, E.F, French. K-R., $992 The cross-section of expect-J mums. Jouma! of Finamx 47,
427-465.
Fama, EF., Ftench, K.R., 1993. Common risk factors in the returns on bonds and stacks Journal of
Financial Economics 33,3- 56.
‘The bootstrap programs usixi in our ~putations were written in Gauss and am available on
req-
S.P. Kothati. J. ShankenlJoumaI of Financial Ecnnomics 44 (1997) I69- 203 203
Irma. E F., Schwert. G.W.. 1977. Asset returns and inflation. Journal of Financial Economics 5,
115-- 146.
Goetzrnann, W.N.. Jorion, P.. 1993. Testing the predict& Egwer of dividend yields. Journal of
Finance 48,663-679.
Hodrick R, 1992. Dividend yields and expected stock tetumsz alternxtiie procedures for i&rencc
and measurement. Review of Financial Studies 5.357-386.
J&S. H, 1957. Tbawy d Probability. 3rd cd. Onford University Press, London.
Kandel S.. ‘%XMocb R, !%ambaugh. RF, 1995. Baytin inrmncc and mobo &ciency.
Review of Financial Studies 8, I -U.
Kaadell S.. strcnbaugk R.F. 1996. C&t the praMaMity of stock CC(UIIK: an asset-allocation
perspective. Journal of Finance 51.385424.
KC& D.B., -b&h. R.F.. 1% Predicting returns in the strzcir and bond marke& Journal of
Finam5ai Econo&cs I :. 357-390.
Kendall. MG, 1954. Note on hias in estimation of autouxreia tioa. Biometrika 41.403 406.
Kothari. SP, !%ankcn. I, 1992 Stock return variatian and expcded dividcnhs a time-series and
cr-1 analysr Journai ol,Finan&i EconrmicJ 31, 177~~210.
Kotbari. SP, Shanken. J, Sloan, R.G., 1995. Anscber iook at the cross-se&on of expected mumt
Journal of Finance 50. I?%224.
Lindky. D., 1957. A statistical paradox. Biomcrrika 44. 187.. 192s
titzaaherger. R.H, Rarnaswaq,. K, 1979. The dfcrt d puxonal (ua and &Mends on capital
asset prices theory and emprkai evidene Joumai of Financial Economics 7. 163- 195.
Lou@tratt.T, 1996. Irtbererbooh-twnarketdkrt? Workingpaper.UniverrityofiowaiowaCity,
IA.
Ma&e&J, Emanud. D.. 1993. -Fact& asset aikxzationz pros and 0~1s Fii Anolystr Journal
NOVMlkr- Duunber. m-43.
Mankiw. N.G, Shapim M.D.. 1986.0 *vc reject tao &en? !5maii sampdc prqcrtvr of tests of
rational expectations inodek iz&wna. ic Letters 20. t39- 145.
Markowita H, Uxmen. N, 1996.7hc iik&:.Jod ofvarious Yaackmarkellwlmdistri~~PartPIUt1:
prMp&sofit&renceJoumaioCRiskardUnccrt&ty13.207-219.
&bra. R, Prescott E.. 1985. The equity ptmniuux a puzzle. Journal of Monetary wrnics 1%
145- 161.
?Wson. C.R.. Kim, MJ, 1993. Pr&ictabie stock returns the mie Oc smali sampie bias. Jouratai of
Fituna 48.64-661.
CM&r. R.R. 1973. The variability Oc the market factor otthe New York Stack Exchan~ Jowui of
B&lsims444j4-453.
Pond J, S&all L, t995. M-to-market as a predictor of market returns Workm8 paper.
university or washingtoll, Seattle.
RoEdt M, 1984. Dividend yields are equity risk premiurnr Journal of Port!&> Management I I.
68-75.
schwcrt. G.W., 1989. Why does stock market voiatillity chasp over tmmc? Journal d FiwM;c 44,
1115-1153.
Sbanken. J., 1987a Muhivariate proxies and asset pricing rdations livinp with the roil critique.
Journat of Financial Economics 18.91 -I IO.
Sbanken. J, 1987b. A Bayesian approach to testing prtfoiio e&ienq. Journal of Financial
Economics 19. 195-215.
Shiiier. RJ, 1984. Stock prim and sxial dynamics. Brookingx Papers on Economic Aciivity 2.
457-498.
Stambaugh, R-F., 1986. B& in regrcssiom with lagged stochastic qsssors. Unpubiisbed manu-
tipt. University of Chicago, Chicago, IL
White, H.. 1980. A heteroskedasticitysonxistent covarianoz matrix estimator and a direct test of
beteroskcdasticity. bnom&rica 48.817-838.