Journal of Banking and Finance 7 (1983) 135-146.

Nort h-Hol l and Publishing Company

A SIMPLE EXAMI NATI ON OF THE EMPI RI CAL RELATI ONSHI P
BETWEEN DI VI DEND YIELDS AND DEVIATIONS FROM THE
CAPM
Edwin ELTON and Martin GRUBER*
New York University, New York, NY 10006, USA
Joel RENTZLER*
City University of New York, New York, NY 10010, USA
Received Oct ober 1981, final version received March 1982
Several papers have been published in recent years dealing with bot h the theoretical and the
empirical impact of dividend yields on security returns. Dividends have been postulated as
affecting stock returns because of tax effects, agency costs and the Wealth Transfer Hypothesis.
In this paper we perform a purely empirical examination of whether and to what extent
deviations from the zero bet a form of the CAPM are explained by divident yields. The paper
demonstrates that dividend yield has a large and statistically significant impact on return above
and beyond that explained by the zero bet a form of the CAPM. This is consistent with the
findings of Litzenberger and Ramaswamy. In addition our results are consistent with the
findings of small firm effects.
1. Introduction
Several papers have been published in recent years dealing with both the
theoretical and the empirical impact of dividend yields on security returns.
Dividends have been postulated as affecting stock returns because of tax
effects [see Elton and Gruber (1970, 1978), Litzenberger and Ramaswamy
(1979)],.information effects [see Aharony and Swary (1980), Taub (1976),
Pettit (1972, 1976) and Watts (1973, 1976)], agency costs [see Jensen and
Meckling (1976)], and the Wealth Transfer Hypothesis [see Kalay (1980)]. 1
In this paper we perform a purely empirical examination of whether and to
what extent deviations from the zero beta form of the CAPM are explained
by dividend yields. The paper demonstrates that dividend yield has a large
and statistically significant impact on return above and beyond that
explained by the zero beta form of the CAPM. This is consistent with the
findings of Litzenberger and Ramaswamy (1979). In addition our results are
*The authors would like to thank Dext er Corporation, Windsor Locks, Connecticut for
funding this research.
IThe Wealth Transfer Hypothesis refers to the ability of stockholders to transfer wealth from
creditors to stockholders by paying out assets in the form of dividends.
0378-4266/83/$03.00
136 E. Elton et al., Dividend yields and CAPM deviations empirical relationship
consi st ent wi t h t he fi ndi ngs of small firm effects whi ch ar e doc ume nt e d by
Roll (1980), Banz (1981) a nd Rei nganum (1981). The next sect i on of t hi s
paper descri bes our pr oc e dur e while t he fol l owi ng sect i on pr es ent s t he
results.
2. Methodology
The s t udy ut i l i zes t he mo n t h l y dat a on di vi dends, pri ces a nd r et ur ns for
New Yor k St ock Exchange securities avai l abl e on t he Uni ver s i t y of
Chi cago' s CRSP t ape. The per i od covered by this s t udy is f r om J a n u a r y 1927
t hr ough De c e mbe r 1976. Since we are i nt erest ed i n exami ni ng t he
r el at i ons hi p bet ween devi at i ons from t he zero bet a CAP M a nd di vi dend
yield, we mus t cal cul at e est i mat es of each of these vari abl es.
Ou r fi rst st ep was t o es t i mat e t he zero bet a CAPM. I n or der t o r educe t he
er r or in me a s ur i ng bet a whi ch is used as i nput in est i mat i ng t he CAP M, t he
por t f ol i o gr oupi ng pr oc e dur e empl oyed by several pr evi ous r esear cher s was
empl oyed [see Bl ack, Jensen and Scholes (1972), and F a ma a nd Ma c Be t h
(1973)]. The bet a for each st ock was cal cul at ed based on five year s of d a t a
(e.g., 1927-1931) a nd usi ng t he CRSP val ue wei ght ed index. 2 All st ocks wer e
r anked a nd gr oupe d i nt o t went y port fol i os on t he basi s of t hese bet as. The
bet a for each por t f ol i o was t hen cal cul at ed usi ng da t a for t he next five year s
(e.g., 1932-1936). Thi s bet a was t hen empl oyed as t he best es t i mat e of t he
bet a of each por t f ol i o for 1937. Thi s pr ocedur e was r epeat ed changi ng each
of t he 5 year per i ods by 1 year so t hat we had bet a est i mat es for 20
por t f ol i os for 40 year l y per i ods (1937-1976). 3 Fo r each of t hese year s t he
aver age r e t ur n on each por t f ol i o j was cal cul at ed and t he fol l owi ng cr oss-
sect i onal r egr essi on was r un:
/~) = 75 + ?~fl~ + e'~, j = 1, . . . , 20, where
/ ~ = aver age r e t ur n on por t f ol i o j in year t,
fl~ = es t i mat ed be t a for por t f ol i o j in year t,
?5, ?~ = cr oss- sect i onal coefficients est i mat ed for year t,
~ = r a n d o m e r r or t e r m in year t.
2The analysis in this paper was also conducted using the CRSP equal weighted index. These
results were indistinguishable from those reported in this paper.
3The composition and size of each of these portfolios changed from year to year. The number
of stocks which had the required dat a over each 11 year period and so were placed into
portfolios varied from a low of 419 to a high of 880 with an average number of 740.
E. El t on et al., Dividend yields and C A P M deviations empirical relationship 137
The excess ret urn a~ for each security i i n year t was then defined as
t ~ - - t " t " t t
+ ~ l f l i ) , i = 1 , . . . , o~i - R i - (Yo N ,
where fl~ is the estimated bet a of the portfolio to whi ch security i belongs, ~ ,
~] are the est i mat ed cross-sectional coefficients for year t. Thi s value for
al pha was used as the best est i mat e of al pha for each of the securities in a
particular portfolio.
Our next step was to obt ai n a forecast of the di vi dend yield for year t
(year 11) where the first forecast of di vi dend yield was made when year 11
was 1937. In order to obt ai n this est i mat e we decided to use the di vi dend
yield on a portfolio of stocks rat her t han on each i ndi vi dual stock. There is
substantial forecast error in the est i mat e of a security' s yield. By utilizing the
dividend yield for a portfolio of stocks rat her t han an i ndi vi dual security we
hoped to reduce this forecast error. Stocks were grouped by their di vi dend
yield in year 9 - - two years before the forecast year. 4 Twent y groups were
formed. The first 19 groups were formed by ordering all stocks which pai d
positive di vi dends from hi ghest to lowest and then dividing these stocks into
19 equal groups. The 20th group cont ai ned all stocks t hat pai d zero
dividends duri ng year t - 2 (year 9).
Placing all stocks with zero di vi dends i nt o one group is i n cont rast with
the procedure used in some pri or studies [e.g., Black and Scholes (1974)].
Our purpose in groupi ng was to reduce forecast error while at the same time
obt ai ni ng portfolios with a wide range of di vi dend yields. In some periods
stocks with zero di vi dend yields were so numer ous that if we had an equal
number of stocks in each portfolio groupi ng, then a large number of these
portfolios woul d have had zero di vi dend yields. Tabl e 1 shows the percentage
of stocks i n our sampl e t hat di d not pay a dividend i n year t - 2 . For
example put t i ng an equal number of stocks in each group would have
resulted i n about hal f the portfolios havi ng zero dividend yield in the early
years and two or three with this zero yield in later years. Havi ng formed the
20 portfolios we t hen had to forecast the di vi dend yield for year t (year 11).
We used as our forecast the act ual di vi dend yield in year t - 1 (year 10). 5
The next step was to relate excess ret urn to di vi dend yield. Fi rst the
average excess return, 0?j, and average forecasted dividend yield, ~, for each
4The arguments behind grouping on dividend yields are directly analogous to the arguments
used in grouping on betas to provide better estimates. In forming groups we defined dividend
yield as dividends in year 9 divided by the ending price. Our purpose was to obtain a set of
portfolios with a wide range of yields in year 10.
5We have employed the simplest possible forecast model in this paper setting next year's
dividend yield equal to last year' s dividend yield. If we had found no dividend yield effect it
might have been due to the use of this naive forecasting model. However, the fact that we do get
strong results with this naive model highlights the importance of the dividend effect. A more
sophisticated forecasting model might lead to even stronger results.
138 E. El t on et al., Dividend yields and C A P M deviations empirical relationship
Table 1
Percent of stocks with zero dividend yield in
grouping year.
Grouping Forecast All Stocks under
year year stocks $5 eliminated
1935 1937 5~4% 45.4%
1939 1941 44.0 32.0
1943 1945 23.6 17.8
1947 1949 13.6 10.6
1951 1953 11.3 9.3
1955 1957 11.5 10.6
1959 1961 14.5 13.8
1963 1965 13.2 11.5
1967 1969 9.1 9.1
1971 1973 16.2 14.8
1974 1976 11.0 6.1
portfolio j was computed. The average was taken over the 40 forecast years
(1937-1976). Then a cross-sectional regression of the form
ffj = Ao + Aa ~ + g~, j = 1, . . . , 20
was run to obt ai n est i mat es of Ao and A t. This averagi ng process was
repeat ed usi ng all eight non-overl appi ng 5-year subperiods in our 40 year
sampl e giving eight ot her estimates of A o and A1.
As will become clear later, there were reasons to suspect t hat the zero
di vi dend gr oup of stocks woul d behave perversely. To test this, a regression
of the fol l owi ng form was run:
~j=A'o+A'~+A'zD+~, j = 1,...,20.
In this regression D is a dummy vari abl e havi ng a value of 1 for the portfolio
of stocks whi ch pai d zero di vi dends in the grouping year and a value of 0 for
all other groups.
Not e t hat this met hodol ogy is designed so that forecasts of bot h the betas
and the di vi dend yields empl oyed in the analysis use only dat a whi ch is
avai l abl e pri or to the dates over whi ch the returns are calculated. Thus, the
existence of an excess ret urn with respect to dividend yield indicates t hat an
i nvest or in t he absence of t ransact i on costs could have acted in a way to
ear n an ext ra return usi ng avai l abl e information. This woul d not necessarily
i ndi cat e t hat he could have earned an excess return because of the presence
of t ransact i on costs and addi t i onal non-market risk incurred. However it
does i ndi cat e an addi t i onal influence on equilibrium returns. Not e also t hat
our procedure is designed so t hat i f there is a bias it is against fi ndi ng a
E. Elton et al., Dividend yields and CAP M deviations empirical relationship 139
dividend influence. If betas are correlated with dividend yields then our two
step procedure would have a bias towards finding no dividend effect even
when one was present. This was a deliberate choice on our part. As has been
discussed in the literature, there is considerable disagreement on the
reasonableness of the procedures used for the j oi nt estimation of dividend
and beta effects. Our procedure, while biased against observing dividend
effects, avoids the estimation controversy. Thus, finding a dividend effect with
this procedure would provide strong evidence in support of such an effect.
3. Re s ul t s
Table 2a summarizes the statistieal relationship between the average yearly
excess return (alpha) and the average dividend yield in each of the 20 groups
where the averages are taken over the entire forecast period 1937-1976. By
examining the first row of dat a in table 2a we see that although there is a
Tabl e 2a
Cross-sect onal regressi ons of aver age excess r et ur n on di vi dend yield over 40 year
forecast per i od (t -val ues in parent heses).
Du mmy
Di vi dend var i abl e
For ecast per i od I nt er cept coefficient coefficient R 2
1937. 1976 - 2 . 2 6 3 0.344 0.188
(2.59) (2.04)
- 4.774 0.794 6.189 0.874
( - 10.87) (9.60) (9.64)
Tabl e 2b
Cross-sect i onal regressi ons of aver age excess r et ur n on di vi dend yield, over five
year per i ods (t -val ues in parent heses).
Du mmy
Di vi dend var i abl e
Forecast per i od I nt er cept coefficient coefficient R 2
193%1941 1.078 - 0.091 - 0.204 0.002
1942-1946 - 13.790 1.076 25.745 0.853
194%1951 - 1.298 0.221 1.211 0.037
1952-1956 - 1.082 0.216 - 0 . 4 1 9 0.077
195%1961 1.425 - 0.337 - 0.299 0.082
1962-1966 - 1.448 0.218 5.588 0.477
1967-1971 - 4 . 9 0 7 1.457 2.059 0.570
1972-1976 - 11.868 2.629 10.727 0.883
Mean - 3.986 0.673 5.551
(2.075) (2.062) (1.862)
140 E. El t on et al., Dividend yields and CAP M deviations empirical relationship
significant positive relationship between excess return and dividend yield
over the entire period, the relationship is small. The R 2 for the relationship is
only 0.188.
An examination of the dat a used in obtaining these regression results
(table 3) shows t hat the weak positive association may be due to the presence
of two distinct relationships. 6 Not e t hat portfolio 20, which was formed two
years before the forecast year using all stocks which did not pay a dividend,
has a high positive excess return (alpha). The return on this group is
considerably higher t han the return on any other group though it has by far
the lowest yield in the year prior to the forecast (year 10) and so the lowest
forecasted yield. These zero dividend stocks constitute an outlier group. This
finding is not surprising and has been noted earlier by Blume (1979). Later in
this section we will examine some possible explanations for the observed high
excess ret urn in this group. Before turning to this discussion we will first
eliminate the effect this group has on the dividend yield-excess return
relationship. This could be accomplished either by deleting this group from
the data or by adding a dummy variable which has a value of one for the
non-dividend portfolio group (number 20) and zero for all other groups.
These two procedures are identical in their effects on the intercept and slope
term. The advantages of using a dummy variable is that it allows a direct
measurement and a test of significance for the difference of behavior of the
excess ret urn in the zero dividend group. The results of including this
dummy variable in the regression for the entire forecast period are shown in
the second line of table 2a. Notice that now the positive relationship between
excess return and dividend yield is both large (0.794) and statistically
significant at the 1% level. Furt hermore the portfolio constructed from stocks
which had paid zero dividends has an excess return beyond what one would
expect given the behavior of the other portfolios. The value of 6.189 is
statistically significant at the 1% level. The inclusion of the dummy variable
also raises the coefficient of determination on the regression from 0.188 to
0.874.
As a further check on the form of this relationship, the regression including
the dummy variable was run for each of the eight five year non-overlapping
subperiods between 1937 and 1976. These results, reported in table 2b, are
supportive of the overall relationship. The relationship between excess return
and dividend yield is positive in six of the eight periods and on average the
regression coefficient is significantly different from zero at the 5% level. The
impact of the dummy variable is positive in five of the eight periods and its
average value is positive and significant at the 5% level.
The high positive excess return associated with zero dividend stocks is
i mport ant and wort hy of additional attention. Is this an effect caused by low
6In examining table 3 recall that group 20 has many more members than other portfolios.
Thus although most groups have negative alphas, the average overall alpha is close to zero.
E. El t on et al., Dividend yields and CA P M deviations empirical relationship 141
f ~
t ~
L
~J
* d
L
<
O ~
~ D
' 4
O ~
r-- ~.
o ~
e4 q
I
I
q ~
1
t r ~
I
o o I " ~
I
t ~ ' 4
I
I
' ,1 q
I
o O ~
I
~ ' 4
I
Q
I
o o
I
142 E. Elton et al., Dividend yields and CA P M deviations empirical relationship
dividends or is there an alternative explanation? The most obvious
alternative is t hat the result is due to a small firm effect. Small firms make up
a much higher proport i on of zero dividend stocks than they do of other
groups and a number of authors have found that small firms on average
have positive excess returns [see Roll (1980), Banz (1981) and Reinganum
(1981)]. Thus the zero dividend effect we observe may be, in part or total, a
small firm effect. At present there is no adequate explanation of the small
firm effect. Roll has an argument that surely explains part of the phenomena.
He shows t hat small firms have many more non-trading days then do large
firms. He then demonstrates that non-trading leads to a downward bias in
the beta. A downward biased beta leads to a positive excess return if a zero
excess return would exist using the unbiased beta. 7
An alternative explanation of the positive excess return on the zero
dividend stocks involves their price. Stocks under $5 in price are treated
differently by brokerage firms and financial institutions and the zero dividend
group of stocks contain a much higher proport i on of these low price stocks
than do the ot her dividend groups. 8 Stocks which sell for less than $5 are
not considered as appropri at e collateral for margin. Thus, in figuring the
amount of assets in an account for purposes of margin, stocks under $5 are
eliminated. Therefore, an investor who borrows or uses margin would find
these stocks less attractive unless they offered a higher return. A similar
argument can be seen by examining fig. 1.
The efficient frontier with borrowing restricted to margin is ABCD. The
position of CD depends on the maximum borrowing that is available. If the
borrowing limit is lowered then, ceteris paribus, the efficient frontier involving
borrowing is lowered. Thus, stocks that reduce the amount that can be
borrowed will be held only if their return compensates for this lowering of
the frontier.
Expected
Return
A
0
~
j . . . .
>
Fig. 1
rt sk
7The use of mont hl y dat a rather than daily reduces the importance of this effect.
8Many financial institutions will not hold low priced stocks due to either transaction costs or
perceived risk.
E. El t on et al., Dividend yields and CAP M deviations empirical relationship 143
Low capitalization stocks in general sell for less t han $5. Thus low price
can serve as a proxy for low capitalization and it is difficult to be sure which
factor might be causing positive excess returns. Roll's argument concerning
infrequent trading of low capitalization stocks holds equally well for st ocks
selling under $5. However, the margin restrictions apply to stocks selling for
less than $5 and not to low capitalization stocks. Thus we decided that, in
order to separate the effect of low priced small capitalization stocks from the
effect of zero dividend payment, it was most relevant to eliminate stocks
selling for under $5 from our population. 9
Tables 4 and 5 are analogous to tables 2 and 3 except t hat all stocks
selling for less than $5 per share were eliminated. Examining the first row in
table 4 shows t hat eliminating such stocks had a major effect. The impact of
dividend yield on excess returns as measured by the regression coefficient has
grown from 0.344 to 0.549 and is statistically significant at the 1% level. The
coefficient of determination between dividend yield and excess ret urn has
changed from 0.188 to 0.471. Eliminating stocks under $5 per share reduced
the unexplained excess return associated with the zero dividend group. This
indicates t hat part of the peculiar behavior of this group was due to the
inclusion of these low priced firms.
However not all of the peculiar behavior of this group is eliminated by
removing $5 stocks from the sample. Table 5 shows that the return on
portfolio 20 which had previously paid a zero dividend is 0.86% whereas
the behavior of the other 19 groups would lead us to expect a negative excess
return for portfolio 20. To test for perverse behavior on the part of the zero
dividend portfolio, the regression was rerun with a dummy variable. The
results again improved. As seen in the second line of table 4a, the t-value of
the intercept and dividend yield coefficient both increase, the coefficient of
determination increases, and the dummy variable enters with a positive and
statistically significant (at the 1% level) coefficient.
Table 4a
Cross-sectional regressions of average excess return on dividend yield over 40
year forecast period with under $5 stocks eliminated (t-values in parentheses).
Dividend Dummy variable
Forecast period Intercept coefficient coefficient R 2
193%1976
- 2.623 0.549 0.471
(3.69) (4.00)
- 4. 398 0.865 4.445 0.841
(8.95) (9.36) (6.28)
9At a number of points in the study we eliminated low capitalization stocks rather than those
selling for less t han $5. This didn' t significantly impact on the results and so these results are not
reported.
144 E. Elton et al., Dividend yields and CA P M deviations empirical relationship
Tabl e 4b
Cr oss- sect i onal regressi ons of aver age excess r et ur n on di vi dend yield over five
year subper i ods wi t h under $5 st ocks el i mi nat ed (t -val ues in parent heses).
Du mmy
Di vi dend var i abl e
For ecas t per i od I nt er cept coefficient coefficient R 2
193%1941 - 0.789 0.369
1942-1946 - 5.054 0.632
1947-1951 - 0 . 5 1 8 0.101
1952-1956 - 0 . 6 1 2 0.111
1957-1961 1.119 - 0 . 2 6 9
1962-1966 0.416 - 0 . 1 9 2
1967-1971 - 4.631 1.437
1972-1976 - 8.869 2.085
Mean - 2.367 0.534
( - 2 . 0 7 2 ) (1.951)
1937-1941 --0. 093 0.246
1942-1946 1.402
1947-1951 - 1.316 0.218
1952-1956 - 1.185 0.201
1957-1961 1.176 - 0 . 2 8 1
1962-1966 - - 1.451 0.260
1967-1971 - 4.838 1.488
1972-1976 2.450
Mean - 3.707 0.748
(2.337) (2.443)
0.044
0.139
0.012
0.016
0.051
0.029
0.602
0.734
- 1.341 0.052
15.063 0.765
1.665 0.043
1.246 0.036
- 0 . 1 3 2 0.052
5.046 0.415
0.733 0.604
8.181 0.845
3.808
(2.102)
Table 4b presents the regression results for all five year non-overlapping
periods with and without the dummy variable. The results continue to
support the earlier findings that there is a strong positive association between
excess returns and forecasts of dividend yield except for the group of stocks
which had previously not paid a dividend. These stocks seem to sell at a real
premium above that which one would expect based on their forecasted beta
and dividend yield.
4. Conclusions
In this article we have examined the relationship between residuals from
the zero beta form of the capital asset pricing model and dividend yields. We
have found a persistent relationship between dividend yield and excess
returns. In particular, except for those stocks which had previously paid zero
dividends, the higher the dividend yield the higher the excess return. One
group of stocks, those which had previously paid no dividends, had excess
returns which were above what we would expect from this relationship. Part
of this differential was shown to be due to the effect of low priced stocks but
E . E l ton et al ., D i vi den d yi el ds an d C A P M devi ati on s empi r i cal r el ati on sh i p 145
. = .
~
p . .
0
0
<
r ' - ~ .
I
I
I
I
t
I
l
"4
' ~ "4
(",1
146 E. El t on et al., Dividend yields and C A P M deviations empirical relationship
an i nf l uence be y o nd t hat still seems t o exist. There seems t o be persi stent
pat t erns i n excess ret urns whi c h are rel ated t o di vi dend yield. Some of these
differences may be due t o tax effects. Ot hers have not as yet been adequat el y
expl ai ned.
Ref erences
Aharony, J. and I. Swary, 1980, Quarterly dividend and earnings announcement s and stock-
holders' returns: An empirical analysis, Journal of Finance 35, no. 1, March, 1-12.
Banz, R., 1981, The relationship between return and market value of common stocks, Journal of
Financial Economics 9, no. 1, March, 3-18.
Black, F. and M. Scholes, 1974, The effects of dividend yield and dividend policy on common
stock prices and returns, Journal of Financial Economics 1, no. 1, March, 1-22.
Black, F., M. Jensen and M. Scholes, 1972, The capital asset pricing model: Some empirical tests,
in: M.C. Jensen, ed., Studies in the t heory of capital markets (Praeger, New York).
Blume, M., 1979, Stock returns and dividend yields: Some more evidence, Rodney L. White
Center for Financial Research, Working paper no. 1-79.
Elton, E. and M. Gruber, 1970, Marginal stockholders' tax rates and the clientele effect, Review
of Economics and Statistics 52, no. 1, Feb., 68-74.
Elton, E. and M. Gruber, 1978, Taxes and portfolio compensation, Journal of Financial
Economics 6, no. 4, Dec., 399-410.
Fama, E. and J. MacBeth, 1973, Risk, return and equilibrium: empirical tests, Journal of
Political Economy 81, no. 3, 607-636.
Jensen, M. and W. Meckling, 1976, Theory of the firm: Management behavior, agency costs and
capital structure, Journal of Financial Economics 3, no. 4, Oct., 305-360.
Kalay, A., 1980, The dividend decision and the conflict of interest between bondholders and
stockholders, Working paper (New York University, New York).
Litzenburger, R. and K. Ramaswamy, 1979, The effect of personal taxes and dividends on capital
asset prices: Theory and empirical evidence, Journal of Financial Economics 7, no. 2, June,
163-195.
Pettit, R., 1972, Dividend announcements, security performance and capital market efficiency,
Journal of Finance 27, no. 5, Dec., 993-1008.
Pettit, R., 1976, The impact of dividend and earnings announcements: A reconciliation, Journal
of Business 49, no. 1, Jan., 86-96.
Reinganum, M., 1981, Misspecification of capital asset pricing: Empirical anomalies based on
earnings' yields and market values, Journal of Financial Economics 9, no. 1, March, 19-46.
Roll, R., 1980, A possible explantion of the small firm effect, U.C.L.A. Graduat e School of
Management, Working paper, Oct.
Taub, M., 1976, On the information content of dividends, Journal of Business 49, no. 1, Jan., 73-
80.
Watts, R., 1973, The informational content of dividends, Journal of Business 46, no. 2 April,
191-211.
Watts, R., 1976, Comment s on ' The informational content of dividends', Journal of Business 49,
no. 1, Jan., 81-85.