Costly dividend signaling : the case of lost firms with negative cash flows

MIT LIBRARIES 3

9080 02618 4603

I I ' .1,

ma

f"

)\r'u,

ii/i i';f

f

^itihf^ I

no

MM^*^'^^

DEWEY

MIT

Sloan

School

of

Management

MIT

Sloan

Working

Paper

4474-04

January

2004

Costly

Dividend

Signaling:

The

Case

of

Loss

Firms

with

Negative

Cash Flows

PeterR.Joos

and

George

A. Plesko

©

2004 byPeter R. JoosandGeorge A.Plesko.

Allrightsreserved.Short sectionsoftext,nottoexceed twoparagraphs,

may

be quoted

withoutexplicitpermission,providedthatfullcreditincluding

©

noticeisgiventothesource. Thispaper alsocan bedownloadedwithoutchargefromthe

SocialScienceResearchNetworkElectronicPaperCollection:

!

MASSACHUSETTS

INSTITUTE

i OF

TECHNOLOGY

DEC

3

im

Costly dividend

signaling:

The

case

of

loss

firms with

negative

cash

flows

PeterJoos

and

George

A. Plesko

FirstDraft:

November

2003

ThisDraft:January

2004

Abstract

We

examine

the dividend-signalinghypothesisin asample offirms for

which

dividend

increases are particularly costly,

namely

loss firms with negative cash flows.

When

compared

tolossfirmswithpositivecashflows,

we

findthepredictive

power

of dividend

increases for future returnonassetstobegreaterforlossfirmswithnegativecashflows, consistentwiththepredictive

power

ofthedividendsignalbeingstronger

when

itscostis

higher.

Our

results provide support for the dividend-signaling hypothesis

and

have broaderimplications sincelossfirms comprisea large

and

increasingshareof publicly-traded firms.

Contactinformation:

Sloan Schoolof

Management

MassachusettsInstituteof

Technology

E52-325,

50

Memorial

Drive Cambridge,

MA

02142-1347

Joos: 617-253-9459,

pioos@mit.edu

(correspondingauthor) Plesko: 617-253-2668.g

plesko@mit.edu

Whether

firms signal futureprospectsthroughdividendchanges hasbeenasource

of

debate

and

researchinthecorporate financeliteraturesince the earlypapers

by

Lintner

(1956)

and

Miller

and

Modigliani(1961). Despite considerableresearch, thedebate over

theempirical validityofthedividend-signalinghypothesisremainsaliveintheliterature.

Nissim

and

Ziv (2001) present evidence consistent with the dividend-signaling

hypothesis

by

showing

dividend increases (but not decreases) relate to fiiture profitability.

However, two

recentpapers

come

todifferentconclusions. Grullonetal. (2003) arguetheresults in

Nissim and

Ziv (2001) follow

from

amisspecificationofthe earnings expectations

model

used topredictexpectedearnings.

They

find theevidence supportingthedividend-signalinghypothesis disappears

when

theearnings expectations

model

accounts fornon-linear patterns in the behaviorofearnings.' In a

more

general

review of dividend policy. Skinner (2003) concludes structural changes in dividend

policy

and

thenatureofcorporateearnings over timeruleout signaling,atleast inrecent

decades.

He

finds dividends have

become

too

smooth and

earnings too volatile for

dividendchangestobe an informativesignal for futureearnings changes.

Although not conclusive, this recentempirical evidence appears to be

moving

towardsrejecting the dividend-signaling hypothesis.^ Inthispaper,

we

contributetothe

debate with a different approach to test the dividend-signaling hypothesis. Instead of

examining

dividend behavior for all firms in the market,

we

examine

the

dividend-signalinghypothesisina setting

where

useof dividendstosignalisparticularly costlyto

thefirm.

Spence

(1973) argues the cost of sending an

economic

signal determines its

increase their dividend

payment

(i.e., cash outflow) while experiencing current losses

caused

by

negativecash flows.^

We

arguethat, since investorscanreadilyobservethe current loss and its components,

management

will need to send a strong

and

credible signal to convincemarketparticipants thatperformancewillimprove.

We

assume

that

increasingcash dividendsatatimethefirmhasanegativecash flowconstitutes astrong

and

costly signal offuture performance for

two

reasons. First, the increase in current

cashdividendswillimmediatelyaffectthe liquidityofthefirm. Second,anincrease in

thecashdividendimpliesa strong

commitment

tomaintainthehigherlevelof dividends

inthefuture,given previousstudies

document

areluctanceof

managers

tocutdividends

(e.g.,Lintner1956

and

Brav

etal.2003).

We

testour hypothesis

by comparing

thepredictive

power of

dividendincreases

between

loss firms with positive

and

negative cash flow

components

for future

performance.

We

collectasample oflossobservationsfrom 1970-2001

and

testwhether

adividendincreaseprovides incremental informationinpredicting firms' return

on

assets

beyond

thatcontainedincurrentearnings

and

a

number

ofcontrol variables.

Our

main

results

show

that, condifioning

on

cash flows, the signaling

power

of dividendincreases for loss firmsexistsonlyfornegativecashflowfirms,consistentwith

thehypothesisthatthecostlierthesignalisthe

more

informationitcontains.

We

verify

our

main

results in additional analyses focusing

on

subsamples offirms with multiple losses for

which

increaseddividend pajmientsareincreasingly costly,

and

on

augmented

specificationsof ourbasicmodel.

Although

some

resultsareconsistentwithalossbeing

sufficient for adividendincreasetoimproveforecastsoffuturereturns irrespectiveofthe sign of the cash flow, all robustness analyses demonstrate the predictive

power

of

dividendincreasesislargerforloss firmswithanegativecash flow thanforlossfirms

withapositivecashflow.

Our

study extends the dividend signaling literature

by

identifying a particular

segment

of firms for

which

we

hypothesize the decision to increase dividends is

particularly costly. Consistent with Spence'scriterion for informative signals,

we

find

dividends helptopredicta firm'sfiitureperformance

when

thedividendsignaliscostly.

We

also extend previous research

on

the relation

between

losses

and

dividends

by

focusing

on

the particularquality oflosses thatrenders the dividend signal costly

and

credible,

namely

thecashflow

component

oftheloss.

In thenextsection,

we

discuss relatedresearch

and

motivateourstudy. Insection

II

we

provide descriptive statistics

of

the

sample

and

present our empirical model.

SectionsIII

and

IV

containour

main

results

and

theresults of robustnessanalyses.

We

concludeinafinalsection.

I.

Background

and

motivation

To

testthedividend-signalinghypothesis,

we

evaluate the predictive

power

of an

increase in cash dividendsforfuture firmperformance ina sample

of

firms thatreport current losses.

We

argue the cost of the dividend signal will determine its informativeness and distinguish

between

losseswith a negative versus a positive cash

flow

component

tocapture thedifferentialcostofthedividendincreaseacrossloss firms.

We

arenot thefirsttoinvestigate the relafion

between

dividends

and

losses.Like

DeAngelo

etal. (1992) [hereafter,

DDS],

we

focus

on

loss firmsto studythe dividend-signaling hypothesis,arguingthatdividendswill have informationcontent

when

current

earnings areanunreliable indicatoroffutureprofitability,

and

thatlossesprovidesucha specialcontext. In asample of 167firmsovertheperiod 1980-1985,

DDS

show

alossis

a necessary, but not sufficient, condition for a firm to decrease dividends.

They

find firms that decrease dividends experience

more

severe

and

more

persistent losses than firmsthat

do

not.Further,unusual

income

items(e.g.,specialitems) area larger factorin

the earningsoflossfirmsthat

do

notreduce dividends thanofthefirms that do. Focusing

explicitly

on

dividend signaling, they find dividend decreases provide incremental informationtopredict futureearnings,althoughtheirforecasting

power

diminishesinthe

presenceof unusual

income

items.

Using

a larger sample over a longtime period. Skinner (2003) finds that

when

firmspayinglargedividends experiencealoss,the lossis

more

likelycaused

by

special items,

and

more

likely toreversethana lossreported

by

a firmthat does not

pay

large dividends. In related work, Joos

and

Plesko (2003)

examine

a large

sample

ofloss

firms,

and

thetiming oflossreversals.

They

show

the lossesoffirms thatcontinueto

pay

dividendsare

more

likely toreverse thanthose of non-dividendpayingfirms,

and

that

eliminating adividendisassociatedwithadecreaseinthelikelihood the loss will reverse

inthe

immediate

future.^

We

extendthis lineofresearch

by comparing

whetherincreasesincash dividends

by

loss firms signal future performance better

when

negative cash flows rather than negative accruals drive theloss.

Whereas

previousresearch focuses specifically

on

the role ofspecial items

when

studying the relation

between

dividends

and

losses,

we

differentiate

between

losseswitha positive

and

negative cash flow

component

tocapture therelativecostofthe decision to increasedividend cashoutflows.

The

evidenceinJoos

and

Plesko (2003)

showing

losses

have

become

more

persistentinrecent years,often

due

to persistent negative cash flows, emphasizes thepotential cost of an increase in cash dividendsforlossfirms.

The most

recent papers

on

the dividend-signaling hypothesis find a reduced

signaling rolefor dividendsovertime, consistentwithfindingsthatfirms have

changed

their dividend-paying behavior (see Skinner

2003)7

In light ofthe evidence ofthe

reducedsignaling role of dividends ina general cross-sectional time-series context,

we

complement

theliterature

by

focusing

on

anarrowersettingthatprovidesa powerfiiltest

ofthedividend-signaling hypothesis.

II.

Sample

construction

and

descriptivestatistics

Our

sample

consists of firm-yearobservations

from

Compustat's Industrial

and

Research

Annual Data

Basesfortheyears 1971-2000. Consistentwith

Hayn

(1995)

we

defineour earningsvariable as

income

(loss)before extraordinaryitems

and

discontinued

operations or

IB

(annual

Compustat

data item #18).

We

define

two main

variables to

capturedividend

payments by

the firm.

We

define

DIVCF

asthetotaldollar

amount

of cash dividends paid

by

the firm (annual

Compustat

data item #21).^

We

focus

on

(changesin) thetotal

amount

paidratherthandividends persharetobeconsistent with our

view

thatthetotaldividend cash outflowestablishes the cost to afirm alreadyfaced with bothaloss

and

a negativecash flowina particular year. Sinceitispossiblethatthe

dividend cash outflow in a particular year increase without dividends per share being

affectedifthefirm's

number

of outstandingshareschanges,

we

define asecondvariable

dataitem#25).

TableIpresents descriptiveinformationforthesample. Panel

A

shows

ourinitial samplecontains206,420 firm-year observations: 146,394profitobservations

and

60,026

loss observations

(29.08%

ofthe total).

The

panel flirther

shows

a relation

between

dividend

payments and

loss occurrence: consistentwith the evidenceinSkinner (2003)

we

find dividend-paying firms are less likelyto incur losses than non-dividend-paying

firms. Focusingfirst

on

thedividend

payments

of ourfirm-year observationsintheyear

prior to the current observation,

we

observe 117,618 firm-year observations with

no

dividends

and

88,802 firm-yearobservations with cashdividends.

Of

thefirms that

pay

no

dividends

44.43%

incur a currentloss;

by

contrast,ofthedividend payingfirmsonly

8.74%

incur a currentloss.

The

contrast

between

dividend-paying

and

non-dividend-payingfirms

becomes

sharper

when

we

focus

on

the

contemporaneous

relation

between

dividend

payments and

firm profitability: ofthe firm-year observations not paying a dividend,

45.90%

have a

contemporaneous

loss,

compared

to only

6.69%

offirms currentlypayingdividends.

Panel

B

provides adescriptionof dividendchanges occurringin our sample. In theftiUsamplethemajorityoffirmsnever

change

theirdividend payments:

measured

as

total cash outflow (or per share)

56.49%

(54.40%) offirm-year observations

do

not change dividends in a given year,

33.00%

(28.85%) increase dividends, and

10.51%

(16.75%) decreasedividends.

The

percentageschange significantly

when

we

partition the sample

between

profit and loss firm-year observations: profitable firms increase

dividends

payments

far

more

oftenthanlossfirms.

By

contrast,the largemajorityofloss firms

do

not

change

theirdividend

payments (86.98%

or

86.04%

depending

on

whether

we

measure

dividendsastotal cash outfloworpershare).

The

high percentagereflects

thefactthatlossfirms are less likelyto

pay

dividends,andthatonlyasmallfractionof

lossfirmsthat

pay

dividendsincreasedividend

payments (4.20%

or

3.80%

depending

on

how we

measure

thedividendvariable).

Panel

C

in Table I cross-tabulates our

two

measures of dividend changes.

The

diagonal percentages in panel

C

show

that in the vast majority ofcases both proxies

reflectthe

same

directionof dividend change.

However,

changesinthe

number

ofshares outstanding, with or without a constant dividend per share, can lead to non-zero

off-diagonalpercentages. For example,

we

observethat in20.54%)ofcases firm'stotalcash dividend

payments

increaseina particular year, yet thedividend pershare variable

shows

adividenddecrease.

Such

a combination istheresult of an increasein the

number

of

sharesinthe

same

year (e.g., as a resultofequityissuances orstock optionexercises).

The two

variablestherefore

complement

eachother asproxies fordividenddecisions

by

management.

While

we

focus primarily

on

the dollarvalueof dividend payments,since

it best captures the

amount

of cash the firm is using,

we

present results using both

variables toillustratethesignaling roleof dividendincreases.

In Table II,

we

present evidence for our

main

variables ofinterest. Since

we

hypothesizethatthesignofthecashflow

component

of negative earningswilldetermine

therelativecostofadividendincrease

we

presentourdescriptivestatistics forasample oflossobservations partitioned

by

thesignofthecashflow

component

ofthelosses.

We

define the cash flow

component

of earnings

(CFO)

as cash flow from operations,

measured

asnet

income

(annual

Compustat

dataitem# 172)lessaccruals.

We

measure

(data item #5)

+ ADebt

in Current Liabilities (data item #34)

+

Depreciation

and

Amortizations(dataitem#14).

Panel

A

of TableII

shows

the

mean,

standard deviation,

and median

for four variablesofinterest. First,

we

define Size asmarket value ofthefirm (stock pricetimes

the

number

of outstanding shares, orannual

Compustat

dataitem#199*data item #25).

Second,

we

define

ROA

as earnings before extra-ordinary items

and

discontinued

operations (or

IB

asdefinedbefore) scaled

by

laggedtotalassets{TA, annual

Compustat

dataitem#6). Third,

we

define

CFO

asbefore. Finally,

we

includeSPI, or specialitems (annual

Compustat

data item #17) scaled

by

lagged assets {TA,.i), in the Panel since

previousresearch singles out

SPI

asthe specific

component

oflosses relatedtoboththe qualityoflosses(i.e.,degreeof

permanence)

andthe dividend-payingbehavioroffirms

(Skinner2003).

Panel

A

reports significant differences

between

the

means and medians

ofthe

two

subsamples (based

on

two-sided t-tests

and

two-sided

Wilcoxon

tests forthe

mean

and

median)

asafiinctionofthesignoftheircash flow component. Generallyspeaking, loss

observations with a positive cash flow are larger

and

exhibit stronger profitability

(smaller losses)than loss observationswith a negative cash flow

component.

Positive

cash flow loss firms

on

average also report less negative SPI, withboth types ofloss

firmshaving

median

SPI of

zero though. Alldifferences

between

means and medians

are

statisticallysignificant.

Panel

B

of Table IIprovides descriptive statistics

on

the incidence of dividend

increases in the sample ofloss observations as a functionofthe signofthe cash flow

thattakes the value of1 ifthe

change

in dividend cash outflow (dividend payout per

share)ispositive,

and

zero otherwise,with

DIVCF

and

DIVPS

asdefinedbefore. Panel

B

shows

a significantlysmallerproportionoflossobservationswithanegativecashflow

component

increases dividends, consistent with a dividend increase being costly:

regardless ofthe dividend variable used, the percentage ofpositive

CFO

loss firms increasing theirdividendsis

more

thantwicethepercentageof negative

CFO

lossfirms.

Table11provides evidenceconsistentwithloss observations beingdifferentasa

fiinction

of

the sign ofthe cash flow

component

ofthe loss: a negative cash flow

component

suggests a greater deterioration in profitability

and

a lower incidence of

dividend increases. In Table III,

we

formally test the relation

between

current

profitability, the signofcurrentcashflows,

and

(changes in) dividend

payments

in the

sample ofloss observations. Specifically,

we

estimate a logisticregression to evaluate the relation of profitability

and

its

components

to the decision to increase current dividends. Focusing

on

our

two

dividend variables

we

estimate the following four specifications:

Prob

(ADIVCFJJP) =

og

+

a,ROA,

+

asCFONEG,

+

as

ROA,*CFONEG,

+

a^LSize,

+

Si (1)

Prob

(ADIVPSJJP) =

bg

+

biROA,

+ b2CFONEG, +

bj

ROA,*CFONEG,

+

b4LSizet

+

82 (2)

where

ADIVCFUP,

ADIVPSUP,

ROA

are asdefinedbefore;

CFONEG

isanindicator

variable equal to

one

ifthe firm reports a negative cash flow,

and

zero otherwise;

variables ofinterest,

ROA,

CFONEG,

and

the interaction

between

both variables,

we

include a control variable for the sizeofthefirmineachspecificationsince

Hayn

(1995) and Joos

and

Plesko (2003)relate the sizeofthe firmto thepersistenceofthe lossand

therefore the potential costofadividendincrease.

Our

sizevariableisLSize, thelogof

themarket value ofthefirm.

Table IIIreports the results ofestimating

models

(1)

and

(2) usingthe

method

detailed

by

Fama

and

Macbeth

(1973). Inboth

models

the coefficient

on

ROA

ispositive

and

highly significant, consistent with a relation

between

higher profitability

and

dividendincreases.

However,

the negative coefficient

on

CFONEG

indicatesthat, on

average, loss firmswith negative cash flows are less likely to increasetheir dividend. Further, the negative and significant coefficient

on

ROA*CFONEG

shows

that the relation

between

higherprofitability

and

dividend increasesinthefull

sample

issmaller

for firms with a negative cash flow.

The

size control variable has a positive

and

significant coefficient, suggesting that larger firms are

more

inclined to increase

dividendsinthecurrentlossyearregardlessofthesignofthecashflow.^

Insum,theanalysesinTablesIthroughIIIsuggesta positive relation

between

a firm's profitability

and

its propensity to increase dividends. Focusing

on

loss

observationsinparticular,

we

findthepresence ofanegativecashflow

component

ofthe

loss reducesthe probabilityofadividend increase,consistentwith negative cash flows fromoperations increasing the costofadividendincrease.

III.

Do

dividend increases forecast futureprofitability?

To

examine

whethercostlydividendincreases constitutestrong signalsoffiature

profitability

we

estimate an earnings forecasting

model

in our sample of loss

observations. Since

we

arguethatincreasesindividend outflows are

more

costly

when

cash flows are negative,

we

predict the decision to increase dividends is a stronger predictoroffutureprofitability fornegative cash flowloss firms than forpositivecash flowloss firms.

We

consider

two

forecasthorizons,

one and

three years,

and

focus

on

futureaccountingprofitability

by

estimating thefollowingparsimonious models:'"

AROA,+r

=

ao+

a,

CFONEG,

+

aj

ROA, +

aj

ADIVCF_UP,+

a4

ADIVCF_UP*CFONEG,

+05

SPI,

+

a^ LSize,

+

83 (3a)

AROA,^,

=

Po+

/^i

CFONEG,

+

yff,

ROA, +

y9j

ADIVPS_UP,+

P4

ADIVPS_UP*CFONEG,

+^5

SPI,

+

_PeLSize,

+

£3 (3b)

We

define futureprofitability as average ftiture

ROA

overthe forecast horizon:

AROA,+^=(Z,+r ROA,+/t,

where

t=1 or 3)

and

estimate

models

(3a)

and

(3b).

The

first

specification (3a) focuses

on

increases in

DIVCF,

and

the second (3b)

on

increasesin

DIVPS.

ROA^

SPI,

CFONEG,

LSizearedefinedas before.

Our

main

variablesofinterest

in

model

(3) are dividend increases

(ADIVCFUP

or

ADIVPSUP),

and

dividend

increases interacted with the negative cash flow indicator variable

CFONEG.

If our

prediction that the decision to increase dividends is a stronger predictor of future

profitabilityfornegative cash flow loss firmsthanforpositive cashflow lossfirms, a^

and

^4willbothbepositive

and

significant.

We

include controlsforcurrentprofitability (ROA), special items,

and

size.

We

include special items {SPI) for the reason

mentioned

earlier,

namely

that previous

researchrelates

SPI

to both the qualityoflosses (i.e., degree of

permanence)

and the

dividend-paying behavior offinns (Skinner 2003).

We

include LSize to control for

potentially omitted variables such as risk or growth ofthe firm.

We

estimate both

specificationsusingthe

Fama-MacBeth

methodology.

Table

IV

presents theresultsofthe estimationof equations(3a)-(3b). All four estimationresults(columns(1)through(4)),focusing

on

differentdividendmeasures

and

forecast horizons,

show

the

same

result for the dividend increase variables: the coefficients as

and

/5j

on

the dividendincrease variables are not significant, indicating

that a dividend increase for loss firms with apositive cash flow

component

does not

signal future profitability controlling for other factors in the model.

By

contrast, the coefficients a^

and

y9^

on

the dividend increase variable interacted with

CFONEG

are positive

and

significant in all specifications. Inuntabulated analysis,

we

also find the

sum

of

Oi+a^ and ^3+^4

arepositive

and

statisticallysignificantinallspecifications.

The

evidence isconsistent with dividend increases signaling futureprofitability, even after

controlling forotherfactors,

when

thecash flow

component

oflossesisnegative. This

findingsupports the hypothesisthat dividendincreases constitutean informative signal

when

the costofthe signalisrelativelyhigh.

Table

IV

also

shows

the coefficients

on

CFONEG

{aj or ^/) are negative

and

significant in all four specifications, consistent with losses with negative cash flows

signaling persistent profitabihtyproblems(see alsoJoos

and

Plesko2003).

By

contrast,

the coefficients

on

ROA

(a2or ^2) are positive

and

highly significantinallspecifications, consistentwiththepreviousfindings

on

theserialcorrelationand

mean

reversionof

ROA

(e.g., Sloan 1996).

The

coefficients

on

SPI

(as or (is) are negative in all four

specifications,butthe levelofsignificance variesdepending

on

the forecast horizon: the coefficients aremarginallysignificantinthe one-year horizon

models

(columns (1)

and

(3)),but highlysignificantoverthethree-yearhorizon (columns(2)

and

(4)),suggesting

special items affect firm profitability

more

over the longer horizon,

and

are less

informative overtheshorterhorizon. Finally, sizepredicts fiitureprofitabilityonlyone year ahead (columns(1)

and

(3)),but not threeyearsahead (columns(2)

and

(4)).

In

summary,

theresultsforbothdividendvariables

and

bothforecasthorizonsare

consistentwitha dividendincrease providing information

on

the futureperformance of

lossfirms only

when

currentcash flows arenegative.

We

interprettheresults toindicate theusefulnessofadividendincreasetosignalfuturefirmperformanceisdirectlyrelated

totheexpectedcostofthedividendincrease.

IV.

Robustness

analyses

We

carryout three (unreported)analysesto testthe sensitivity

and

robustnessof ourfindings. Inourfirstanalysis,

we

focus

on

a

subsample

offirmswith

more

than

one

sequential loss, omitting observations for

which

the current loss is preceded

by

a profitable year.

We

assume

that forthesefirms the costto increasedividendsshouldbe

greaterthanforfirmsexperiencingafirstloss."

We

findchangesindividendsvaryasa

fianctionof whethera lossisthefirstloss(i.e.,the prior year'searnings

were

positive) or

whether the loss is

one

in a sequence (a repeat loss):

10.38%

(9.08%) of first loss

observations increase dividend cash outflows (dividend cash outflows pershare) versus

1.44%

(1.27%) ofrepeat loss observations. This finding is consistent witha string of

losses revealingcontinuingprofitabilityproblems,

making

it

more

difficult forthe firmto

increase dividend payments.

The

proportion ofrepeat losses with negative cash flows (74.90%) is also largerthan the corresponding proportion offirst losses with negative

cash flows (58.60%). Repeat losses with negative cash flows also exhibit lower

profitabilitythanthosewitha positivecashflowcomponent.

The

descriptiveevidencesuggests the decisiontoincreasedividends

when

afirm faces repeat losses constitutes apowerfulsignalregardlessofthe signofthecash flow.

The

existence of such a strong signal forall multiple loss firms could diminish the signalingvalue ofa dividendincrease forloss firms with negativecash flows.

We

re-estimate the predictiontestsin the sample ofrepeat lossobservations

and

observethat,

consistent with our conjecture that increasing dividends

when

facing repeat losses is

costly evenforpositive

CFO

firms,the coefficients aj

and

_P}

become

significant(atthe

10%

level) in the one-year

models

but not in the three-yearmodels.

More

important though, consistentwith our previousresults, the coefficients a4

and

_P4in(3a)

and

(3b)

remainpositive

and

significantinallspecifications. Allotherresultsremainqualitatively thesame.

In a secondanalysis,

we

include additional control variables in (3a) and(3b)to

capture the levelofliquidity

and

recentgrowth ofthe firm.

We

include cash

and

short-termsecuritiesscaled

by

assets(annual

Compustat

dataitem#1 scaled

by

dataitem #6)

as a

proxy

for liquidity

and

the logofSales/Salesi.i (whereSales isannual

Compustat

dataitem #12)as aproxyforgrowth.

When

we

includebothvariablesinthemodels,

we

findthe coefficient

on

liquidityisinsignificantinallspecifications,whilethe coefficient

on

the

growth

proxyispositive

and

significant(atthe

5%

level) inallspecifications.

As

in theprevious robustnesstest, the coefficientsaj

and

fij

become

significant atthe

5%

level intheone-year horizon

models

(correspondingto

columns

(1)

and

(3) inTableIV), butremain insignificant inthe three-yearhorizonmodels. Throughout,the coefficients a4

and

^4 remain positive

and

significant in all specifications, reinforcing the stronger predictive role for fiiture profitability of dividend increases

when

the cash flow

component

oflossesisnegative. Allotherresultsremainqualitativelythesame.

In a final test,

we

re-estimate (3 a)

and

(3b) in separate samples of loss

observations determined

by

the sign of cash flows. That is, ratherthan incorporating

CFONEG

and

the interaction term,

we

estimate a simplifiedversionof(3a)

and

(3b) in

separate samples, allowing the coefficients

on

ROA,

SPI,

and

LSize to vary in both samples.

The

procedure allows usto evaluatewhetherthe coefficients

on

the dividend

increase variable

and

on

the interactiontermin

Models

(3a)

and

(3b)capturedifferential

forecasting

power

oftheother variablesincludedintheequationwith'fixed'coefficients.

Our

estimations

show

thatthe coefficients a^

and

^2

on

ROA

varyasafunctionofthesign

ofthecash flow

component

oflosses;the coefficients

on

SPI and

LSize

however

arenot

different across the subsamples.

The

coefficient

on

the dividend increase variable is

significantatthe

5%

levelin allspecifications,indicatingthatdividendincreases signal futureprofitabilityirrespectiveofthe signofthecash flow

component

oftheloss

when

we

estimate the

models

inseparatesubsamples.

Most

importantlythough,in supportof ourprediction

and

previous results, the magnitude ofthe coefficient

on

the dividend

increase variable remains significantly larger in the subsample of negative

CFO

loss

firms thaninthesubsample ofpositive

CFO

losses.

Summarizing,theresultsoftheadditionalanalysesinsubsamplesoflossfirms or

for different specifications of

Models

(3a)

and

(3b) are all consistent with dividend

increases

by

loss firmsbeing

more

informativeabout futureprofitability

when

thecash

flow

component

ofthe lossisnegative than

when

itispositive.

We

interpretourresults

toindicatethatthe

more

costlyadividendsignalis,the

more

informativethedividendis aboutthe firm'sftitureperformance.

V.

Conclusion

This paper provides

new

evidence

on

the role of dividends in signaling firms' future performance.

We

examine

whether firms that report a current loss

and

have a negativecashflowsignalftitureperformance ofthefirmthroughcostlyincreasesincash

dividends.

We

distinguish

between

losses determined

by

negative cash flows versus

negative accruals to capture the cost ofa current dividend increase.

We

argue that

increasingdividend

payments

when

thefirmisalready losing

money

constitutesa strong signaloffiatureperformancefor

two

reasons. First,theincreaseincurrentcash dividends

affectsthecurrentliquidityofthefirm. Second,theincreaseincashdividendsimplies a strong

commitment

to an increased level of dividend cash outflows in the fiature since

previousresearch

documents

ahighreluctanceof

managers

tocutdividends.

The

evidencein thepaperstrongly supports thehypothesis thatcostly dividend

increases

by

loss firmswith negative cash flowsconsistently predict future measures of performance betterthan dividendincreases

by

other loss firms.

While

recent empirical

resultshave discounted the role of dividendsas a signaling

mechanism

in large cross-sectional samples(Benartzi etal. 1997, Grullon etal. 2003, Skinner 2003), ourresults

suggest costly dividend increases are informative for a

narrow

group of firms. For

negativecashflowlossfirms,theuseof cashto

pay

a dividend,ratherthantoreinvestin

theongoingoperationsofthefirm,suggests

management

judgesthe prospectsofthefirm

tobe good, even thoughcurrentearningsarenot.

Our

focus

on

lossfirmshasbroaderimplications as research

shows

thatlossfirms

comprisea large

and

increasingshareofpublicly-tradedfirms(e.g.,

Hayn

1995,Joos

and

Plesko 2003, Skinner 2003). Therefore, an increasingly larger set of

managers

is

confi-onted with reporting negative earnings that are generally

much

less informative aboutfutureperformance ofthe firm.

As

aresult,theyface the

need

torely

on

additional

mechanisms beyond

reportedprofitabilitymeasurestoprovideinvestorswith information

aboutthefirm'sprospects.

References

Aharony, Joseph,

and

Itzhak Swary, 1980. Quarterly dividend

and

earnings

announcements and

stockholders' returns:

An

empirical analysis.Journal

of

Finance,

35, 1,1-12.

Asquith, Paul,

and David

W.

MuUins,

Jr, 1983.

The

impact of initiating dividend

payments on

shareholders' wealth.Journal

of

Business, 56, 1,77-96.

Bemartzi,

Shlomo,

Roni

Michaely,

and

RichardThaler, 1997.

Do

changes individends

signalthe future or the past?Journal

of

Finance,52,3,1007-1043.

Brav, Alon, John R.

Graham, Campbell

R. Harvey,

and

Roni

Michaely, 2003, Payout

policyinthe21stcentury.

Working

paper.

Duke

University.

DeAngelo,

Harry, Linda E.

DeAngelo, and Douglas

J. Skinner, 1992. Dividends

and

losses.Journal

of

Finance,47, 1837-1863.

DeAngelo,

Harry, Linda E.

DeAngelo, and Douglas

J. Skinner, 2003.

Are

dividends

disappearing?dividendconcentration

and

theconsolidationofearnings.

Forthcoming

Journal

of

FinancialEconomics.

Fama, Eugene

F.,

and

Kenneth

R.French, 2001. Disappearingdividends:changingfirm

characteristicsorlower propensitytopay? Journal

of

FinancialEconomics,60, 3-43.

Fama,

Eugene

F.,

and James

D.

MacBeth,

1973. Risk, return

and

equilibrium:Empirical

tests.Journal

of

Political

Economy,

81,607-636.

Grullon, Gustavo,

Roni

Michaely,

Shlomo

Benartzi

and

Richard H. Thaler, 2003. Dividend changes

do

not signalchanges infutureprofitability.

Working

paper.Rice

University,CornellUniversity,Universityof Chicago.

Hayn,

C,

1995.

The

information content of losses. Journal

of

Accounting

and

Economics,20, 125-153.

Healy, Paul. M.,

and

Krishna G. Palepu, 1988. Earnings information

conveyed by

dividendinitiations

and

omissions.Journal

of

FinancialEconomics, 21,2, 149-175.

Joos,Peter,

and

George

A. Plesko,2003. Lossreversals,

and

earnings-basedvaluation.

Working

paper,

MIT.

Lintner, J., 1956. Distribution of

incomes

ofcorporations

among

dividends, retained earnings,

and

taxes.

American

Economic

Review,46, 97- 113.

Miller,Merton,

and

Franco Modigliani, 1961.Dividendpolicy, growth andthevaluation

Miller, Merton,

Kevin

Rock, 1985, Dividend policy under asymmetric information Journal

of

Finance,40, 1031

-

1051.

Nissim, Doron,

and

Ziv,Amir, 2001. Dividend changes

and

futureprofitability.Journal

of

Finance, 56,2111-2133.

Sloan,Richard, 1996.

Do

stock pricesfully reflectinformationinaccruals

and

cash flows

for futureearnings?

The Accounting

Review, 71,3, 289-315.

Skinner,

Douglas

J., 2003.

What

do

dividendstellus aboutearnings quality?

Working

paper, University of

Chicago

Graduate School of Business

and

University of

Michigan

BusinessSchool.

Spence, Michael, 1973.Jobmarketsignaling.

The

QuarterlyJournal

ofEconomics,

87,3,

355-374.

Watts,Ross, 1973.

The

informationcontentofdividends.Journal

of

Business, 46,2,

191-211.

Earlier

work by

Benartzietal. (1997)discusses

how

previous empirical

work

provides evidencethemarkettreatschangesindividendsas

newsworthy

(see

Aharony and Swary

1980, Asquith

and

MuUins

1983).

They

also point out it is not clear that dividend changessignal futureearnings changes,as thehypothesispredicts.

The

authorsconclude changes in dividends

summarize

information about the past,

namely

past earnings increases,ratherthanthefuture,i.e.,

upcoming

earnings increases.

Other

examples of

studies thatfailtofindevidenceor findevidencethatis

mixed on

the question whether dividend changes

map

intofuture earningschanges areWatts (1973),

Healy and

Palepu(1988),

DeAngelo

etal.(1992).

" _{Previousresearch}

proposescostsassociatedwithdividend payments,for

example

Miller

and

Rock

(1985) arguethesignaling cost

of

dividendsisforgone investment.

However,

we

know

of

no

researchthatstudies the

power

ofthedividendsignal asaftinctionofthe

relativelevelofthecost.

Skinnerdefines largedividend-payingfirms as firms

whose

dividendisinthe top quartileof dividends paidineach decade.

^

_Note

_{the results inJoos}

_and

Plesko (2003)

and

Skinner (2003) are at odds withthe findinginBenartzi etal. (1997)"thatdividendcuts reliably signal anincreasein fiiture

earnings."(p. 1031-1032, emphasisintheoriginal).

DDS

reportthatdividendreductionsoccurlessoften

when

the lossesincludeanaccrual for special items, but

do

not differentiate whether the underlying loss is driven

by

accruals or cash flows. Further, the results ofBenartzi et al. (1997) suggest that if

dividendshave

any

signalingpower,itisthrougha reduction,ratherthananincrease.

^

_The

_findingsof

_Fama

_and

_{French (200}

1)

who

report adecreaseintheproportion of

dividend-payingfirmsinthe

US

over time

and

DDS

(2003)

who

show

a large increasein

theconcentrationof dividend

payments

over timealsounderline thechangesin

dividend-paying behavior of

US

firms.

^

_We

_{winsorizeallvariablesofinterest atthe}

1st

and

99' percentile.

' _{In unreported analyses,}

_we

_{estimate additional specifications of}

_models

(1) and (2).

Specifically,

we

also include SpecialItems inone ofthe specifications

and

findthatthey obtain a significant negative coefficient, whereas the other resuhs remain qualitatively

unchanged. Similarly,

we

includeanindicatorvariable todistinguish

between

firstand

repeat losses

and

find a positive coefficient on this variable, with all other results

remainingqualitativelyunchanged. Finally,

we

alsoestimate a versionof

models

(1)

and

(2) that includes a control variable forthedividendpolicyofthe firm sinceunreported

analysis

shows

theincidenceof dividendincreasesrelates towhetherthefirmpreviously paidadividendornot:

we

includeanindicatorvariablethattakes

on

thevalueof one if

the firmpays adividend(i.e., annual

Compustat

data item #21 islargerthanzero),and

zerootherwise

and

findtheresults

do

notchangequalitatively.

We

alsoestimatetwo-year horizon

models

with results similarto the one

and

three-yearhorizon models. Also, consistentwith previousresearch

we

focusonthe prediction

of an accounting variable only (e.g.,

Nissim and

Ziv 2001). Benartzi et al. (1997) furthermorepoint outthatthe relation

between

dividendincreases

and

futureaccounting

variables or future returnsisdistinct:whereasthey find

no

evidenceofa relation

between

dividend changes

and

future earningschanges, they observepositiveexcess returnsfor

the three yearsfollowingadividendincrease.

'

'Joos

and

Plesko (2003)

document

thattheprobabihtyofalossfirm returningto

profitabilityishigherfor firms incurring afirstlossthanforfirmswithrepeatlosses.

'^

_We

_also

_compare

_{the descriptive}

statisticsinTableIItodescriptivestatisticsforrepeat losses

and

find the repeat loss observations generally exhibit significantly lower

profitabilitythan thetotalgroup oflossfirms.

TABLE

I

Sample

Information

Ourinitialsampleconsistsoffirm-year observations fromCompustaVsIndustrialand ResearchAnnual Data Basesforthe years 1971-2000.

We

defineour earnings variableasincome(loss)before extraordinary items anddiscontinued operations orIB(annualCompiistat data item #18).

We

define

ADIVCF

as the

changeintotal dollaramount ofdividends paidbythefirm (annualCompustatdataitem #21).

We

define

ADIVPS

asthechangeintotaldollaramount ofdividends paidbythe firm per share (annualCompustat

data item #21/ annualCompustatdataitem #25).

Panel

A:

Full

Sample:

Profitvs.

Loss

Observations

IB,>0

N

%

IB,<0

N

%

TOTAL

Full

Sample

146,394

70.92%

60,026

29.08%

206,420

Profitvs.Loss

and

Past Dividends

Div,.,

=

65,356

55.57%

52,262

44.43%

117,618

Div,.,>0

81,038

91.26%

7,764

8.74%

88,802

Profit

and

Loss

and

Current Dividends

Div,

=

63,771

54.10%

54,105

45.90%

117,876

TABLE

II

Loss

Firm

Sample:

DescriptiveStatistics

Ourinitialsampleconsists offirm-yearobservationsfromCompustafs Industrialand ResearchAnnual DataBasesforthe years 1971-2000. Thetableshowsdescriptivestatisticsfor lossfirm-year observations

(N=60,026). Sizeis marketvalue (priceorannual Compustatdata item#199* numberofoutstanding shares or data item #25).ROA, =IB/TA,.],wherewedefineour earnings variableasincome(loss)before extraordinary itemsanddiscontinued operations orIB(annualCompustatdataitem #18).

We

define the cash flowcomponent ofearnings or

CFO

ascashflowfi-omoperationsscaledbylaggedassetswherewe

measurecashflowfromoperationsasnetincome(annualCompustatdataitem #172)-accruals. Accruals

is(ACurrentAssets (data item #4)-ACash(dataitem #1)-ACurrentLiabilities(dataitem #5)+ADeblin

CurrentLiabilities(dataitem#34) +DepreciationandAmortizations(dataitem #14).SPIisspecialitems (annualCompustatdata item#17)scaledbylaggedassets {TA,.i). Thetablereportstestsofdifferences

betweenthemeans and medians ofthetwo subsamplesdeterminedbythesignofthe

CFO

variable (two-sidedt-testsandtwo-sidedWilcoxontestsforthemeanandmedian).

ADIVCFUP

(ADIVPSUP)

isan indicatorvariablethattakes thevalueof1 ifthechangeindividend cashoutflow (dividendpayoutper

share), and zero otherwise, where dividends are annual Compustat data item #21 and number of

outstanding sharesisannualCompustatdataitem #25.

TABLE

III

LogisticRegression

Model

ofProbability of

Dividend

Increase

Thetablecontains theresultsofa logisticregressionwithasindependent variable an indicator variablethat

takes the valueof1ifthecompanyincreasesitsdividends(measuredasatotalcashflow(DIVCF)orcash flow per share(DIVPS)) andzero otherwise.

Prob(ADIVCFJJP)

=ao

+ a,ROA, +OiCFONEG, +a,

ROA,*CFONEG,

+a^LSize,+e, (1)

Proh

(ADIVPSUP)

=bg+ b,ROA, +bsCFONEG,+b,

ROA,*CFONEG,

+b^LSize,+£2 (2)

ADIVCFJUP (ADIVPSUP)

isan indicator variablethattakesthevalueof1 ifthechangeindividend cash

outflow (dividendpayoutpershare),andzero otherwise,wheredividendsareannualCompustaldata item #21 andnumberofoutstanding sharesisannualCompustatdataitem #25. ROA, =IB/TA,.i,where IBis

income(loss)before extraordinary itemsanddiscontinued operations (annualCompustatdataitem #18)

and TAistotalassets (annualCompustatdataitem#6);

CFONEG

isanindicator variablethattakesonthe valueofoneifthe firmreportsa negative cash flow(CFO), andzero otherwise;

ROA

*CFONEG

isthe

interactionvariablebetween

ROA

and

CFONEG;

LSizeislog(market valueof the firm)where

we

define

market value as closing price (annual Compustat data item #199) times shares outstanding (annual

Compustatdata item#25)atfiscal-yearend. Ourinitialsampleconsistsoffirm-yearobservationsfrom CompustafsIndustrialand ResearchAnnual DataBasesfortheyears 1971-2000, and

we

carry outour analysis in asubsample of 60,026 loss firm-year observations.

We

report the resultsofthe logistic

Table

IV

Prediction Tests

Thetablereportstheresultsofthe

OLS

regressions,estimatedusingtheFama-Macbethmethodology: AROA,^,

=

ao+a,

CFONEG,

+a^ROA, +a,

ADIVCFJJP,+

a^

ADIVCF_UP*CFONEG,

+asSPI,

+«« LSize,+Eia (3a)

AROA,^,=_fio+

P,

CFONEG,

+y3,ROA, +P3AD1VPS_UP,+fi^

ADIVPS_UP*CFONEG,

+P,SPI,

+P6 LSize,+Eib (3b)

whereaveragefuture

ROA

overtheforecasthorizonisAROA,+,=(£,+r ROAi+/t) whereROA, =IB/TA,.i,IB

is annual Compustatdataitem #18, TA istotal assetsorannual Compustatdataitem#6 (t=1 or 3);

CFONEG

isan indicator variablethattakes thevalueof1 if

CFO

isnegative,andzero otherwise;

we

measureCFO,orcashflowfromoperationsasnetincome(annualCompustatdataitem#172)-accruals.

We

measureaccruals or

ACC

as {ACurrentAssets (data item #4) - ACash (dataitem #1) - ACurrent

Liabilities(dataitem #5)+ ADebtinCurrentLiabilities(dataitem#34)+DepreciationandAmortizations (dataitem #14);SPIisspecial items(annualCompustatdataitem#17) scaledby laggedassets {TA,.i);

ADIVCFUP

{ADIVPSJUP)isanindicatorvariablethattakesthevalueof1 ifthechangeindividend cash outflow (dividendpayoutpershare),andzero otherwise,wheredividendsareannualCompustatdataitem #21andnumberofoutstanding sharesisannualCompustatdataitem #25; LSizeislog(market valueofthe firm)where

we

define marketvalueas closing price (annualCompustatdata item#199)times shares outstanding(annualCompustatdataitem#25)atfiscal-yearend. Ourinitialsampleconsistsoffirm-year observationsfrom Compustat'sIndustrialand Research Annual Data Basesfor theyears 1971-2000,and

we

carry out the analysis reportedinthetable inasampleconsistingof 60,026lossfirm-year observations

MITLIBRARIES