Kelle , Tobias; Glaum, Ma in; Bausch, And eas; Bunz, Tho s en
A icle — Published Ve sion
The “CEO in con ex ” echnique e isi ed: A eplica ion and
ex ension o Hamb ick and Quigley (2014)
S a egic Managemen Jou nal
P o ided in Coope a ion wi h:
John Wiley & Sons
Sugges ed Ci a ion: Kelle , Tobias; Glaum, Ma in; Bausch, And eas; Bunz, Tho s en (2022) : The
“CEO in con ex ” echnique e isi ed: A eplica ion and ex ension o Hamb ick and Quigley (2014),
S a egic Managemen Jou nal, ISSN 1097-0266, John Wiley & Sons, L d., Chiches e , UK, Vol. 44, Iss.
4, pp. 1111-1138,
h ps://doi.o g/10.1002/smj.3453
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/287826
S anda d-Nu zungsbedingungen:
Die Dokumen e au EconS o dü en zu eigenen wissenscha lichen
Zwecken und zum P i a geb auch gespeiche und kopie we den.
Sie dü en die Dokumen e nich ü ö en liche ode komme zielle
Zwecke e iel äl igen, ö en lich auss ellen, ö en lich zugänglich
machen, e eiben ode ande wei ig nu zen.
So e n die Ve asse die Dokumen e un e Open-Con en -Lizenzen
(insbesonde e CC-Lizenzen) zu Ve ügung ges ell haben soll en,
gel en abweichend on diesen Nu zungsbedingungen die in de do
genann en Lizenz gewäh en Nu zungs ech e.
Te ms o use:
Documen s in EconS o may be sa ed and copied o you pe sonal
and schola ly pu poses.
You a e no o copy documen s o public o comme cial pu poses, o
exhibi he documen s publicly, o make hem publicly a ailable on he
in e ne , o o dis ibu e o o he wise use he documen s in public.
I he documen s ha e been made a ailable unde an Open Con en
Licence (especially C ea i e Commons Licences), you may exe cise
u he usage igh s as speci ied in he indica ed licence.
h p://c ea i ecommons.o g/licenses/by-nc-nd/4.0/
RESEARCH ARTICLE
The “CEO in con ex ” echnique e isi ed: A
eplica ion and ex ension o Hamb ick and
Quigley (2014)
Tobias Kelle
1
| Ma in Glaum
1
| And eas Bausch
2
|
Tho s en Bunz
2
1
WHU –O o Beisheim School o
Managemen , Vallenda , Ge many
2
Jus us-Liebig-Uni e si ä Giessen,
Gießen, Ge many
Co espondence
Ma in Glaum, WHU –O o Beisheim
School o Managemen , Bu gpla z 2,
56179 Vallenda , Ge many.
Email: [email p o ec ed];www.
whu.edu/ acc .
Abs ac
Resea ch Summa y: Hamb ick and Quigley's (2014)
“CEO in con ex ”(CiC) echnique leads o a much
la ge CEO e ec han adi ional ANOVA o mul i-
le el modeling. We eplica e H&Q's s udy, apply hei
CiC echnique o a much mo e comp ehensi e U.S.
sample, and assess he sensi i i y o he model indings
o a ia ions in me hod and da a. We gene ally con i m
H&Q's inding o a high CEO e ec , bu ind a smalle
indus y e ec and a la ge i m e ec in ou much
la ge sample. Applying he CiC echnique wi h
adjus ed R
2
s has only a mode a e impac on yea ,
indus y, and i m e ec s, bu ma kedly educes he
CEO e ec . We also documen ha CiC model indings
a e sensi i e o sample cha ac e is ics, namely i m size
and CEO enu e.
Manage ial Summa y: Hamb ick and Quigley (2014)
in oduced a new me hod o analyze he in luence o
CEOs on i m pe o mance. The s udy's empi ical anal-
ysis ocused on la ge U.S. i ms. We eplica e he o igi-
nal s udy and ex end i o a much la ge ,
comp ehensi e sample o U.S. i ms ha is composed
o 33,996 i m-yea obse a ions, compa ed o 4,866
Recei ed: 27 Sep embe 2015 Re ised: 31 July 2022 Accep ed: 18 Augus 2022 Published on: 19 Sep embe 2022
DOI: 10.1002/smj.3453
This is an open access a icle unde he e ms o he C ea i e Commons A ibu ion-NonComme cial-NoDe i s License, which pe mi s
use and dis ibu ion in any medium, p o ided he o iginal wo k is p ope ly ci ed, he use is non-comme cial and no modi ica ions o
adap a ions a e made.
© 2022 The Au ho s. S a egic Managemen Jou nal published by John Wiley & Sons L d.
S a Mgm J. 2023;44:1111–1138. wileyonlinelib a y.com/jou nal/smj 1111
i m-yea s in he o iginal s udy. Con olling o he
numbe o a iables used in he es ima ions, he model
a ibu es abou a hi d o he o al a iance o i m pe -
o mance (ROA) o he CEO. Fu he analyses show
ha he model indings di e o i ms o di e en size
and CEO enu e; he la ge he i ms and he longe
he CEO enu es, he smalle ends o be he pe cen -
age o a iance explained by he CEO.
KEYWORDS
CEO e ec , CEOs, i m pe o mance, manage ial disc e ion,
a iance pa i ioning
1|INTRODUCTION
The impac o op execu i es on o ganiza ional ou comes has been a opic o scien i ic inqui y
o decades (e.g., C ossland & Hamb ick, 2007; Hamb ick & Finkels ein, 1987; Liebe son &
O'Conno , 1972). Some heo is s ha e a gued ha op execu i es ha e a conside able impac on
o ganiza ional ou comes (Child, 1972; Hamb ick & Mason, 1984). In con as , o he s ha e
sugges ed ha he impac o execu i es is negligible because hei ac ions a e g ea ly con-
s ained by ex e nal p essu es (DiMaggio & Powell, 1983; Hannan & F eeman, 1977).
Va ious schola s ha e a emp ed o analyze he in luence o execu i es empi ically by pa -
i ioning o al a iance in i m pe o mance and a ibu ing i o di e en le els: he indus y,
he i m, and, inally, he CEO. Ea lie s udies apply sequen ial ANOVA (e.g., Liebe son &
O'Conno , 1972); la e wo k mos ly elies on mul ile el modeling (e.g., C ossland &
Hamb ick, 2011). Mo e ecen ly, Hamb ick and Quigley (2014; he ea e H&Q) p oposed he
“CEO in con ex ,”o CiC, echnique o mo e accu a ely con ex ualize he impac o CEOs on
i m pe o mance. A key c i icism o cus oma y a iance pa i ioning me hods is ha he com-
mon use o indus y and i m indica o s (o g and a e ages on hose le els) leads o mis-
speci ica ions. To cap u e hese con ex ual in luences mo e accu a ely, H&Q in oduce
benchma k a iables ha aim o sepa a e he e ec s o he CEO om hose o he i m o indus-
y. Based on da a o la ge U.S. i ms o he yea s 1992–2011, hei s udy yields es ima es o
he CEO e ec ha a e much highe han hose o mos p e ious s udies using ANOVA o mul-
ile el modeling.
In his a icle, we eplica e H&Q's (2014) o iginal wo k; ex end he s udy o a much la ge ,
comp ehensi e sample o U.S. s ock-lis ed i ms; and analyze he sensi i i y o he s udy's in e -
ences o an impo an change in i s me hod, he use o adjus ed R
2
, and o wo impo an sam-
ple cha ac e is ics, i m size and CEO enu e. Ou wo k esponds o a call o eplica ion
s udies in s a egic managemen (Be is, E hi aj, Gamba della, Hel a , & Mi chell, 2016; Hub-
ba d, Ve e , & Li le, 1998). Re isi ing and es ing he gene alizabili y o ex an indings is pa -
icula ly ele an o CEO e ec s udies, o se e al easons: he impo ance o op execu i es is
one o he cen al ques ions in s a egic managemen ; he ques ion con inues o a ac con o-
e sy (e.g., Fi za, 2014,2017; Quigley & G a in, 2017); and empi ical esul s a y wi h measu e-
men echniques and samples. We ocus on H&Q (2014) because o he s udy's impo ance. I
1112 KELLER ET AL.
in oduces a sophis ica ed and a guably supe io echnique o isola e he CEO's impac on i m
pe o mance, and i gene a es a high CEO e ec es ima e, which, acco ding o he au ho s, is
“much mo e in line wi h wha would be expec ed om accep ed heo y abou CEO in luence
on pe o mance”(p. 475).
We s a ou wo k by p o iding wha Be is, Hel a , and Sha e (2016) call a “na ow epli-
ca ion”o H&Q's s udy (also see Tsang & Kwan, 1999), using he same da a and he same
esea ch design as H&Q (2014). Subsequen ly, we p oceed wi h wha Be is e al. call “quasi-
eplica ions”— ha is, we assess he sensi i i y and obus ness o H&Q's indings o a ia ions
in he me hod and he da a. Mo e speci ically, H&Q ocused on a sample o la ge i ms,
“ oughly he 1,500 la ges U.S. co po a ions”(Hamb ick & Quigley, 2014, p. 480). This aises
he ques ion o whe he hei indings also hold o a b oade sample ha is mo e ep esen a i e
o he economy as a whole—mo e speci ically, a sample ha also includes smalle i ms, which
end o ha e a highe pe o mance a iance han la ge i ms ( o a simila line o a gumen see
Chang & Sing, 2000). Thus, ollowing he guidelines de eloped by Be is, Hel a , and Sha e
(2016), we s ep by s ep enla ge H&Q's sample and analyze he e ec s on he model es ima ions.
Ou la ges sample, which exploi s as comple ely as possible he cu en ly a ailable da a on
CEOs o s ock-lis ed i ms in he U.S., comp ises 33,996 i m-yea obse a ions, 1,983 unique
i ms, and 5,191 CEOs, as compa ed o H&Q's 4,866 i m-yea s, 315 i ms, and 830 CEOs. Ou
sample also spans a longe ime ho izon (50 yea s s. H&Q's 20 yea s), includes mo e indus-
ies, and e lec s mo e di e si y in i m size and p o i abili y.
In he second pa o ou in es iga ion, building on con ibu ions by Fi za (2014) and
Quigley and G a in (2017), we assess he sensi i i y o he CiC model o he use o adjus ed R
2
.
Fi za (2014) a gues ha he commonly applied a iance decomposi ion me hods o e s a e he
CEO e ec because pa o he a iance in pe o mance ha is a ibu ed o he CEO is e ec-
i ely andom. In a eply o Fi za (2014), Quigley and G a in (2017) a gue, in e alia, ha he
“ andom chance”elemen o a iance decomposi ion can be accoun ed o by basing in e ences
on adjus ed R
2
, ins ead o unadjus ed R
2
. Thus, while H&Q (2014) in oduced hei CiC me hod
wi h unadjus ed R
2
s, in he second pa o ou empi ical analysis we in es iga e in de ail how
using adjus ed R
2
s a ec s he indings om hei model.
In he hi d pa o ou in es iga ion, we examine he impac o i m size and CEO enu e
on he CiC model indings, ollowing H&Q's own sugges ion o assess “when and whe e CEOs
ma e mos (and leas )”(Hamb ick & Quigley, 2014, p. 488). We do his mainly by spli ing ou
o al sample in o qua iles o each o he wo cha ac e is ics and applying he CiC me hod o
he esul ing subsamples. This pa o ou analysis bene i s om ou comp ehensi e da ase ,
which includes a la ge and highly di e se se o i ms, and a la ge numbe o CEOs, among
hem many CEOs wi h long enu es.
Ou indings can be summa ized as ollows. Fi s , while we gene ally con i m H&Q's main
inding o a high CEO e ec , ex ending H&Q's o iginal sample does a ec he con ex ual e ec s
o he CiC model. Mo e speci ically, H&Q's sample ocused on he la ges i ms in he U.S. s ock
ma ke . When we add da a o smalle and mo e ola ile i ms, he CEO e ec emains high in
all s eps o ou analysis, bu he indus y e ec s a e e y small and he i m-speci ic e ec
becomes mo e impo an han in Hamb ick and Quigley (2014).
Second, applying H&Q's CiC model wi h adjus ed R
2
s ins ead o unadjus ed R
2
s does no
ha e a s ong impac on he es ima es o he yea , indus y, and i m e ec s. In con as , he
CEO e ec is ma kedly lowe wi h adjus ed R
2
s, because o he la ge numbe o CEO indica o
a iables ha a e in oduced in he las s age o he CiC model. Howe e , compa isons o yea ,
KELLER ET AL.1113
indus y, i m, and CEO e ec s ac oss samples a e no ma e ially a ec ed by he adjus men o
R
2
s, because he adjus men lea es di e ences ac oss samples la gely in ac .
Thi d, we documen ha he CiC model indings a e sensi i e o sample cha ac e is ics,
namely i m size and CEO enu e. Mo e conc e ely, he CEO e ec is la ge , and bo h i m and
indus y e ec s a e smalle , in smalle companies han in la ge ones. As ega ds enu e, ou
in es iga ion e eals ha CEO enu e is nega i ely ela ed o bo h he CEO and he i m e ec s.
Some o he e ealed associa ions be ween he sample cha ac e is ics and he model indings
a e concep ually ounded and in ui i e. O he associa ions a e less in ui i e and may be d i en
by echnical aspec s o he CiC me hod, o example, he size-weigh ing o he indus y
benchma ks.
2|STUDIES INVESTIGATING THE CEO EFFECT
O e he pas decades, nume ous s udies ha e a emp ed o empi ically gauge he CEO's in luence
on i m pe o mance, by using a iance pa i ion me hods ha a ibu e he o al a iance o a
dependen a iable (usually e u n on asse s [ROA]) o ac o s such as he yea , he indus y, he
i m, and he CEO. We p o ide an o e iew o he mos impo an CEO e ec s udies in Table S1.
1
In an ea ly pape , Liebe son and O'Conno (1972) ind ha he CEO accoun s o 14.5% o
he a iance in pe o mance. Howe e , subsequen s udies ha e gene a ed e y di e se indings,
wi h CEO e ec es ima es anging om abou 2% (Be and & Schoa , 2003) o mo e han 40%
(Weine & Mahoney, 1981). Possible easons o he di e se esul s could be he di e en ime
pe iods and samples o he s udies, bu also he di e en es ima ion me hods (analysis o a i-
ance [ANOVA], maximum likelihood es ima ion, o mul ile el modeling [MLM]).
Hamb ick and Quigley (2014) c i icize ano he common aspec o adi ional CEO e ec
s udies, he use o simple indica o (“dummy”) a iables o measu e he con ex ual e ec s o
indus ies and i ms:
Nominal indica o s o con ex do no speci y he pe inen , p oximal condi ions in
which indi idual CEOs a e loca ed, which […] causes subs an ial blu ing o con-
ex ual e ec s and CEO e ec s. The use o nominal p edic o s is especially p ob-
lema ic because i ea s some o he CEO's own impac as pa o he con ex in
which he o she is ope a ing, hus sys ema ically unde es ima ing o e all CEO
in luence. (Hamb ick & Quigley, 2014, p. 474).
As we explain in de ail below, H&Q ins ead in oduce yea -speci ic benchma ks o he pe o -
mance o i ms' indus ies and CEO-speci ic a iables o i ms' “inhe i ed”heal h and pe o -
mance. Applying hei e ined model o a sample o la ge U.S. i ms o he yea s om 1992 o
2011 yields a CEO e ec o 38.5%, conside ably highe han hei esul s om sequen ial
ANOVA (16%) o MLM (20%). In he i s pa o ou own empi ical analysis, we eplica e he
o iginal H&Q s udy and hen in es iga e whe he hei indings also hold in a much b oade
sample ha also includes medium-sized and smalle i ms.
1
In a ela ed s eam o esea ch, s a ing wi h Schmalensee (1985) and Rumel (1991), s udies using a iance
decomposi ion in es iga e he ela i e impo ance o indus y, business line, and co po a e e ec s. Fo an o e iew o
ea lie s udies, see Bowman and Hel a (2001); o mo e ecen s udies, see McGahan and Po e (2002), Adne and
Hel a (2003), Misangyi, Elms, G eckhame , and Lepine (2006), and Guo (2017).
1114 KELLER ET AL.

In a c i ical assessmen o he CEO e ec li e a u e published in pa allel o Hamb ick and
Quigley (2014), Fi za (2014) a gues ha pa o he a iance in pe o mance ha a iance-decompo-
si ion s udies a ibu e o he CEO is in ac andom. As Fi za u he poin s ou , he CEO e ec is
he mo e in la ed he sho e he CEOs' enu es, and he lowe he numbe o indi idual CEOs in a
sample. In a eply, Quigley and G a in (2017) a gue ha he “ andom chance”elemen o a iance
decomposi ion can be con olled bybasingin e encesonadjus edR
2
, ins ead o unadjus ed R
2
.
2
We
ake accoun o his me hodological deba e by ca e ully examining, in he second pa o ou empi i-
cal in es iga ion, how using adjus ed R
2
s a ec s he indings om H&Q's CiC model.
O e he yea s, se e al au ho s ha e a gued ha esea che s should no con en hemsel es
wi h he ques ion o whe he CEOs, in gene al, ha e an impac on i m pe o mance, bu
should also in es iga e unde which se s o condi ions CEOs ma e mo e, o less (e.g.,
Fi za, 2017; Hamb ick & Quigley, 2014; Quigley & G a in, 2017; Wasse man, Anand, &
Noh ia, 2010). Fo example, Liebe son and O'Conno (1972) and Wasse man e al. (2010) poin
ou ha he magni ude o he CEO e ec di e s ac oss indus ies; C ossland and
Hamb ick (2007,2011) iden i y coun y di e ences; and he indings o Quigley and Hamb ick
(2015) sugges ha he impo ance o CEOs has inc eased o e ime. In he hi d pa o ou
own empi ical in es iga ion, we examine mo e closely how i m size and CEO enu e mode a e
he CiC model indings. We ocus on hese wo ac o s because Hamb ick and Quigley (2014)
limi ed hei o iginal s udy o la ge U.S. i ms and because enu e is closely connec ed o he
CEO e ec and plays an impo an ole in he a gumen a ion o Fi za (2014).
3|METHOD
Ou s udy eplica es and ex ends H&Q's 2014 s udy. H&Q de ine a se o nes ed equa ions o
analyze i m pe o mance and o isola e he po ion o i s a iance ha is a ibu able o he
CEO. As in mos s udies on he CEO e ec , he pe o mance a iable is ROA, calcula ed as ne
income di ided by o al asse s o each i m-yea .
Like p e ious esea che s, in a i s s ep H&Q use indica o a iables o calenda yea s o cap-
u e mac oeconomic e ec s [see Equa ion (1) below]. P e ious s udies also used indica o a iables
o measu e he in luence o he indus y. I is a key insigh o H&Q ha his app oach is impe ec
because he i m i sel con ibu es o he mean pe o mance in i s indus y,andbecausei assumes
ha he indus y in luence on pe o mance is cons an o e he ime ho izon o he s udy. The e-
o e, H&Q eplace he indus y dummies wi h an indus y benchma k, calcula ed as he size-
weigh ed mean ROA pe indus y in a gi en yea , excluding he ocal i m (and hence i s CEO).
3
Simila a gumen s apply a he i m le el. Tha is, jus as he i m i sel con ibu es o he mean pe -
o mance in i s indus y, each CEO con ibu es o he i m's a e age pe o mance ac oss he s udy's
da a panel. In addi ion, H&Q poin ou ha using i m indica o a iables o he en i e ime span
o a s udy no only assumes cons ancy o he i m e ec bu also implies ha CEOs a e assessed
agains an a e age i m pe o mance ha includes he yea s beyond hei enu es (Hamb ick &
Quigley, 2014, p. 479). To add ess hese issues, H&Q eplace he i m dummies wi h wo a iables
mean o accoun o he condi ion o he company a he s a o each CEO's enu e: “inhe i ed
2
Mo e ecen ly, Fi za (2017) sugges s ha CEO e ec es ima es in a iance decomposi ion s udies can be a ec ed no
only by andom noise, bu also by he au oco ela ion o i m pe o mance.
3
In acco dance wi h H&Q, when calcula ing he indus y benchma ks we include all i ms in Compus a o which he
necessa y basic da a (i.e., ne income, o al asse s) a e a ailable.
KELLER ET AL.1115
p o i abili y,”p oxied by he mean ROA o he 2 yea s p eceding he CEO's enu e, and “inhe i ed
heal h,”measu ed as he company's ma ke - o-book a io (MTB) di ided by he median MTB a io
o he indus y excluding he ocal i m, a he s a o he CEO's enu e. In he ou h and las o
he eg ession equa ions, in acco d wi h p e ious s udies, H&Q (2014) use indica o a iables o
indi idual CEOs o cap u e hei agg ega e e ec .
The ollowing equa ions summa ize Hamb ick and Quigley's (2014) CiC model:
ROAyiko =XDYEARy+ε1
yiko ð1Þ
ROAyiko =XDYEARy+I–BENCHyik +ε2
yiko ð2Þ
ROAyiko =XDYEARy+I–BENCHyik +INPROFko +INHEALTHko +ε3
yiko ð3Þ
ε3
yiko =XDCEOo+ε4
yiko,ð4Þ
whe e subsc ip s y,i,k, and oindica e he yea , indus y, i m, and CEO, espec i ely.
The yea e ec is es ima ed as he R
2
om Equa ion (1). The indus y and i m e ec s a e es i-
ma ed as he inc emen al R
2
so Equa ions(2)and(3) espec i ely. The CEO e ec is calcula ed in
wo s eps
4
: i s , by calcula ing he R
2
om Equa ion (4), which indica es how much o he emaining
unexplained pe o mance a iance in Equa ion (3) is explained by he CEO indica o a iables; and
second, by mul iplying his R
2
alue by he pe cen age o unexplained a iance om Equa ion (3),
which yields he inc emen al o al a iance explained by he CEO indica o a iables.
5
In he i s pa o ou own analysis, we “na owly” eplica e and hen ex end H&Q's 2014
s udy by p og essi ely enla ging hei sample in i e s eps, which we desc ibe in de ail in Sec-
ion 4. As we enla ge he sample, we compa e yea , indus y, i m, and CEO e ec s ac oss he
s eps. We use Fishe 's z- ans o ma ion o es ima e 95% con idence in e als o all e ec es i-
ma es, and in ou discussion, we ocus on s ep- o-s ep changes ha a e clea ly s a is ically dis-
inguishable because he con idence in e als o he wo e ec s do no o e lap.
6
4
Acco ding o H&Q (2014, p. 482), “using he esiduals om Equa ion (3) as he dependen a iable a he han adding
he CEO dummy a iables as u he explana o y a iables o Equa ion (3), assu es ha he ixed e ec coe icien s o
indi idual CEOs […] can be meaning ully in e p e ed as each CEO's ne e ec a e comple ely con olling o
con ex ual ac o s.”They u he explain, “adding he CEO dummies o he ull model gene a es he same amoun o
a iance explained, o R-squa ed, bu yields less s able es ima es o he indi idual CEOs.”
5
We ollow H&Q (2014) and winso ize con inuous a iables in all eg ession es ima ions a he op and bo om 2.5
pe cen iles. To ake accoun o po en ial se ial co ela ion in he panel da a, H&Q use gene alized es ima ing equa ions
(GEE). Thei speci ic GEE model assumes a no mal dis ibu ion and a i s -o de au o eg essi e s uc u e o he
wi hin-panel co ela ion ( o de ails o GEE model es ima ion, see Ha din & Hilbe, 2013). In ou own empi ical
in es iga ion, we mainly use OLS es ima ion. The eason o his is ha a majo pa o ou analyses is conce ned wi h
he use o adjus ed R
2
, and we a e no awa e o an es ablished me hod o compu e adjus ed R
2
s o GEE.
6
We a e g a e ul o one o he e iewe s o sugges ing he use o Fishe 's z- ans o ma ion. Thus, ollowing Quigley and
Hamb ick (2015), we in e p e yea , indus y, i m, and CEO e ec s as pa ial R
2
s and use Fishe 's z- ans o ma ion o
es ima e con idence in e als. We use he S a a command co cii wi h he ishe op ion o calcula e Fishe 's z (see
Cox, 2008). In line wi h gene al SMJ policy, we use he con idence in e als desc ip i ely and no o iden i y speci ic
cu o le els o s a is ical signi icance. Focusing on con idence in e als ha do no o e lap is a conse a i e app oach,
because some di e ences in R
2
s ac oss s eps could be “s a is ically signi ican ”in he adi ional sense e en when
con idence in e als o e lap ( o de ails see Schenke & Gen leman, 2001). Fo de ailed discussions on he es ima ion o
con idence in e als o R
2
, see Olkin and Finn (1995) and Zou (2007).
1116 KELLER ET AL.
In he second pa o ou analysis, building on he wo k o Fi za (2014) and Quigley and
G a in (2017), we in es iga e how using adjus ed, a he han unadjus ed, R
2
s a ec s he
model's indings. In he hi d and inal pa o ou analysis, we in es iga e how sensi i e he
H&Q model is o impo an sample cha ac e is ics, namely i m size and CEO enu e. We do
his mainly by spli ing ou sample in o qua iles and unning he model sepa a ely o he
esul ing subsamples. We again use con idence in e als based on Fishe 's z- ans o ma ion o
assess di e ences be ween he qua iles.
4|DATA
The s a ing poin o ou analysis is H&Q's da a on CEOs o la ge s ock-lis ed U.S. co po a-
ions o he yea s 1992 o 2011 om he Execucomp da abase, ma ched o key inancial da a
o hese i ms om he Compus a da abase: 4,866 i m-yea obse a ions, pe aining o 830
CEOs and 315 unique i ms in 44 indus ies (4-digi SIC), wi h a leas ou i ms in each
indus y.
7
We ollow H&Q (2014) and exclude inancial ins i u ions, public sec o en i ies,
and unclassi ied i ms, as well as i ms wi h only one CEO and CEOs who s ayed only 1 yea
o less.
8
We hen ex end he sample in s eps. In S ep 1, we ocus on he same i ms as H&Q (2014)
and use he longes ime se ies o which we can ind da a in Execucomp and Compus a
(1971–2019). In S ep 2, we add da a o all u he a ailable i ms in he same indus ies as
in H&Q (2014), and in S ep 3 we ex end he sample o addi ional indus ies, ha is, we
make ull use o he a ailable da a in Execucomp and Compus a .
9
In S ep 4, we b oaden
ou sample u he by adding da a om he Re ini i Eikon da abase ( o me ly
ThompsonReu e s Eikon), which p o ides bo h CEO da a and inancial da a (i.e., i includes
he o me ly sepa a e da abases Da as eam and Wo ldscope). A his s age, he sample has
29,318 i m-yea obse a ions and comp ises da a o 4,515 CEOs, 1,725 unique i ms, and
164 indus ies.
While we hus ha e ma kedly expanded ou sample in compa ison o H&Q (2014), we a e
s ill cons ained by equi ing a leas ou benchma k i ms in each indus y, measu ed a he
ou -digi SIC le el, he ines a ailable indus y classi ica ion.
10
We could a ain a s ill b oade
7
We hank Tim Quigley o ha ing made a ailable o us he inal da ase o H&Q's (2014) s udy as well as his S a a code
o es ima ing he a iance pa i ioning.
8
In line wi h H&Q (2014), we use he Execucomp a iable “COPEROL” o iden i y he CEOs.
9
He e and in wha ollows, in o de o i ms o be included in ou sample we equi e hem o ha e nonnega i e equi y
and a leas US$ 1 m in asse s.
10
A ac o limi ing he es ima ion o indus y e ec s in a iance decomposi ion s udies in gene al is he quali y o he
da a, ha is, he i ms’indus y classi ica ion p o ided in da abanks such as Compus a and Re ini i Eikon. Like H&Q
(2014) and p e ious esea che s, we assign all i ms in ou samples o he indus ies o hei espec i e p ima y p oduc
segmen s. Howe e , he assignmen o i ms o indus ies can be complica ed and may lead o e o s in he da abases
(Jacobs & O'Neill, 2003; Kahle & Walkling, 1996). Mo eo e , i ms’p oduc po olios change o e ime, so using s a ic
SIC indus y a ilia ions, as is common in a iance pa i ioning s udies, is p oblema ic, especially o e longe ho izons.
Also p oblema ic is he indus y classi ica ion o mul isegmen i ms, whe e he da a p o ide s assign “p ima y codes”
o indica e he i ms’majo a eas o ac i i ies. Fo all o hese easons, indus y e ec es ima es om a iance
pa i ioning s udies need o be in e p e ed wi h ca e.
KELLER ET AL.1117
ep esen a ion o i ms and CEOs i we conduc ed he analysis on he b oade h ee-digi SIC
le el ins ead.
11
Thus, in an in e media e s ep, we examine whe he he CiC model es ima ions
a e sensi i e o he g anula i y o he indus y ca ego iza ion by es ima ing he model o he h ee-
digi SIC le el, using he same i m-yea obse a ions as in S ep 4. Finding ha he es ima ion
esul s a e la gely he same as hose based on ou -digi SIC indus y benchma ks, in he i h and
las s ep we include all i ms o which CEO and inancial da a a e a ailable and o which we can
ind a leas ou i ms in he same indus y, de ined a SIC le el 3. The esul ing sample, he la ges
possible sample o which we can apply he H&Q (2014) CiC model, has 33,996 i m-yea obse a-
ions, wi h 5,191 indi idual CEOs o 1,983 i ms in 140 SIC le el-3 indus ies.
5|FINDINGS
5.1 |Pa 1: Na ow eplica ion o H&Q (2014) and s epwise sample
ex ensions
The indings o he i s pa o ou analysis a e summa ized in Table 1. The i s column o
Table 1p esen s H&Q's (2014) o iginal es ima ion. The nex wo columns p esen ou na ow
eplica ion o hei es ima ion, and he subsequen columns p esen he indings om quasi-
eplica ions in which we ex end ou sample s ep by s ep. Panel A desc ibes he samples; Panel
B shows he es ima ion esul s o he pe o mance a iance componen s. We p esen poin es i-
ma es o all yea , indus y, i m, and CEO e ec s, and we p esen con idence in e als based
on Fishe 's z- ans o ma ions o all OLS es ima ions. Panel C p esen s desc ip i e s a is ics on
sales, o al asse s, ROA, and CEO enu e o he samples.
We begin by eplica ing H&Q's o iginal es ima ion as closely as possible, using H&Q's se o
sample i ms and me hod (GEE), bu ou own da a and calcula ions (S a a code). As he esul s
in column 2 o Table 1documen , ou es ima ion yields nea ly he same esul s as hose o
H&Q— he indus y e ec is 7.22% ( s. 6.9% in H&Q [2014]), he i m e ec is 11.51% ( s.
12.1%), and he CEO e ec is 39.01% ( s. 38.5%). The mino di e ences be ween ou poin es i-
ma es and hose o H&Q may be ela ed o he indus y benchma ks, ha is, I–BENCHyik in
Equa ions (2) and (3). To calcula e hese benchma ks, H&Q use all i ms a ailable in Com-
pus a , and we ollow hei p ocedu e. Bu i is likely ha he se o i ms co e ed by Compus a ,
and he indus y a ilia ion o some i ms, ha e changed o e ime.
12
In he second column, we p o ide he indings om OLS es ima ions wi h he same da a,
ha is, he o iginal H&Q sample. These indings a e e y simila o hose wi h GEE. In ac , he
i m e ec is now sligh ly highe han wi h GEE, and he CEO e ec sligh ly lowe , b inging
he es ima es ye close o he o iginal H&Q (2014) esul s. Ha ing hus es ablished ha he e is
11
Weine (2005) has shown ha he deg ee o homogenei y o i ms belonging o he same indus y acco ding o
Compus a o Da as eam Wo ldscope (now Re ini i Eikon) is e y simila o he ou -digi and he h ee-digi SIC
classi ica ion le els. Weine a gues ha na ow indus y classi ica ions, such as he ou -digi SIC le el, lead o small
popula ions, so ha esul s can easily be biased h ough ou lie s. B oade indus y de ini ions do no su e as much
om ou lie s bu end o become inhomogeneous. Examining his ade-o u he , Weine concludes ha he h ee-
digi SIC le el yields he mos accu a e alue p edic ions. Fo a mo e o mal ea men o he ade-o be ween he
e iciency gains om pooling and he possible bias esul ing om he e ogenei y see Wang, Zhang, and Paap (2019).
12
Using Tim Quigley's inal da ase and his S a a code yields exac ly he esul s epo ed by H&Q (2014). Fu he mo e,
we also ge exac ly he same esul s when we use his da ase and ou code.
1118 KELLER ET AL.
be e y small.
23
The same holds o he nex s ep, which adds wo p oxy a iables o i ms’
inhe i ed p o i abili y and heal h: he impac o adjus ing R
2
on he i m e ec will also be
small. The si ua ion is di e en in Equa ion (4), whe e he esiduals om Equa ion (3) a e
eg essed on a se o (o−1) CEO dummy a iables. He e, he numbe o a iables is e y high
in ela ion o he sample sizes. Hence, adjus ing R
2
will ha e a p onounced impac on he CEO
e ec .
In Table 2and Figu e 1, we p esen he CiC model indings o H&Q's o iginal (2014) sam-
ple and ou i e sample ex ensions, using bo h unadjus ed R
2
s (Panel A) and adjus ed R
2
s (Panel
B).
24
As shown in Table 1, we p esen poin es ima es and con idence in e als o he e ec s,
using Fishe 's z- ans o ma ion. Panel C o Table 2p esen s he di e ences be ween he poin
es ima es. Figu e 1g aphically summa izes he main esul s om Table 2, p esen ing he con i-
dence in e als o he CiC e ec es ima es based on unadjus ed R
2
s nex o hose o es ima es
based on adjus ed R
2
s. The ou e ec s o he CiC model a e p esen ed in ows, wi h he
unadjus ed e ec es ima es on he le and he adjus ed e ec es ima es on he igh .
When we compa e he esul s based on adjus ed R
2
s wi h hose calcula ed wi h unadjus ed
R
2
s, he di e ences a e consis en wi h ou easoning abo e. We ocus i s on he poin es i-
ma es. The yea e ec s based on adjus ed R
2
s a e somewha lowe han hose based on
unadjus ed R
2
s, bu he di e ences a e small, a leas in absolu e e ms. As one would expec ,
he di e ences be ween he adjus ed and unadjus ed yea e ec s ge smalle as we expand he
sample in he la e s eps by adding mo e obse a ions o an (almos ) cons an numbe o yea s.
The es ima es o indus y and i m e ec s a e e y simila o unadjus ed and adjus ed R
2
s
(mean di e ences ac oss samples <0.1%).
25
In con as , and p edic ably, he CEO e ec based
on adjus ed R
2
s is ma kedly lowe han he es ima es based on unadjus ed R
2
s— he es ima es
now ange be ween 27.81% (S ep 2) and 34.02% (S ep 4), whe eas hose based on unadjus ed R
2
a e be ween 34.92% (S ep 1) and 40.72% (S ep 4). On a e age, ac oss he H&Q (2014) sample
and ou sample ex ensions, he CEO e ec es ima es based on adjus ed R
2
s a e 6.99 pe cen age
poin s lowe han hose based on unadjus ed R
2
s.
An impo an insigh , pa icula ly appa en in Figu e 1, is ha he R
2
adjus men lea es he
di e ences be ween e ec es ima es ac oss sample ex ensions (i.e., ou i e s eps) la gely in ac :
he es ima es a e e ec i ely shi ed downwa ds in pa allel. Hence, whene e con idence in e -
als do no o e lap be ween subsamples wi h unadjus ed R
2
s, hey also do no o e lap o
23
One could a gue ha I-BENCH, while echnically only one a iable, is simila o a se o indus y benchma ks, ha is,
indus y-yea dummy a iables o indus y-yea ROA means. This is because excluding he ocal i m om he yea ly
ROA mean leads o alues ha a e i m-yea -speci ic, bu none heless e y simila wi hin a gi en indus y-yea . This
holds especially in la ge indus ies, whe e he emo al o one obse a ion does no ma e ially a ec he mean. The mo e
he ex- ocal- i m benchma k co ela es wi h he benchma k including he i m, he mo e i is compa able o a se o
indus y-yea indica o s. To he bes o ou knowledge, he e is no guidance in he li e a u e on how o adjus R
2
in such
a con ex , and we he e o e use he “s anda d”app oach o conside ing I-BENCH as a single eg esso . Howe e , i I-
BENCH we e in e p e ed as a se o indica o s, in oducing i in o he eg ession equa ion would use up a highe
numbe o deg ees o eedom, and he adjus men o R
2
would ha e a g ea e impac on he indus y e ec .
24
The e ec es ima es based on unadjus ed R
2
in Table 2, Panel A, a e iden ical o hose p esen ed in Table 1.We
include hem he e again o ease di ec compa isons wi h he esul s based on adjus ed R
2
s.
25
As appea s in Panel C o Table 2, mos o he di e ences be ween he indus y and i m es ima es based on unadjus ed
R
2
s and hose based on adjus ed R
2
s a e nega i e. Howe e , hese di e ences a e all e y close o ze o in absolu e e ms.
Fu he mo e, he es ima ion o he indus y and he i m e ec s ollows he es ima ion o he yea e ec s, and he R
2
-
adjus men educes he yea e ec s mo e s ongly, lea ing mo e a iance po en ially o be explained in he subsequen
s ages o he CiC model. The esul ing inc ease in he indus y and i m e ec s appea s o o e compensa e o he e y
small adjus men o he pa ial R
2
s.
KELLER ET AL.1125

FIGURE 1 Legend on nex page.
1126 KELLER ET AL.
subsamples wi h adjus ed R
2
s. This is he case o he indus y and i m e ec s in S eps 2 and 4,
and o he CEO e ec in S ep 4.
To sum up, CEO e ec es ima es a e in la ed when hey a e based on unadjus ed R
2
s. Thus,
in line wi h Quigley and G a in (2017), he H&Q (2014) CiC model should be used only wi h
adjus ed R
2
s. As ou analysis shows, doing so ma kedly lowe s he CEO e ec , bu does no
s ongly al e he yea , indus y, and i m e ec s, no he compa isons o e ec s ac oss samples.
5.3 |Pa 3: Sensi i i y o H&Q (2014) model indings o sample
cha ac e is ics
In he hi d and las pa o ou analyses, using only adjus ed R
2
s, we conduc addi ional es s o
examine in de ail he sensi i i y o he H&Q (2014) CiC me hod o wo sample cha ac e is ics,
i m size, and CEO enu e. This analysis also esponds o a call, exp essed se e al imes in he
CEO e ec li e a u e, o in es iga e in mo e de ail unde wha condi ions and ci cums ances
CEOs ma e mo e, o less (e.g., Fi za, 2017; Hamb ick & Quigley, 2014; Quigley &
G a in, 2017). We examine i m size because H&Q (2014) ocused on la ge U.S. co po a ions,
and in he s epwise sample enla gemen s in he i s pa o ou empi ical analyses we saw ha
adding smalle i ms o H&Q's sample had a ma ked impac on he CiC es ima ions. And we
in es iga e CEO enu e because i is closely connec ed o he cen al objec o ou in es iga ion,
he CEO e ec , and because CEO enu e plays a cen al ole in he a gumen a ion o Fi za
(2014). We conduc he ollowing es s on ou o al sample, he one comp ising 33,996 i m-yea
obse a ions.
5.3.1 | Fi m size
To examine he impac o i m size mo e sys ema ically, we di ide ou la ges sample in o qua -
iles, using sales as a size p oxy.
26
Figu e 2p esen s CiC es ima ion indings o he ou qua -
iles; we also p esen he es ima ion esul s and desc ip i e s a is ics o he qua iles in
Table S2.
The g aph documen s ha i m size has non i ial implica ions o he CiC model e ec s.
Mos s iking is he impac on he indus y e ec — he pe o mance o small i ms ends o be
idiosync a ic, and consequen ly, in he i s qua ile he e is p ac ically no indus y e ec a all
(and e en he yea e ec is signi ican ly smalle han in he o he size classes). As i ms become
la ge , hei pe o mance ge s mo e simila o he pe o mance o hei espec i e indus ies;
ha is, he indus y e ec es ima es inc ease mono onically om one size qua ile o he nex ,
FIGURE 1 Con idence in e als o yea , indus y, i m, and CEO e ec es ima es om he CiC model,
based on OLS es ima ion and unadjus ed R
2
s (le column) and adjus ed R
2
s ( igh column), o he o iginal
H&Q (2014) sample and s ep-by-s ep sample ex ensions. The “handleba s” ep esen 95% con idence in e als,
based on Fishe 's z- ans o ma ions, o he e ec es ima es om he CiC model es ima ions, ha is, he lowe
ba indica es he lowe alue o he con idence in e al, and he uppe ba indica es he uppe alue. The g aphs
use di e en scalings on he e ical axes. Desc ip i e s a is ics o he a ious samples a e p esen ed in Table 1
26
The esul s a e quali a i ely e y simila i we use o al asse s ins ead.
KELLER ET AL.1127
and in he ou h size qua ile he indus y e ec explains a li le mo e han 10 % o he i ms'
pe o mance a iance (10.83%).
27
A echnical eason ha helps o explain he p onounced
impac o i m size on he indus y e ec is ha H&Q use size-weigh ed ROAs as indus y
benchma ks. In his analysis, we spli ou sample acco ding o size and hen compa e he i ms
in each qua ile wi h he comple e, “unspli ”indus y benchma ks, ha is, he en i e popula-
ion o i ms in he same indus y a ailable in Compus a and Re ini i Eikon. Ob iously, he
size-weigh ed benchma ks explain he pe o mance o la ge i ms be e han ha o smalle
i ms. The qua ile eg essions p esen ed in Figu e 2 hus b ing o ligh a speci ic aspec o he
FIGURE 2 CiC model es ima ions o qua ile subsamples based on i m size (sales). The igu e p esen s
esul s om H&Q CiC model es ima ions o subsamples—qua iles based on i m size, as measu ed by sales
e enues—o ou la ges sample (n=33,996). The es ima ion indings o he smalles i ms a e p esen ed on he
le , and he es ima ion indings o he la ges i ms a e on he igh . The ba s ep esen he es ima ed yea ,
indus y, i m, and CEO e ec s as well as he emaining unexplained a iance om he CiC model. All e ec s
a e calcula ed using OLS es ima ion and adjus ed R
2
s. The es ima ion esul s and desc ip i e s a is ics o he
a ious subsamples a e summa ized in Table S2
27
He e, and gene ally in he hi d pa o ou analyses, we ocus on c oss-qua ile di e ences o which he con idence
in e als de i ed om Fishe 's z- ans o ma ions do no o e lap. Fo example, he con idence in e als indica e ha he
indus y e ec es ima es a e clea ly di e en om qua ile o qua ile. Analogously, he i m e ec s in qua iles one,
wo, and h ee a e di e en om hose in qua iles wo, h ee, and ou , espec i ely. And he CEO e ec in he i s
qua ile is di e en om ha in he second. The CEO e ec s in he second and hi d qua iles a e no di e en om
each o he , bu he es ima o in he hi d qua ile is again di e en om ha in he ou h. The con idence in e als o
all e ec es ima es in all qua iles a e epo ed in Table S2.
1128 KELLER ET AL.
CiC me hod, because, in he “no mal”applica ion o he me hod in he o al sample, smalle
sample i ms a e also compa ed o size-weigh ed indus y benchma ks. Figu e 2 u he e eals
ha he i m e ec is mo e p onounced o e y small (11.58%) and e y la ge i ms (13.07%)
and less p onounced o medium-sized ones (7.44% and 10.17%). I seems ha inhe i ed pe o -
mance and heal h pe sis longe in small and la ge i ms han in medium-sized i ms. Finally,
he CEO e ec dec eases mono onically as i m size inc eases. I is 35.75% o he smalles size
qua ile and only 25.79% o he la ges qua ile—10% poin s smalle in he la ges qua ile. I is
in ui i ely ob ious, and in line wi h p e ious s udies (Finkels ein & Hamb ick, 1990), ha
CEOs ha e ela i ely less disc e ion and hus less impac on i m pe o mance as i ms ge
la ge .
28
5.3.2 | CEO enu e
We p esen wo se s o analyses o examine he impac o CEO enu e on he H&Q (2014) CiC
model indings. Fi s , we again spli ou sample in o qua iles, his ime by he i ms' maximum
CEO enu es.
29
The indings a e summa ized in Figu e 3.
30
Second, o p o ide ye deepe
insigh s we un a se ies o 24 CiC model es ima ions in which we successi ely inc ease he min-
imum enu e equi ed o including a CEO in he samples. The “e ec lines” esul ing om
hese es ima ions a e p esen ed in Figu e 4.
The wo igu es documen ha CEO enu e p o oundly in luences he es ima ion esul s.
Figu e 3shows ha in i ms wi h sho e CEO enu es, ha is, i ms in he i s and he second
qua ile, he indus y e ec does no explain i m pe o mance a all. Tha is, he pe o mance
o hese i ms is qui e idiosync a ic, dissimila o ha o hei indus y pee s.
31
A he same
ime, i ms wi h sho e CEO enu es ha e a p onounced i m e ec . In ac , he i m e ec is
high in he i s h ee qua iles, a 19.57% in he i s qua ile, 17.45% in he second, and 19.84%
in he hi d qua ile, be o e d opping o only 9.11% in he ou h qua ile. Figu e 3 u he indi-
ca es ha he CEO e ec declines mono onically wi h inc easing CEO enu e. I is 34.68% in
he i s qua ile and only 26.42% in he ou h qua ile.
32
28
We also no e ha i m size is ela ed o CEO enu e. The a e age enu e inc eases mono onically wi h size; mean
(median) enu e is only 5.31 (4) yea s in he i s size qua ile and 7.17 (6) yea s in he ou h qua ile (Table S2). The
leng h o CEOs’ enu es may i sel be ela ed o he s eng h o he CEO e ec ; we examine his ela ion below. Mo e
gene ally, we acknowledge ha he uni a ia e analysis in his pa o ou s udy does no allow us o isola e he “pu e”
impac o a gi en i m cha ac e is ic (e.g., size) on he CiC model e ec s, by holding all o he ac o s cons an .
29
As in he p e ious in es iga ion on i m size, we spli he sample in o qua iles based on a i m cha ac e is ic (he e,
maximum enu e) and assign all obse a ions o a gi en i m o he same qua ile. Thus, he numbe o i ms is
oughly he same in all ou qua iles. Howe e , in he p esen in es iga ion, he numbe s o i m-yea obse a ions a e
o cou se highe in he qua iles wi h longe maximum enu es.
30
Fo an o e iew o es ima ion esul s and desc ip i e s a is ics, see Table S3.
31
Acco dingly, he s anda d de ia ion o ROA is nega i ely ela ed o CEO enu e; i is 46.9% in he i s qua ile and
only 12.9% in he ou h qua ile; see Table S3 o de ails; also see B ookman and This le (2009) on he ela ion be ween
CEO enu e and i m cha ac e is ics.
32
These numbe s gauge he impo ance o he CEO ela i e o he a iance o ROA in he espec i e subsamples. In
absolu e e ms, CEOs “ma e ”much mo e in he i s qua ile because he pe o mance o i ms in his qua ile is
much mo e he e ogeneous. Among he i ms in he i s qua ile, CEOs explain 36.36% o a s anda d de ia ion o ROA
o 46.9% age poin s, whe eas in he ou h qua ile CEOs explain 26.34% o a s anda d de ia ion o ROA o only 12.9%
poin s.
KELLER ET AL.1129
The e ec lines in Figu e 4show in g ea e de ail ha yea (b own line) and indus y e ec
( u quoise line) es ima es inc ease s eadily wi h longe CEO enu es, albei on low absolu e
le els. The i m e ec (g een line) and he CEO e ec ( ed line) dec ease as we inc ease he en-
u e equi emen , un il minimum enu es o 13 and 11 yea s, espec i ely. The nega i e associa-
ion be ween CEO enu e and he i m e ec may be a leas pa ly due o a echnical eason.
The CiC model p oxies o he i m e ec , inhe i ed p o i abili y (INPROF
ko
), and inhe i ed
heal h (INHEALTH
ko
), a e CEO-speci ic and ge “upda ed”wi h each new CEO. While hese
p oxies yield g ea e insigh han he simple i m dummies used in adi ional a iance decom-
posi ion s udies, elying on hem implies ha he i m i sel a ec s pe o mance o e he en i e
cou se o he CEO's enu e solely h ough inhe i ance. Pu di e en ly, changes on he i m le el
ha a ec pe o mance a e a ibu ed en i ely o he CEO; he es o he o ganiza ion is
assumed no o ma e a all. I should also be no ed ha , in gene al, wi h inc easing enu es, i
becomes inc easingly di icul o disen angle s a is ically he i m e ec and CEO e ec .
Tu ning o he CEO e ec , one migh expec ha CEOs wi h long enu es in luence hei
i ms' pe o mance mo e p o oundly. Figu e 4indica es he opposi e: he CiC model es ima es
o he CEO e ec become smalle and smalle as we impose highe and highe minimum enu e
es ic ions. The e a e se e al possible explana ions o his phenomenon. Fi s , o e ime CEOs
FIGURE 3 CiC model es ima ions o qua ile subsamplesbasedonleng ho CEO enu e(maximumCEO
enu es). The igu e p esen s esul s om H&Q CiC model es ima ions o subsamples—qua iles based on he leng h
o he CEOs' enu es, as measu ed by i ms' maximum CEO enu es—o ou la ges sample (n=33,996). The
es ima ion indings o i ms wi h he sho es CEO enu es a e p esen ed on he le , and he indings o i ms wi h
he longes CEO enu es a e on he igh . The ba s ep esen he es ima ed yea , indus y, i m, and CEO e ec s as well
as he emaining unexplained a iance om he CiC model. All e ec s a e calcula ed using OLS es ima ion and
unadjus ed R
2
s. The es ima ion esul s and desc ip i e s a is ics o he a ious subsamples a e summa ized in Table S3
1130 KELLER ET AL.

may lose some o hei ene gy and become less adap able o changing en i onmen s (e.g.,
Mille , 1991; on he dynamics o manage ial capabili ies, also see Adne & Hel a , 2003; Hel a
& Ma in, 2015).
33
Second, i ms wi h longe enu es may ope a e in s able ma ke s, in which
i ms may need (and allow) minimal CEO disc e ion (Hamb ick & Ab ahamson, 1995).
34
Thi d,
he associa ion be ween CEO enu e and he s eng h o he CEO e ec may also be d i en by
o he a iables ha a e ela ed o bo h enu e and he CEO e ec — o example, i m size,
which inc eases, and i m pe o mance, which imp o es, o e he CEO enu e qua iles
FIGURE 4 CiC model e ec lines, condi ional on CEO minimum enu e. The igu e p esen s es ima ion
esul s om a se ies o 24 CiC model es ima ions in which we successi ely inc ease he minimum enu e ha we
equi e o he inclusion o he CEOs in he espec i e samples. The b own line ep esen s he esul s o he
yea e ec s, he u quoise line he indus y e ec , he g een line he i m e ec , and he ed line he CEO e ec .
The g ay line a he op o he igu e shows he emaining unexplained a iance. All o he e ec es ima es a e
based on OLS es ima ions and adjus ed R
2
s. The poin s on he e y le o he lines a e based on a minimum
enu e o 2 yea s, ha is, he same equi emen as in ou main analysis. As we mo e along he lines o he igh ,
we inc ease he minimum enu e o 3, 4, and mo e yea s, and elimina e all CEOs wi h sho e enu es, un il we
each a minimum enu e o 25 yea s. The hickness o he e ec line indica es he size o he sample
33
Also see Simsek (2007) on he ela ion be ween CEO enu e and op managemen 's isk- aking p opensi y.
34
Acco dingly, we ind ha indus y disc e ion sco es om Hamb ick and Ab ahamson (1995) and Finkels ein,
Hamb ick, and Cannella (2009) dec ease mono onically o e he ou qua iles.
KELLER ET AL.1131
(Table S3).
35
Fou h, he use o CEO dummies assumes a cons an CEO e ec o e he whole
enu e. Mo e p ecisely, he dummies measu e a ime-in a ian di e ence in ROA o a gi en
CEO compa ed o o he , ea lie , and la e , CEOs a he helm o he same i m. Bu he ROA di -
e en ial migh no be cons an . Fo example, du ing a c isis a CEO may incu w i e-o s and
es uc u ing losses
36
; i he u na ound is success ul, ea nings eco e in la e pe iods. Thus, a
s ong CEO ac ion may lead, in he CiC model (as well as in o he a iance decomposi ion
models) o a a he weak CEO e ec .
Gi en ha he wo mos po en e ec s in he CiC model, he i m e ec and he CEO e ec ,
lose explana o y powe wi h highe minimum enu es, i is clea ha he unexplained esidual
a iance mus inc ease. In Figu e 4, he esidual a iance ( ep esen ed by he g ay line) is
45.49% a he e y le , ha is, in ou ull sample wi h he wo-yea minimum enu e, and i
inc eases o mo e han 60% wi h minimum enu es o 11 o mo e yea s.
Finally, Figu e 4 e eals ha he ela ions be ween CEO enu e and yea , indus y, i m,
and CEO e ec s become less s able on he igh -hand side o he g aph, wi h sudden kinks and
e e sals. One eason is ha he e e -s ic e enu e equi emen s lead o changes in he indus-
y composi ion o he sample, which in u n can p o oke sudden changes in e ec es ima es.
37
Ano he eason is ha he size o he sample dec eases owa ds he igh -hand side, and emo -
ing u he i ms om he sample hus can ha e a s onge impac on he o e all esul .
6|DISCUSSION
In he ollowing, we summa ize and discuss he main indings om ou empi ical analyses.
(1) Ou eplica ion gene ally con i ms H&Q's main inding o a high CEO e ec on i m pe o -
mance. We con i m he high CEO e ec in H&Q's o iginal sample and in ou s epwise sample
ex ensions. We again documen s ong CEO e ec s, albei shi ed downwa ds o a somewha lowe
absolu e le el, when we ees ima e he CiC model using adjus ed R
2
s (33.04% o he la ges sample,
see Table 2). And we also gene ally ind p onounced CEO e ec s when we spli ou sample by i m
size o maximum CEO enu e, s ill using adjus ed R
2
s. The las esul s in ou analyses sugges ha
a lowe -bound es ima e o he CEO e ec , among U.S. CEOs wi h long enu es, may be 25% (Fig-
u es 3and 4). While lowe han he 38.5% o iginally epo ed by H&Q, ou es ima es a e co ec ed
o he numbe o a iables used in he es ima ion equa ions, and hey a e s ill la ge han he CEO
e ec s epo ed in mos p e ious s udies ha used adi ional ANOVA o MLM a iance pa -
i ioning me hods in mos p e ious s udies ha used adi ional ANOVA o MLM a iance pa -
i ioning me hods (see Table S1 in he online appendix).
35
CEO enu e may also a y ac oss indus ies. Hence, he i ms in he ou qua iles may di e no only in CEO enu e,
bu also in indus y composi ion and pe haps o he ela ed cha ac e is ics.
36
See, o example, S ong and Meye (1987). Guan, W igh , and Leikam (2005) ind e idence ha sugges s ha
incoming CEOs o en ac i ely dep ess ea nings by engaging in “big ba h”ea nings managemen .
37
Ou ull sample comp ises i ms om 140 indus ies (SIC 3). When we impose minimum CEO enu e equi emen s,
up o a minimum o 6 yea s he numbe o indus ies wi hin he sample emains cons an . Wi h minimum enu es
be ween 7 and 10 yea s, he sample s ill comp ises 139 indus ies, ha is, only one indus y d ops ou . Howe e , a
minimum enu es o 12 (15) [18] yea s, only 124 (104) [81] indus ies a e le , and he d op-ou o se e al indus ies in
each o hese s eps can ha e a ma ked impac on he es ima es. I should be no ed ha he sample sizes, in e ms o
i m-yea obse a ions, emain ela i ely la ge e en a highe minimum enu e le els. Fo example, a a minimum CEO
enu e o 16 yea s he sample is made up o 5,001 i m-yea obse a ions, mo e han in H&Q's o iginal (2014) s udy.
1132 KELLER ET AL.
While he CiC me hod consis en ly gene a es high CEO es ima es, he indus y e ec es i-
ma es in ou sample ex ensions and in mos o he sample spli s a e lowe (and i m e ec es i-
ma es a e highe ) han hose ound by H&Q (2014).
38
The e may be se e al easons o his.
Fi s , ou o al sample comp ises no only la ge i ms, as did he o iginal H&Q (2014) sample,
bu also medium-sized and smalle i ms, whose pe o mance ends o be mo e idiosync a ic.
Wi h he addi ion o he smalle i ms o he sample, he wi hin-indus y a iance becomes
la ge ela i e o he be ween-indus y a iance, ende ing he indus y e ec s less use ul. Sec-
ond, he H&Q CiC model uses size-weigh ed indus y means as benchma ks. This was jus i ied
in H&Q's o iginal s udy, wi h i s ocus on he la ges U.S. i ms, bu may no be app op ia e o
s udies a ge ing samples ha also include medium-sized and smalle i ms. Thi d, unlike he
simple indus y indica o a iables in adi ional a iance pa i ioning s udies, he CiC model's
indus y benchma ks (I-BENCH) by cons uc ion exclude he ocal company and, in his sense,
gene a e “ou -o -sample”es ima es, o which i is much ha de o achie e high R
2
sco es han
o pu e wi hin-sample explana ion. The weak indus y e ec lea es mos o he pe o mance
a iance unexplained in Equa ion (2) o he CiC model, allowing s onge explana o y powe
o he i m-speci ic a iables INPROF and INHEALTH in Equa ion (3).
(2) The CEO e ec is ma kedly lowe wi h adjus ed R
2
s. The eason is ha Equa ion (4)o
he CiC model includes a la ge numbe o CEO dummy a iables—in H&Q's o iginal s udy,
829. Gi en he sample size o 4,866 i m-yea obse a ions, he e a e only 5.87 obse a ions o
each addi ional eg ession pa ame e o be es ima ed. Simila ly, in ou la ges sample, wi h
5,191 CEOs and a sample size o 33,996 i m-yea obse a ions, he e a e 6.55 obse a ions o
each addi ional pa ame e es ima e. Rules o humb in he econome ics li e a u e ypically
equi e 20, 15, o a leas 10 obse a ions pe p edic o in linea eg ession es ima ions.
39
In
o he wo ds, while bo h H&Q's sample and ou own may appea la ge in absolu e e ms, hey
a e ac ually no , in ela ion o he numbe o model pa ame e s o be es ima ed. In echnical
e ms, he numbe o deg ees o eedom in he es ima ion o Equa ion (4) o he CiC model is
qui e low ( he same holds o he es ima ion o he CEO e ec in o he a iance decomposi ion
s udies), and as a esul , he CEO e ec is likely o be o e es ima ed because he model a i i-
cially desc ibes pa o he andom e o in he da a. Because o his “o e i ing” he unadjus ed
R
2
s om he es ima ion do no e lec he model's “ ue”p edic i e abili y in he gene al popu-
la ion.
40
To co ec o he o e i ing, in line wi h Quigley and G a in (2017) he H&Q (2014)
CiC model should be used wi h adjus ed R
2
s.
41
38
In H&Q's (2014) sample, he indus y e ec was 6.9%. In ou la ges sample i is only 0.37%. Ea lie CEO s udies, as
well as s udies ocusing on indus y, pa en company, and business uni e ec s (e.g., Bowman & Hel a , 2001; McGahan
& Po e , 2002; Sho , Ke chen, Palme , & Hul , 2007), mos ly a i e a indus y e ec s be ween 10% and 20%.
39
See, o example, Ha ell (2015, pp. 72–74). In a mo e de ailed analysis, G een (1991) sugges s ha eg ession analysis
should be based, a a minimum, on a base sample o 50 obse a ions plus eigh addi ional obse a ions pe explana o y
a iable. Also see Aus in and S eye be g (2015) and Jenkins and Quin ana-Ascencio (2020).
40
A i s glance, one could pe haps a gue ha ou da ase comes close enough o he gene al popula ion o CEOs o U.S.
s ock-lis ed companies du ing ou sample pe iod so ha he e migh be no need o make in e ences; ins ead, i migh su ice
o in e p e he obse ed e ec s. Howe e , such an a gumen would miss he poin ha , e en i we could indeed obse e all
CEOs o all U.S. i ms du ing a gi en sample pe iod, he unadjus ed R
2
s om es ima ions o he CiC model, while e lec ing
s a is ical associa ions wi hin his sample, will always be in la ed by andom noise and hus canno be in e p e ed as
measu ing he e ec s o CEOs on i m pe o mance. Fu he mo e, he obse ed unadjus ed R
2
s a e no p edic i e o new
da a gene a ed by i ms in yea s a e he sample pe iod, o o i ms and CEOs in coun ies ou side he U.S. Thus, i may be
use ul o in e p e ou sample as being d awn om a hypo he ical in ini e “supe popula ion”(see La akas, 2008).
41
Al e na i ely, one could assess he model i in he popula ion also by using ou -o -sample p edic ion o c oss
alida ion; see, o example, Ha ell (2015).
KELLER ET AL.1133
(3) The CiC model indings a e sensi i e o he sample cha ac e is ics i m size and CEO
enu e. Some o he associa ions be ween hese cha ac e is ics and he model esul s a e concep-
ually ounded and in ui i ely plausible. Fo example, CEO e ec s a e la ge , and i m e ec s
a e smalle , in smalle (and mo e ola ile) companies. In o he cases, ou analyses ha e e -
ealed associa ions ha a e no immedia ely ob ious. One example is he link be ween i m size
and he s eng h o he indus y e ec , a link ha , in e alia, likely e lec s he use o size-
weigh ed indus y means as benchma ks. Ano he example is he nega i e ela ion be ween
CEO enu e and he s eng h o he CEO e ec . One could pe haps a gue ha longe CEO en-
u es allow o a mo e p ecise es ima ion and ha he lowe CEO e ec es ima es in samples
wi h longe CEO enu es a e he e o e mo e accu a e. On he o he hand, he median CEO en-
u e in ou mos comp ehensi e sample is 5 yea s. Res ic ing he sample o long enu es would
hus igno e a la ge pa o he popula ion and would po en ially bias he indings, o he ex en
ha sho CEO enu es indeed di e sys ema ically om longe ones. As we explain abo e,
ocusing on long CEO enu es also aises ano he issue, he assumed cons ancy o he es ima ed
e ec s. The CiC me hod gene a es a e age, “ ixed”e ec s; in his sense, like mos o he a i-
ance pa i ioning models, i is “c oss-sec ional in design and s a ic in i s logic”(Hende son,
Mille , & Hamb ick, 2006, p. 447). Howe e , esea ch sugges s ha CEOs a i e in hei posi-
ions well equipped and lea n e en mo e in he ea ly phases o hei enu es, bu la e become
“ i ed, ensh ined and s ale”(Mille , 1991, p. 41). Mo eo e , he speed o his obsolescence
depends on he dynamism o he indus y (Hende son e al., 2006). The CiC model and mos
o he a iance decomposi ion models mask hese empo al e ec s, as well as possible indus y
di e ences.
42
7|CONCLUSION
Using a e ined and a guably supe io me hod o pa i ion i m pe o mance a iance and o
isola e he CEO's impac on i m pe o mance, and using da a o la ge U.S. i ms o he yea s
1992–2011, Hamb ick and Quigley (2014) es ima ed he CEO e ec a 38.5%, much highe han
mos p e ious es ima es based on adi ional ANOVA o mul ile el modeling. In he p esen
pape , we eplica e H&Q's s udy, apply hei CiC echnique o a much mo e comp ehensi e U.
S. sample, and assess he sensi i i y and obus ness o hei indings o a ia ions in he me hod
and he da a.
We gene ally con i m H&Q's inding o a high CEO e ec , bu ind a smalle indus y e ec
and a la ge i m e ec in ou la ge sample. We show ha applying he CiC echnique using
adjus ed R
2
s changes yea , indus y, and i m e ec s only mode a ely, bu ma kedly educes he
CEO e ec . And we documen ha he CiC echnique is sensi i e o sample cha ac e is ics,
namely i m size and CEO enu e.
While ou eplica ion s udy speaks di ec ly o he alidi y and gene alizabili y o H&Q's
(2014) a gumen , we also mo e gene ally add o he discou se on CEO e ec s and he impo -
ance o op manage s, and we hope ha ou indings will p o ide impulses o u u e esea ch.
Fo example, u he esea ch could e isi he ques ion “when and whe e CEOs ma e mos
42
Guo (2017) employs longi udinal mul ile el modeling o analyze s able and dynamic pa s o pe o mance a iance,
albei o business uni , co po a ion, indus y, and yea e ec s, no o CEO e ec s.
1134 KELLER ET AL.