ArticlePDF Available

Explaining article influence: capturing article citability and its dynamic effects

Authors:

Abstract and Figures

Researchers from diverse disciplines have exam-ined the many factors that contribute to the influence of published research papers. Such influence dynamics are in essence a marketing of science issue. In this paper, we propose that in addition to known established, overt drivers of influ-ence such as journal, article, author, and Matthew effects, a latent factor "citability" influences the eventual impact of a paper. Citability is a mid-range latent variable that captures the changing relationship of an article to a field. Our analysis using a discretized Tobit model with hidden Markov processes suggests that there are two states of citability, and these dynamic states determine eventual influence of a paper. Prior research in marketing has relied on models where the various effects such as author and journal effects are deemed static. Unlike ours, these models fail to capture the continuously evolving impact dynamics of a paper and the differential effect of the various drivers that depend on the latent state a paper is in at any given point of time. Our model also captures the impact of uncitedness, which other models fail to do. Our model is estimated using articles published in seven leading marketing journals during the years 1996–2003. Findings and implications are discussed.
Content may be subject to copyright.
ORIGINAL EMPIRICAL RESEARCH
Explaining article influence: capturing article citability
and its dynamic effects
Shibo Li &Eugene Sivadas &Mark S. Johnson
Received: 10 October 2013 /Accepted: 25 April 2014/Published online: 28 May 2014
#Academy of Marketing Science 2014
Abstract Researchers from diverse disciplines have exam-
ined the many factors that contribute to the influence of
published research papers. Such influence dynamics are in
essence a marketing of science issue. In this paper, we propose
that in addition to known established, overt drivers of influ-
ence such as journal, article, author, and Matthew effects, a
latent factor citabilityinfluences the eventual impact of a
paper. Citability is a mid-range latent variable that captures the
changing relationship of an article to a field. Our analysis
using a discretized Tobit model with hidden Markov processes
suggests that there are two states of citability, and these
dynamic states determine eventual influence of a paper. Prior
research in marketing has relied on models where the various
effects such as author and journal effects are deemed static.
Unlike ours, these models fail to capture the continuously
evolving impact dynamics of a paper and the differential effect
of the various drivers that depend on the latent state a paper is
in at any given point of time. Our model also captures the
impact of uncitedness, which other models fail to do. Our
model is estimated using articles published in seven leading
marketing journals during the years 19962003. Findings and
implications are discussed.
Keywords Citation .Citability .Matthew effect .
Uncitedness .Scientometrics .Hidden Markov model .Tobit
Introduction
The perceived quality of research is important both for individual
scholars and for journals as they are intrinsically linked to the
standing of both and are at essence a marketing of scienceissue
(Stremersch et al. 2007; Varadarajan 2003).Twowidelyused
approaches to assess quality are to look at either an input measure
(the perceived quality or ranking of the journal the article was
published in) or output measures (quality assessment via the
influence and impact of an articles accrued citations) (Bergh
et al. 2006; Garfield 1979;Medoff2006). Both these approaches
exemplify the peer recognition of an academicsresearch
which is central to the academic reward system (Medoff 2006).
However, looking at journal reputation, rankings, or
journal impact factor is problematic as a measure of
individual article influence as many articles published in
influential journals may not be influential, and influential
articles are also published in relatively less prestigious
journals. As Seglen (1997) and Woodside (2009)argue,jour-
nalimpactfactormetricsareunrepresentativeofmostofthe
articles published in academic journals and a poor proxyfor
the actual impact a specific article might generate. The San
Francisco Declaration on Research Assessment (DORA) ini-
tiated by the American Society of Cell Biology in 2012 also
argues that the Journal Impact Factor is a poor indicator of
scientific quality of an individual article (http://am.ascb.org/
dora/). The San Francisco declaration further urges readers to
assess research on its own merit.
In this paper, we propose that to completely capture and
understand the reasons for an articles influence, one needs to
assess article citability. We develop a method to gauge this
Mark S. Johnson is deceased
S. Li
Kelley School of Business, Indiana University, 1309 East Tenth
Street, Bloomington, IN 47405, USA
e-mail: shili@indiana.edu
E. Sivadas (*)
Milgard School of Business, University of Washington, Tacoma,
1900 Commerce St, Campus Box 358420, Tacoma, WA 98402, USA
e-mail: sivadas@u.washington.edu
J. of the Acad. Mark. Sci. (2015) 43:5272
DOI 10.1007/s11747-014-0392-7
latent construct. We suggest that continually evolving
citability explains article influence measured using citation
counts, and we shed light on the dynamic change of article
citabilityon impact over time after publication. We capture the
moderating effect of aggregate article citability on various
known drivers of impact and suggest that citability of a paper
dynamically changes over time; this citability determines
overall article influence.
Our research builds upon extant work and contributes to the
understanding of marketing scholarship in specific and academ-
ic scholarship in general in the following ways. We introduce to
the literature a construct we term article citability.The
citability construct captures a disciplines and aggregate
scholarslevel of interest in a paper. We treat this interest as
dynamic and ever changing and therefore develop a dynamic
model that can capture this latent construct. We treat article
citability as a mid-range latent variable that taps into the rela-
tional embededness of a paper to a discipline, as well as the
interest a paper might generate within a discipline or sub-
discipline due to its brand equity or quality. Notions of quality
and brand equity are therefore subsumed in the citability con-
struct. To capture article level dynamics we adopt a modified
version of the count modela discretized Tobit model with a
hidden Markov process proposed by Li et al. (2005). Our model
incorporates key influencers of influence and controls for time
and unobserved article heterogeneity not encapsulated by pre-
vious research (Burrell 2003;Stremerschetal.2007).
We posit that an articles influence dynamics are driven by
the latent citability of the work and the disciplines relationship
with a paper, which dynamically changes over time. By cap-
turing these dynamics, we demonstrate that various effects that
drive impact (such as author, article, and journal) are not static
but change over time. We demonstrate that our model outper-
forms the commonly used NBD model, a static model which
assumes that the various effects are static and do not change
over time. We demonstrate that various influencers of impact
change according to the latent citability states, and citability
states of an article also changes dynamically over time.
There has been some recognition amongst scientometric
scholars that the works of scholars and articles with greater
visibility are more likely to be noticed and cited. This phe-
nomenon is termed the Matthew effect (Merton 1968).
However, prior research in marketing (e.g., Stremersch et al.
2007) has conceptualized the Matthew effect narrowly, simply
accounting for the halo of author prestige and its effect on
citations based impact. We expand on the conceptualization of
the Matthew effect and suggest that the Matthew effect is not
simply a famous author effect but a much larger brand equity
signaling a famous paper effect mechanism. The visibility of
the work itself or what may be characterized as the famous
paper effect is captured in our model. We find that famous
articles have a significant Matthew effect independent of the
famous author effect that generates a halo and results in
citations that go above and beyond what can be explained
by the quality of the work itself.
We provide new and deeper insights into assessing article
citability and capturing the dynamic nature of such an enter-
prise. Additionally, we capture the effect of uncitedness that
most models fail to capture. We next introduce our conceptual
framework. We then discuss our data and model, and provide
empirical results and outline implications and theoretical con-
tributions. Finally, we conclude with research limitations and
future directions.
Conceptual framework
Not surprisingly, the topic of quality and influence of published
researchscientometrics and bibliometricshas attracted the
attention of a large and diverse group of scholars from a variety
of disciplines: social, information, life, and engineering
sciences. Journal publishers and editors, librarians in charge
of making collections decisions, and individual scholars as well
as administrators and others evaluating them for hiring, tenure,
and promotion are interested in issues surrounding the quality
and impact of published research. Thus, it is not surprising that
research on this topic, while widespread and multidisciplinary,
is rather fragmented.
In the marketing area, Stremersch et al. (2007) suggest that
the influence an article generates can be explained by three
differing theoretical viewpoints. The first, which they term the
universalist perspective,suggests that an article is cited
because of its content and what it says. The universalist
perspective comprises two dimensions: article quality and
article domain. The second perspective is termed the social
constructivist perspectiveand suggests that articles are cited
because of the fame of the authors and personal promotion of
articles by them. The third perspective, presentation perspec-
tive,indicates that an article is influential for howthe
authors say what they say (p. 174). The presentation perspec-
tive suggests that the article title, keywords, and expositional
clarity would influence citations. They found support for the
universalist perspective and partial support for the social con-
structivist perspective but very little support for the presenta-
tion perspective.
Our conceptual framework (Fig. 1) draws upon Stremersch
et al.s universalist and social constructivist perspectives and
integrates it with the notion of citability, a latent construct we
develop using the literature on relationship marketing, brand
equity, and research on the Matthew effect. As Fig. 1indi-
cates, we suggest that the impact of an article is driven by an
unobserved time-varying construct article citabilitywhich
determines the latent citation score of the article. When the
citation score is positive, the article will receive a positive
number of citations; otherwise, it will obtain zero citations.
Our framework is consistent with the literature in that
J. of the Acad. Mark. Sci. (2015) 43:5272 53
citability is informed by (1) journal effects, (2) author effects,
and (3) article effects, which are all moderated by the latent
article citability.
Article citability is a dynamic construct that captures the
changing relationship of an article to a field. As Baldi (1998)
notes, academic knowledge is cumulative and the references
in a journal article highlight the relationship of an article to
previous studies (Sivadas and Johnson 2005). Samiee and
Chabowski (2012, p. 368) point out that articles influence a
field only if they are heavily cited by others.Our conceptu-
alization of citability is inspired by the extensive work in the
area of relationship marketing (cf. Dwyer et al. 1987; Morgan
and Hunt 1994). Berry (1983, p. 25) noted that relationship
marketing is all about a process of attracting, maintaining,
and enhancing customer relationship.In a similar vein, we
suggest that an articles citability is conditioned on its ability
to draw and maintain attention amongst the community of
scholars and embed itself in a relational network. As
Morgan and Hunt (1994,p.22)note,relationshipmar-
keting refers to activities directed toward establishing,
developing, and maintaining successful relational ex-
changes.When a paper is cited, the citer is acknowl-
edging a debt of intellectual influence, whilst those
cited get recognition in exchange (Garfield 1979). This
subsumes notions of trust and interdependence, two
concepts central to conceptualization of relationships.
When one cites another, one indicates a certain level
of trust in their ideas and the quality of their reasoning
or conclusions.
Also, as Dwyer et al. (1987) and others have noted, rela-
tionships lie on a transactional/relational continuum with a
transactional or a relational orientation. Similarly, citation
practices or the relationship of individual scholars with an
individual paper can thus lie on such a continuum, but the
collective citations reflect the relationship of the article to the
discipline. While each citation event may be construed as a
separate transaction,a series of transactions (citations) can
create a relationship between the paper and the field at large
(Netzer et al. 2008). Researchers have recognized that rela-
tionships go through different states and such transitions can
be triggeredby a series of discrete encounters (Li et al.
2011; Netzer et al. 2008). We propose that article citability
reflects a finite setof relational states of a paper with a field.
That relationships go through different phases or states is well
documented in the literature (Luo and Kumar 2013). The
hidden Markov process we utilize can capture transitions
between relational states and the inherently dynamic nature
of relationships. Some papers never resonate with a discipline
and are never able to develop a relationship; they start out and
remain obscure. Others may start out slow but eventually
embed themselves in a relational network, whilst others may
start out strong and their relationship with a discipline may
strengthen or weaken over time. Fundamentally, relationships
are ever-changing and dynamic (Dwyer et al. 1987; Luo and
Kumar 2013).
The latent citability of an article is also in part the
aggregate-level average assessment of quality by the scholars
in a given field which is subsumed in the relational state of a
paper with the field. Initial quality judgments are made by
journal reviewers and editors duringthe reviewprocess. Then,
once published, quality assessments may initially typically be
made based on the perceived quality of the journal the article
is published in or author reputation.
Our goal is to capture the influence dynamics and build a
model that can better explain the influence of an article. When
a scholar reads an article, he/she may form his/her perceptions
about quality of the work, which will determine whether he/
she cites the article in his/her own work since citations are a
form of acknowledging intellectual inspiration (Cronin 1981;
Garfield 1979). Sometimes a scholar may not read the article
before citing it and may simply base the citation decision on
the quality perception he/she obtains indirectly from other
studies that cite the original article. Thus the relational
embededness of an article assumes greater importance. As
time goes by, more new information about the published
article may emerge. For instance, the article may win some
prestigious awards, become famous over time, or become
obsolete over time. Therefore, a disciplinesrelationshipwith
an article may move from state to state and dynamically
change by going up or down depending on the new informa-
tion received. Our focus is not on individual scholarscitation
decisions. We suggest that the main effects such as article,
author, or journal effects are affected by the fields aggregate
article citability state. Additionally, a fieldsrelationshipwith
an article can change dynamically over time from high to low
or low to high given the changes in the status of the paper over
time.
We measure citability through a variety of known
(citations, author standing, journal reputation) and latent var-
iables. We define article citability as the relational state of a
paper with the field. Such a relational state in part captures
article brand equityand will be driven by the visibility,
quality, and ability to elicit the correct response amongst
readers (Keller 2012). We suggest that by understanding arti-
cle citability we can explain the extent to which papers be-
come influential and capture the ever-evolving levels of influ-
ence. We explicitly account for the dynamics and the latent
citability states of articles using a hidden Markov model
(HMM) and allow the journal, author, and article effects to
dynamically change over time according to their citability
states. We also allow the starting probability of the HMM
states to be affected by the author effects and journal effects in
the publication year, and the transition probability and waiting
time of the HMM to be driven by the article-level Matthew
effect, uncitedness effect, and other time-varying variables.
We next discuss several variables in detail.
54 J. of the Acad. Mark. Sci. (2015) 43:5272
Journal effects
Much of the research on scientometrics has focused on journal-
level impact and journal rankings. As Baumgartner and Pieters
(2003) have noted, the perceived quality and status of the
journal significantly influences an articles impact. Scholars
have to make judgment calls about what to read in order to
use time efficiently, and consequently work published in higher
status journals is likely to draw greater attention. More presti-
gious journals are also more accessible and widely available
(Sivadas and Johnson 2005). Furthermore, as article citability
captures a papers current and changing relationship to a field
and reflects global assessment and brand equity of the paper, the
impact of the journal status may depend on the unobserved
article citability state such that high citability state of the article
enhances the journal effects as opposed to if the article were to
be in a low citability state (Netzer et al. 2008;Lietal.2011). To
put it differently, the citation score of a paper is driven by both
its citability and its journal effects. Enhanced citability of an
article also magnifies the journal effects. While an article in a
low citability state also benefits from stronger journal effects,
the effects are muted because of the inherently low citability
state of the article. When that same article moves into a high
citability state, the journal effects also give it a greater boost.
Author effects
Our author effects dimension draws upon Stremersch et al.s
social constructivist perspective. This essentially taps into the
halo the work of famous scholars can have, i.e., the fact that
their work is more likely to be noticed and cited (Merton 1968;
Cole and Cole 1973;Medoff2006). This greater than propor-
tional (proportional to article quality or contribution) citation
of the work of well-known scholars is commonly called the
(author-level) Matthew effect (Merton 1968). The phrase
Matthew effectis derived from the Gospel of Matthew
(25:29), for unto every one that hath shall be given, and he
shall have abundance; but from him that hath not shall be taken
away even that which he hath.The Stremersch et al. concep-
tion of the Matthew effect was limited to famous authors.
Merton (1968) suggested that better regarded scholars get
greater recognition for equivalent quality work as compared
to less known or less established scholars. As Tol (2009,p.420)
puts it, famous works are more easily noted, and authority
lendsweighttoanargument.In addition, the impact of the
author reputation may also depend on the unobserved article
citability state such that high citability of the article may
strengthen the author effects. We argue that a paper needs to
be in a high citability state for it to have greater traction, and
articles that have such traction benefit more from author repu-
tation. The effect of author reputation is muted or considerably
diluted when the article is in a low citability state.
Article effects
Most research on scientometrics has focused on journal-level
effects rather than individual article-level effects. As Chow
et al. (2007) suggest, articles should be evaluated on their own
merit and publication in a top journal is not necessarily a good
proxy for the quality of the article. Similarly the subject area of
the article has been shown to influence impact, with certain
subject areas such as relationship marketing and services
marketing tending to be more cited than other subject areas
such as sales (Stremersch et al. 2007; Bettencourt and
Houston 2001). Subject area and article quality fall within
what Stremersch et al. identified as the universalist perspective
Zero
Cite
Citation
Score
(CS)
Journal Effects
Author Effects
Article Effects
Interaction Effects
Uncitedness Effect
Positive
Cites
CS<=0
CS>0
Matthew Effect
Controls: persistence,
unobserved
heterogeneity
Latent Citability
States
Others: time,
no of awards
Starting Prob
Waiting Time
Fig. 1 Conceptual framework
J. of the Acad. Mark. Sci. (2015) 43:5272 55
on citations, and contribute to overall article citability. Thus,
similar to the journal and author effects, the extent of article
effects may also depend on the unobserved article citability
state such that high citability of the article may strengthen the
article effects. We measure article effect through article awards
and subject area. We suggest that articles that are in the higher
citability states benefit more from awards and from relating to
hottersubject areas than articles that are in low citability
states. Thus, articles in higher citability states benefit more from
article effects than do articles in lower citability states.
As shown in Fig. 1, we also incorporate the interaction
between the journal effects and author effects to see if author
fame is less important when a paper is published in an A-level
marketing journal and how this interaction effect is moderated
by article citability. In addition, we control for the persistence
effect of previous number of citations on current citations as
well as unobserved article heterogeneity. It is important to
note that although journal effects, article effects, and author
effects on citation have been studied in prior research
(Stremersch et al. 2007), it has all been in cross sectional
studies (i.e., across articles) in a static setting. This is different
from our examination of these effects on the dynamics of
article citations (i.e., both within and across articles) that
allows these effects to dynamically change over time accord-
ing to the latent citability state in which the articles are. Given
the latent nature of article citability, a discrete-state HMM is
appropriate to capture the citability states since it has been
employed to successfully model competitive promotions, cus-
tomersunobserved life stages, and relationship states in the
marketing literature (Du and Kamakura 2006;Moonetal.
2007;Netzeretal.2008;Lietal.2011).
Starting probability of the HMM
When an article is in its first publication year, the scholars
initial citability state is likely to be driven by the author and
journal effects, which may serve as quality signals (Spence
1973). However, the subject area of the article is unlikely to
serve as a signal since anyone can choose a topic to work on.
Therefore, we allow the author and journal effects but not
article effects to affect the starting probabilities of the HMM as
shown in Fig. 1.
Transition probability and waiting time of the HMM
After the publication year of the article, as time goes by, the
article may win some prestigious awards, more scholars have
access to it, it could accumulate some fame, or it may go uncited
for a long time. Therefore, we allow these time-varying vari-
ables to affect the transition probabilities and waiting time of the
article citability states in the HMM as shown in Fig. 1.The
transition probabilities will govern which HMM state the article
will jump into from its current state if a jump occurs, and the
waiting time of the latent citability states determines how long
thearticleisgoingtostayinthestateitisin.
Article-level Matthew effect on the transition probability
and waiting time of the HMM
The Matthew effect, while mostly conceptualized as an author
effect (Stremersch et al. 2007), also exists for institutions
(Medoff 2006), for countries (Bonitz et al. 1997), and for
famous papers as well (Tol 2009). Some papers authored by
less known scholars may initially draw attention for their
content or for their novelty (the universalist perspective on
citations) and over time become famous papers. and as Tol
(2009, p. 423) points out, often-cited papers are cited more
often.So, we suggest that the Matthew effect is not simply a
famous author effect, but it holds true for famous papers as
well, i.e., citations beget more citations. The Matthew effect
can come into effect for famous papers (written by not-so-
prominent authors) or for papers published by less known
authors with prestigious institutional affiliation (Medoff
2006). Thus, in our framework we separate the article-level
Matthew effect from mere author prominence and incorporate
the former into the HMM process. For the article-level
Matthew effect, we propose a new behavioral rationale on
how it works. Specifically, we posit that the Matthew effect
strengthens the relationship of an article with a discipline by
increasing the transition probability of the article jumping
from a lower citability state to a higher state, and/or decreasing
the articles time staying in the lower state and increasing its
time in the higher state. These in turn result in more citations.
Uncitedness effect on the transition probability and waiting
time of the HMM
Weale et a l. (2004) suggest that the rate of non-citations to
articles published in journals can also be used as a measure of
(lack of) journal quality. Hendrix (2008) views the percentage
of non-cited articles as one metric for assessing research
quality. Garfield (2005) estimates that nearly half the articles
indexed in ISIs database have zero citations. As Stern (1990,
p. 193) put it, in the absence of citations there is no firm and
readily available evidence that a publication has contributed to
the advancement of scholarship.Bibliometric scholars have
examined various factors that may contribute to uncitedness
such as the influence of the disciplinary citation practices of
number of references per article or the journal in which an
article was published (Seglen 1992;Stern1990). Seglen
(1992) and others have noticed the practice of concentration
of citations with a small fraction of articles accounting for
nearly half of the citations accrued by a journal.
Different from these studies, our focus is on the issue of what
happens to a papers long-term impact if it does not get cited
early on. We focus on short-term uncitedness here. This is to be
distinguished from long-term uncitedness, which is defined by
56 J. of the Acad. Mark. Sci. (2015) 43:5272
Stern (1990) as papers that go uncited eight years after publi-
cation. A paper that may not draw citations early on (within the
first two years) may get some citations over the first 8 years;
however, we focus on the consequences of not generating
citations early on and suggest that not generating citations early
on has a detrimental effect on the long-term impact of an article.
Consequently, zero-cited articles become less and less likely to
get cited over time. We term this the uncitedness effect. As Tol
(2009) points out, citations beget more citations. When an
author cites a scholar, they draw attention to that scholars work.
Thus a citation can serve as an advertisement or a signal of
quality (Spence 1973). It may also serve the function of slowing
down the obsolescence of published work by keeping it fresh in
the mind of scholars working in the area, as when they read
more recent articles in the area they are referred to older articles
that are cited therein. Our model (unlike prior models) can
explicitly account for zero citations and estimate the effect of
early uncitedness on the rate of later citations. Similar to the
article-level Matthew effect, we expect that early uncitedness
decreases citability of the article by increasing the transition
probability of the article jumping from higher citability state to
lower state, and/or by increasing the articles time staying in the
lower state and decreasing its time in the higher state. These in
turn result in lower citability and less citations over time.
Data
Traditionally, the Journal of Marketing,Journal of Marketing
Research,andJournal of Consumer Research are ranked as
the major journals in the field, with Marketing Science as a
more recent addition to this group (Lehmann 2005; Seggie
and Griffith 2009). This study also includes other recognized
journals with a broad rather than specialized coverage of
marketing topics namely, the Journal of the Academy of
Marketing Science,International Journal of Research in
Marketing,andJournal of Retailing. Various studies suggest
these seven journals are ranked in the top ten in an interna-
tional context (Mort et al. 2004;Guidryetal.2004). In the
business disciplines, journals with a more general focus tend
to have higher citations than do specialized journals (Zinkhan
and Leigh 1999).
These journals were also selected for our study because of
their breadth, rather than specialization, in marketing topics.
Davis (1998) describes a number of difficulties entailed in mak-
ing apples to orangescomparisons between the general interest
and the specialized journals with respect to journal citation
impact. While it may not be universally viewed as a general
marketing journal, the Journal of Retailing is included in this
group of general marketing journals rather than as a specialized
journal because it has a broad substantive and methodological
focus that includes services marketing, econometric models,
advertising, sales promotions, supply chain management, and
consumer behavior. Additionally, in terms of structural
influence of journal intercitations, Baumgartner and Pieters
(2003)rateJR as a core marketing journal rather than a marketing
application journal. The other six journals, with the exception of
JCR, also fall into the sub-area of core marketing journals as
opposed to managerial, application, and education journals in
Baumgartner and Pietersstudy.
This study analyzes a total of 1,591 articles consisting of all
articles published during the calendar years 19962003 for the
seven selected journals. For this study, article citation data
were downloaded from Scopus, a citation tracking service
developed by Elsevier as a competitor to Thomson ISI (i.e.,
data source used in previous studies; see Stremersch et al.
2007). Elsevier began the development of Scopus in 2002 and
launched the product in 2004 after extensive testing. Scopus
has worked backwards to create coverage of social science
journals to 1996, which is the starting point of this study.
Scopus does not include journals from the arts and human-
ities (journals that seldom cite marketing articles), but the
coverage of the sciences, social sciences, and business is
greater than ISI. Scopus has a current database of over
21,000 titles from 5,000 publishers including 20,000 peer-
reviewed journals, and conference proceedings, trade
publications, and book series. Scopus is also regarded as
having greater international coverage than Thomson ISI.
Consequently, Scopus may record more citations than ISI for
these marketing articles. For example, Zinkhan (2005) lists the
15 most cited JAMS articles published between 1998 and
2004. Zinkhan reports these articles have a total of 416 cita-
tions in the ISI database as of December 2004; in the Scopus
database these same articles have a total of 511 citations as of
2004, representing an increase of 22.8%. The higher level of
Scopus citations is consistent across all of the articles with one
exception.The reason issimple. ISI is based on the premise of
scientometry,that useful knowledge in any given discipline
is only contained in a small subset of journals. Thus ISI tracks
only a smaller subset of journals in each discipline. ISI tracked
citations to only 20 marketing journals in 2004, and in 2012 it
tracked 30 marketing journals. Thus, ISI does not record
citations (i.e., examine and record references) to articles pub-
lished in other marketing journals besides the small number of
journals they track. Papers cited in articles published in mar-
keting journals such as Journal of Personal Selling & Sales
Management,Journal of Product & Brand Management,and
Journal of Consumer Marketing (to name a few) are not
recorded. Thus, ISI undercounts the number of citations that
accrue to specific articles in comparison to Scopus.
A second problem that led us to use the more comprehensive
Scopus database is that ISI has only recently started tracking
citations for some of the respected marketing journals in our list
such as Journal of the Academy of Marketing Science and
International Journal of Research in Marketing, and data do
not go back to the mid-1990s for these journals. Sivadas and
J. of the Acad. Mark. Sci. (2015) 43:5272 57
Johnson (2005) suggest that it takes up to 6 years post-
publication for citations to a specific article to peak, and we
wanted to allow a sufficient timeframe from publication to data
downloading. Thus, for these two reasons we choose the more
comprehensive Scopus. We would like to emphasize that
Scopus tracks all the marketing journals tracked by ISI plus
several more and is thus more comprehensive. Furthermore,
Archambault et al. (2009)findthereisaveryhighdegreeof
consistency between Scopus and ISIs SSCI databases.
Scopus and ISI also face competition from the citation
tracking capability of Google Scholar, which is a freely avail-
able internet product. A large number of academic publishers
provide Google Scholar with article abstracts and reference
lists (but not full-text). Consequently Google Scholar may
record even more citations for these marketing articles.
However, Google scholar may also record citations to unpub-
lished articles as well that are posted on individualswebsites.
Hence, we chose to go with Scopus.
We downloaded citations at the end of 2006. Thus we
allowed for a minimum of three years post-publication and a
maximum of 10 years from date of publication to date of
citation count download. As Katerratanakul et al. (2003)point
out, it typically takes about 2 years from date of publication
for citations to start accruing. For this study, year-by-year
citation data were downloaded for each of the 1,591 articles
published by the seven journals during the period of 1996
2003. Annual citations for all articles were downloaded
through the end of 2006.
We cross-checked the Scopus database against the table of
contents for the seven journals and found only one substantial
misclassificationthe entire March 2003 issue of JCR was
mistakenlycoded as 2002. This was corrected.We also did not
include five erratum pages from MKS in our analysis.
We found that there exist big citation differences across
the seven journals. JM,JMR,andJCR have the highest
mean or median number of citations, highest percentage of
articles with 20 or more cites, and lowest percentage of
zero-cited articles. Further, different articles seem to have
different citation dynamics. The articles with high total
citations at the end of the data period tended to grow much
faster over time than those with low total citations (see
Appendix A).
Model setup
To capture article-level citability dynamics, we adopt a mod-
ified version of the count modela discretized Type I Tobit
model with a hidden Markov process proposed by Li et al.
(2005). This model has several advantages over the common-
ly used NBD model for counting data in our context (Burrell
2003; Mingers and Burrell 2006; Stremersch et al. 2007).
First, the latent construct in the Tobit model is suitable for
capturing the unobserved dynamic citation score and can
easily accommodate the various effects on citation score in
our conceptual framework in Fig. 1. Second, the Tobit model
can explicitly account for the dynamics of zero citation and
positive citation over time. Third, the hidden Markov process
can capture the latent citability states among scholars and their
dynamic changes over time. Although past studies
(Stremersch et al. 2007) have included time as a covariate in
the NBD model to capture the time impact, the parameters
(i.e., the author, journal, and article effects) in their models are
all static and do not change over time as we allow in the HMM
process. Given the count nature of article-level citations, we
discretized a Type I Tobit model and take log transformation
of the number of new citations at time tto capture the long tail.
We also compare the proposed model to the NBD model in the
empirical analysis section and show that the former
outperforms the latter. The modification from Li et al.
(2005) is that while the HMM is homogenous in their study,
we allow the HMM to be heterogeneous and be affected by the
Matthew effect, uncitedness effect and other time-varying
factors, which will shed light on the dynamics of article
citations over time. We model the number of new citations
(Z
it
) for article iat time tas follows
Zit ¼k¼Floor exp Z
it

if Z
it >0andlnkZ
it <ln kþ1ðÞ
0 otherwise
ð1Þ
Z
it ¼γ0
isXit þεits;εits N0;σ2
s

;with probability pssuch thatXS
s¼1ps¼1
ð2Þ
where the articles latent citability state is indexed by s, which
we explain in the next sub-section. Z
it
*
is a latent dynamic
variable that captures the articles citation score, k is a positive
integer, and Floor(Y) is the integer component of Y where the
discretization occurs. As depicted in the conceptual frame-
work, there is a threshold of zero for an article to ever get a
positivecite. That is, when its citation count is greater than
zero (Z
it
*
>0), the article will obtain positive cites (Z
it
)and
zero cite otherwise. X
it
includes the article, author, journal
effects, and the interaction effects shown in the conceptual
framework. We also take log transformation for the de-
pendent variable Z
it
(i.e., the exp() operator in Eq. 1). γ
is
(including an intercept) is a vector of article- and
citability-state-specific parameters to be estimated and cap-
tures the impact of the various effects. σ
s
2
is the latent
citability state specific variance of unobserved disturbance
in citation score to be estimated as well.
An articles latent citability state
The parameters of our discretized Tobit model (Eq. 2)are
indexed by state sat each time period. This state is meant to
capture an articles latent citability which governs the dynamic
58 J. of the Acad. Mark. Sci. (2015) 43:5272
evolvement of its citations over time. We assume an article can
be allocated to one of Slatent states at each time period and the
total number of states Swill be determined empirically. The
transition among these states is governed by a first-order
continuous-time discrete-state hidden Markov model
(HMM) (Montgomery et al. 2004;Lietal.2011). For brevity
we interpret our latent states as an indicator of the articles
citability state among scholars. Our interpretation of states is
based upon a comparison of the estimated coefficients differ-
ent across states and summary statistics. However, our inter-
pretation and labeling of citability states are not unique, just as
a label for a segment in cluster analysis or factor in factor
analysis is not unique (Li et al. 2011).
A hidden Markov model of latent citability states
We use an S x S matrix M
it
to denote the probabilities for
article ito transition to another state at time t:
Mit ¼
0Pit12 Pit1S
Pit21 0Pit2S
⋮⋮
PitS1PitS20
2
6
6
4
3
7
7
5
:ð3Þ
Each element in the transition matrix P
itmn
represents arti-
cle is probability of transiting from state mat t-1 to state nat
time t.Hence,0P
itmn
1, and the row sum is one.
The diagonal elements of M
it
are zeroes since we do not allow
same-state transitions. Instead, we capture persistence within a
state as a waiting time for the state, which is the duration an
article stays in one particular state. We define W
it
(s) as the waiting
time in state sand assume it follows a gamma distribution in a
continuous time domain (Li et al. 2011):
Pr Wit sðÞjλit sðÞ;kisðÞðÞ¼
kisðÞ
λit sðÞ
Γλ
it sðÞðÞ
Wit sðÞ
λit sðÞ1eWit sðÞkisðÞ
:
ð4Þ
λ
it
(s) is the shape parameter and k
i
(s) is the inverse scale
parameter for state s.Noticeifλ
it
(s)=1wehaveanexponen-
tial distribution. Being article specific, λ
it
(s)andk
i
(s)deter-
mine how long article istays in state s. More specifically, the
expected waiting time until the next state equals the ratio of
the shape parameter to the inverse scale parameter:
EW
it sðÞ½¼
λit sðÞ
kisðÞ
:ð5Þ
Unlike the homogeneous HMM Du and Kamakura (2006),
Li et al. (2005) and Moon et al. (2007) use, we adopt a
heterogeneous HMM and allow the articles waiting time
(e.g., the shape parameter λ
it
(s)inEq.4)tobeaffectedby
the articles award status, time impact, the Matthew effect, and
the uncitedness effect. Specifically, we assume λ
it
(s)followsa
log-normal distribution:
log λit sðÞðÞNλ¯it sðÞ;σ2
λ

;ð6Þ
where its mean λit sðÞ is a function of the articlesaward
status, time impact, the Matthew effect, and the uncitedness
effect:
λ¯it sðÞ¼α0is þα1isTimeit þα2is Time2
it þα3isAwardit
þα4islog rankðÞ
it1þα5islog2rankðÞ
it1
þα6islog years nociteðÞ
it1
þα7islog2years nociteðÞ
it1
:
ð7Þ
The coefficient α
0is
captures an articlesintrinsic
tendency to stay in citability state s. Time denotes the
number of years passed. α
1is
and α
2is
capture the time
impact and its potential non-linear effect on waiting
time, respectively. Award refers to the number of major
journal awards won by the article and α
3is
capture the
award effect. log(rank)andlog(years_nocite)referto
the logarithm of the citation ranking of the article and
the cumulative number of years with zero cites up to
the end of previous year. α
4is
and α
6is
capture the main
impact of these two variables on waiting time, while
α
5is
and α
7is
denote the Matthew effect and uncitedness
effect, respectively.
Initial citability state probabilities of the hidden Markov
model
We define the initial state probabilities of article iresiding in
state sfor s=1,,Sat time 0 as a vector
i
=(π
i
(1),,π
i
(S))
. The row vectors of the transition matrix and the vector of
initial starting probabilities are assumed to follow a Dirichlet
distribution:
PitjDτitj

;ΠiDηis
ðÞ;ð8Þ
Where P
itj
denotes the j
th
row of the transition matrix P
it
,
and τ
itj
and η
is
refer to the hyper-parameters for the transition
and starting probabilities, respectively. Similar to the specifi-
cation of the waiting time intensity, we assume τ
itj
and η
is
follow a log-normal distribution:
log τitj

Nτ¯itj;σ2
τ

;log ηis
ðÞNη¯is;σ2
η

:ð9Þ
In order to take into account the impact of the author and
journal effects on an articles starting probabilities in state s,
we define η
is as a function of an articles author effect, journal
effect, and article effect at time 0 (e.g., year of article publi-
cation). That is,
η¯is ¼ω0is þω1isVi0;ð10Þ
where V
i0
consists of the author and journal effects as in Eq. 2
at time 0. Coefficients ω
1is
measure how these variables at
time 0 affect the probability that an article starts in state s.
J. of the Acad. Mark. Sci. (2015) 43:5272 59
Article heterogeneity and estimation
We model the article-level unobserved heterogeneity follow-
ing a random-coefficient approach in a hierarchical Bayes
framework (Rossi et al. 1996). That is,
γisMVN γ¯s;Λs
 ð11Þ
Given the model specification above, we have the follow-
ing conditional likelihood function for article iat time t:
Pr ZitjsðÞ¼
Z
γis
Pr Zit ¼kjsðÞðÞ
IZ
it>0ðÞ
Pr Zit ¼0jsðÞðÞ
IZ
it¼0ðÞ
fγis
ðÞdγis ð12Þ
where Pr(Z
it
=k|s)=TN
[lnk,ln(k+1))
[γ
is
X
it
,σ
s
2
] and Pr(Z
it
=0|s)=
TN
(−∞,0]
[γ
is
X
it
,σ
s
2
]. I() is an indicator function, and TN
(a, b)
denotes the truncated normal distribution between values a
and b. f(γ
is
) is the heterogeneity distribution specified in
Eq. 11. The unconditional likelihood and identification of
the hidden states are presented in Appendix B.
Given the high-dimensional integrals in the likelihood
function, a Hierarchical Bayes approach is demonstrated to
be a good choice for estimation (Rossi et al. 1996). We use the
Gibbs Sampler and the Slice Sampler (for truncated Normal
distributions; see Damien et al. 1999) to obtain draws from
the full conditional distributions of the parameters (Chib and
Greenberg 1995). Additionally, using the Data Augmentation
approach (Tanner and Wong 1987), we treat the unknown
utilities Z
it
*
as parameters and make draws for them from their
own full conditional distributions. We estimate the empirical
model using a program coded in C++. The chain for the
Gibbs sampler was run for a total of 50,000 iterations. The
first 40,000 iterations were discarded as burn-inbefore
convergence was attained (Gelfand and Smith 1990). The
remaining draws were used for inference.
Empirical analysis and results
Measures
We randomly divide the total sample into two parts: Part I with
three quarters of the data (1,193 articles and 8,851 observa-
tions) constitutes the estimation sample, while the remaining
quarter forms the holdout sample (398 articles and 2,961
observations) for model comparison purpose. The simple
statistics are presented in Table 1.
Dependent variable Since we are interested in article
citability and the dynamics of the citation, we operationalize
the dependent variable as the number of new citations that the
article receives in year t. Citations are shown to be an objective
measure of influence, impact, or attention (Pieters and
Baumgartner 2002). As seen in Table 1, the annual number
of citations per article is 3.27 with median 2 and a standard
deviation of 0.05.
Independent variable: article-level Matthew effect Merton
(1968) defines the Matthew effect as the phenomenon that
fame breeds fame in the form of citations. To capture the
potential article-level Matthew effect, we adopt the approach
proposed by Tol (2009) on the basis of the theory of growth of
firms such that the log of the firm size is proportionalto the log
of the rank of the firm (Gibrats law) (Ijiri and Simon 1974;
Simon 1955). Tol (2009) uses the following empirical test:
ln(citations) = α+β×ln(rank) + γ×ln
2
(rank), where cita-
tionsis the total number of citations an article receives at a
particular time and rankis the ranking of an article (i.e.,
article with most previous citations = rank 1) based on the
number of the previous citations to that article. In this equa-
tion, γis used to measure the article-level Matthew effect
(second-order effect) such that if γis significantly less than
zero, there are increasing returns toscale andhence a Matthew
effect (i.e., it is due to the fact that the ranking is reverse-
scored such that the article with most previous citations = rank
1). Otherwise, there is no Matthew effect.
In our study, we also capture the potential article-level
Matthew effect through the coefficient of the square of log
of article rank based on the cumulative citations of all papers
in the sample in previous year.
1
To capture the first-order
effect of article rank, we also incorporate the variable ln(rank)
as an independent variable. When we rank the articles in the
sample, we exclude those with no citations in the previous
year. In other words, the value of this variable will be zero for
those articles without any citations. There are two reasons why
we do that. First, it is consistent with the definition of the
Matthew effect such that often-cited papers get cited more
often. If an article has not been cited at all and hence has no
recognition, the Matthew effect does not apply. Second, we
are also interested in the potential existence of uncitedness
effect such that zero-cited papers become less and less likely
to be ever cited over time. Therefore, we separate articles with
positive citations from those without citations to identify the
Matthew effect and uncitedness effect respectively. From
Table 1, we can see that the average log rank of an article
with positive cites is 3.64 with standard deviation 0.78.
Stremersch et al. (2007) measured the author-level
Matthew Effect by looking at authorseditorial board mem-
bership, number of publications in five journals, and ranking
of the business school they were affiliated with. We capture
editorial board membership in our author effect variable. One
strength of our model is that our ability to capture unobserved
variables (see below) makes it more easily implementable.
1
We also tried the square of article rank without log transformation. The
results remain the same but with less model fit.
60 J. of the Acad. Mark. Sci. (2015) 43:5272
Independent variable: uncitedness effect As discussed in the
conceptual framework section, we want to empirically test for
the existence of uncitedness effect. Similar to the
operationalization of the article-level Matthew effect, we con-
struct an independent variable of log of the number of years
with zero citations for those articles that never got cited (first-
order effect) as well as its square term (second-order effect).
Note that the value of these variables will be zero for those
articles with positive citations. Thus, the coefficient of the
square term can be used to test the uncitedness effect. In the
sample, the log of number of years with zero citations is 0.64
on average with standard deviation 0.49.
Independent variables: article effects author effects, and jour-
nal effects We capture the article effects on paper citation
through two measures: number of journal awards received
and subject area of the article. First, the awards we consider
include the following best article awards at various journals:
Best Article Award (IJRM), Best Article Award (JCR), Harold
H. Maynard Award (JM), MSI/H. Paul Root Award (JM), Paul
E. Green Award (JMR), William F. ODell Award (JMR), and
John D.C. Little Award (MKS), Sheth Foundation best paper
award (JAMS), best article award (JR). Since these awards are
chosen by editorial boards of the corresponding journal and
may be considered the choice of the highest-quality article by
leading scholars in marketing, the number of journal awards
received should be a good indicator of article quality. The
average number of awards received in our sample is 0.06 with
standard deviation 0.26 and the maximum of two awards.
Second, we adopt the categorization scheme of subject areas
used by Stremersch et al. (2007) and come up with 16 subject
areas (therefore 15 dummy variables as the last one otheras
the default case) as shown in Table 1. The subject area is the
subject on which an article focuses, and an article may belong
to multiple subject areas. Stremersch et al. (2007)haveshown
that the categorization is reliable and the subject areas have
Tabl e 1 Simple statistics Effects Variables Mean Std. Dev. Min Median Max
Dependent variable No of new cites 3.27 0.05 0 2 109
Author effects No of authors 2.30 0.93 1 2 8
Editbrdjm 0.35 0.58 0 0 3
Editbrdjmr 0.30 0.53 0 0 3
Editbrdjcr 0.28 0.55 0 0 4
Editbrdmks 0.24 0.53 0 0 3
Journal effects JM 0.16 0.36 0 0 1
IJRM 0.12 0.32 0 0 1
MKS 0.12 0.32 0 0 1
JMR 0.19 0.39 0 0 1
JAMS 0.14 0.35 0 0 1
JCR 0.16 0.37 0 0 1
Article effects No of awards 0.06 0.26 0 0 2
Sub_newprod 0.06 0.25 0 0 1
Sub_b2b 0.06 0.24 0 0 1
Sub_relationship 0.07 0.26 0 0 1
Sub_prodbrand 0.12 0.32 0 0 1
Sub_ad 0.06 0.25 0 0 1
Sub_pricing 0.08 0.27 0 0 1
Sub_promotion 0.06 0.23 0 0 1
Sub_retailing 0.10 0.30 0 0 1
Sub_strategy 0.09 0.29 0 0 1
Sub_sales 0.04 0.20 0 0 1
Sub_method 0.16 0.37 0 0 1
Sub_services 0.08 0.28 0 0 1
Sub_cb 0.17 0.37 0 0 1
Sub_international 0.05 0.23 0 0 1
Sub_ecom 0.03 0.17 0 0 1
Matthew effect Log(Rank last year) 3.64 0.78 0 3.81 4.58
Uncitedness effect Log(no of years with zero cites) 0.64 0.49 0 0.69 2.30
J. of the Acad. Mark. Sci. (2015) 43:5272 61
important impact on article citations. The subject areas of the
articles in our sample are fairly broad with slightly more
articles on consumer behavior, methodology, product and
brand management, retailing, and strategy.
Two measures are used to capture the author effects of an
article on its citations. The first one is number of authors for a
particular article. We can see that in the sample the average
number of authors per article is 2.3 with minimum 1 and
maximum 8. The second measure is the number of editorial
board membership among the authors for a particular article
for the four top marketing journals: JM,JMR,JCR,andMKS.
Table 1shows that the mean of the number of editorial board
membershipper article for the four journals is0.35, 0.30, 0.28,
and 0.24, respectively.
We incorporate the journal effects using six journal dummy
variables with JR as the default journal. In the sample, the
frequency of articles across these journals is roughly the same
with slightly higher frequency for JMR,JM,andJCR.
Independent variables: interaction effects We are interested
in the interaction between author effects and journal effects to
see if author fame is less important when a paper is published
in an A-level marketing journal. To simplify the analysis, we
use the following variables to capture the author effects or
journal effects: the total number of editorial board member-
ships across the four top marketing journalsJM,JMR,JCR,
and MKS (author effect)and whether it is an A-level jour-
nalJM,JMR,JCR,andMKS (journal effect).
Independent variables: controls To control for the time im-
pact on article citations, we include a lag log of cumulative
citations as a covariate. This variable may capture the persis-
tence of the article citations (Seetharaman 2004). To control
for any other missing article-specific variables, we incorporate
it through the article-specific intercept. Unobserved article
heterogeneity is accounted for using the random-coefficient
approach in the heterogeneity equation (Eq. 11). We also
check the correlations among the covariates and find that
multicollinearity is not an issue in this study.
Model comparison
To validate the proposed modeling framework, we compare it
to several benchmark models. This first benchmark is the
aggregate NBD model with the same covariates of the pro-
posed model but without accounting for article heterogeneity.
The second benchmark model is the heterogeneous NBD
model, which adds article heterogeneity to the first model,
both of which are non-nested within the proposed model. Both
NBD model setups are standard and available upon request
from the authors. The other models are nested benchmark
models of the proposed model. We estimate the aggregate
version of the proposed model without article heterogeneity,
and the proposed model with HMM of one state, two states,
three states, and four states, respectively, and report the results
in Table 2. To test if allowing the citability states switch up
only in the HMM fits the data better, we also estimated the
proposed model with HMM switch up only of one state, two
states, three states and four states, respectively. Due to space
constraint, we report only the best model out of the four
versionsthe two-state discretized model with switch up only
in HMM. The log marginal density (Chib 1995), mean abso-
lute error (MAE), and square root of mean square error
(RMSE) for each of the seven models for both the estimation
sample and holdout sample are presented in Table 2.
The results show that the heterogeneous NBD model out-
performs the aggregate NBD model with higher log marginal
density and lower MAE and RMSE for both the estimation
sample and holdout sample, demonstrating the importance of
accountingfor article heterogeneity. The aggregate discretized
Tobit model outperforms the aggregate NBD model, and the
one-state discretized Tobit model (i.e., the heterogeneous
discretized Tobit model) outperforms the heterogeneous
NBD model, indicating the appropriateness of the proposed
discretized Tobit framework. Out of the four versions (one
state, two state, three state, and four state) of the proposed
model with HMM, the two-state discretized Tobit is the best
model in terms of higher log marginal density, and lower
MAE and RMSE for both the estimation sample and holdout
sample. This model also outperforms the best model among
the four versions of the proposed model with HMM with
switch up onlythe two-state discretized model with switch
up only in HMM, demonstrating the possibility of HMM
states switching both up and down over time. Since the two-
state proposed model is the best model overall, we will focus
on the discussion of the estimation results of this model. Also,
to show the potential estimation bias in the heterogeneous
discretized Tobit model without HMM, we also report its
estimation results in Table 3(i.e., the results for the One-
State Model in the table). The significant estimates (i.e., zero
does not lie in the 95% posterior probability interval of the
estimate) are highlighted in bold.
Estimation results of the two-state model
From Table 3, we can see that the intercept of State 2 is higher
than that of State 1 following the identification condition of
HMM, indicating articles in the second state tend to receive
more new citations. Articles in State 2 also have slightly
higher persistence effect than those in State 1.
Author effects, journal effects, and article effects Interestingly,
we find that the author effects, journal effects, and article
effects vary significantly across the HMM states. The number
of authors significantly decreases the number of new citations
in the second state while it has insignificant impact in State 1,
62 J. of the Acad. Mark. Sci. (2015) 43:5272
possibly because it is considered as a negative signal in State 2
but not in State 1 by scholars. Articles authored by those with
prestigious editorial board membership at JM,JMR,JCR,and
MKS significantly attract more new citations in State 2.
However, in State 1 only the editorial board membership at
JM and JCR out of the four top marketing journals has
significantly positive impact on the number of new citations.
In the first state, the U.S.-based journals (all the journals in this
study except IJRM) tend to attract more new citations com-
pared to the base journal, JR, especially the top four marketing
journals (JM,JMR,JCR,andMKS). The more internationally
oriented journal IJRM attracts less citations compared to JR,in
both states. However, in State 2, only JM and JCR attract more
new citations compared to JR.
In terms of the subject areas, in the first state, we find that
articles in the areas of relationship marketing, services, con-
sumer behavior, international marketing, or ecommerce tend
to be popular topics and receive more new citations, while
articles on pricing, promotion, sales, or methodology attract
considerably less citations. But in State 2, the story is some-
what different. In the second state, articles on new products
also tend to attract more new citations, while those on busi-
nessto-business marketing, product and brand management,
advertising, pricing, retailing, sales, or methodology attract
considerably less citations. This may be due to the fact that the
popularity of subject areas in different states is different.
Lastly, regardless of the states, surprisingly, publishing in
A-level marketing journals (journal effects) tends to mitigate
the positive impact of prestigious editorial board membership
at JM,JMR,JCR,andMKS, with even more negative inter-
action effect in State 2. This indicates that author fame is less
important when a paper is published in an A-level marketing
journal, especially in the high citability state. The variance of
new citations in State 2 seems to be larger than that in State 1,
indicating more variation of citations in State 2.
Results of the HMM We find that in the year of publication, an
article on average is likely to start inthe first state with a much
higher probability of 0.78 (0.02) compared to the probability
of 0.22 (0.02) in the second state. (The numbers in parentheses
are the posterior standard deviations.) The average waiting
time of State 1 and 2 is 5.54 (0.06) and 5.61 (0.05) years,
respectively, which indicates slightly longer duration in the
higher state once in that state.
Table 4presents the estimation results on what drives the
starting probability of the HMM. To interpret the results, we
need to point out that the starting probability of the HMM
follows a Dirichlet distribution with its hyper-parameters as
functions of covariates shown in Table 4. The impact of one
covariate on the expected starting probability in one particular
state will be determined by the relative ratio of the estimated
coefficients for the covariate across the two states. For instance,
take the intercept as an example: it has insignificant impact on
the hyper-parameter in State 1 but negative impact on the hyper-
parameter in State 2. This means that overall, intrinsically an
article is more likely to start in the first state than in State 2.
Similarly, we find that the number of authors, editorial board
membership at JMR, or submitting to the journals IJRM,MKS,
or JMR (compared to submission to the base journal JR)in-
creases probability of the article starting in the low stateState
1. In contrast, the editorial membership at JM or submitting to
JCR results in higher starting probability of State 2, possibly
due to the relatively higher citability from these actions.
Interestingly, we also find that editorial board membership at
JCR or MKS, or submitting to the journals JM or JAMS (com-
pared to submission to the base journal JR) has no significant
impact on the articles starting probability in either state.
Table 5summarizes the estimation results for the waiting
time equations. Based on the estimated intercepts, we find that
once in State 1, intrinsically an article is likely to stay for
shorter time in the state and more likely to jump to the other
state compared to in State 2, which indicates higher persis-
tence in the high state. As time goes by, the article is less likely
to stay long in both states. However, there is a U-shaped
curve-linear impact of time on the waiting time in State 1
but not in State 2. This means that as time passes by and after
certain point, the article in the first state tends to stay in the
state longer and longer, while this trend does not exist for the
Tabl e 2 Model comparison Model Estimation sample Holdout sample
Log marginal density MAE RMSE MAE RMSE
Aggregate NBD model 22812.00 11.98 31.44 16.71 32.30
Heterogeneous NBD model 22382.30 11.38 21.77 12.26 22.88
Aggregate discretized Tobit 22752.16 11.93 25.96 14.71 27.13
One-state discretized Tobit 13795.06 10.94 15.34 12.12 17.86
Two-state Tobit with switch up only 6556.74 2.72 7.25 9.88 12.91
Two-state discretized Tobit 6614.95 2.67 6.63 9.39 11.67
Three-state discretized Tobit 6843.09 10.61 12.53 11.98 16.82
Four-state discretized Tobit 7162.75 11.24 15.27 12.12 21.91
J. of the Acad. Mark. Sci. (2015) 43:5272 63
second state. As expected, winning prestigious awards leads
the article to stay for less time in the low state and jump to the
high state, as well as stay longer in the second state due to
higher perceived quality from the scholars.
The estimate for the log rank is significantly positive in
State 1 but not in State 2, which shows that low-ranking
articles (with high log rank value) tend to stay in the low state
longer, possibly due to lower perceived quality. Consistent
with our expectation, we find there is a significant Matthew
effect for marketing articles in the data. Specifically, the
posterior mean estimate for the coefficient of the square of
log rank is significantly positive in State 1 while negative in
State 2, indicating the longer duration in the high state versus
low state for the reverse-scored high ranking papers. That is,
when an article becomes more famous with higher rank (i.e.,
lower value of log rank), it results in higher perceived quality
and hence stays in the high state longer, which in turn
Tabl e 3 Estimation results for
the discretized Tobit Model with
HMM
Effects Variables One-state model Two-state model
State 1 State 2
Intercept 0.619 (0.059) 0.333 (0.155) 0.001 (0.181)
Persistence effect Lag of log no of cites 0.991 (0.001) 0.813 (0.001) 0.825 (0.001)
Author effects No of authors 0.002 (0.005) 0.006 (0.004) 0.025 (0.005)
Editbrdjm 0.104 (0.016) 0.082 (0.018) 0.096 (0.017)
Editbrdjmr 0.127 (0.033) 0.046 (0.027) 0.129 (0.026)
Editbrdjcr 0.092 (0.025) 0.056 (0.025) 0.123 (0.021)
Editbrdms 0.056 (0.028) 0.046 (0.033) 0.063 (0.025)
Journal effects JM 0.306 (0.051) 0.460 (0.040) 0.278 (0.042)
IJRM 0.257 (0.059) 0.151 (0.039) 0.451 (0.051)
MKS 0.205 (0.059) 0.269 (0.063) 0.051 (0.051)
JMR 0.223 (0.052) 0.310 (0.036) 0.071 (0.040)
JAMS 0.078 (0.049) 0.149 (0.032) 0.032 (0.037)
JCR 0.278 (0.056) 0.347 (0.048) 0.180 (0.041)
Article effects Sub_newprod 0.165 (0.069) 0.094 (0.070) 0.158 (0.057)
Sub_b2b 0.130 (0.065) 0.060 (0.074) 0.166 (0.060)
Sub_relationship 0.285 (0.066) 0.295 (0.041) 0.216 (0.052)
Sub_prodbrand 0.034 (0.030) 0.032 (0.037) 0.079 (0.032)
Sub_ad 0.128 (0.061) 0.073 (0.059) 0.207 (0.055)
Sub_pricing 0.079 (0.056) 0.115 (0.059) 0.203 (0.045)
Sub_promotion 0.123 (0.068) 0.158 (0.065) 0.138 (0.071)
Sub_retailing 0.002 (0.048) 0.071 (0.050) 0.135 (0.040)
Sub_strategy 0.113 (0.041) 0.027 (0.045) 0.024 (0.035)
Sub_sales 0.169 (0.096) 0.175 (0.083) 0.292 (0.082)
Sub_method 0.048 (0.029) 0.067 (0.025) 0.076 (0.025)
Sub_services 0.264 (0.049) 0.303 (0.044) 0.116 (0.035)
Sub_cb 0.095 (0.025) 0.118 (0.034) 0.018 (0.033)
Sub_international 0.098 (0.078) 0.158 (0.066) 0.028 (0.076)
Sub_ecom 0.495 (0.105) 0.463 (0.090) 0.452 (0.076)
Journal effect*author effect 0.066 (0.010) 0.024 (0.011) 0.085 (0.010)
Variances in Tobit 2.183 (0.083) 0.200 (0.023) 0.339 (0.030)
Tabl e 4 Estimation results for the starting probability equations
Variables State 1 State 2
Intercept 18.134 (17.538) 28.615 (8.921)
No of authors 75.454 (7.799) 59.110 (6.646)
Editbrdjm 61.943 (9.033) 71.011 (23.265)
Editbrdjmr 57.288 (10.158) 0.923 (24.826)
Editbrdjcr 29.026 (20.385) 14.206 (11.106)
Editbrdmks 10.091 (12.705) 9.108 (11.456)
JM 15.600 (11.672) 3.225 (19.236)
IJRM 71.181 (12.922) 49.984 (15.815)
MKS 76.177 (14.181) 53.519 (6.920)
JMR 60.513 (10.108) 35.640 (13.313)
JAMS 9.921 (9.628) 11.644 (7.598)
JCR 7.315 (7.479) 20.358 (9.701)
64 J. of the Acad. Mark. Sci. (2015) 43:5272
disproportionately attracts more excess citations because of
fame alone.
The estimatefor the long number of years with zerocitation
is significantly negative in State 1 but not in State 2. It
indicates that zero-cited papers tend to stay shorter in the first
state and jump to the high state with the passage of time.
However, we find presence of significant uncitedness effect
for the marketing articles we analyzed. The posterior mean
estimate for the coefficient of the square of log number of
years with zero cites is significantly positive in State 1 but not
in State 2, indicating that as time passes by, after certain point,
the zero-cited papers tend to stay in the low state longer and
longer. That is, zero-cited papers are more and more likely to
be perceived to be relatively low quality and hence may be
less and less likely to get cited over time.
For comparison purposes, we also include the main esti-
mation results of the one-state discretized Tobit model in
Table 3, which is the same as the heterogeneous discretized
Tobit model without accounting for the latent citability states.
It is clear from Table 3that there exist significant estimation
biases in the one-state model compared to the two-state model,
which may lead to incorrect managerial implications (see
Appendix C).
Dynamics of HMM states
Next, we examine how the HMM states dynamically change
over time and how the predicted and actual citations differ by
the HMM states over time. Figure 2shows the estimated
average probability of an article being in the high citability
stateState 2 based on the estimation sample as well as the
logarithm of actual new citations over time. It is interesting to
see that as the actual new citations keep increasing over time,
an article typically has a low probability to start in the high
state (about 0.22) in the first two years, then jumps to 0.60 in
the third year and stays relatively stable afterwards. This
indicates the importance for authors of getting the article into
the high citability state early on.
Figure 3shows the actual and predicted new citations by
the HMM states over time. It is clear from this figure that the
citation dynamics are quite different across the two HMM
states. In the high state, the annual new citations of articles
tend to grow much faster than those in the low state. The
predicted new citations in the two states seem to match the
actual citations quite well. Again, this demonstrates the im-
portance of both across and within article heterogeneity (i.e.,
the dynamics of HMM states) in explaining the citation dy-
namics of articles.
Based on the estimation and predication results, we next
outline the theoretical contributions and implications. We
focus on the interesting insights on dynamic citation predic-
tion and the article citation dynamics which are not examined
in the marketing literature.
Summary
As Clark et al. (2013) and Stremersch et al. (2007) point out,
marketing scholars need to pay more attention to article im-
pact since it is in essence a marketing of scienceproblem.
Citations are an important measure widely used to assess the
quality and influence of an article. Although there exists much
research on citation measures and drivers in the marketing and
scientometric literature (Stremersch et al. 2007), to the best of
our knowledge, there is little recognition or research on the
dynamic nature of citation influence. Our model allows insti-
tutions, scholars, journals, and scientometric scholars to un-
derstand the citation dynamics and estimate the future impact
of a published article. The implications of our model are multi-
disciplinary in that the model can be used by scholars in varied
disciplines.
Theoretical contributions
Relying on the literature on relationship marketing, we intro-
duce to the literature the construct of article citability and
suggest that article influence is determined by its relational
embededness and the relational state an article is in. We
empirically demonstrate the influence exerted by this latent
factor on eventual article influence. The empirical results
show that the discretized Tobit model with two-state HMM
process outperforms all other competing models including the
widely used NBD models in our context. We find that an
article typically starts (with high probability) in the relatively
low citability state in the publication year. On average, both
latent states have high persistence such that an article will stay
in either state for more than five years before jumping to
another state, with slightly longer duration in the high state.
Different drivers of citations exert differential influence on
citability in the two states, and these influence dynamics
change over time. Article citability moderates the influence
of various known effects on final citation count.
Most models of citations such as the NBD model assume
that the influence drivers are static and do not change over
time with changes in an articles latent state. We demonstrate
Tabl e 5 Estimation results for the waiting time equations
Variables State 1 State 2
Intercept 21.695 (11.446) 14.339 (8.665)
Time 35.468 (7.146) 45.446 (10.256)
Square of Time 18.705 (3.989) 14.069 (10.001)
No of Awards 43.827 (4.448) 22.703 (9.270)
Log(rank) 36.522 (13.552) 9.984 (18.690)
Log
2
(rank) 74.763 (6.020) 42.328 (7.182)
Log(no of years with zero cites) 56.313 (14.742) 4.218 (7.240)
Log
2
(no of years with zero cites) 56.715 (8.784) 24.317 (18.742)
J. of the Acad. Mark. Sci. (2015) 43:5272 65
the shortcoming of this assumption and identify why it is
critical to capture these dynamic changes. We recommend
that researchers use our dynamic HMM model as it outper-
forms the NBD model and more accurately reflects citation
dynamics.
Our study not only confirms that the trend of increasing
yearly citations to the highly cited articles suggests that cita-
tions beget more citations, it but also provides a behavioral
explanation for why the article-level Matthew effect exists.
That is, highly cited articles result in a higher citability state
and hence lead to more citations. This is different from the
disproportionate attention rationale proposed by Small (2004)
which suggests first citations reflect expert judgments of the
contribution of an article: The operation of this feedback
mechanism, once set into motion will increase the inequality
of citations by focusing attention on a smaller number of
selected sources, and widening the gap between symbolically
rich and poor(p. 74). Our findings suggest that the article-
level Matthew effect has a positive effect that improves latent
citability states of the article. More importantly, we demon-
strate that the Matthew effect is not merely a famous author
effect but there is a famous papereffect as well. Our model
outperforms the commonly used NBD model and can account
for zero citations explicitly. Furthermore, to the best of our
knowledge, no study has investigated the existence and dy-
namics of an uncitedness effect and the article-level Matthew
effect on the citation dynamics of articles. To fill this
knowledge gap, we proposed a conceptual framework of
article citability and adopted a modified version of the
model proposed by Li et al. (2005) to model individual article
citations over time. Based on the citation literature, we
propose appropriate measures to quantify both the article-
level Matthew effect and the uncitedness effect.
Zinkhan (2005) suggests zero citation is a pattern charac-
teristic of the marketing literature: Over time, a few articles
emerge as influential and frequently cited and the majority of
articles sink into obscurity(p. 253). However, zero citation is
not unique to marketing. Garfield (2005) estimates that ap-
proximately 50% of the articles in the entire ISI database have
zero citations. The NBD model and other models that purport
to uncover the dynamics of citation fail to account for the
impact of uncitedness. We view this as a critical limitation of
those models.
Implications
All seven of the journals in our sample are highly regarded
journals. Our findings suggest that articles start out in one of
two citability states and these two states are not simply a case
of elite (A-level) journals starting out in higher citability states
in comparison to the near-elite (A-minus) journals. Citability
states and article influence are not merely about the ranking of
journals.
We find that quantitatively oriented articles (those pub-
lished in MKS,JMR,andIJRM) tend to start and stay in the
lower citability state. One important implication for marketing
journals (particularly MKS,JMR,IJRM) will be examining
why articles published at these outlets are in the relatively low
citability state in the year of the article publication compared
to the base journalJR. Or, tostate this differently, JR articles
have a higher citability than the more highly regarded MKS
and JMR. We conjecture that the more quantitatively oriented
MKS and JMR may have more articles using sophisticated
techniques that fewer researchers fully comprehend. Clark
et al. (2013) comment that the declining influence of market-
ing scholarship could be attributed in part to methodological
sophisticationtaking precedence over substantive issues.
Editors at MKS and JMR should pay closer attention to the
substantive and theoretical aspects of papers being submitted
to their journals.
To further explore this we did some post-hoc analysis on
JR. While JR is not positioned as a highly quantitative/
modeling journal per se, it does publish modeling papers
frequently. So we compared the citations accrued to modeling
versus all other papers at JR. A t-test indicated that modeling
papers in JR were less influential than non-modeling papers
published there. This leads us to conclude that journals such as
JMR and MKS that focus on highly sophisticated techniques
should explore ways to make their findings more accessible
by publishing an expanded non-technical summary of the
findings and paying closer attention to the theoretical contri-
butions of work submitted there. The greater citability state
and higher probability of starting in the higher state of JCR
compared to all other leading marketing journals is something
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
1234567891011
Year
Prob of Being in State
2
Log of Actual New
Cites
Fig. 2 Dynamics of HMM states over time
0
2
4
6
8
10
12
14
16
18
1234567891011
Year
Actual New Cites in
State 1
Predicted New Cites
in State 1
Actual New Cites in
State 2
Predicted New Cites
in State 2
Fig. 3 Actual and predicted new cites by HMM states
66 J. of the Acad. Mark. Sci. (2015) 43:5272
that needs to be explored further. JCR is positioned as an inter-
disciplinary journal and may perhaps draw attention from a
broader array of researchers. The greater influence dynamics
of JAMS relative to MKS and JMR is also worth noting.
Further, based on the estimated main effects of the journals
on new citations, it is important to note that only JM and JCR
have significant impact in both high and low citability states,
while MKS,JMR,andJAMS increase citations in the low state
only, and IJRM decreases citations in both states compared to
JR. Therefore, the latterfour journals may again want to boost
their relative citability states in scholarsminds. IJRM is
accorded an A-minus status in many institutions but seems
to have considerably less citability than its peer journals.
We find the presence of a strong article-level Matthew
effect and uncitedness effect for marketing articles in that they
increase or decrease the articles citability, which leads to more
or less influence for the article, respectively. There are signif-
icant author effects, journal effects, and article effects on
articlesinfluence dynamics, and these effects dynamically
change over time according to the latent citability state in
which the article is. The effect of author reputation is muted
or considerably diluted when the article is in a low citability
state. Also, for papers published in A-level marketing journals
or for papers starting out in the high citability state, author
editorial board membership and author fame are less critical
for eventual influence. This indicates that author fame is less
important when a paper is published in an A-level marketing
journal, especially in the high citability state. Further, we find
that editorial board membership at JM or publishing in JCR
results in a higher citabilitystate. And articles published in JM
or JCR seems to be in relatively higher citability states com-
pared to other leading marketing journals. Contrary to our
expectations, enhanced citability state of an article mitigates
the journal effects. When an article is in a low citability state, it
benefits from stronger journal effects compared to being in
high citability state. When that same article moves into a high
citability state, the journal effects become weaker possibly due
to some ceiling effect.
We find that the number of authors significantly decreases
the number of new citations in the high citability state while it
has insignificant impact in the low citability state. We also find
in our context that an article typically has a low probability of
starting in the high state (about 0.22) in the first 2 years, then
jumps to 0.60 in the third year and stays relatively stable
afterwards. In the high state, the annual new citations of
articles tend to grow much faster than those in the low state.
This indicates the importance for authors of getting the article
into the high citability state early on. Our research indicates
that if a paper is not cited early, while it may not go uncited in
the future, it may result in low relational state in scholars
minds, and hence its long-term influence is in great doubt.
Therefore, our study has important managerial implications
for journals, institutions, and scholars.
Limitations
We note the relative newness of Scopus in comparison to the
ISI database. There is a large literature that has investigated
problems in the ISI database such as the numerator/
denominator discrepancy, but Scopus was launched in 2004
so there is no research indicating what potential biases and
anomalies may be present in this database. There is also a
problem of an absence of comparable Scopus data; previous
studies using ISI data such as Zinkhan and Leigh (1999)
provide statistics from earlier years that can be used for
comparative purposes. Another difficulty is that Scopus
citations for social science journals have been extended back
only to 1996, which is insufficient for determining the time in
which citations to these marketing articles may decline and
exhibit ageing or obsolescence.
Another limitation of this study is the focus on a small
group of seven core marketing journals. Based on an analysis
of economic journals, Davis (1998) argues there are substan-
tial differences in citation impact patterns for general versus
specialized journals within a discipline. Further research is
needed to investigate patterns characteristic of specialized
marketingjournalssuchastheJournal of Consumer
Psychology,Journal of Advertising,orIndustrial Marketing
Management. Davis (1998) also argues for citation impact
differences between disciplinary and interdisciplinary
journals. Interdisciplinary business journals such as Harvard
Business Review,Sloan Management Review,and
Management Science may have different citation impact char-
acteristics than the marketing-focused journals. While we did
analyze 1,591 articles published in seven journals, we ac-
knowledge that these articles probably represent around 15
20% of all articles published in marketing journals during this
time frame.
2
The article ranking used to capture the article-level
Matthew effect in this study is based on the sample. The
Matthew effect may be even stronger with larger sample with
more articles or journals included. Further, due to data un-
availability, we do not observe self-citation information for the
articles in this study. Future research could investigate the
dynamics of self-citations and the difference from non-self-
citations over time. While we find that the number of co-
authors is a net positive for eventual influence, it does raise
interesting questions regarding the relative contributions of
individual authors. Sahu and Panda (2014) suggest that more
co-authors can foster interdisciplinary work and boost the
overall quality of the paper. But on the other hand, Persson
and Glanzel (2014) raise the specter of honorific co-authoring
2
Hult et al. (1997) ranked 29 marketing journals, ten business journals,
and two proceedings. The 7 journals we studied were among the 39
journals that were ranked. We conservatively estimate that the number
of articles analyzed would be about 1520% of articles published during
19962003 in the 39 journals ranked by Hult et al.
J. of the Acad. Mark. Sci. (2015) 43:5272 67
that might emerge when authorship is credited to those mak-
ing the most minimal of contributions. We commend re-
searchers to investigate this.
While citation analysis has been viewed as superior and
more objective than simply collating opinions, it has its lim-
itations and this needs to be acknowledged. Sometimes cita-
tions may not measure intellectual influence but may reflect
criticisms of a paper or temporary faddish interest in certain
topics (Hofacker et al. 2009). Our paper focused on aggregate
scholar/fields citation practices and did not focus on individ-
ual reader behaviors (Hofacker et al. 2009). Finally, our focus
was on the impact of academic research on the literature. In
applied disciplines, scholarship can also have impact on prac-
tice (Reibstein et al. 2009; Schultz 2012). Citation based
analysis does not measure this type of impact.
Appendix A
Journal differences and citation dynamics
Table 6presents the citation measure summary for the
seven journals in the data. Clearly, there exist big dif-
ferences across the journals. JM,JMR,andJCR have
the highest mean or median number of citations, highest
percentage of articles with 20 or more cites, and lowest
percentage of zero-cited articles, followed by JR and
JAMS,thenMKS and IJRM.
To see if different articles have different citation dynamics,
we divide the articles into two groups based on the average
total number of citations received at the end of the data period:
articles with higher than average total citations versus those
with lower. Then we plot the annual average new citations
over time for the two groups in Fig. 4and the annual change of
average new citations by year in Fig. 5. From Fig. 4,itisclear
that the citation dynamics for these two sets of articles are
dramatically different. Those articles with high total citations
in the end tend to grow much faster over time than those with
low total citations. For the latter, it seems that their new
citations initially increase and quickly become somewhat
stable after 4 or 5 years of publications with even slightly
declining trend after that. Figure 5further confirms the differ-
ent citation dynamics for the two sets of articles. Also, we can
clearly see from Fig. 5that the within-article citation
change varies dynamically over time. For instance, for
the articles with high total citations, the annual change
of citations first increases in the first 3 years, then
decreases afterwards with some bounce back in year
10. All these indicate that it is important to account
for article heterogeneity and the citation dynamics both
within and across articles.
Appendix B
Likelihood and identification of the proposed model
The unconditional likelihood is given by (Liechty et al. 2003):
Prob ZðÞ¼
i
s
πisðÞðÞ
IDi0¼s
fg
t
Prob Zit jDit
ðÞ
⋅∏
l
Pitsl IDit¼s&Ditþ1ðÞ
¼l
fg

⋅∏
k
fW
itk Dit ;λit;κi
j
ðÞ
;
where Zis the whole observed citation sequence across arti-
cles and time, π
i
is the starting probability at time 0, and P
itsl
is
the transition probability from state sto lat time tdefined in
the HMM process. I{} is an indicator function. D
it
is the
realized latent citability state for article iat time t. f(W
itk
|D
it
,-
λ
it
,κ
i
) is the density of the k-th waiting time of article iat time
tconditional on the realized hidden state D
it
and Gamma
shape and scale parameters λ
it
and κ
i
, which is given by
fW
itk Dit;λit ;κi
j
ðÞ¼ζkX
m¼0
κiλik Dit
ðÞ
Γλ
ik Dit
ðÞðÞ
Witk Dit
ðÞτkþm
ðÞ
λik Dit
ðÞ1exp κiWitk Dit
ðÞτkþm
ðÞ
fg
Iτkþm<Witk Dit
ðÞτkþmþ1
fg:
Tabl e 6 Citation difference
across journals Influence measure JM JMR JCR MKS JAMS JR IJRM
# of articles 96-03 241 302 282 189 212 180 185
Mean number of Citations 32.25 17.63 16.03 11.45 15.66 16.40 8.46
Median number of citations 19.0 12.0 11.0 7.0 10.0 10.0 5.0
Maximum number of cites 423 108 149 68 139 110 94
Percent uncited 1.6 1.0 2.8 4.4 1.9 3.6 5.8
Percent with 20 or more cites 49.0 29.5 27.8 15.8 23.3 25.5 9.0
68 J. of the Acad. Mark. Sci. (2015) 43:5272
Here λ
ik
(D
it
) is the Gamma shape parameter associated with
the state of an article when the k-th jump time begins and with the
realized state Dat time t,andτ
k+m
is the m-th time after τ
k
the
k-th jump time of the articlethat Dchanged states, with τ
k+0
=
0. ζ
k
is a normalizing constant of the density.
To ensure identification of the hidden states, we restrict the
average new citations to be non-decreasing in the states (Li et al.
2011; Netzer et al. 2008). That is, we impose this restriction at
the intercepts of Eq. 2in the text such that γ
0i1
γ
0i2
≤⋯≤γ
0iS
.
We also refer scholars interested in the estimation procedure to
Li et al. (2005), Liechty et al. (2003) and Li et al. (2011).
Appendix C
Comparison to the one-state model results
For comparison purposes, we include the main estimation
results of the one-state discretized Tobit model in Table 3in
the text. Note that the one-state model is the same as the
heterogeneous discretized Tobit model without accounting
for the citability states. It is clear from Table 3that there exist
significant estimation biases in the one-state model compared
to the two-state model, which may lead to incorrect
managerial implications. For instance, without incorporating
the dynamics of citability states of articles, the one-state model
tends to over-estimate the persistence effect (p-value = 0).
Regarding the author effects, based on the one-state model,
authors may incorrectly conclude that the number of co-
authors does not matter in attracting new citations but it
actually may hurt when the article is in the high citability
state. The one-state model also tends to over-estimate the
impact of the editorial board membership at JM,JMR,JCR,
and MS compared to those in the low state in the two-state
model while under-estimating their impacts compared to those
in the high state. For the journal effects, there exist significant
under-estimation biases when compared to those in the low
state but over-estimation biases when compared to those in the
high state in the two-state model. Similarly, for the article
effects, significant estimation biases are also present in the
one-state model. For example, the topics of new products and
business to business marketing are shown to have significant
impacts on new citations in the one-state model while their
effects are only significant in the high state in the two-state
model. Many subject areas (i.e., product and brand manage-
ment, advertising, pricing, promotion, retailing, sales, meth-
odology, international marketing) are shown to have insignif-
icant impact in the one-state model but have significant impact
0
1
2
3
4
5
6
7
8
9
10
1234567891011
Year
Articles with Hight
Total Cites
Articles with Low Total
Cites
Fig. 4 Average new cites by year
-0.5
0
0.5
1
1.5
2
2.5
12345678910
Year
Articles with Hight
Total Cites
Articles with Low
Total Cites
Fig. 5 Annual change of average
new cites by year
J. of the Acad. Mark. Sci. (2015) 43:5272 69
either in low state or high state in the two-state model. Lastly,
the one-state model seems to over-estimate the variance ofthe
new citations compared to those in the two-state model. These
estimation biases in the one-state model highlight the impor-
tance of accounting for the dynamics of latent citability states
of articles over time.
Appendix D
Concentration of citations to the top five articles by journal
The effects of hitpapers are also indicated in statistics on
concentration of citations to the top five articles of the year.
Table 7presents an overview of the extent of concentration by
year for these selected journals. MKS in 2000 records the
highest concentration of citations (72.2%), reflecting the suc-
cess of a special issue on marketing science and the Internet.
Concentration in this case is the percentage of citations for the
five most cited articles of the year relative to total citations for
all articles in the year. The results suggest the journals with
lower prestige may be consistently high in their concentration
percentages or dependencies on their most cited articles. For
example, concentration for IJRM ranges from 49.3 to 66.2%
over the period of 19962003.
There is also significant variability year-to-year reflecting
the publishing of hitarticles or the lack of hitsin a year.
For example, the top five 1999 articles in MKS account for
only 39.9% of cites. However, in 2000 the concentration
jumps to 72.2%, driven by the success of three articles on
internet shopping from a special issue of MKS (Novak et al.
2000; Lynch and Ariely 2000 and Haubl and Trifts 2000).
JCR has the lowest average concentration of these seven
journals. However, JCR concentration jumps from 27.8% in
1997 to 47.3% in 1998. The year 1998 includes the only three
JCR articles with more than a hundred citations in this dataset
(Bettman et al. 1998; Fournier 1998 and Steenkamp and
Baumgartner 1998).
References
Archambault, É., Campbell, D., Gingras, Y., & Larivière, V. (2009).
Comparing bibliometric statistics obtained from the web of science
and scopus. Journal of the American Society for Information
Science and Technology, 60(7), 13201326.
Baldi, S. (1998). Normative versus social constructivist process in the
allocation of citations: a network-analytic model. American
Sociological Review, 63(December), 829846.
Baumgartner, H., & Pieters, R. (2003). The structural influence of mar-
keting journals: a citation analysis of the discipline and its sub-areas
over time. Journal of Marketing, 67,123139.
Bergh, D. D., Perry, J., & Hanke, R. (2006). Some predictors of
SMJ article impact. Strategic Management Journal, 27,81100.
Berry, L. L. (1983). Relationship marketing. In L. Berry, G. L. Shostack,
&G.D.Upah(Eds.),Emerging perspectives on services marketing
(pp. 2558). Chicago: American Marketing Association.
Bettencourt, L. A., & Houston, M. B. (2001). The impact of
article method type and subject area on article citations and
reference diversity in JM, JMR, and JCR. Marketing Letters,
12(4), 327340.
Bettman, J. R., Luce, M. F., & Payne, J. W. (1998). Constructive con-
sumer choice process. Journal of Consumer Research,
25(December), 187217.
Tabl e 7 Concentration of citations to the top five articles by journal and year of publication
Year of article publication JM JMR JCR MKS JAMS JR IJRM
Citations to top five of 1996 1016 269 204 316 227 192 299
% of total citations of 1996 articles 44.1% 35.3% 28.6% 52.9% 56.0% 53.8% 57.6%
Citations to top five of 1997 794 462 279 218 440 233 135
% of total citations of 1997 articles 54.5% 40.7% 27.8% 42.1% 68.6% 49.8% 49.3%
Citations to top five of 1998 554 212 549 270 163 187 126
% of total citations of 1998 articles 34.6% 24.9% 47.3% 56.5% 50.3% 62.8% 51.9%
Citations to top five of 1999 448 336 200 212 258 94 59
% of total citations of 1999 articles 29.1% 34.8% 33.7% 39.9% 42.3% 49.0% 54.6%
Citations to top five of 2000 299 191 161 346 237 203 131
% of total citations of 2000 articles 33.6% 35.4% 24.2% 72.2% 32.5% 59.2% 66.2%
Citations to top five of 2001 192 227 162 113 119 119 55
% of total citations of 2001 articles 32.6% 38.0% 26.5% 51.6% 57.5% 46.7% 56.7%
Citations to top five of 2002 157 98 104 81 104 68 58
% of total citations of 2002 articles 28.6% 34.6% 24.2% 39.3% 30.8% 58.6% 61.1%
Citations to top five of 2003 91 78 69 65 50 36 41
% of total citations of 2003 articles 32.6% 40.4% 28.9% 34.9% 45.0% 54.5% 64.1%
Mean concentration of citations to top five 1996-2003 36.2% 35.5% 30.2% 48.7% 47.9% 54.3% 57.7%
All the citation numbers in the table are as of 2006
70 J. of the Acad. Mark. Sci. (2015) 43:5272
Bonitz, M., Bruckner, E., & Scharnhorst, A. (1997). Characteristics and
impact of the Matthew effect for countries. Scientometrics, 40,407422.
Burrell, Q. L. (2003). Predicting future citation behavior. Journal of the
American Society for Information Science and Technology, 54(5),
372378.
Chib, S. (1995). Marginal likelihood from the Gibbs output. Journal of
the American Statistical Association, 90(432), 13131321.
Chib, S., & Greenberg, E. (1995). Understanding the Metropolis Hastings
algorithm. American Statistician, 49,327335.
Chow, C. W., Haddad, K., Singh, G., & Anne, W. (2007). On using
journal rank to proxy for an articles contribution or value. Issues in
Accounting Education, 22(3), 411427.
Clark, T., Key, T., Hodis, M., & Rajaratnam, D. (2013). The intellectual
ecology of mainstream marketing research: an inquiry into the place
of marketing in the family of business disciplines. Journalofthe
Academy of Marketing Science.doi:10.1007/s11747-013-0362-5.
Cole, J. R., & Cole, S. (1973). Social stratification in science. Chicago:
University of Chicago Press.
Cronin, B. (1981). The need for a theory of citing. Journal of
Documentation, 37,1624.
Damien, P. S., Wakefield, J., & Walker, G. (1999). Gibbs sampling for
Bayesian nonconjugate and hierarchical models using auxilliary
variables. Journal of Royal Statistical Society, Series B, 61(Part 2),
331344.
Davis, J. B. (1998). Problems in using the social sciences citation index to
rank economic journals. American Economist, 42(Fall), 5965.
Du, R., & Kamakura, W. (2006). Household life cycles and lifestyles
in the United States. Journal of Marketing Research, 43(1), 121
132.
Dwyer, F. R., Schurr, P. H., & Sejo, O. (1987). Developing buyer-seller
relationships. JournalofMarketing,51(April), 1127.
Fournier, S. (1998). Consumers and theirbrands:developing relationship
theory in consumer research. Journal of Consumer Research,
24(March), 343373.
Garfield, E. (1979). Citation indexingits theory and application in
science, technology, and humanities. New York: John Wiley &
Sons.
Garfield, E. (2005). The agony and the ecstasythe history and the
meaning of the Journal Impact Factor. http://garfield.library.upenn.
edu/papers/jifchicago2005.pdf.
Gelfand, A., & Smith, A. (1990). Sampling-based approaches to calcu-
lating marginal densities. Journal of the American Statistical
Association, 85,398409.
Guidry, J. A., Guidry, B. N., Holler, L. J., Tanner, J. R., & Veltsos,
C. (2004). Surveying the cites: a ranking of marketing journals
using citation analysis. Marketing Education Review, 14(Spring),
4559.
Haubl, G., & Trifts, V. (2000). Consumer decision making in online
shopping environments: the effects of interactive decision aids.
Marketing Science, 19(Winter), 421.
Hendrix, D. (2008). An analysis of bibliometric indicators, National
Institutes of Health Funding, and faculty size at Association of
American Medical Colleges, medical schools. Journal of Medical
Library Association, 96(4), 324334.
Hult, G., Tomas, M., Neese, W. T., & Edward Bashaw, R. (1997). Faculty
perceptions of marketing journals. Journal of Marketing Education,
19(1), 3752.
Hofacker, C., Gleim, M. R., & Lawson, S. (2009). Revealed reader
preference for marketing journals. Journal of the Academy of
Marketing Science, 37(2), 238247.
Ijiri, Y., & Simon, H. A. (1974). Interpretations of departures from the
Pareto curve firm-size distributions. Journal of Political Economy,
82,315331.
Katerratanakul, P., Han, B., & Hong, S. (2003). Obective quality ranking
of computing journals. Communications of the ACM, 46(October),
111114.
Keller, K. L. (2012). Understanding the richness of brand relationships:
research dialogue on brand as intentional agents. Journal of
Consumer Psychology, 22,186190.
Lehmann, D. R. (2005). Journal evolution and the development
of marketing. Journal of Public Policy & Marketing, 24(1), 137
142.
Li, S., Liechty, J. C., & Montgomery, A. L. (2005). Modeling category
viewership of web users with multivariate count models, working
paper. Bloomington: Kelley School of Business, Indiana University.
Li, S., Sun, B., & Montgomery, A. L. (2011). Cross-selling the right
product to the right customer at the right time. Journal of Marketing
Research, 48(4), 683700.
Liechty, J., Pieters, R., & Wedel, M. (2003). Global and local covert
visual attention: evidence from a bayesian hidden Markov model.
Psychometrika, 68(4), 519541.
Luo, A., & Kumar, V. (2013). Recovering hidden buyer-seller relation-
ship states to measure the return on marketing investment in
business-to-business marketing. Journal of Marketing Research,
Vo l. L ,143160.
Lynch, J. G., Jr., & Ariely, D. (2000). Wine online: search costs affect
competition on price, quality, and distribution. Marketing Science,
19(Winter), 83103.
Montgomery, A. L., Li, S., Srinivasan, K., & Liechty, J. C. (2004).
Modeling online browsing and path analysis using Clickstream
Dat.Marketing Science, 23(4), 579595.
Moon, S., Kamakura, W. A., & Ledolter, J. (2007). Estimating promotion
response when competitivepromotions are unobservable. Journal of
Marketing Research, 44(3), 503515.
Morgan, R. M., & Hunt, S. D. (1994). The commitment-trust theory of
relationship marketing. Journal of Marketing, 58(July), 2038.
Medoff, M. H. (2006). Evidence of a Harvard and Chicago Matthew
effect. Journal of Economic Methodology, 13(4), 485506.
Merton, R. K. (1968). The Matthew effect in science. Science,
159(January 5), 5663.
Mingers, J., & Burrell, Q. L. (2006). Modeling citation behavior in
management science journals. Information Processing and
Management, 42,14511464.
Mort, G. S., McColl-Kennedy, J. R., Kiel, G., & Soutar, G. N. (2004).
Perceptions of marketing journals by senior academics in Australia
and New Zealand. Australasian Marketing Journal, 12(2), 5161.
Netzer, O., Lattin, J. M., & Srinivasan, V. (2008). A hidden Markov
model of customer relationship dynamics. Marketing Science, 27(2),
185204.
Novak, T. P., Hoffman, D. L., & Yung, Y.-F. (2000). Measuring the
customer experience in online environments: a structural modeling
approach. Marketing Science, 19(Winter), 2242.
Persson, O., & Glanzel, W. (2014). Discouraging honorific authorship.
Scientometrics, 98(2), 14171419.
Pieters, R., & Baumgartner, H. (2002). Who talks to whom? Intra- and
interdisciplinary communication of economic journals. Journal of
Economic Literature, 40(2), 483509.
Reibstein, D. J., Day, G., & Wind, J. (2009). Guest editorial: is
marketing academia losing its way? Journal of Marketing,
73(4), 13.
Rossi, P. E., McCulloch, R. E., & Allenby, G. M. (1996). The value of
purchase history data in target marketing. Marketing Science, 15(4),
321340.
Sahu, S. R., & Panda, K. C. (2014). Does the multi-authorship trend
influence the quality of an article. Scientometrics, 98(3), 2161
2168.
Samiee, S., & Chabowski, B. (2012). Knowledge structure in
international marketing: a multi-method bibliometric analysis.
Journal of the Academy of Marketing Science, 40(2), 364
386.
Schultz, D. E. (2012). The academic silly season. Marketing News,
September 30, 14.
J. of the Acad. Mark. Sci. (2015) 43:5272 71
Seetharaman, P. B. (2004). Modeling multiple sources of state depen-
dence in random utility models: a distributed lag approach.
Marketing Science, 23(2), 263271.
Seggie, S. H., & Griffith, D. A. (2009). What does it taketo get promoted
in marketing academia? Understanding exceptional publication pro-
ductivity in the leading marketing journals. Journal of Marketing,
73(January), 122132.
Seglen, P. O. (1992). The skewness of science. Journal of the American
Society for Information Science, 43(9), 628638.
Seglen, P. O. (1997). Whythe impact factor of journals should not be used
for evaluating research. British Medical Journal, 314(February 15),
498502.
Simon, H. A. (1955). On a class of skew distribution functions.
Biometrika, 42,425440.
Sivadas, E., & Johnson, M. S. (2005). Knowledge flows in marketing: an
analysis of journal article references and citations. Marketing
Theory, 5(4), 339361.
Small, H. (2004). On the shoulders of Robert Merton: towards a norma-
tive theory of citation. Scientometrics, 60(1), 7179.
Spence, M. A. (1973). Job market signaling. Quarterly Journal of
Economics, 87,355374.
Steenkamp, J.-B. E. M., & Baumgartner, H. (1998). Assessing measure-
ment invariance in cross-national consumer research. Journal of
Consumer Research, 25(June), 7890.
Stern, R. E. (1990). Uncitedness in the biomedical literature. Journal of
the American Society for Information Science, 41(3), 193196.
Stremersch, S., Verniers, I., & Verhoef, P. C. (2007). The quest for
citations: drivers of article impact. Journal of Marketing, 71(July),
171193.
Tanner, M. A., & Wong, W. H. (1987). The calculation of posterior
distributions by data augmentation. Journal of the American
Statistical Association, 82(398), 528540.
Tol, R. S. J. (2009). The Matthew effect defined and tested for the 100
most prolific economists. Journal of the American Society for
Information Science and Technology, 60(2), 420426.
Varadarajan, P. R. (2003). From the editor: Journal of the Academy of
Marketing Science, 2 2000 to 2003. Journal of the Academy of
Marketing Science, 31(4), 365367.
Weale, A. R., Bailey, W., & Lear, P. A. (2004). The level of non-citation
of articles within a journal as a measure of quality: a comparison to
the impact factor. BMC Medical Research Methodology, 4,14.
Woodside, A. G. (2009). Journal and author impact metrics: an editorial.
Journal of Business Research, 62(1), 14.
Zinkhan, G. M. (2005). From the editor: scientific status and knowledge
use: two perspectives. Journal of the Academy of Marketing
Science, 33(Summer), 251253.
Zinkhan, G. M., & Leigh, T. W. (1999). Assessing the quality ranking of the
Journal of Advertising. Journal of Advertising, 28(Summer), 5170.
72 J. of the Acad. Mark. Sci. (2015) 43:5272
... To better understand the factors that contribute to such scholarship, this longitudinal study examines the effects of coauthorship, database indexing, and article length on citation counts, a widely used measure of scholarly impact (Li, Sivadas, & Johnson, 2015). To be clear, it is important to note that other methods exist for determining the impact of an academic's work (Howard & Garland, 2015). ...
... Critics, however, have argued that the impact of a journal represents a problematic method for assessing the impact of the individual articles featured in a journal (Li et al., 2015). For instance, journal impact factors-perhaps the most widely used measure of journal impact-are based upon the mean number of citations the av-erage article receives over a given time period. ...
... The shift toward a more direct measurement of scholars' work is not unique to social work (Li et al., 2015). The evaluation of individual outcomes is largely a university-driven phenomenon (Holosko & Barner, 2016). ...
Article
Objective Disseminating high-impact scholarship is a critical task for many social work academics. Although the factors that contribute to this process have been investigated in other disciplines, there is a paucity of equivalent research in social work. This longitudinal study addresses this gap in the literature by examining the effects of coauthorship, database indexing, and article length on subsequent citation counts, a widely used measure of scholarly impact. Based upon the extant research, we hypothesized that all three factors would be associated with a greater number of citations 5 years after publication. Method The sample consisted of 3,066 articles, published inclusively from 2005 to 2009 in 18 disciplinary social work journals. Multilevel negative binomial regression was used to model the effects of each factor on 5-year citation counts. Results The findings generally supported the hypotheses. Articles were more likely to be cited in subsequent scholarship if they were (a) written by 3 or more authors, (b) retrievable from more databases, and (c) longer. Conclusions The results raise the possibility that authors interested in high-impact scholarship might benefit from working in authorship teams to create longer papers containing more original ideas, and then submitting the resulting manuscripts to journals that are indexed in multiple electronic databases.
... Impact measures pertaining to public uptake, such as Alternative metrics (Altmetrics), often take the sum of article-related press releases, case studies, public policy documents, and patents, as well as public, social, and alternative media (Altmetric 2017;Bornmann, Haunschild, and Adams 2019;Costas, Zahedi, and Wouters 2015;Gumpenberger, Glänzel, and Gorraiz 2016;Mukherjee, Subotić, and Chaubey 2018;Ozanne et al. 2017;Thelwall et al. 2013). Prior research into what influences research articles' academic impact shows that university reputation, affiliation, and journal ranking matter (Li, Sivadas, and Johnson 2015;Stremersch, Verniers, and Verhoef 2007). The impact of service research articles also is fundamentally driven by their content and style. ...
... We controlled for several external features of the articles that previous research has linked to their impact: the number of authors, whether the first author has a U.S. affiliation (i.e., employed by a U.S. university or institution ¼ 1 or not ¼ 0), whether (¼ 1) or not (¼ 0), the article was a JSR Best Paper award winner, and the age of the article as the number of years since its publication (e.g., Li, Sivadas, and Johnson 2015;Stremersch and Verhoef 2005). Third, in line with Humphreys (2010) and Trusov, Bucklin, and Pauwels (2009), we include a press release measure of the number of press releases containing the full title and journal information within a year of the article's publication date, obtained from the Dow Jones Factiva database (https://global-factiva-com). ...
Article
Full-text available
For service researchers, contributing to academic advancement through academic publications is a raison d’être. Moreover, demand is increasing for service researchers to make a difference beyond academia. Thus, service researchers face the formidable challenge of writing in a manner that resonates with not just service academics but also practitioners, policy makers, and other stakeholders. In this article, the authors examine how service research articles’ lexical variations might influence their academic citations and public media coverage. Drawing on the complete corpus of Journal of Service Research ( JSR) articles published between 1998 and 2020, they use text analytics and thereby determine that variations in language intensity, immediacy, and diversity relate to article impact. The appropriate use of these lexical variants and other stylistic conventions depends on the audience (academic or the public), the subsection of this article in which they appear (e.g., introduction, implications), and article innovativeness. This article concludes with an actionable “how-to” guide for ways to increase article impacts in relation to different JSR audiences.
... it is widely believed that the number of citations received by a journal or an article measures (or reflects) its influence, impact, and/or intrinsic quality (Baumgartner and Pieters 2003;Eisend and Lehmann 2016;Li et al. 2015;Stremersch et al. 2007). ...
... In marketing, there are to date three major studies that investigated the drivers/factors/ causes of article citations (i.e., Li et al. 2015;Stremersch et al. 2007Stremersch et al. , 2015. One of these studies indicates that article length has a significant positive effect on article citations (see Stremersch et al. 2007, p.180). ...
Article
To date, limited studies have examined the citations of articles published in predatory journals, and none appears to have been done in marketing. Using Google Scholar (GS) as a citation source, this study aims to examine the extent of citations of (articles published in) 10 predatory marketing journals. Citation analyses indicate that the most cited predatory marketing journal gathered 6,296 citations since it was first published in 2008. Four of the 10 predatory journals gathered over 732 citations each since they were launched (i.e., highly cited). Three other journals were cited between 147 and 732 times (i.e., moderately cited). The three remaining journals received below 147 citations each (i.e., trivially cited). Findings show that the 1,246 articles published in these 10 predatory journals, and which are visible to GS, received 10,935 citations with 8.776 citations per paper. About 11.624% of these 1,246 articles were cited 13 times or more. The most cited article received 217 citations, of which 21 are from journals indexed in Clarivate Analytics’ Social Sciences Citation Index. Based on these findings, this study concludes that the conventional marketing literature has been already contaminated by predatory marketing journals.
... On the other hand, perceived research weighting is less likely to be different between genders. Specifically, unlike teaching and service, research output is not a within-institution metric and is "credited" outside of a university and across academic circles and networks through impact factors (e.g., Li, Sivadas, & Johnson, 2015;Theuβl, Reutterer, & Hornik, 2014). In fact, research performance is an expectation for most tenure track and tenured professors throughout business schools. ...
... We did not hypothesize differences in PPW-R across genders because universities are more likely to clearly enunciate research expectations within and across disciplines. Even though these expectations may vary with different business disciplines and subdisciplines, impact factors, journal rankings, and citation counts tend to be an extremely salient and highly researched metric for research performance (e.g., Li et al., 2015;Theuβl et al., 2014). On the other hand, these types of third-party metrics are not readily available for service and teaching performance. ...