Professional Documents
Culture Documents
1. Introduction
Recent research has reopened the classic sociological problem of how systems of
categories shape social structures (Durkheim and Mauss 1969 [1903]). Much of this
research works in the spirit of what Garfinkel (1968) called breaching experiments.
It seeks to understand the role of categorical boundaries by examining the conse-
quences to actors who ignore them. In this context, ignoring category boundaries
means claiming membership in multiple categories.
Category spanners adopt patterns of feature values that fit more than one cat-
egory and, sometimes, claim membership in more than one category at one time
or over time in a sequence of categorical affiliations. Sociological studies show that
categorical boundaries are indeed consequential in diverse domains: combining cat-
egories generally generates some form of direct or indirect devaluation (Hannan
2010; Negro, Koçak, and Hsu 2010b).
The line of research on boundary crossing has not reached its full potential
because it has not considered the structure of the space of categories. In the interest
of tractability, researchers treat all pairs of categories alike, meaning that all kinds
of spanning are expected to have the same consequences. Yet some combinations of
categories are much more difficult to interpret than others. Therefore the impact
of combining categories depends on the socio-cultural distance separating them.
This paper brings distance into sociological models of categories. It builds mod-
els that take account of categorical distances and proposes an empirical strategy
for using commonly available data to estimate them. Before turning to these mat-
ters, we briefly sketch the range of sociological applications for which our proposed
models and methods have relevance.
Date: November 4, 2011.
We appreciate the helpful comments of Glenn Carroll, Greta Hsu, and Gaël Le Mens. We thank
the University of Lugano, Stanford Graduate School of Business, and the Stanford GSB Faculty
Trust for generous financial support for this project.
1
2 BALÁZS KOVÁCS AND MICHAEL T. HANNAN
hard-to-interpret producers and devalue such producers when they do transact with
them.
Here we specialize the arguments to contexts that do not supply a categorical
focus. That is, we assume that audience members perceive a set of label assign-
ments, and they try to make sense of the producers in light of the assigned labels.
If an offering bears only one label (e.g., a film is labeled only as a “Comedy”) then
the simple label provides a categorical focus. But suppose that two labels have
been applied (e.g., a film is tagged as “Comedy” and “Science Fiction”). How do
audiences make sense of the combination? We propose that answering this ques-
tion requires attention to the distances among categories, specifically the distances
among an audience member’s schemas.
Schemas are cognitive representations of what it means to be a full-fledged mem-
ber of a category. Specifically schemas are combinations of feature values. In the
fuzzy-membership approach, a producer/offer that fits fully the constraints ex-
pressed by an audience member’s schema for a label has grade of membership
(GoM) of one in that agent’s meaning of the label. Partial fits produce partial
memberships. In other words, GoMs in schemas tell degrees of typicality.
The socio-cultural distance between a pair of categories depends on how much
their associated schemas overlap. If the overlap is considerable, then the categories
lie close to each other in the socio-cultural space. If the schemas do not overlap,
then the schemas are distant. In the latter case, a producer cannot fit both schemas
well. Any offering that partly fits a pair of distant schemas must be a very atypical
instance of the labels associated with these schemas.
Based on this reasoning, we propose that an offering’s typicality in any label falls
with (1) the number of labels used to describe it and (2) the distances among the
schemas associated with these labels. When many (distant) labels are applied to a
producer, atypicality characterizes the producer as a whole: not only is it atypical
for some label, it is atypical for all of the applied labels. We use the term overall
typicality to refer to such a global assessment of the degree to which a producer fits
any label.
This reasoning implies that producers that do not fit any applied label, those
with high overall typicality, have low intrinsic appeal. We now spell out some of
the implications of this idea using a nonmonotonic logic. The theory on which
we build expresses postulates as rules with exceptions, a formulas quantified by a
nonmonotonic quantifier N (Pólos and Hannan 2002, 2004). This quantification
expresses what “normally” holds, with the proviso that more specific information
can overrule the normal case.
We use the predicate nf(y), which reads as “each producer bears a set of category
labels, the audience member y associates non-empty schema with each label, and the
context does not focus the audience’s attention to one of them.” We incorporate
this predicate in the antecedents of the formulas in the following postulates and
propositions to make clear that the argument applies to this special case.
As noted above, the argument builds on a premise about the relationship between
(overall) typicality and intrinsic appeal.
Postulate 1. A producer’s intrinsic appeal increases with its overall typicality.
N x, x0 , y [nf(y) ∧ (t(x, y) ≥ t(x0 , y)) → α(x, y) > α(x0 , y)],
6 BALÁZS KOVÁCS AND MICHAEL T. HANNAN
where t(x, y) is a real-valued function that tells the overall typicality of the producer
x to the audience member y, and α(x, y) is a real-valued function that records the
intrinsic appeal of the producer x to the audience member y.
Next we tie this postulate to more specific ones that relate categorical niche
width to typicality and category contrast to typicality.
Categorical Niche Width and Intrinsic Appeal. Category generalists are
confusing—they are “neither fish nor fowl.” Such generalism can be expressed
well in terms of the width of the categorical niche. Producers associated with one
category—category specialists—have a niche width of zero. Niche width increases
as a producer gains a broader and more diverse set of categorical affiliations.
Postulate 2. A producer’s typicality declines with its categorical niche width.
N x, x0 , y [nf(y) ∧ (w(x, y) ≥ w(x0 , y)) → t(x, y) < t(x0 , y)],
where w(x, y) denotes a non-negative, real-valued function that gives the categorical
niche width of the producer x from the perspective of the audience member y. (Below
we discuss ways to measure niche width.)
A testable implication of the argument follows immediately from Postulates 1
and 2.
Proposition 1. A producer’s intrinsic appeal decreases with its categorical niche
width.5
P x, x0 , y [nf(y) ∧ (w(x, y) ≥ w(x0 , y)) → α(x, y) < α(x0 , y)].
Proof. This follows from a cut-rule applied to Postulates 1 and 2.
Contrast and Intrinsic Appeal. The second part of the argument concerns a
category-level variable: contrast. According to Hannan et al. (2007), categories
with very fuzzy boundaries exert less social power than crisper ones. A relatively
crisp category stands out from the social background and likely serves as a basis of
enduring expectations about those who bear the category label. The main intuition
holds that membership in a high-contrast category conveys greater advantage than
membership in a fuzzy category. This is because audience members generally find
the crisper label to be a more stable basis on which to form expectations. This
intuition has been found to hold in diverse contexts, as noted in the Introduction.
Negro, Hannan, and Rao (2011) argue that pervasive spanning by members of
a label, especially long-distance spanning, lowers its contrast and thereby reduces
the appeal of all of its members. Lowered contrast likely reduces the appeal of
all offerings in a category in two ways. One involves the relationships among cat-
egories. Fuzziness implies a loss of distinctiveness of a category relative to the
others, raising questions about what comparisons are appropriate for the members
of a category. With increasing fuzziness, clusters of producers/offerings become
less salient and elicit lower attention. Previous research shows that comparisons
become more difficult; audience members have trouble using distinct descriptors,
5In the nonmonotonic logic used in this line of theory building, the implication of a set of rules
with possible exceptions is a formula quantified by another nonmonotonic quantifier P. This
quantification expresses what “presumably” follows from the premises of the current stage of a
theory. Future theory stages might incorporate more specific considerations might destroy a
provisional implication from an earlier theory stage.
CATEGORY SPANNING, DISTANCE, AND APPEAL 7
Again we have a testable implication of the argument, which in this case follows
immediately from Postulates 1, 3, and 4.
Proposition 2. When categorical focus does not hold, the intrinsic appeal of a
producer generally increases with the maximum contrast of the categories assigned
and decreases with secondary contrast.
concepts such as “fruit” and “furniture” have an internal structure: apples and or-
anges are viewed as typical fruits and olives and pineapples as atypical fruits and
so forth (Rosch 1975; Rosch and Mervis 1975; Hampton 2007). One useful way to
represent such internal structure is to view the label as referring to a fuzzy set, a set
whose membership function admits partiality (Hannan et al. 2007). Our research
implements this view.
Recent research on categories attempts to construct meaningful GoMs in labels
(such as film genres and product categories) from sparse data that does not allow
measurement of schemas.6 The now-common study design obtains assignments to a
predetermined list of category (or genre) labels. In most previous research settings,
as well as ours, some market intermediary (such as a the managers of publications
or websites that post reviews) assigns the labels. The analyst lacks information
about how individual audience members would apply the labels. This means that
using such data to test arguments stated at the level of the audience member, as
above, requires an assumption of homogeneity. If members of the audience use
the domain language in idiosyncratic ways, there is little hope of finding systematic
relationship between combinations of the (externally given) labels and the responses
of the audience members. What some agents will see as spanning will not be so for
others, and so forth. So we must assume that the audience uses the language in
a homogenous manner, that they associate similar schemas with the labels of the
domain.
The basic data on label assignment can be represented as a vector that assigns
to a producer a value (say one) for each label assigned and a value (zero) for those
that are not. The analyst wants to calculate a GoM in each label for each market
participant as a first step in characterizing producers and categories.
Suppose the language of the relevant domain contains L labels for producers.
Throughout we make reference to the label function (of an unspecified party7) as
given by (
1 if label i is assigned to x;
li (x) =
0 otherwise.
The basic analytical question asks how to use a label profile to make inferences
about memberships in the categories/genres that correspond to the labels. We next
consider some answers to this question.
A Qualitative Approach. Much recent work follows what we call the qualitative
approach (for instance, Pontikes 2008; Hsu et al. 2009; Pontikes 2009; Carroll et al.
2010; Negro et al. 2010a; Kovács and Hannan 2010; Negro et al. 2011). It works as
follows. In the first (largely implicit) step, the analyst assumes that, because the
schemas for the various labels differ (they impose different constraints), producers
and products with only one label generally fit better the schema for that label than
those assigned two labels. For instance, a film classified as “Comedy” and “Horror”
likely lacks the typical features of either genre. Similarly a restaurant labeled as
6Obviously we would prefer to have access to data that tell what schemas the audience members
associate with the relevant labels. Then categories could be represented as sets in a space of the
values of categorically relevant features and relations. Questions about combining categories could
then be addressed in terms of positions in the feature space.
7Because the labeling is not done by the members of the audience, we do not assign an “audience-
member slot” to this function.
10 BALÁZS KOVÁCS AND MICHAEL T. HANNAN
“Mexican” and “Thai” can hardly typify either label.8 The reasoning then makes a
similar assertion about two-label versus three-label entities, and so forth.
Following this reasoning, a market participant’s GoM in any label generally
declines with its number of labels. In particular this reasoning suggests that the
GoM function in any assigned label decreases monotonically with the number of
labels assigned: µi (x) = g(lx ), with g 0 (lx ) < 0 subject to the condition that g lies
in the unit interval: 0 ≤ g(lx ) ≤ 1.
Hsu et al. (2009) proposed the following functional form for relating label as-
signments and GoM that satisfies these desiderata:9
li (x)
(1) µQi (x) = ,
lx
where lx denotes the number of labels applied to x. For example, if three labels
are applied to a producer or product, then its GoM in each of these labels is set to
1/3, and its GoM in each of the other labels is set to zero.
We refer to this approach as qualitative because it does not use any information
distance, about the space of the categories. Specifically it does not adjust for the
fact that some pairs of categories are closer than others. Our proposal seeks to
rectify this limitation by bringing distance into the picture.
A Metric Approach: Incorporating Distance. We want to measure GoMs
in labels in a way that reflect a market participant’s typicality for each label. As
noted above, a producer whose feature values cause it to be assigned to two distant
labels is generally an atypical member of each. But one that gets assigned two close
labels might be quite typical of each. We propose that GoMs in labels be defined
in a way that incorporates information about the distances among labels:
l (x)
(2) µD
i (x) = P i ,
1 + j∈L lj (x) dij
where dij denotes the distance from the label i to the label j with (dii ≡ 0).
This definition, like the qualitative one (eqn. 1), sets µi (x) = 1 if i is the only
label assigned to x, and it sets µj (x) = 0 for j 6= i in such cases. When two labels,
i and j, are assigned, then it sets µi (x) = (1 + dij )−1 and µj (x) = (1 + dji )−1 , and
so forth. The addition of each label lowers GoMs, but it does so much more when
the added label lies far from the others.
The qualitative procedure for calculating GoMs from label assignments (eqn. 1)
can be viewed as a special case of our proposed measure (eqn. 2) that constrains all
of the categories to stand at a distance of unity from one another. That is, eqn. (1)
results from imposing the constraint dij = 1 for all i and j in eqn. (2).
and perceived distances are asymmetric, that, say, label i is more similar to label
j than vice versa. So in general we refer to the distance from one label to another.
We denote the distance from i to j as dij and the distance from j to i by dji . Of
course if distance is symmetric, then dij = dji .
Following the foundational work of Shepard (1987), we posit a negative expo-
nential relationship between perceived socio-cultural distance and similarity:
(3) s(i, j) = exp(−γ dij ), γ > 0.
Although we vary the measure of similarity, we leave unchanged the relation be-
tween similarity and distance in eqn. 3.
This paper proposes that the relatedness of categories is reflected by their ten-
dency to co-occur in systems of classification. For example, if “Western” films also
tend to be classified as “Drama,” we infer that these labels have similar meanings.
Such a frequentist approach enables researchers to map out the relationships among
categories as we show below.
Symmetric Similarity. The standard assumption that distance is a metric en-
tails the assumption of symmetry. So we begin with this case and later relax the
assumption of symmetry.
We use a simple and widely used symmetric measure of category similarity due
to Jaccard (1901).10 The Jaccard similarity of a pair of labels amounts to a simple
calculation on their extensions.11 Let i denote the extension of li , that is, i = {x |
li ∈ l(x)}. Then the Jaccard similarity of labels li and lj can be defined as the
ratio of the number of producers/products that are categorized as both li and lj
to the number that are categorized as li and/or lj . Formally, if |i ∩ j| denotes the
cardinality of the set of producers that are categorized as both li and lj , and |i ∪ j|
denotes the cardinality of the set of producers that are categorized as li and/or lj ,
then
|i ∩ j|
(4) SimJ (i, j) = .
|i ∪ j|
This index takes values in the [0, 1] range, with 0 denoting perfect dissimilarity and
1 denoting perfect similarity. For example, the dataset on restaurants analyzed
below contains nine restaurants labeled as “Malaysian” and eleven “Singaporean.”
Four of these restaurants are assigned both labels. Thus the Jaccard similarity of
“Malaysian” and “Singaporean” in these data is 3/(9 + 11 − 4) = 0.25.
Asymmetric Similarity. An influential line of research in cognitive psychology,
stimulated largely by Tversky (1977), questions the symmetry of similarity judg-
ments. Research shows that more prominent objects are judged less similar to less
prominent ones than vice versa (e.g., Tversky found that subjects judged North
Korea to be more similar to Red China than the reverse). More generally objects
for the subject knows more feature values are judged as less similar to those about
which they know fewer feature values, e.g., a portrait is more similar to its subject
than the reverse.
10For a detailed discussion of alternative similarity and dissimilarity measures, see Batagelj and
Bren (1995). Some preliminary results show that the main findings of this paper apply to other
measures as well, but we leave this direction of investigation for further research.
11In the usual language of logic and linguistics, the extension of a label refers to the set of objects
that bear the label.
12 BALÁZS KOVÁCS AND MICHAEL T. HANNAN
While previous formalizations assumed that niche width increases simply with
the number of labels assigned, we now need a measure that also incorporates the
structure of the categories. Specifically, we want a measure to have the following
properties:
(1) be non-negative,
(2) have a minimal value of zero (if an agent gets assigned to a single category),
(3) increase with the number of categories assigned and with the distances
among them.
We consider two measures that meet these desiderata. Both use the total pairwise
distance among the labels assigned:
P P
(9) Dx = i∈L j∈L li (x) lj (x) dij ,
and the average of the pairwise distances among the labels assigned:
Dx
(10) d¯x = .
lx (lx − 1)
The first measure sets categorical niche width to the product of the number of
labels assigned and the average of the pair-wise distances between them.12 For
reasons explained below, we call this a constant-increment measure of width and
denote it by WC .
(
0 if lx = 1;
(11) WC (µx ) = ¯
lx dx otherwise.
The qualitative behavior of this measure can be seen more clearly when we replace
d¯x according to eqn. (10) with the restriction that lx > 1:
WC (µx ) = Dx (lx − 1)−1 , lx > 1.
With this measure, niche width increases with total distance at a constant positive
rate (when the number of labels assigned is held constant):
∂WC
= (lx − 1)−1 > 0, lx > 1.
∂Dx
In other words, if the number of applied labels remains constant but the total
pairwise distance among them increases (for instance, if one label is replaced with
another that lies further away from the rest), then the niche grows wider. But,
holding total distance constant, niche width decreases if the number of labels as-
signed increases (for instance if one label far from the others is replaced with two
closer labels):
∂WC
= −Dx (lx − 1)−2 < 0, lx > 1.
∂lx
We think this is a desirable property for a measure of niche width, because more
labels with the same total distance means that the applicable labels cluster more
tightly in the socio-cultural space. A market participant that spans a closely packed
set of labels has a narrower niche than one that spans fewer more-distant labels.
One property of WC has less substantive appeal. The effect of increasing total
pairwise distance does not depend on the level of niche width. The same given
12We also explored setting niche width equal to the total distance among the labels assigned, the
numerator in equation 10. However, this measure performed poorly as compared with the one
based on the product of the number of labels and average distance.
14 BALÁZS KOVÁCS AND MICHAEL T. HANNAN
increase in D has the same implications when D ≈ 0 and when D 0. From the
perceptual perspective, these two situations ought to be quite different. In the for-
mer, a producer that had a sharply defined categorical niche as a specialist in one
label starts getting hard to comprehend. In the later, an already extremely confus-
ing instance is getting more so. We think that something like a perceptual ceiling
effect operates here—there is a limit on how confusing a producer’s categorical
position can be.
So we also propose an alternative measure with the property that niche width
increases more with a given increment of distance at low levels of total distance
spanned than at higher levels. In other words, if a producer with a narrow niche
adds a label at a fixed distance, its niche width increases more than if it initially
had a broad niche. We refer to this measure as the nonproportional measure, which
we denote it by WN . We represent this idea as follows:
(
0 if lx = 1;
(12) WN (µx ) = ¯ −1
1 − (1 + lx dx ) otherwise.
Category Contrast. Contrast refers to the degree to which a set stands out from
the background, the clarity of its boundary. Hannan, Pólos, and Carroll (2007)
defined contrast as the average GoM in the category among those with positive
GoM. Using our notation for label functions and extensions, we have
P
(14) C(li ) = x∈i µi (x)/|i|.
Obviously the qualitative approach for measuring GoM from label assignments
understates the contrasts of categories that lie close to others in the space of cat-
egories and overstates contrasts for categories that overlap distant categories. In
our empirical analysis below, we show that adjusting for distances in calculating
contrasts does indeed make a difference.
CATEGORY SPANNING, DISTANCE, AND APPEAL 15
%#!!!"
%!!!!"
$#!!!"
$!!!!"
718"9,18:;<:'"
#!!!"
='<">8?)@)<"
!"
&' 01 /"
+) /"
&' 01 /"
+) /"
&' 01 /"
+) /"
&' 01 /"
+) /"
&' 01 /"
+) /"
&' 01 /"
+) /"
0 !/"
"
$/
.
!
,-!
() 2-!
,-!
() 2-!
,-!
() 2-!
,-!
() 2-!
,-!
() 2-!
,-!
() 2-$
,-$
-$
12
+)
*
*
()
&'
(not just the restaurant labels in the combined data).14 This classification uses the
average clustering method applied to all labels that contain five or more members,
in which each step joins two clusters if the average distance between the members
of the clusters is smaller than any other possible combinations of clusters at that
point of the classification.15 The numbers at the top of the figure indicate the
average distance between the members in that joint cluster. Consider the branch
at the bottom of this figure. At the first level it combines “Tea rooms” with the
set consisting of “Moroccan,” “Turkish,” “Middle Eastern,” “Greek” and “Mediter-
ranean.” At the next branching point, it breaks out “Moroccan,” and so forth. The
main branch ends with the pair “Greek” and “Mediterranean;” these are very close
according to this analysis.
Note that several pairs of labels have highly overlapping extensions and thereby
are quite close to each other in the nonmetric space in Figure 2. The include
“Spanish–Basque,” “Greek–Mediterranean,” “Pakistani–Indian,” and “Japanese–Sushi
bar.” We note these pairs because taking account of the distance spanned makes
the biggest difference for them.
For example, the qualitative measure assigns a GoM of 0.5 in each label to a
restaurant categorized as both “Spanish” and “Basque.” What happens under the
alternative that pays attention to distance spanned? Answering this question re-
quires that we supply a value for the free parameter γ that relates distance and
similarity in equation 3. In the analyses reported below we experimented with dif-
ferent values of γ and found that model fits were highest with γ = 1. So we use
this value in this illustration (and below in our regression analyses). Because the
Jaccard similarity of “Spanish” and “Basque” in the combined data equals 0.48, the
distance between these categories equals − ln(0.42)/1 = 0.73. So our measurement
strategy assigns GoM of 1/(1+0.73)=0.54 in both labels. What about dissimilar
labels? Consider “Chinese” and “French,” whose Jaccard similarity is 0.0007. Using
the Shepard transformation, this gives a distance 7.2, which means that a restau-
rant in the intersection has GoM= 1/(1+7.2)=0.12 in each genre according to our
definition of GoM. Recall that the qualitative measure assigns a GOM of 0.5 to
such cases.
0 5 10 15 20
Burmese
Indonesian
Singaporean
Malaysian
Portuguese
German
Fondue
Russian
Modern European
Argentine
Tapas/Small Plates
Creperies
Belgian
Fish & Chips
Irish
Gastropubs
British
Cafes
Ethiopian
African
Tapas Bars
Spanish
Basque
French
Brasseries
Soul Food
Cluster Dendrogram
Cajun/Creole
Caribbean
Cuban
hclust (*, "average")
Cambodian
Himalayan/Nepalese
Persian/Iranian
Halal
Indian
Pakistani
Filipino
Moroccan
Kosher
dist
Chicken Wings
Latin American
Peruvian
Steakhouses
Seafood
Brazilian
Buffets
Chinese
Taiwanese
Hawaiian
Korean
Barbeque
Japanese
Sushi Bars
Asian Fusion
Thai
Turkish
Middle Eastern
Mediterranean
Greek
Tex−Mex
Mexican
Burgers
Fast Food
Italian
Pizza
American (New)
American (Traditional)
Breakfast & Brunch
Diners
Soup
Vietnamese
Delis
Sandwiches
Food Stands
Hot Dogs
Vegetarian
Vegan
Gluten−Free
Live/Raw Food
Contrast
San Francisco Los Angeles
Q D
Genre using µ using µ using µQ using µD
Asian Fusion 0.709 0.538 0.659 0.476
Barbeque 0.731 0.570 0.696 0.550
Chinese 0.914 0.862 0.882 0.815
Italian 0.803 0.715 0.766 0.675
Indian 0.698 0.681 0.676 0.619
Japanese 0.687 0.664 0.724 0.666
Modern European 0.583 0.416 0.524 0.265
Pakistani 0.509 0.542 0.467 0.457
Steakhouses 0.747 0.603 0.718 0.590
Thai 0.939 0.905 0.912 0.863
Vegetarian 0.522 0.321 0.505 0.284
result, contrast generally rise even for these categories when we adjust for distances
as can be seen in Table 2. Only for “Pakistani” in San Francisco does contrast rise.
However, it declines only slightly for “Chinese,” “Indian,” “Italian,” “Japanese,”
and “Thai.” At the other extreme, contrast falls considerably with the distance
correction for “Asian Fusion,” “Barbeque,” and “Vegetarian.” What matters here is
not the exact magnitudes of changes in contrasts (as these depend on the specific
distance measure used) but that incorporating distance alters the order of categories
in terms of contrast.
Categorical Niche Width. Adjustments for distance also affect measures of cat-
egorical niche width. Table 3 compares the three measures (WQ , WC , and WN )
for some restaurants with three or four label assignments (the measures do not
differ for category specialists). Of course, WQ , does not discriminate cases with
the same number of labels assigned. However, the distance-based measures do
discriminate. The combination in the first row of Table 3, “Chinese,” “Japanese,”
“American (New),” and “Hawaiian” spans a very considerable distance, and it re-
ceives have values for WC , and WN . However, the restaurant in the fourth row,
which also bear four labels, combines more similar categories: “Persian/Iranian,”
“American (New),” “Middle Eastern,” and “Mediterranean.” Consequently the val-
ues of distance-based measures of niche width are lower. The table illustrates a
similar pattern among restaurants assigned three labels.
The comparisons of the extreme cases in Table 3 also show that the range of
variation of WC is very large relative to that of the constrained WN . Over the
ranges in this table, the ratio of the maximum to the minimum of WC is 1.65,
while the same calculation for WN gives 1.03. So these two measures differ greatly
for extreme values, even though they mostly agree on the ordering. The Pearson
correlation of the two distance-based measures is 0.85 for the multi-label cases.
The correlations of the qualitative measure with the two distance-weighted mea-
sure are modest for the multi-label cases. The correlation of WQ and WC is 0.43;
20 BALÁZS KOVÁCS AND MICHAEL T. HANNAN
WQ WC WN Labels
0.750 24.0 0.959 Chinese, Japanese, American (New), Hawaiian
0.750 19.2 0.950 Chinese, Japanese, Korean, Thai
0.750 17.5 0.936 Middle Eastern, Greek, Barbeque, Med.
0.750 14.7 0.941 Persian/Iranian, American (New), Middle Eastern, Med.
0.667 18.3 0.948 Mexican, Korean, Burgers
0.667 15.1 0.938 Japanese, Italian, Asian Fusion
0.667 7.54 0.883 Russian, Modern European, German
0.667 4.84 0.829 Singaporean, Malaysian, Indonesian
and the correlation of WQ and WN is 0.28. Note that in some cases the ordering
switches, for example the three-label restaurant “Mexican,” “Korean,” “Burgers”
have a wider niche according WC than the four-label “Persian/Iranian,” “American
(New),” “Middle Eastern,” “Mediterranean” restaurant.
We also pay attention to another aspect of the niche. Any restaurant that
engages outside the very broad food category has a very wide niche. We constructed
a dummy variable to capture this dimensions of niche width, labeled “any non-food
category,” equals one for restaurants with an assignment to a non-food category
such as “Gas station” and equals zero for those whose category assignments all
come from the food domain.
Maximum and secondary contrast. We measure maximum contrast for a restaurant
as the maximum over the labels assigned to it of the label contrasts (average GoM in
a label for those restaurants with positive GoM in the label), and we set secondary
contrast to the value for the next-highest contrast label assigned. Of course, the
values of these two variables depend on which method is used for calculating GoMs
in labels.
The distributions of these theoretically relevant variables can be found in Table 4.
Note that we discriminate among the various combinations of measurements of
interest. This table shows that organizational niche width is much more sensitive
to choice of alternative measures than contrast. We will see below that this means
that choice of measure has more impact on estimated effects of niche width on
appeal.
Controls We control for variation among reviewers in engagement in the category
and the website using (the natural log of) the number of reviews posted for restau-
rants bearing the category label, following Koçak, Hannan, and Hsu (2009). We
refer to this variable as the reviewer’s activism. We also control for the producer’s
prominence (on the website), measures as the natural log of the number of reviews
it receives. We include the date of review to control for secular trends in appeal.
Finally, we control for price levels, which can take four ordered values: coded one
(for “cheap”) to four (“splurge”).
Hypotheses. In forming hypothesis for test with data on appeal, we must make an
assumption about producers’ engagement with the audience. Our theory pertains to
intrinsic appeal and the ratings reflect actual appeal, which, according to Hannan
et al. (2007), depends on intrinsic appeal and engagement. It seems clear that
the restaurants have engaged the audience, but we do not know much about the
22 BALÁZS KOVÁCS AND MICHAEL T. HANNAN
Two other hypotheses concern the expected difference in effect of maximum and
secondary contrast on appeal:
Hypothesis 2. The actual appeal of restaurants increases with the maximum
(distance-weighted) contrast of the categories assigned.
Hypothesis 3. The actual appeal of restaurants decreases with secondary (distance-
weighted) contrast of the categories assigned.
Results. At a broad level, our analyses reveal three clear tendencies. First, the
results mainly support the main study hypotheses in a robust way: the hypothesized
patterns hold across the various ways of measuring GoMs, similarity, and niche
width (with the exception of the results that bear on Hypothesis 2 using data for
LA). Second, the specifications built on distance-related measures fit substantially
better than those that build on the distance-independent qualitative measures.
Third, the distance-based specifications generally yield more consistent evidence
in support of the hypotheses. Introducing some adjustment for category structure
CATEGORY SPANNING, DISTANCE, AND APPEAL 23
WQ WC WN
SimQ SimJ SimT SimJ SimT
(1) (2) (3) (4) (5)
San Francisco
ln L −378,503 −377,883 −377,888 −378,092 −378,133
Los Angeles
ln L −306,829 −306,794 −308,802 −306,776 −306,779
matters more for substantive conclusions than the choice among the alternative
ways of making these adjustments.
Recall that nearly three-quarters of the organizations are category specialists—
they get assigned only one label. We have done the analysis with these specialists
included and with them excluded. The general pattern of results is quite similar.
However, analysis restricted to the restaurants with two or more labels is more
informative for our purposes, because the specialists do not vary in niche width
or secondary contrast. The results for the category specialists, in the Appendix,
reveal a strong and significant effect of category contrast on appeal, in support of
Hypothesis 2.
Here we concentrate on the multiple-category restaurants. Because the compar-
isons of interest involve many different measures, we begin by considering model fits
(without providing all of the parameter estimates) in Table 5. We are especially in-
terested in comparing the fits of the specifications built on the qualitative measures
with those based on distance-weighted measures. It is clear that the qualitative
approach yields specifications (columns 1 and 4 in Table 5) that fit much less well
than those that make adjustments for the distances among categories.
The best fit partly differs by city. For SF, the best fits come from the specification
with proportional niche width, but the reverse is the case for LA. With that choice
set to the best fit, there are only small differences in fit stemming from the choice
between the Jaccard and Tversky measures of similarity. For both cities, the fits
are marginally better with the Jaccard measure than with the asymmetric Tversky
measure.
The key findings that bear on our hypotheses appear in Table 6, which reports
the parameter estimates for the qualitative approach along with the those based on
proportional niche width for SF and nonproportional niche width for LA.
Consider the effects of niche width, beginning with SF. Here we see an impor-
tant substantive consequence of taking account of distances among categories. The
qualitative approach gives results that run counter to Hypothesis 1: the estimated
effects of niche width in column 1 in Table 6 are positive for all price levels (and
significant for all but the third price category). According to these results, the
audience prefers restaurants with broad categorical niches. However, the estimates
of parameters of specifications that build on the distance-weighted measures, in
columns 2 and 3, tell the opposite story. They show that a broad niche increases
appeal at the lowest price level but beyond that price point, increasing niche width
24 BALÁZS KOVÁCS AND MICHAEL T. HANNAN
is associated with lower appeal, with the effect becoming stronger and more signif-
icant at higher price levels. In other words, once account is taken of the distances
spanned, it becomes clear that having a broad categorical niche makes a restaurant
less appealing to this audience.
For LA, the qualitative approach and metric (distance-weighted) approaches
agree in yielding a pattern in which niche width increases appeal in the lower price
categories. According to the results built on the qualitative approach, the effect of
niche width on appeal remains positive (and significant) at the third price range
and only become negative at the highest price range. The distance-based estimates
indicate the effect of niche width on appeal is negative and significant at both of
the higher price ranges, with the absolute value of the effect greatest at the highest
price.
Overall we see that the parameter estimates with distance-weighted measures
provide stronger and more systematic support for Hypothesis 1 than do those built
on the qualitative measure of similarity/distance.
Next consider maximum contrast. The first-order effect on appeal is positive for
SF, but negative and insignificant for LA in the better fitting specifications. So
only the results for SF support Hypothesis 2. Figure 3 plots the implied effects
of maximum contrast for SF for the qualitative, Jaccard, and Tversky measures
(over the varying ranges of maximum contrast produced by the different measures).
Note that the effect of maximum contrast is considerably weaker (flatter) with the
qualitative measure and strongest with the Tversky measure.
Finally, Hypothesis 3 receives strong support. For both cities, appeal declines
with increasing secondary contrast; and this effect is significant in all cases for the
distance-based approach (but only for LA with the qualitative approach). The
better-fitting models show that, net of the effect of maximum contrast, higher
secondary contrast lowers appeal, as hypothesized. Figure 4 illustrates the joint
effect of maximum and secondary contrast on appeal using the estimates in the
second column in Table 6. The peak of the function (the effect on appeal) in
the upper right equals approximately 0.35. Moving to the left from that point
(increasing secondary contrast) drives the function toward zero, the value when
secondary contrast equals one. The decline in appeal with falling maximum contrast
is steeper. At the left-hand face of the graph of the function, the effect on appeal
falls from 0.35 to roughly −0.35 as maximum contrast falls over its observed range.
We explored whether the effects of the theoretically relevant variables differ for
the more and less active reviewers. Again the answer differs by city. For SF it is
clear that activists are less sensitive to issues of interpretability. For all three of
the theoretically relevant effects there is a statistically significant interaction with
activism whose sign opposes the sign of the main effect. For instance the coefficient
of the interaction of niche width and activism is 0.066 in column 2 in Table 6. At
the mean of (ln) activism for SF, 3.86, the interaction effect is 0.25; at the 75th
percentile, 4.29, it is 0.28; at the maximum, 7.29, it is 0.48. These differences are
small relative to the main effects of niche width. The total effect of niche width
(including the main effect and the interaction) is negative even at the observed
maximum of activism for the three highest price levels. So, although activists react
less to boundary crossing, they find broad spanners less appealing than restaurants
that do not span categories or that span close by categories. The situation for the
interactions of activism with maximum and secondary contrast in SF is similar. The
26 BALÁZS KOVÁCS AND MICHAEL T. HANNAN
−0.1
−0.2
Effect on appeal
−0.3
−0.4
−0.5 Qualitative
Jaccard
Tversky
−0.6
−0.7
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Maximum contrast
7. Discussion
Our point of departure is a general pattern emerging in contemporary research:
organizations affiliated with multiple categories suffer diminished appeal in the
eye of audience members (Zuckerman 1999; Dobrev, Kim, and Hannan 2001; Hsu
et al. 2009). We argue that better understanding of the consequences of category
spanning requires consideration of the structure of the underlying categories space.
We made two arguments. First, the socio-cultural distance of the categories be-
ing combined affects evaluations: the less similar are the spanned categories, the
greater is the confusion of identity (the overall atypicality) resulting from combining
them. Such confusion lowers appeal to audience members. Second, the contrasts of
categories also influence the consequences of combination. Because high-contrast
categories come with stronger expectations and norms, we expect that combining
CATEGORY SPANNING, DISTANCE, AND APPEAL 27
0.4
0.3
0.2
Effect on appeal
0.1
−0.1
−0.2
−0.3
0.9
0.8
0.7 1
0.6 0.9
0.8
0.5 0.7
0.4 0.6
Secondary contrast 0.3 0.5
0.4
0.2 0.3 Maximum contrast
0.1 0.2
0.1
Our goal to take category structure into account led us to rework this approach.
We proposed that inter-category distances be built directly into the GoM function
in a way that yields lower GoMs for combining more distant categories. We then
use the new measure of GoM to calculate category contrast. Thus we arrive at a
measure of contrast that builds on the category structure of the domain. We also
introduced two alternative measures to categorical niche width that incorporate the
similarity structure of the categories.
In the empirical part of the paper we analyzed customers’ evaluations of restau-
rants in SF and LA (many of which cross categorical boundaries, such as “Mexican”–
“Vietnamese”). We found generally strong support for our theoretical propositions.
We also found that the proposed measures provide a better model fits than the
approach used previously, suggesting that they provide a more precise description
of the data. In other words, incorporating information about category structure
indeed leads to a better understanding of the consequences of category spanning.
Using labels to infer category membership, however, could be more tenuous
than we have assumed. For instance, multiple labels do not differentiate between
“foodcourt” and “fusion” situations (Baron 2004). That is, a “Mexican”-“French”
restaurant can be either such that on side of its menu it offers Mexican dishes
while on the other side it offers French dishes; or, it might serve only dishes that
fuse elements of the two cuisines. These are qualitatively distinct cases of category
spanning, and these restaurants would attract different audiences. We cannot,
however, tell these cases apart in our data. Future work is needed to address this
distinction both theoretically and empirically (for example, by analyzing restaurant
menus).
Our proposed approach could be useful in diverse empirical settings. Organi-
zational examples include law firms (some practices are closer to others), wineries
(looking at different blends of varietals to calculate distances among varietals),
movies (certain movie genres are closer to others), financial organizations such as
hedge funds (some stocks and financial instruments are closer to each other than
others, thus hedge funds differ in their focus). Other possible applications include
category combination in work (Leahey 2007), innovation (Carnabuci et al. 2011),
or culture (Goldberg 2011).
The measures proposed for estimating distance of categories could aid researchers
in studying the evolution of categories over time. Categories that are distant in one
time period might become much less so in subsequent ones. Adjusting for the dis-
similarity of categories seems particularly useful when the categorical structure is
in flux. Indeed the finding that distance among categories and the contrast of cate-
gories influence producers’ evaluation might have interesting dynamic consequences.
In the case of categorical contrast, if producers in low-contrast categories are more
likely to cross boundaries, then the contrasts of these categories further decrease;
and, similarly, we expect that the contrast of high-contrast domains would further
increase or at least remain stable. Pontikes and Hannan (2011) find evidence of
such a pattern in the software industry. These processes would imply a tendency
toward the macro-level polarization of categories’ contrasts.
Another possibly interesting dynamic links category similarity and the distances
among categories. On one hand, distance affects the prevalence of spanning. Cate-
gory combination, however, influences how audiences perceive the distance among
categories: research in cognitive psychology and linguistics show that categories
CATEGORY SPANNING, DISTANCE, AND APPEAL 29
that tend to occur together are perceived similar (e.g., Church and Hanks 1990).
This feedback loop between spanning and category distances would imply the polar-
ization of pairwise distances of categories: categories that are initially similar will
get combined more often thereby increasing their similarity; and dissimilar pairs
will rarely be combined, keeping their similarity low.
How do the processes that we study intersect with the phenomenon of cultural
omnivore (Peterson 1992; Peterson and Kern 1996; Goldberg 2011)? Although
we controlled for individual differences in reviewing histories, we did not study
how individual differences affect reactions to category combination. The literature
on structural differences in omnivorousness would indicate, however, that these
differences are present and would make a difference. A main finding in this literature
states that individuals in high-status occupations more are more likely than others
to be involved in a wide range of cultural activities, for example they attend both the
opera (high-brow) and rock concerts (low-brow) (see, e.g., Peterson and Kern 1996).
This pattern suggests that high-status individuals are more open to trying a variety
of cuisines and their combinations. The negative effects of category spanning,
then, would depend on the audience structure, specifically the composition of each
organization’s audience in terms of omnivorousness.16
Status likely plays a role in these processes as well. Clearly, categories differ in
status (Sharkey 2010), and specific to our empirical domain, among cuisines (for
example, French and sushi are traditionally considered as high-brow). Organiza-
tions could be punished for combining status-incoherent cuisines. Although status
differences would be picked up by our relational approach to category distances,
future research could scrutinize the impact of status differences on category span-
ning. Some relevant research has addressed this topic. For instance, research on
middle-status conformity investigates the antecedents of category spanning: orga-
nizations with either low or high status are more likely than middle-status ones
to deviate from (non-fundamental) category codes (Phillips and Zuckerman 2001).
Rao et al. (2005) demonstrates that high-status chefs are more likely to cross bound-
aries. There is less research, however, on the relationship between status and the
consequences of crossing category boundaries.
Another potential extensions could be to use the proposed approach to study
the diversity of organizations entering a given category (McKendrick and Carroll
2001). The process of legitimation would likely change if de alio entrants come from
a set of industries that are close versus distant from one another. Future research
might look at how, when novel forms first emerge, their distance relative to others
affects legitimation (Ruef 2000).
Future research could also explore alternative measurement approaches. For
example, one could put more emphasis on the analysis of audience structures and
on the taste of audience members regarding category spanning by using the novel
approach of relational class analysis (Goldberg 2011). Future research that uses
review data should also consider the selection problems that arise in such data:
while this paper analyzed how category spanning influences whether the restaurants
get high or low ratings, we did not model the chance that these restaurants are
16The audience member’s omnivorousness is not necessarily indicative of her willingness to accept
category spanning though. It is not evident that an omnivore is open to this combination. For
example, even if she likes Mexican and French cuisines, she might care about the authenticity
of the experience and would only appreciate authentic instances of either genre but not their
combination.
30 BALÁZS KOVÁCS AND MICHAEL T. HANNAN
Appendix
Here we present the estimated effect of maximum contrast (the contrast of the
only applicable category) for restaurants in the two cities that bear only one cate-
gory label.
References
Baron, James N. 2004. “Employing Identities in Organizational Ecology.” Industrial
and Corporate Change 13:3–32.
Batagelj, Vladimir and Matevz Bren. 1995. “Comparing Resemblance Measures.”
Journal of Classification 12:73–90.
Carnabuci, Gianluca, Balázs Kovács, and Filippo Wezel. 2011. “A Breach of Dis-
cipline? Category Contrast and the Impact of Knowledge Recombination.” Pre-
sentated at the Meetings of the Nagymaros Group on Organizational Ecology,
Lugano, Switzerland.
Carroll, Glenn R., Mi Feng, Gaël Le Mens, and David G. McKendrick. 2010. “Orga-
nizational Evolution with Fuzzy Technological Formats: Tape Drive Producers
in the World Market, 1951–1998.” Research in the Sociology of Organizations
31:203–234.
Carroll, Glenn R. and Dennis R. Wheaton. 2009. “The Organizational Construction
of Authenticity: An Examination of Contemporary Food and Dining in the U.S.”
Research in Organizational Behavior 29:255–282.
Church, Kenneth W. and Patrick Hanks. 1990. “Word Association Norms, Mutual
Information, and Lexicography.” Computational Linguistics 16:22–29.
Dobrev, Stanislav D., Tai-Young Kim, and Michael T. Hannan. 2001. “Dynam-
ics of Niche Width and Resource Partitioning.” American Journal of Sociology
106:1299–1337.
Durkheim, Emile and Marcel Mauss. 1969 [1903]. Primitive Classification. London:
Routledge.
Garfinkel, Harold. 1968. Studies in Ethnomethodology. Polity Press.
Goldberg, Amir. 2011. “Mapping Shared Understandings Using Relational Class
Analysis: The Case of the Cultural Omnivore Reconsidered.” American Journal
of Sociology 116:1397–1436.
Griswold, Wendy. 1987. “The Fabrication of Meaning: Literary Interpretation in
the United States, Great Britain, and the West Indies.” American Journal of
Sociology 92:1077–1117.
Hampton, James A. 2007. “Typicality, Graded Membership, and Vagueness.” Cog-
nitive Science 31:355–383.
Hannan, Michael T. 2010. “Partiality of Memberships in Categories and Audiences.”
Annual Review of Sociology 36:159–181.
Hannan, Michael T., László Pólos, and Glenn R. Carroll. 2007. Logics of Orga-
nization Theory: Audiences, Codes, and Ecologies. Princeton, N.J.: Princeton
University Press.
Hsu, Greta. 2006. “Jacks of All Trades and Masters of None: Audiences’ Reac-
tions to Spanning Genres in Feature Film Production.” Administrative Science
Quarterly 51:420–450.
Hsu, Greta, Michael T. Hannan, and Özgeçan Koçak. 2009. “Multiple Category
Memberships in Markets: An Integrative Theory and Two Empirical Tests.”
American Sociological Review 74:150–169.
Hsu, Greta, Michael T. Hannan, and László Pólos. 2011. “Typecasting, Legitima-
tion, and Form Emergence: A Formal Theory.” Sociological Theory 29:97–123.
Jaccard, Paul. 1901. “Étude Comparative de la Distribution Florale dans une
Portion des Alpes et des Jura.” Bulletin de la Société Vaudoise des Sciences
Naturelles 37:547–579.
CATEGORY SPANNING, DISTANCE, AND APPEAL 33
Koçak, Özgeçan, Michael T. Hannan, and Greta Hsu. 2009. “Enthusiasts and the
Structure of Markets.” Presented at the Meeting of the Society for the Study of
Socioeconomics, Paris.
Kovács, Balázs. Forthcoming. “A Monte Carlo Permutation Test for Co-Occurrence
Data.” Quality and Quantity .
Kovács, Balázs and Michael T. Hannan. 2010. “The Consequences of Category Span-
ning Depend on Contrast.” Research in the Sociology of Organizations 31:175–
201.
Kruskal, Joseph B. and Myron Wish. 1978. Multidimensional Scaling. Sage.
Leahey, Erin. 2007. “Not by Productivity Alone: How Visibility and Specialization
Contribute to Academic Earnings.” American Sociological Review 72:533–561.
Leung, Ming D. 2011. “Apples to Oranges: How Category Overlap Faciliates Com-
mensuration in an Online Market for Services.” Presented at the Meetings of the
Academy of Management.
Leung, Ming D. and Amanda Sharkey. 2009. “Out of Sight Out of Mind: The Mere
Labeling Effect of Multiple-Category Memberships in Markets.” Unpublished
Manuscript, Stanford University.
McKendrick, David G. and Glenn R. Carroll. 2001. “On the Genesis of Organiza-
tional Forms: Evidence from the Market for Disk Drive Arrays.” Organization
Science 12:661–683.
Negro, Giacomo, Michael T. Hannan, and Hayagreeva Rao. 2010a. “Categorical
Contrast and Audience Appeal: Niche Width and Critical Success in Winemak-
ing.” Industrial and Corporate Change 19:1397–1425.
Negro, Giacomo, Michael T. Hannan, and Hayagreeva Rao. 2011. “Categorical
Challenge and Response: Modernism and Tradition in Italian Wine Production.”
Organization Science in press.
Negro, Giacomo, Özgeçan Koçak, and Greta Hsu. 2010b. “Research on Categories
in the Sociology of Organizations.” Research in the Sociology of Organizations
31:1–35.
Peterson, Richard A. 1992. “Understanding Audience Segmentation: From Elite
and Mass to Omnivore and Univore.” Poetics 21:243–258.
Peterson, Richard A. and Roger M. Kern. 1996. “Changing Highbrow Taste: From
Snob to Omnivore.” American Sociological Review 61:900–907.
Phillips, Damon J. and Ezra W. Zuckerman. 2001. “Middle-status Conformity: The-
oretical Refinement and Empirical Demonstration in Two Markets.” American
Journal of Sociology 107:379–429.
Pólos, László and Michael T. Hannan. 2002. “Reasoning with Partial Nnowledge.”
Sociological Methodology 32:133–81.
Pólos, László and Michael T. Hannan. 2004. “A Logic for Theories in Flux: A
Model-theoretic Approach.” Logique et Analyse 47:85–121.
Pontikes, Elizabeth G. 2008. Fitting In or Starting New? An Analysis of Invention,
Constraint, and the Emergence of New Categories in the Software Industry. Ph.D.
thesis, Stanford University.
Pontikes, Elizabeth G. 2009. “Leniency and Audience Evaluations.” Presented at
the the meetings of the Nagymaros Group on Organizational Ecology.
Pontikes, Elizabeth G. 2011. “Two Sides of the Same Coin: How Category Leniency
Affects Multiple Audience Evaluations.” Technical report, University of Chicago.
34 BALÁZS KOVÁCS AND MICHAEL T. HANNAN
Stanford University