How Do Fast-Fashion Copycats Affect the Popularity of Premium Brands? Evidence from Social Media

Traditional high-end fashion brands such as Gucci, Prada, and Louis Vuitton have maintained a strong position within the industry, backstopped by increasing demand from developing economies. At the same time, fast-fashion (low-end) brands such as Zara and H&M have been storming the globe with versatile styles and low prices (Coughlan and Kumar 2005; Ergin, Gumus, and Yang 2022). The success of these fast-fashion brands is controversial. Every year, large fast-fashion chains spew close-to-the-runway originals at lightning speed. High-end brands, believing the copycats will steal their customers and hurt their profitability, spare no effort in fighting back with lawsuits (e.g., Hays 2017). Meanwhile, copyright laws have done little to protect high-end designs because lawmakers tend to view the utilitarian nature of clothing and fashion as more important than its artistic and stylistic purposes (e.g., The Fashion Law 2017).

Despite the growing tension around the enforcement of fashion copyrights and the purported negative impact of copycats, the effect of fashion copycats on high-end brands remains empirically unclear. High-end brands may not face any threat if their consumers belong to a different segment, with different valuations of brands and styles, than the consumers of fast-fashion brands. Despite the enormous size of the fashion industry, research on this critical question has been surprisingly limited, due to two factors: First, there is no easy way to quantify or measure fashion. Second, consumer-level data on fashion choices have not been widely available. We overcome these two challenges by studying microlevel consumer choices of fashion goods, and our results provide insights for the fashion industry and copyright policy makers.

To overcome the challenge of quantifying fashion, we build on theories from sociology and psychology that explain how fashion consumption satisfies dual needs: differentiation and conformity (e.g., Brewer 1991; Simmel 1904). We develop two high-level style measures that are relevant for fashion goods: compatibility, which quantifies how well the items in an outfit match each other, and distinctiveness, which measures how visually different one fashion item is from others in the same category. We apply deep learning techniques to extract the two features from a large set of images.

To overcome the data challenge, we obtain data on the fashion choices of individual consumers by analyzing user-generated content from a fashion-focused online community. Users of the community post their own fashion “looks,” tag the brands of the items in each look, and receive feedback in the form of “likes” from peers. ¹ The data comprise 10,262 users and 64,681 fashion posts over four years. We analyze these users’ choices of brands and styles to investigate the impact of low-end fashion copycats on the popularity of high-end brands in this social media context. We acknowledge that social media users cannot represent the whole population, and their social media posts cannot fully capture their purchases. ² Nevertheless, we argue that the choices of fashion-conscious social media users are worth studying, given the global prevalence of social media usage and the influence of social media trends on the entire market. Social media users are a large segment of fashion consumers: an estimated one-third of the world's population uses social media (Statista 2023). Furthermore, many social media users are not only consumers but also trendsetters who influence the choices of other fashion consumers (Strugatz 2016). ³ For companies that use social media as a marketing tool to reach customers and drive demand (Farivar, Wang, and Yuan 2019; Müller, Mattke, and Maier 2018; Peng et al. 2018), it is important to understand what social media users choose to post online. We take advantage of the granular data available for fashion consumers on social media to derive new insights into the ongoing exploration of consumers’ choices between copycat and high-end fashion options. As there has been no systematic analysis in the field, our exploratory approach represents a significant first step that could foster future projects to work toward a cogent and deeper understanding. In the rest of the article, we use the terms “consumers” and “users” interchangeably to refer to the fashion-conscious consumers who post on social media.

Model-free evidence suggests both static and dynamic patterns in the use of fast-fashion and high-end items in social media posts. The static pattern is that many users mix and match fast fashion with high-end items to make a complete ensemble, suggesting that the items are complementary. The dynamic pattern is that users typically start by posting fashion looks composed of fast-fashion brands exclusively; then, the affordable, good styles from fast-fashion brands enable users to build popularity (measured as the number of likes received from peers). As users get more popular, they tend to post more high-end items. Based on these findings, we reason that fast-fashion copycats could have both a negative cannibalization effect and a positive market expansion effect on high-end brands, as summarized in Figure 1. The negative cannibalization effect occurs because copycats are substitutes for high-end brands and therefore reduce the popularity of high-end brands. For the positive market expansion effect, we identify three mechanisms, one static and two dynamic. First, in the “mix-and-match” mechanism (static), fast-fashion brands offer styles that are distinctive but compatible with high-end brands, so consumers can mix and match fast-fashion and high-end items to create an affordable, satisfactory outfit. Many consumers could not afford full outfits of high-end brands, so mixing and matching results in more posting of high-end brands. Second, in the “value enhancement” mechanism (dynamic), the good styles provided by fast-fashion brands enable users to build their popularity over time. Most users derive some utility from being popular, and an increase in popularity leads to an increase in the user's valuation of high-end brands, so users incorporate more high-end items as they become more popular. Third, in the “cost reduction” mechanism (dynamic), users who become very popular may receive sponsorship opportunities, which relax their financial constraints and enable them to purchase more high-end items. The three market expansion mechanisms are novel and distinct from the promotional effect documented in the counterfeit literature (e.g., Qian 2014). ⁴

OPEN IN VIEWER

Figure 1.

Impacts of Fast-Fashion Copycats on High-End Brands.

We develop a dynamic structural model of users’ fashion choices that incorporates all the previously mentioned mechanisms. We account for consumer heterogeneity and estimate the structural model following a hierarchical Bayesian framework, and we conduct counterfactual analyses to examine the probable impacts of copyright policies that outlaw fast-fashion copycats.

Our counterfactual results show that the impact of prohibiting copycats depends on the threshold of similarity. If only extremely similar fast-fashion styles are prohibited, then high-end brands benefit from the policy. The result indicates that the cannibalization effect must dominate the market expansion effect among extremely similar copycats, so high-end brands perform better in their absence. If the ban covers moderately similar fast-fashion items, however, then high-end brands suffer. This somewhat surprising result suggests that the market expansion effect must dominate the cannibalization effect among moderately similar fast-fashion items, so high-end brands perform worse in the absence of moderately similar copycats. We repeat the analysis while excluding professional fashion bloggers, whose behavior may not be representative of ordinary consumers, and the results are qualitatively similar to the main results.

Our results have actionable implications for both managers and policy makers. With an understanding of how fashion consumers value brands and styles on social media, managers can make better decisions about social media strategies, branding, and product design. For managers of high-end brands, our results suggest that copycats may not always be harmful; instead, in some cases, fast-fashion copycats may increase the popularity of high-end brands. These insights can guide product strategies and alleviate some stress about the threat from fast-fashion copycats. Policy makers also should be aware of the potential benefit provided by some fast-fashion copycats. If policy makers amended copyright policies to prohibit copycats, the change may decrease total welfare due to a negative overall impact on the firm side and potentially fewer style options for consumers.

Literature Review

Within the literature on marketing, economics, and information systems, our study relates to four streams of research: copycats, counterfeits and piracy, conspicuous consumption, and computer vision methods in machine learning.

Of the few studies that examine the market response to copycats, the most related is that of Appel, Libai, and Muller (2018), which examines three monetary impacts (acceleration, substitution, and overexposure) of knockoff designs on the original fashion design. Whereas Appel, Libai, and Muller use simulation analysis to examine each of the three predefined mechanisms, we directly model consumer choices and discover and disentangle fashion copycats’ effects from the choice data. Moreover, Appel, Libai, and Muller's predefined effects can be captured in our distinctiveness measure and cannibalization effect, while our proposed market expansion effect (and associated three mechanisms) are novel. Other related studies include those of Van Horen and Pieters (2012), Hofstetter, Nair, and Misra (2018), and Wang, Li, and Singh (2018). Van Horen and Pieters conduct lab experiments and survey studies in the grocery context to demonstrate how copycats can gain or lose from their resemblance to the original brands, but the authors do not investigate how copycats affect the original brands. Hofstetter, Nair, and Misra investigate agent behaviors on an online crowdsourcing platform where participants (“solvers”) who enter into contests later tend to imitate designs that were submitted earlier. The authors find that the online crowdsourcing market seems to reward originality and disincentivize imitation. Wang, Li, and Singh examine the aggregate impact of copycat mobile apps on the demand for the original app. They find that deceptive and low-quality copycat apps may positively affect the demand for the original app, consistent with a promotional effect. Fashion goods are fundamentally different from groceries and mobile apps, however, in that consumers mix and match clothing items, and peer feedback plays an important role in affecting users’ utility. Furthermore, we use a microlevel study to identify novel mechanisms—distinct from the traditional promotional effect—by which copycats can benefit high-end brands.

The literature on counterfeits and piracy (e.g., Ma, Montgomery, and Smith 2016; Oberholzer-Gee and Strumpf 2007; Qian 2014; Smith and Telang 2009) provides evidence that counterfeits (pirated goods) have both cannibalization and promotional effects on the originals. However, copycats and counterfeits are fundamentally different. Counterfeits involve the duplication of both style/content and the trademark (i.e., the brand logo), thereby violating trademark law, whereas copycats do not copy brand logos and typically are legal. While counterfeits can benefit the original brand by improving awareness of it (i.e., promotional effect), copycats do not directly give consumers information about the original brand.

Our research also relates to the literature on conspicuous consumption. Various theoretical works have investigated fashion firms’ strategies for information disclosure (e.g., Yoganarasimhan 2012) and competitive pricing (e.g., Amaldoss and Jain 2005). These studies recognize a unique feature of the fashion industry: consumers have dual needs for differentiation and conformity (Brewer 1991; Simmel 1904). Specifically, when making fashion product choices, consumers trade off between expressing their own individuality and conforming to others’ preferences, as explained by the psychology theory on private and public self-consciousness (Fenigstein, Scheier, and Buss 1975). Bernheim (1994) shows that despite the heterogeneity of underlying preferences, people tend to conform to a single standard of behavior when social status is more important than intrinsic utility.

To extract and quantify fashion styles, we leverage machine learning methods for image analytics. Specifically, we apply the Faster R-CNN (Ren et al. 2015), U2-Net (Qin et al. 2022), ResNet (He et al. 2016), Siamese CNN (Hadsell, Chopra, and LeCun 2006; Vasileva et al. 2018), support vector machine (SVM), and support vector regression (SVR) methods to extract clothing style features (compatibility and distinctiveness) and user appearance features (facial attractiveness, gender, and body mass index [BMI]).

Data

Our research context is the world's largest fashion-specific social media platform, lookbook.nu. ⁵ Users of the community post photos that feature their fashion outfits. The fashion-specific social media platform in our research context is similar to other photo-based social media platforms except that the content is restricted to fashion, and users are encouraged to tag the brand of each item in each posted “look.” ⁶ The brand information is clearly displayed next to the fashion look, visible to all users. Figure 2 shows an example of a fashion post on the website. We collect individual-level historical data from August 2013 to August 2017. The data set contains the entire history of the fashion content generated by a random sample of 10,262 users who registered after August 2013, of whom 9,915 users posted at least once (hereinafter called “active users”). For each fashion post, we collect the image, brand(s), time stamp, and the number of likes. For each user, we collect their age from the brief biography. Since user profiles do not explicitly specify a gender, we predict it from their pictures (see the “User's Appearance” subsection). Table 1 shows the summary statistics for the active users and their fashion posts. For measures such as the number of posts, the standard deviation is large compared with the mean, and the measures have skewed distributions; a few users post very often and receive many likes, while many users post infrequently and receive few likes. This observation is similar to that of most social media platforms.

OPEN IN VIEWER

Figure 2.

Example of a Fashion Post.

OPEN IN VIEWER

Table 1.

Descriptive Statistics.

A: User-Level Summary Statistics
	No. of Posts	No. of Cumulative Likes	No. Following	No. of Fans	Age in Years
M	6.5236	168.8566	33.4380	19.9187	23.4032
SD	17.1523	1,453.5293	183.1134	216.2042	4.7575
Min.	1	0	0	0	4
Median	3	25	6	4	24
Max.	561	91,620	12,781	16,372	99
Obs.	9915	9,915	9,915	9,915	9,915

B: Post-Level Summary Statistics
	M	SD	Min.	Median	Max.	Obs.
No. of likes	25.8841	52.1091	0	10	1,747	64,681

Notes: “No. Following” is the number of other users that the focal user follows, whereas “No. of Fans” is the number of other users who follow the focal user.

Brand Categorization

Following Ha et al. (2017), we group the fashion brands into three categories: fast fashion (Level 1; e.g., Zara, H&M), designer (Level 2; e.g., Kate Spade, Coach, Michael Kors), and mega couture (Level 3; e.g., Prada, Chanel). ⁷ The categorization was determined by domain experts in the fashion industry and is based on brand identity and price ranges.

Feature Extraction from Images

For each fashion look, we examine two key visual aspects—the clothing style and the user's appearance—that together constitute the fashion look and therefore affect the user's utility derived from posting the look. ⁸

Clothing styles

Our choice to focus on compatibility and distinctiveness is grounded in the extant literature that fashion satisfies consumers’ social needs for group cohesion and differentiation (e.g., Brewer 1991; Simmel 1904; Yoganarasimhan 2012). Compatibility is the extent to which the clothing items in an outfit match each other (e.g., a shirt with matching pants). Distinctiveness is the extent to which each item is visually different from others within the same category (e.g., how different the focal shirt is from all other shirts). We abstract away from more granular factors (e.g., color, texture) and capture these high-level styles because the high-level style features lead to clearer managerial implications.

Note that distinctiveness is the item-level tension between conformity and differentiation; a high score indicates that the item is fairly differentiated from (rather than conforming with) other items in the category. Meanwhile, compatibility is the ensemble-level tension between conformity and differentiation; a high score means that the items in the ensemble conform with each other fairly closely. We include both the item-level and ensemble-level measures because the “art” of a fashion look typically involves multiple items put together. For example, an item in a fashion ensemble may not look special within its own category (i.e., low distinctiveness), but it may contribute to an unusual look when combined with another item (e.g., a suit top and lounge pants; low compatibility). Thus, in our study of fashion looks, we need to capture both types of tension between differentiation and conformity. In Figure 3, we list three steps for extracting the two style features from the fashion looks. Figure 4 provides a visual illustration of the style feature extraction process.

Step 1: Detect clothing items (U2NET). We start by detecting and identifying the clothing items. ⁹ For each fashion look, we extract only the clothing items, which are the most visually dominant items in a fashion picture. ¹⁰ We adopt a pretrained network architecture of U2-Net (Qin et al. 2022), which classifies the pixels into three categories: top item, bottom item, and full-body item. As a result, we can obtain each detected item as a separate image.

Step 2: Embed clothes in a style space. We adopt a deep learning approach to learn a feature transformation from images of clothing items into a latent space that represents styles. We follow the approach of Vasileva et al. (2018) to jointly learn the notions of item similarity (the opposite of differentiation) and compatibility in an end-to-end way. We briefly explain the method here to illustrate the rationale and validity of the approach, and readers can refer to the original paper for more details.

The model is built on ResNet (He et al. 2016) and Siamese CNN (Hadsell, Chopra, and LeCun 2006). The model learns a feature transformation, f : I → X, from the image space I (the raw representation of the images) to the style space X (a representation that captures the style features). Moreover, the architecture learns two kinds of spaces: a general embedding space (denoted by X_G) and type-specific embedding spaces (denoted by X_T), explained subsequently. Each fashion item's embedding in the general space is an overall representation of the item's visual traits, and it is used to measure the fashion item's distinctiveness (described in Step 3). The general embedding is trained using a visual-semantic loss between the image and a text description of the corresponding item, ensuring that semantically similar items are projected nearby in the general space. In the Web Appendix, we provide some visualization of the style space to verify the validity of the embeddings of clothing styles.

The type-specific embedding space is defined at the pair level; that is, the model learns an embedding space for each pair of clothing types: top–bottom, top–full, and bottom–full. Then, the appropriate embedding space is used to score the compatibility of two item types in a pair. The embedding of items in both the general and type-specific spaces are 64-dimensional vectors (which we use to develop measures of compatibility and distinctiveness in the next step). Vasileva et al. (2018) trained the model on a data set collected from the Polyvore fashion website, where users create images of “sets” of clothing items that are highly compatible (positive labels); meanwhile, incompatible pairs are generated by randomly sampling items of each type (negative labels). This data set is widely used and accepted in the state-of-the-art machine learning literature for learning compatibility, namely, evaluating how well clothing items match each other (e.g., Han et al. 2017; Tan et al. 2019; Vasileva et al. 2018). Therefore, we follow this stream of literature and adopt the trained model from Vasileva et al., which performs as well as or better than models from other recent papers in scoring compatibility.

Step 3: Develop measures of compatibility and distinctiveness.

Compatibility. Following the literature, we quantify the compatibility of two fashion items using the Euclidean distance between the embeddings of the two items in the corresponding type-specific style space X_T), where compatible items have smaller distances.

Let I m i denote an item m of type i ∈ { top , bottom , full } and x 1 i _ j denote its embedding in the type-specific space with type j. Consider a fashion look H, consisting of three items, I 1 top , I 2 bottom , and I 3 full , and their corresponding embeddings in the type-specific spaces, x 1 top _ bottom , x 1 top _ full , x 2 bottom _ top , x 2 bottom _ full , x 3 full _ bottom , and x 3 full _ top . The L² distance between items I 2 bottom and I 1 top is L ( I 2 bottom , I 1 top ) = ‖ x 1 top _ bottom − x 2 top _ bottom ‖ 2 , that between I 2 bottom and I 3 full is L ( I 2 bottom , I 3 full ) = ‖ x 3 full _ bottom − x 2 full _ bottom ‖ 2 , and that between I 1 top and I 3 full is L ( I 1 top , I 3 full ) = ‖ x 1 full _ top − x 3 full _ top ‖ 2 . We take the average of the three distances as the overall distance among the items in look H; that is, L ( look H ) = L ( I 2 bottom , I 1 top ) + L ( I 2 bottom , I 1 top ) + L ( I 2 bottom , I 1 top ) 3 . As higher distance means lower compatibility, we calculate the final compatibility score for follow-up analysis as max i ≠ j , m ≠ n L ( I m i , I n j ) − L ( look H ) . If an outfit contains only a top item and a bottom item, we can feed the two items into the neural network and calculate the compatibility score based on the embeddings in only the top–bottom space. If an outfit contains only one full-body item, we split the item into an upper half and a lower half and calculate compatibility as if the outfit contained a top item and a bottom item.

Distinctiveness. We measure the distinctiveness of a clothing item by calculating how visually different the item is from all other items in the same category (top, bottom, or full-body). Specifically, we use the embedding in the general style space X_G to represent each clothing item's style. We consider the embeddings of only the items that were posted in the three months prior to the focal post because a style could be distinctive when posted but become common later on, or vice versa. We calculate the distinctiveness of one item as the average of the distances between its style embedding and the embedding of each of the other items of the same category. The distance between two embeddings is widely accepted as a measure of the similarity between two visual items; a larger distance indicates less similarity (and, in our context, higher distinctiveness). Distinctiveness is also conceptually similar to (the opposite of) “prototypicality” in Toubia and Netzer (2017). For a single-piece look (i.e., a full-body item), the distinctiveness score reflects how different the single piece is from all other single pieces. When there are multiple items, we average the distinctiveness scores of all items in the look.

OPEN IN VIEWER

Figure 3.

Steps of Clothing Style Feature Extraction.

OPEN IN VIEWER

Figure 4.

Illustration of Feature Extraction.

User's appearance

When peers judge a fashion outfit, their perceptions may be affected by the user's appearance as well as the clothing styles and brands. We include the user's BMI (to capture the user's body features) and facial attractiveness (an overall evaluation of facial features) to control for the possibility that the user's appearance will confound the effects of the styles and brands on peer likes. Figure 5 provides the steps of the analysis. First, we detect and crop the face from each picture. Second, we generate a low-dimensional embedding of the face using the deep neural network implementation from the Open Face Project (Amos, Ludwiczuk, and Satyanarayanan 2016), which was trained for facial recognition and generates a 128-dimensional intermediate layer that represents a low-dimensional embedding of any face image. Third, we develop measures of appearance features using the face embeddings.

Facial attractiveness. We train a supervised learning model to predict attractiveness. Our training data consist of 3,000 images for which attractiveness is scored on a ten-point scale, with higher values indicating that the face is more attractive. Each image was labeled by five workers on Amazon Mechanical Turk, and we take the average as the image's final rating (see the Web Appendix for more details). We train an SVR model to learn the relationship between the 128-dimensional face embedding and the attractiveness rating. The model achieves a high prediction accuracy with a mean absolute error of .66 on the test sample.

Body features. We train an SVR model on a data set of 4,206 face images that are labeled with the true BMI, made available by Kocabey et al. (2017). The images were collected from Reddit posts linked to the imgur.com service and depict people with BMIs ranging from 11.5 to 50.8. We crop the faces in the training data set to get the 128-dimensional embedding, which we use as input for the SVR model. Our trained model achieves a mean absolute error of 2.45. Prior research shows that, on average, a person's BMI must decrease by 2.49 to make their face more attractive (Re and Rule 2016), and our mean absolute error is below this threshold, so we believe our trained model generates measurements that are accurate enough to control for users’ basic attractiveness and body features. Using this model, we predict BMI for a given fashion look according to the face detected from the photo.

Gender. About half of the users did not list their gender on the website, so we employ SVM to predict gender from the cropped face images based on the 3,000 human-labeled images (see the Web Appendix). This approach yields gender information for 97% of the users, with 92.89% accuracy on the out-of-sample prediction.

OPEN IN VIEWER

Figure 5.

Steps of Extracting the User's Appearance.

In Figure 6, we show examples of images that received low scores (bottom 1%) and high scores (top 1%) for BMI, compatibility, and distinctiveness. Table 2 provides the summary statistics for the features extracted from the fashion looks in our data set.

OPEN IN VIEWER

Figure 6.

Example Photos of the Extracted Features.

OPEN IN VIEWER

Table 2.

Summary Statistics for the Extracted Features.

Variable	M	SD	Min.	Max.	Obs.
Facial attractiveness	4.8201	.6121	.3348	7.1819	64,681
BMI	27.2535	3.6385	11.4966	50.7738	64,681
Compatibility	.5693	.1264	.0250	1.1857	64,681
Distinctiveness	1.4061	.1013	1.0908	2.0794	64,681

Exploratory Data Analysis

We conduct an exploratory data analysis to shed light on the mechanisms underlying consumers’ choices of styles and brands.

Evidence of the static mechanism: “Mix-and-match.”

After extracting the style features, we exaamine the distributions of the extracted style features among the three brand levels. The box plots in Figure 7 and the summary statistics in Table 3 show that the styles of the three brand levels have similar distributions of compatibility and distinctiveness, perhaps as a consequence of fast-fashion copycats that intentionally mimic the style of high-end brands. Therefore, for financially constrained users, fast-fashion brands seem to represent viable substitutes for mega couture brands.

OPEN IN VIEWER

Figure 7.

Box Plots of Distinctiveness and Compatibility for the Three Brand Levels.

OPEN IN VIEWER

Table 3.

Summary Statistics for the Three Brand Levels.

	Level 1: Fast-Fashion Brands (Adopted in 99.79% of Looks)		Level 2: Designer Brands (Adopted in 3.05% of Looks)		Level 3: Mega Couture Brands (Adopted in 2.00% of Looks)
	M	SD	M	SD	M	SD
No. of items	1.57	.85	1.10	.29	1.15	.35
Face	4.82	.61	4.85	.55	4.91	.62
BMI	27.26	3.64	27.19	3.56	26.94	3.61
Age	24.68	5.23	24.90	3.42	26.35	5.24
Gender (percentage female)	81.25%	—	78.08%	—	89.22%	—
Compatibility	.57	.13	.57	.13	.57	.13
Distinctiveness	1.41	.10	1.40	.10	1.40	.10
No. of likes	30.37	59.76	57.91	106.76	55.90	89.50
Total no. of items	1.66	.96	2.67	.92	2.66	1.02
Obs.	64,545		1,973		1,294

Notes: “No. of items” is the number of items of the focal brand level per look that contains the focal brand level. “Total no. of items” is the number of items of all brands per look that contains the focal brand level.

Table 4 shows that at least one fast-fashion item appears in 80.15% of the looks that contain a mega couture brand and 86.40% of the looks that contain a designer brand. Thus, many users combine high-end brands with more affordable options to make a satisfactory ensemble, in other words, a “mix-and-match” strategy. Users may engage in mixing and matching because of financial constraints and because of the availability of fast-fashion items that match well with high-end items.

OPEN IN VIEWER

Table 4.

Evidence of Mixing and Matching in Fashion Looks.

A: For Looks with Mega Couture Brands
	Only Mega Couture	Mega Couture and Fast Fashion	Mega Couture and Designer	All Three
Percentage	4.6%	80.15%	.37%	14.89%
B: For Looks with Designer Brands
	Only Designer	Designer and Fast Fashion	Designer and Mega Couture	All Three
Percentage	3.58%	86.40%	.24%	9.77%

Evidence of dynamic mechanisms involving popularity

Fashion is a social phenomenon: other people's opinions can profoundly affect our decisions about what to wear. The utility a consumer derives from a fashion look depends on the feedback received from others. For example, a consumer enjoys a new dress much more when many people compliment it than when nobody notices it. In the context of social media, we measure the intensity of positive external feedback as the number of likes; we consider a user to be more popular if the user has attracted more cumulative likes (SumLike) from their peers/audience. An alternative measure of popularity is the number of followers, ¹¹ but it was not tracked over time in our data, though we observe a high correlation (.85) between the number of followers and SumLike at the time of data collection. ¹² Throughout this research, we calculate popularity as ln(1 + SumLike_i,t−1).

We conduct five separate regressions to test how a user's popularity relates to their choices of brands and styles in the current period. We treat one week as a period because 99% of the posting occasions across all active users involve at most one post per week, and most likes are received within one week of posting. For the few users who post more than once per week, we average the features across the looks posted within the week. We examine both the full sample and a subsample of the most recent one year. The results, shown in Table 5, indicate that as users become more popular, they post more mega couture brands. There is also a marginal decrease in designer brands in the full sample. Furthermore, popularity seems to be associated with style choices; there is a significant positive association between popularity and compatibility in the one-year sample.

OPEN IN VIEWER

Table 5.

Popularity Correlates with Style and Brand Decisions.

	DV:Brand 1(Fast Fashion)		DV:Brand 2(Designer)		DV:Brand 3(Mega Couture)		DV:Compatibility		DV:Distinctiveness
	Recent Year	Full Sample	Recent Year	Full Sample	Recent Year	Full Sample	Recent Year	Full Sample	Recent Year	Full Sample
Popularity	−.0010	−.0053	−.0001	−.0021*	.007**	.003***	.0049***	.0001	−.0017	.0003
	(.0110)	(.0058)	(.003)	(.0013)	(.003)	(.0011)	(.0018)	(.0007)	(.0016)	(.0006)
Time fixed effects		Yes		Yes		Yes		Yes		Yes
User fixed effects		Yes		Yes		Yes		Yes		Yes
Obs.	3,956	19,280	3,956	19,280	3,956	19,280	3,956	19,280	3,956	19,280
R2	.713	.639	.372	.278	.542	.394	.229	.168	.294	.224

*p < .1.

**p < .05.

***p < .01.

Notes: The dependent variables (DVs) of the first three linear regressions are the number of corresponding brands adopted in each fashion look. The DVs of the last two linear regressions are the extracted compatibility and distinctiveness measurements of each fashion look. An observation corresponds to a post; that is, the analysis does not include the user-week observations where the user did not post. Robust standard errors are clustered at the individual level.

In Figure 8, we observe that users make different brand choices at different popularity levels. In Panel A, the average post by users with low popularity contains more than two fast-fashion items; however, as popularity increases, users on average may post one or fewer fast-fashion items. This indicates that one's popularity may negatively correlate with one’s choices of fast-fashion items. Meanwhile, in Panel B, few users with low popularity post mega couture brands; only those with intermediate to high popularity post one or two mega couture items. The data patterns indicate that users’ fashion choices may differ when popularity changes.

OPEN IN VIEWER

Figure 8.

Brand Choices Across Popularity.

In online discussions and articles, we find anecdotal evidence that social media users are forward-looking about the future utility of gaining popularity. For instance, a wikiHow article titled “How to Get Popular on Instagram” states that one needs to be forward-looking to gain popularity: “Sit down at the start of each week or month to identify upcoming events. … Plan in advance so you have great posts ready when those days arrive” (Whitehair 2021). A Forbes article titled “Want to Be More Popular on Social Media? Try Doing These Simple Things” advises, “We tend to post about our own accomplishments. … That's human nature, but on social media, it's important to fight against that tendency. Instead, choose to promote what other people are doing and saying. … If you promote others, they might do the same for you” (Brandon 2021). In other words, to gain popularity on social media (long-term goal), a user may post something different from the content that brings the user the highest utility in the short term. Examples like these are prevalent.

Given the widespread interest in becoming popular on social media, it is plausible that users are forward-looking about the utility gain from future popularity. We explore this possibility further by regressing the user's current-period choices of brands and styles on their future change in popularity, controlling for current popularity. Table 6 presents the results for the full sample and for the most recent year. We find that users who gain more popularity in the future tend to post looks with more high-end brands, more compatible ensembles, and less distinctive clothing items in the current period. The results provide correlational support for forward-looking behavior. ¹³

OPEN IN VIEWER

Table 6.

Evidence of Forward-Looking Behavior.

	DV: Brand 1 (Fast Fashion)		DV: Brand 2 (Designer)		DV: Brand 3 (Mega Couture)		DV: Compatibility		DV: Distinctiveness
	Recent Year	Full Sample	Recent Year	Full Sample	Recent Year	Full Sample	Recent Year	Full Sample	Recent Year	Full Sample
ΔPopularityt+1,t	−.0641	−.0189	−.0342**	−.0034	.0177	.0127**	.0180**	.0009	−.0140*	−.0018
	(.0480)	(.0246)	(.0162)	(.0047)	(.0143)	(.0051)	(.0078)	(.0024)	(.0075)	(.0019)
Popularityt	−.0465	−.0173	−.0241**	−.0044	.0191*	.0111***	.0175***	.0005	−.0116**	−.0009
	(.0349)	(.0188)	(.0122)	(.0037)	(.0111)	(.0038)	(.0056)	(.0017)	(.0057)	(.0014)
Time fixed effects	Yes		Yes		Yes		Yes		Yes
User fixed effects	Yes		Yes		Yes		Yes		Yes
Obs.	3,956	19,280	3,956	19,280	3,956	19,280	3,956	19,280	3,956	19,280
R2	.713	.639	.373	.278	.545	.395	.230	.168	.295	.224

*p < .1.

**p < .05.

***p < .01.

Notes: The DVs of the first three linear regressions are the number of corresponding brands adopted in each fashion look. The DVs of the last two linear regressions are the extracted compatibility and distinctiveness measurements of each fashion look. Robust standard errors clustered at the individual level are in parentheses.

In sum, the model-free evidence supports our hypothesis that a user's desire for popularity has dynamic effects as the user makes brand and style choices with consideration of both current utility and future utility. Specifically, as a user becomes increasingly popular over time, two mechanisms may come into play. First, an increase in popularity may lead to higher utility from any given fashion post, so users may invest in building popularity in the current period instead of myopically maximizing current utility. This is also supported by the regression results. Second, an increase in popularity may lead to opportunities for sponsorship, which increase future utility by decreasing the user's future cost of posting; ¹⁴ the possibility of future payoffs may motivate a user with lower popularity in the current period to focus on building popularity today. This is supported by the finding that the posting of high-end brands increases with popularity.

Model

According to the exploratory data analysis in the “Exploratory Data Analysis” subsection, a model of the user's decision-making process in our context needs to capture both static and dynamic mechanisms. Moreover, the model should incorporate the timing of events related to decision making, as summarized in Figure 9.

OPEN IN VIEWER

Figure 9.

Timing of Events.

Specifically, in each period (i.e., one week):

1. The user observes the popularity gained with previous posts (if any).

2. The user decides whether to post a fashion look. If so, the user chooses the brand(s) and style. If not, the user goes to the next period.

3. The number of peer likes on the new post (if any) realizes.

The Basic Model

Let P_it be the binary decision of whether to post; it equals 1 if the user posts and 0 otherwise. ¹⁵ If posting, the user decides on the style St_it and brand Br_it. The per-period utility of user i in period t is

u it ( S t it , B r it , P it | S it ; α i 1 , ρ i , η i ) = 1 { P it = 1 } × F it ( S it ; η i ) × R it ( S t it , B r it ; α i 1 , ρ i ) + ϵ it ,

(1)

The total number of posting decisions made by each user is equal to the number of periods (weeks) from the registration date up to August 2017. R it ( S t it , B r it ; α i 1 , ρ i ) captures the intrinsic utility that is attributable to brand Br­_it and style St_it. By “intrinsic preference,” we mean the self-gratification utility derived from the composition of a fashion look for the individual users apart from utility derived from other users’ feedback or view. In other words, it captures how much the user is self-gratified by the fashion look, whether it be visually pleasing or a sense of achievement. F it ( ⋅ ) captures the valuation of a post apart from the intrinsic utility derived from the brand and style choices (e.g., from peer feedback), and ϵ it is the brand and style choice-specific random error, following a Type I extreme value distribution. S_it is the state of user i in period t, as specified in the “State Variables” subsection. We explain R it ( S t it , B r it ; α i 1 , ρ i ) and F it ( S it ; η i ) in the following two subsections.

Intrinsic utility R i t ( S t i t , B r i t ; α i 1 , ρ i )

We assume that users make optimal allocations of their available resources (e.g., time, money) when choosing styles and brands, the input factors for each fashion look. This is akin to firms making production decisions by determining the optimal allocation of resources to the input factors (e.g., capital and labor), subject to the marginal value, cost, and elasticity of substitution of all inputs. Therefore, we adopt the production function specification pioneered by the seminal papers of Dixit and Stiglitz (1977) and Solow (1956) to characterize users’ style and brand decisions. We allow the utility gain from the fashion post to follow a nested constant elasticity of substitution (CES) specification:

R it ( S t it , B r it ; α i 1 , ρ i ) = [ α i 1 B r it ρ i 1 + ( 1 − α i 1 ) S t it ρ i 1 ] 1 ρ i 1 ,

(2)

where the elasticity of substitution between style St_it and brand Br_it is r i 1 = 1 1 − ρ i 1 , and α_i1 is the share parameter that determines the brand's share of the contribution to utility. Consistent with the marketing literature on branding (e.g., Borkovsky et al. 2017; Datta, Ailawadi, and Van Heerde 2017; Keller and Lehmann 2006), we regard the brand value as the incremental value associated with the brand name that is not related to other product characteristics (i.e., style attributes in our context). For example, apart from any style features, people choose high-end brands because they are status symbols and provide hedonic value.

We adopt the CES functional form for two main reasons. First, the CES utility function provides a clear picture of the extent to which users treat styles and brands (the two input factors) as substitutes; the answer has implications for the effect of fast-fashion copycats on the popularity of high-end brands. Second, if brands and styles were linearly additive in the utility function, then a post could generate utility from only brands (without styles) or from only styles (without brands), which is invalid in our context. In other words, a fashion look must contain at least one brand and some style features to be a valid fashion look, and this property is upheld by CES.

For brand choice, users choose brands from three categories: fast fashion (Level 1), designer (Level 2), and mega couture (Level 3), denoted by l ∈ { 1 , 2 , 3 } , respectively. For each brand level, the user decides how many items to include in a post. The brand choice for a fashion post is characterized as follows: ¹⁶

B r it = exp ( ∑ l = 1 3 γ i , l x it , l + ∑ l ≠ k γ i , l k x it , l x it , k ) ,

(3)

where x_it,l denotes the number of clothing items of brand category l in the fashion look posted by user i in period t, and γ_i,l captures the utility gained by including an additional item of brand l. To allow for the possibility that users gain extra utility from mixing multiple brands, we include the interaction terms ∑ l ≠ k γ i , l k x it , l x it , k , where γ_i,lk is the utility gained from matching items from brand categories l and k.

For style choices, users make choices about the distinctiveness of the items in their outfits and the compatibility of their outfits as a whole (as described in the “Clothing Styles” subsection). The style, as a component of the whole fashion look, is incorporated as a subnest of the CES utility function:

S t it = [ α i 2 f it , 1 ρ i 2 + ( 1 − α i 2 ) f it , 2 ρ i 2 ] 1 ρ i 2 ,

(4)

where α_i2 is the share parameter, and f_it,1 and f_it,2 denote the choices about compatibility and distinctiveness, respectively. The elasticity of substitution between the two style features is r i 2 = 1 1 − ρ i 2 . Note that consumers face the same style choices at any brand level, as evidenced by the similar ranges of compatibility and distinctiveness in all three brand levels (see Figure 7). In other words, brand and style choices are not correlated; the styles of high-end brands do not have symbolic value (i.e., they are not more distinctive), possibly due to the prevalence of fast-fashion copycats. Therefore, we let the support of the style choices be the same across the three brand levels, using the empirical bounds as the upper and lower limits of the choice sets. As we elaborate in the counterfactual analysis, when fast-fashion copycats are prohibited, consumers have different fast-fashion and high-end style options.

Popularity affects utility through F i t ( S i t ; η i )

F it ( ⋅ ) in Equation 1 captures the user's valuation of a post apart from the intrinsic utility derived from the brand and style choices. In a CES production function, the multiplier F is the factor productivity, capturing the effect on production driven by factors apart from labor and capital, such that the factors contribute to the product but cannot generate value alone (e.g., technology facilitates production but does not generate value if no production happens). Similarly, a fashion post is a product that has value for the user who posted the look, so F it ( ⋅ ) captures factors that amplify/shrink the value of the fashion look for the user but cannot generate value for the user on its own. A user's popularity plays such a role.

Intuitively, a user's valuation of a post may depend on popularity; as a user becomes more popular, the user may derive a higher utility from posting because an increase in popularity leads to more social interactions through the post (Lee et al. 2015). This is supported by the regression results in Table WA2 in the Web Appendix: users who are more popular are also more likely to post. The value of popularity exists only on top of the posted content; if the content is bad, one's popularity brings less value to the user. We also allow for the possibility that the effect of popularity on utility could be negative for some users (e.g., those who seek escapism via social media posting; see the Web Appendix). We capture the effect of popularity with the following specification for F it ( ⋅ ) :

F it ( S it ; η i ) = η i 0 + η i 1 ln ( 1 + SumLik e i , t − 1 ) ,

(5)

where SumLike_i,t−1 is the cumulative number of likes on all of user i’s previous fashion posts, η_i1 characterizes the value of popularity, and η_i0 is the base value when popularity is zero.

Budget and cost

A user's choices of styles (f_it,1 and f­_it,2) and brands (the number of clothing items x_it,l, l ∈ { 1 , 2 , 3 } ) come with costs, such as the time required to search for a good style and the money required to purchase the clothing item. We specify the following budget constraint:

∑ j = 1 2 t ij f it , j + ∑ l = 1 3 t ix , l x it , l ≤ y i ,

(6)

where y_i is the budget constraint for user i, t_ij characterizes the cost of the effort to obtain style j, and t_ix,l measures the cost of purchasing an item of brand l. Each style choice f_it,j is bounded within the range [0, max i , t f it , j ^ ], where f it , j ^ refers to the empirical counterpart of f_it,j (i.e., user i's choice of style j in period t). Users in our context (and on many other social media platforms) may lower their costs by obtaining sponsorships from fashion companies. In fact, many fashion bloggers are paid millions of dollars every year by fashion companies (Nerney 2023). A user's likelihood of receiving sponsorship opportunities increases with the user’s popularity, which reflects the magnitude of the user’s potential impact on prospective consumers. ¹⁷ We account for the possible financial benefit of popularity by allowing the cost of purchasing a clothing item to vary with SumLike_t−1:

t ix , l = t ~ i , l 1 + δ i × ln ( 1 + SumLik e i , t − 1 ) ,

(7)

where l ∈ { 1 , 2 , 3 } , t ~ i , l is the baseline cost of obtaining an item of brand l, and δ_i measures the decreasing effect of popularity on the cost of brand choices. Though we cannot perfectly observe the cost for each posted item, in order to separately identify the cost parameters ( t ~ i , l ) from the utility parameters (γ_i,l and γ_i,lk), we utilize the price information available on e-commerce sites to proxy t ~ i , l . Specifically, we collected price information from Amazon's publicly available data sets, including products listed from 1996 to 2018 (Ni 2018; Ni et al. 2019). We examine only the clothing product category within the period that overlaps with our analysis, yielding a sample of 1,165,941 unique clothing product titles (98.08% fast fashion, 1.5% designer brands, and .42% mega couture brands). The proportions are comparable to the distribution of brands in our social media posting data (fast fashion in 99.79% of the looks, designer brands in 3.05%, and mega couture brands in 2%). We use the average price of an item in each brand level ($48 for fast fashion, $140 for designer, $222 for mega couture) as the cost (i.e., t ~ i , 1 = 48 , t ~ i , 2 = 140 , t ~ i , 3 = 222 ). ¹⁸

User i's budget constraint, y_i, is the highest cost user i is willing to pay for fashion consumption in each period. We set the budget constraint as the empirical highest-ever cost since the user started posting. Note that, for identification purposes, we treat the financial benefit of popularity as a decreasing cost, which is observationally equivalent to an increasing budget. We assume a fixed budget constraint over weeks as we would not expect the user's allocation of time and pecuniary resources to other regular activities (e.g., working, entertainment) to change much in a short period of time.

It is noteworthy that we study posting behaviors, not purchasing behaviors; we remain agnostic about a user's decision on the time of purchasing the item posted. We acknowledge that users may not post items in the week of purchase; for instance, users may purchase a bunch of clothes at once and post them over the next several weeks. In this case, budget constraint y_i corresponds to the brand and style resources allocated to (not procured in) each period.

Incorporating the static and dynamic mechanisms

To better explain how the model incorporates the static and dynamic mechanisms that drive the effects of copycats on premium brands, we provide a graphical illustration in Figure 10 and a breakdown of the model components in Table 7. Figure 10 shows the isoquant curves for our CES utility model. The x-axis is the brand (Br) choice, and the y-axis is the style (St) choice; brands and styles are somewhere between perfect substitutes and perfect complements.

OPEN IN VIEWER

Figure 10.

Graphical Illustration of the Mechanisms.

OPEN IN VIEWER

Table 7.

Model Components Associated with Each Hypothesized Effect of Fast-Fashion Copycats on High-End Brands.

Impact of Copycats on High-End Brands	Corresponding Model Components	Explanation	Impact on the Posting of High-End Brands
Cannibalization effect	Equation 2: R it ( S t it , B r it ; α i 1 , ρ i ) and budget constraint	If ρi > 0, brands and styles are strong substitutes, so copycats may decrease the posting of high-end brands.	Negative (depending on whether brands and styles are substitutes or complements)
Market expansion (static): mix-and-match mechanism	Equation 5: S t it ( f 1 , f 2 )	Mix-and-match options provide larger choice sets for f1, f2	Positive
Market expansion effect (dynamic): value enhancement mechanism	Equation 3: F it ( S it ; η i )	Popularity increases the utility derived from each fashion post	Positive
Market expansion effect (dynamic): cost reduction mechanism	Equation 7: tix,l	Popularity decreases the cost of adopting each fashion item	Positive

In the dynamic mechanisms, suppose U₁ is a user's original isoquant curve. As the user’s popularity increases over time, the dynamic mechanisms shift the curve upward and to the right. In Figure 10, we show the cost reduction mechanism shifting the curve to U₂ and the value enhancement mechanism further shifting the curve to U₃. The optimal choice changes from A₁ to A₃, which corresponds to an improvement in the user's style and brand choices. In other words, as the user becomes more popular, the user is increasingly likely to post high-end brands and styles that are more compatible and/or distinctive.

In the static mechanism, suppose a user's budget constraint limits the user’s brand choice below Br₁. When the style options also are very limited (e.g., below s₁) because fast-fashion copycats are not available, then A₁ is the optimal choice. But when style options improve (e.g., to s₂) because copycats enable better mix-and-match options across brand levels, then the optimal choice is A₃, resulting in more posts that feature high-end brands.

State Variables

The state variables are S it = { fac e i , bod y i , ag e i , gende r i , SumLik e i , t − 1 , f j ¯ , f j _ , ϵ it } , where SumLike_i,t−1 is time-varying; that is, SumLike_i,t = SumLike_i,t−1 + Like_i,t, and others are static for the same user. Let f j ¯ and f j _ be the upper and lower bounds for each style choice j ∈ { 1 , 2 } . We set the bounds for each style choice as the empirical maximum and minimum in our observation window across all users, and we assume they do not change. The state transition is

Lik e it = g ( SumLik e i , t − 1 , fac e i , bod y i , ag e i , gende r i , P it , f i t , x i t ; b ) ,

(8)

where g ^ ( ⋅ ) is estimated by a negative binomial regression with parameter b, f_it,1 denotes compatibility, f_it,2 denotes distinctiveness, f i t = ( f it , 1 , f it , 2 ) , and x i t = ( x it , 1 , x it , 2 , x it , 3 ) . In our data set, we observe S ^ it = { fac e i , bod y i , ag e i , gende r i , SumLik e i , t − 1 } . The state transition regression results are shown in the “Results” section.

Intertemporal Trade-Off

The exploratory data analysis indicates that users face intertemporal trade-offs when deciding whether to post and, if so, which brands and styles to include. When deciding whether to post, the user may post even if it is worse than not posting in the current period (because the value of posting is lower than the monetary and time costs) because posting can build popularity, enabling the user to gain much more in the future. In other words, the discrete choice of whether to post may be dynamic. Once a user decides to post, the brand and style choices also involve intertemporal trade-offs. A user's brand and style choices affect the number of peer likes that are received on the post, but the best choices for attracting likes (and building popularity) may not optimize the user’s own per-period intrinsic preferences ( R it ( S t it , B r it ; α i 1 , ρ i ) in the utility function). A user with low current popularity might focus more on attracting likes and focus less on personal intrinsic tastes, whereas a popular user might be less constrained by budget concerns and can focus more on expressing intrinsic tastes.

We formulate these insights by allowing each user to decide on an infinite sequence of decision rules, { f i t , x i t , P it } t = 1 ∞ , to maximize the expected discounted utility. Substituting the brand and style choices (Equations 4 and 5) into the per-period utility function (Equation 1), we have

max { f it , x it , P it } t = 0 ∞ E { S it } t = 0 ∞ { ∑ t = 0 ∞ β t U t ( f i t , x i t , P it | S it ; θ i ) | x i 0 , f i 0 , P i 0 , S i 0 } ,

where

U t ( f i t , x i t , P it | S it ; θ i ) = [ η i 0 + η i 1 ln ( 1 + SumLike t − 1 ) ] × { α i [ exp ( ∑ l = 1 3 γ i , l x it , l + ∑ l ≠ k γ i , l k x it , l x it , k ) ] ρ i 1 + ( 1 − α i ) [ β i f 1 it ρ i 2 + ( 1 − β i ) f 2 it ρ i 2 ] ρ i 1 ρ i 2 } 1 ρ i 1 × 1 { P it = 1 } + ϵ it .

Let V(S_it) denote the value function:

V ( S it ) = max { f i τ , x i τ , P i τ } τ = t ∞ E { S i τ } τ = t ∞ { U t ( f i t , x i t , P it | S it ; θ i ) + ∑ τ = t + 1 ∞ β τ − t U τ ( f i τ , x i τ , P i τ | S i τ ; θ i ) | x it , f it , P it , S it } .

The Bellman Equation (Bellman 1957) for the dynamic optimization problem is

V ( S it ) = max f it , x it , P it E S it + 1 { U t ( f i t , x i t , P it | S it ; θ i ) + β V ( S it + 1 ) | x i t , f i t , P it , S it } .

All the decisions (that is, f_it, x_it, P_it) are dynamic in that the current period's decisions affect the next period's state through SumLike_t−1, which further affects the cost and utility of the user's brand and style choices. In other words, the users face intertemporal trade-offs regarding the utility derived from the post today versus its impact on the utility of posting tomorrow.

Heterogeneity

Users may be heterogeneous in their responses to peer feedback. For example, some users care a lot about the number of likes they receive, while others care primarily about expressing themselves. Similarly, some users base their utility heavily on the brands they wear, while others care more about styles. We employ a hierarchical Bayesian framework to account for heterogeneity. All structural parameters θ i ∈ Θ i = { α i 1 , α i 2 , ρ i , η i , t i , γ i , δ i } have a random coefficient specification. ¹⁹ The prior distribution is normal for η_i and γ_i and log-normal for α_i1, α_i2, r_i, t_i, and δ_i, where r i = 1 1 − ρ i . The prior distribution is specified as N ( μ θ , σ θ 2 ) . We use diffuse hyperprior distributions for all parameters.

Identification and Estimation

Identification

In our model, the unknown parameters include those in the state transition process (i.e., b from Equation 8) and the primitives in the utility function ( Θ i ). We briefly explain the identification of the key parameters as summarized in Table 8. First, the transition function parameters can be identified by the variation in the number of likes (Like_i,t) and the corresponding fashion brands, style features, and user demographics (i.e., SumLike_i,t−1, face_i, body_i, age_i, gender_i, f_it, x_it). The discount factor β cannot be separately identified from the utility parameter (the individual fixed effect in our context), as acknowledged in the literature (e.g., Rust 1987), so we fix the weekly discount at .998, consistent with the literature (e.g., Hartmann and Nair 2010). ²⁰

OPEN IN VIEWER

Table 8.

Summary of the Parameters.

Notation	Explanation
αi1	Share parameter for brand and style
αi2	Share parameter for the style features (compatibility and distinctiveness)
ρi	Elasticity of substitution, ρi  = {ρi1, ρi2}, for brand versus style and between style features
ηi	Governs the efficiency or productivity; ηi  = {ηi0, ηi1}, where ηi0 captures the individual fixed effect and ηi1 measures the effect of the cumulative likes
ti	Cost parameters for style choices; ti  = {tix2, tix3}
γi	Utility gained from posting an item of a certain brand level; γi  = {γi1, γi2, γi3, γi12, γi13, γi23}
δi	The cost reduction parameter of the cumulative likes

The elasticity of substitution between brand and style can be identified by the variation in the brand (Br_it) and style choices (St_it) across time. Similarly, the elasticity of substitution between compatibility and distinctiveness can be identified by the variation in the two style choices across periods. The brand and style choices help us back out the underlying cost of each user. We can identify the separate effects of SumLike on the utility (η_i1) and on the cost (δ_i) because the cost reduction mechanism applies only to brand choices, whereas η_i1 affects both brand and style choices. So, the differently evolving patterns of brand and style choices can help us identify η_i1 and δ_i. In summary, the structural parameters to be estimated are Θ i = { α i 1 , α i 2 , ρ i , η i , t i , γ i , δ i } .

Estimation Methods

To estimate the infinite horizon dynamic structural model, we consider several methods, including the conditional choice probability estimation (Aguirregabiria and Mira 2007; Arcidiacono and Miller 2011; Hotz and Miller 1993), the simulated method of moments (McFadden 1989), and the Bayesian estimation method or IJC method (Imai, Jain, and Ching 2009). We consider two primary criteria: computational efficiency and the ability to capture rich heterogeneity in responses across individuals. The simulated method of moments matches data moments with simulated moments; it is computationally inefficient because it requires a full solution to the dynamic optimization problem, and it cannot capture rich heterogeneity. The conditional choice probability method has better computational efficiency but also is limited in its ability to capture heterogeneity. By contrast, the IJC method serves our goal of capturing rich individual responses, and the method requires evaluating the value function only once in each iteration, making it computationally efficient. The IJC method is designed for discrete choice models, but our model includes a continuous choice (the style decision) as well as discrete choices (the brand and post decisions). We follow the approach of Ching and Osborne (2020) to discretize the choice space of style decisions. We refer the reader to the Web Appendix for details of our estimation procedure.

Results

Model Comparison

We use a hold-out sample of 12 weeks (i.e., one season) to compare our proposed model (Column 5 in Table 9) with four benchmark models. We estimate each model, and Table 9 lists the hit rates for the posting decision and choice of each brand level.

OPEN IN VIEWER

Table 9.

The Proposed Model vs. Alternative Models.

	(1) No Style	(2) No Utility from Popularity	(3) Linear Utility Function	(4) No Forward-Looking	(5) Proposed
Hit rate: post	58.25%	52.47%	70.19%	72.75%	76.21%
Hit rate: fast-fashion choice	54.49%	55.42%	66.72%	71.11%	72.69%
Hit rate: designer choice	85.27%	91.64%	95.44%	93.56%	96.85%
Hit rate: mega couture choice	87.66%	97.67%	97.39%	96.51%	98.94%

The first benchmark model (Column 1) does not involve style decisions, as if researchers could not quantify styles. The hit rates are much lower than the models that incorporate styles, demonstrating the importance of incorporating style measures when studying fashion choices. In the second benchmark model (Column 2), users do not derive utility directly from popularity (the term η i 1 ln ( 1 + SumLik e i , t − 1 ) in the F it ( S it ; η i ) component of the utility function of Equation 1). When there is no direct utility gain from popularity, the only effect of popularity is a decrease in the cost of posting. The third benchmark model (Column 3) uses a simple linear per-period utility instead of the CES framework. Specifically, Equation 1 becomes u it ( S t it , B r it , P it | S it ; θ i ) = 1 { P it = 1 } ⋅ G it ( S t it , B r it | S it ; θ i ) + ϵ it , where G it ( . ) is a linear additive function of {popularity_i,t−1, f_it,1, f_it,2, x_it,1, x_it,2, x_it,3, x_it,1x_it,2, x_it,1x_it,3, x_it,2x_it,3}. In the fourth benchmark model (Column 4), we treat users as myopic instead of forward-looking; that is, they do not consider the impact of today's choice on tomorrow's state (i.e., popularity).

Our proposed model (Column 5) outperforms all benchmarks on all four hit rates, demonstrating the gain in model fit from including style considerations, allowing users to gain utility directly from popularity, using the CES functional form, and treating users as forward-looking. In the Web Appendix, we conduct a simulation study with 200 users and 30 periods to confirm our estimation is implemented correctly so that the true parameters can be accurately estimated.

Parameter Estimates

Table 10 shows the transition process estimated with a negative binomial regression with the number of likes as the dependent variable. We can see that popularity does have a substantial positive effect on the peer likes for a new fashion post. We can also see that having more items of any brand helps attract likes, and the positive effect magnifies at a higher brand level. The average audience seems to prefer higher facial attractiveness, lower BMI, younger users, and female users. In addition, the average audience seems to prefer more compatible ensembles and less distinctive clothing items. The negative coefficient of the interaction between fast-fashion brands and mega couture brands shows that people do not respond favorably to outfits that mix and match items from the two brand levels. However, the positive effect of including either a fast-fashion brand or a mega couture brand dominates the negative interaction effect.

OPEN IN VIEWER

Table 10.

Negative Binomial Regression for State Transition with Number of Likes as the DV.

Variables	Estimates	SE
Popularity	.2758***	.0022
No. of fast fashion (Level 1)	.2172***	.0086
No. of designer (Level 2)	.3730***	.1250
No. of mega couture (Level 3)	.5496***	.1279
Face	.1727***	.0135
BMI	−.0174***	.0021
Age	−.0096***	.0015
Gender (male)	−.0999***	.0187
Compatibility	.1509**	.0667
Distinctiveness	−.4585***	.0827
Level 1 × Level 2	−.0416	.0492
Level 1 × Level 3	−.1753**	.0529
Level 2 × Level 3	−.1891	.1284
Constant	2.3102***	.1724
Dispersion	.9532***	.0093
Observationsa	20,656

*p < .1.

**p < .05.

***p < .01.

^a

The number of observations in Table 5 or Table 6 is smaller because controlling for fixed effects loses the observations of users who posted only once.

Notes: Observations are collapsed at the weekly level.

Table 11 reports the estimation results of the structural parameters. We ran 10,000 iterations and report the results using the last 1,000 iterations after burn-in; the estimation converged after around 4,000 iterations. ²¹ Figure 11 shows histograms of three structural parameters of individual users.

OPEN IN VIEWER

Figure 11.

Distribution of Structural Parameters Across Users.

OPEN IN VIEWER

Table 11.

Structural Model Estimation Results.

Parameter	Interpretation	M (μθ)	95% Credible Interval for μθ	SD (σθ)	95% Credible Interval for σθ
η0i	Fixed effect	.665	(.599, .724)	.578	(.571, .589)
η1i	Popularity effect on utility	.667	(.652, .697)	.598	(.592, .606)
ρ1i	Substitution parameter	.535	(.531, .540)	.266	(.240, .309)
ρ2i		−.248	(−.265, −.224)	.843	(.772, 1.011)
γ3i	Gain from brand choices	6.053	(5.156, 6.824)	8.913	(6.730, 11.042)
γ2i		2.383	(2.350, 2.417)	1.604	(1.498, 1.707)
γ1i		1.710	(1.686, 1.734)	1.066	(1.006, 1.131)
αi1	Share parameter	.747	(.743, .751)	.099	(.098, .101)
αi2		.518	(.515, .522)	.107	(.106, .109)
t1i	Cost of style choices	1.981	(1.918, 2.045)	1.663	(1.500, 1.818)
t2i		1.580	(1.550, 1.611)	1.135	(1.065, 1.249)
δi	Cost reduction effect	1.448	(1.419, 1.479)	.981	(.905, 1.068)
γ12,i	Gain from interaction of brand levels	.162	(.144, .178)	.525	(.518, .533)
γ13,i		.184	(.162, .200)	.501	(.494, .508)
γ23,i		.066	(.057, .079)	.491	(.481, .501)

The mean substitution parameter estimate is .535, corresponding to an elasticity of substitution of r_1i = 1/(1 − ρ_1i) = 2.15, meaning that a 1% increase in the marginal rate of substitution between brand and style would require the average user to increase the ratio of Br_it and St_it by 2.15% to achieve the same utility. In other words, for most users, styles and brands are good substitutes, so it is relatively easy to derive the utility associated with a high-end brand by using a good style from a fast-fashion brand instead. ²² As shown in Figure 11, Panel A, ρ₁ > 0 for 92.52% of users, so they treat styles and brands as substitutes. For these consumers, high-end brands may need to worry about the cannibalization effect of copycats. For the other 7.48%, styles and brands are complements, so they are unlikely to be attracted to good styles at the expense of good brands, and the copycat problem is less worrisome.

The estimates of η_1i indicate that most users (70.48%, those with η₁ > 0) derive some positive utility from popularity, but substantial heterogeneity exists across individuals as shown in Figure 11, Panel B (standard deviation of .598). According to Lee et al. (2015), people post on social media for a variety of reasons including self-expression, social interaction, archiving, and escapism. We find that 29.52% of users derive negative utility from popularity. We speculate that this group of users may post for the purpose of archiving or escapism, which requires little or no attention from others, and may even feel uncomfortable or distressed when more people are watching and paying attention to them.

The estimation result γ_3i > γ_2i > γ_1i indicates that, on average, high-end brands both attract the most likes (see Table 10) and provide the most intrinsic utility to users. The estimate of the share parameter α_i2 indicates that both compatibility and distinctiveness contribute positively to the average user's intrinsic utility of a fashion post. Together, the results in Table 10 and Table 11 suggest that higher compatibility both attracts more likes and increases the user's intrinsic valuation of the fashion look, whereas higher distinctiveness does not attract more likes but increases the intrinsic valuation. Finally, the estimation results for δ_i, shown in Figure 11, Panel C, provide evidence that popularity has a significant cost reduction effect, consistent with the phenomenon of sponsorships from fashion companies.

Counterfactual Studies: If Fast Fashion Is Prohibited from Copying Mega Couture Styles

We explore what would happen to consumers’ choices if copyright laws provided more protection for fashion designs, that is, if fast-fashion companies were prohibited from producing styles similar to the original styles of mega couture brands. As a result, consumers may have different style options across brand levels. Such a policy would involve a threshold of similarity; fast-fashion items whose similarity to existing high-end styles was above the threshold would be banned, while those below the threshold could continue to sell.

To gauge the similarity between original designs and copycats, we collect an external data set from the website “Fashion Copycats” ²³ and the Instagram account “Diet Prada,” ²⁴ containing a total of 1,380 original–copycat pairs with corresponding images. For each pair, we feed the images into the embedding extraction pipeline (see the “Clothing Styles” subsection), resulting in two embeddings. The Euclidean distance between the two embeddings reflects the visual similarity between the two items (with smaller distances indicating higher similarity). We obtain 1,380 measures of the similarity between an original item and a copycat. The minimum, mean, and maximum are .03, .36, and .85, respectively. We use the measures to set thresholds of similarity for two counterfactual analyses: Counterfactual policy A bans only extremely similar designs, so Threshold A = .03, the smallest distance among all the pairs. Counterfactual policy B bans moderately similar designs, so Threshold B = .36, the mean distance.

To conduct the analysis, we first calculate the distance between each mega couture item and all fast-fashion items, using the style embeddings. Then, for each policy, we take the last 100 Markov chain Monte Carlo draws of the individual-level structural parameter estimates and conduct the simulation for each draw under the counterfactual policy. In Table 12, we report both the average change in consumer choices and the 95% credible intervals for the 100 counterfactual results.

OPEN IN VIEWER

Table 12.

Results of Counterfactual Studies.

	(1)	(2)	(3)
	Prohibit Extremely Similar Copycats (Policy A)	Prohibit Moderately Similar Copycats (Policy B)	No Professional Users Under Policy B
Δ No. of posts	−4.32%	−11.29%	−10.04%
	(−.0593, −.0224)	(−.1263, −.1000)	(−.1191, −.0842)
Δ No. of fast fashion	−5.53%	−15.19%	−13.77%
	(−.0798, −.0263)	(−.1723, −.1366)	(−.1671, −.1204)
Δ No. of designer	−5.35%	6.12%	9.34%
	(−.2662, .2300)	(−.0957, .1868)	(−.1873, .3232)
Δ No. of mega couture	6.75%	−17.09%	−15.99%
	(.0137, .1532)	(−.3846, −.0109)	(−.4883, .1445)
Δ No. of mix-and-match	−6.30%	−28.14%	−27.90%
	(−.2514, .1755)	(−.4345, −.1039)	(−.5197, −.0295)

Notes: 95% credible intervals are in parentheses. “No. of mix-and-match” is the number of looks that feature a fast-fashion item plus a designer or mega couture item.

Counterfactual Policy A: Prohibit Only Extremely Similar Designs

In the style choice space, we replace the prohibited styles (similarity < Threshold A, .03) with the average of the remaining fast-fashion styles, and we calculate the highest distinctiveness for the remaining fast-fashion styles and the compatibility between fast-fashion and high-end items. ²⁵ We find that the compatibility between the remaining fast-fashion styles and mega couture styles is similar to that before the policy change, which implies that consumers can still create compatible ensembles by mixing and matching items from different brand levels. However, the distinctiveness of the remaining fast-fashion items is lower than the distinctiveness of the mega couture items (while the styles are comparable across brand levels in the absence of a policy; see Figure 7).

In Column 1 of Table 12, we find that the policy change caused a 4.32% decrease in posting and a 5.53% decrease in the number of fast-fashion items posted, likely because users who like distinctive styles found fewer good fast-fashion options after the exclusion of extremely similar copycats. At the same time, the number of mega couture items posted increased by 6.75%, indicating that some users would switch from fast-fashion brands to mega couture brands when extremely similar copycats are no longer available. The results indicate that among extremely similar fast-fashion copycats, the cannibalization effect dominates the market expansion effect, so mega couture brands would be better off in a counterfactual world in which extremely similar copycats are prohibited.

Counterfactual Policy B: Prohibit Moderately Similar Designs

As before, we replace the prohibited styles (similarity < Threshold B, .36) with the average of the other fast-fashion styles in the style space. ²⁶ Now, the maximum compatibility between the remaining fast-fashion styles and mega couture styles is only 82% of that before the policy change. In addition, the highest distinctiveness of the remaining fast-fashion items is only 75% of that of the mega couture items. In Column 2 of Table 12, we find that the posting probability drops by 11.29%, and there are decreases in both the number of fast-fashion items posted (−15.19%) and the number of mega couture items posted (−17.09%). Why are the high-end brands worse off when moderately similar copycats are prohibited? We compare consumers’ choices in the counterfactual and real worlds, and we find three mechanisms, detailed subsequently, that contribute to the decreased posting of mega couture brands.

Mix-and-match mechanism

Many users combine clothing items from multiple brand levels to make a complete outfit: in Table 4, about 80% (86%) of fashion looks that featured a mega couture (designer) brand also included one or more fast-fashion items, further demonstrating the economic significance of the mix-and-match mechanism in the fashion market. Consumers may mix and match if they cannot afford to create ensembles that feature high-end brands exclusively. In the counterfactual world, where style choices are restricted for fast-fashion brands, consumers are subject to higher financial pressure to buy high-end brands to get a satisfactory ensemble. Consumers who previously coped with budget constraints by mixing and matching will be dissatisfied under the counterfactual policy: the remaining fast-fashion items are less distinctive and less compatible with high-end items (at least, under counterfactual policy B), while high-end items remain unaffordable. As a result, some consumers may value the purchase/post option lower than the outside option (i.e., buying nothing), so they end up buying and posting fewer clothes from both brand levels. The mechanism is supported by the decrease in the number of mix-and-match posts (the last row of Table 12).

We illustrate the mechanism with the real-life example of a copycat of Céline from Zara, shown in Figure 12. Suppose a consumer has a $3,000 budget for fashion spending and wants to get a set of clothes of the Céline style (Figure 12, Panel A). Unfortunately, the full set costs $4,750. However, Zara offers a highly similar copycat (Figure 12, Panel B), so the consumer can make a satisfactory and well-matched ensemble by pairing the Zara copycat skirt with the Céline shirt (outlined in rectangles). In the counterfactual world, the consumer would not find a frugal alternative to the Céline skirt and thus may end up buying nothing from Céline.

OPEN IN VIEWER

Figure 12.

Example of an Original Céline Outfit and the Zara Copycat.

Value enhancement mechanism

Some users may not value high-end brands much at the beginning of their fashion blogging experience, but they can accumulate popularity over time by wearing attractive styles from fast fashion. As they become more popular, the increase in the attention received on each fashion post increases the utility derived from the posted fashion look (at least, among the majority of users for who enjoy being popular). Therefore, they will be more likely to post fashion looks in general, including looks that feature high-end brands.

Cost reduction mechanism

Many users cannot afford mega couture brands in the beginning, so they rely on similar styles from fast-fashion brands to build popularity. If they become popular enough to receive sponsorships, then they can afford more high-end brands.

The three mechanisms comprise the market expansion effect of fast-fashion copycats on high-end brands. Our findings demonstrate that among the moderately similar copycats (targeted by policy B), the market expansion effect dominates the cannibalization effect, so mega couture brands would be worse off in a counterfactual world in which moderately similar copycats are prohibited.

We conduct additional analysis to separate the effects of static mechanism (mix-and-match) and the dynamic mechanisms (value enhancement and cost reduction) on mega couture brands. Specifically, under policy B, we run several more counterfactual studies. First, we turn off both dynamic mechanisms (by setting η₁ and δ to zero) so that only the static mechanism plays a role. Second, we turn off only the value enhancement mechanism (by setting η₁ to zero), leaving only the remaining mix-and-match and cost reduction effects to operate. We simulate consumer choices in each case using the last hundred Markov chain Monte Carlo parameter estimates. The results are respectively shown in the first and second row of Table 13, and the last row corresponds to the case where all the three mechanisms exist. Columns 1 and 2 of Table 13 display the average number of mega couture items adopted per user, under no policy change (Column 1) and counterfactual policy B (Column 2). The last column shows the percentage change due to the counterfactual policy (i.e., Column 2 − Column 1 Column 1 ).

OPEN IN VIEWER

Table 13.

Breakdown of Policy B's Effect on Mega Couture Brands.

	(1)	(2)	(3)
	No Policy Change	Counterfactual Choices	Change
Only mix-and-match mechanism	.0192	.0179	−6.32%
	(.0040, .0351)	(.0021, .0285)	(−.1223, −.0015)
Mix-and-match + cost reduction	.0553	.0409	−11.43%
	(.0470, .0684)	(.0322, .0531)	(−.2344, −.0495)
Mix-and-match + cost reduction + value enhancement	.0627	.052	−17.09%
	—	(.0385, .0620)	(−.3846, −.0109)

Notes: 95% credible intervals are in parentheses. The percentage changes are calculated for each simulation, and Column 3 reports the mean and credible intervals across the 100 draws rather than calculating the change from the means in the first two columns. The value in Column 1 of Row Centre to column head 3 corresponds to the real choices in the data, so there is no credible interval.

Of the overall decrease of 17.09% in the posting of mega couture brands, the mix-and-match mechanism contributes an average of −6.32%; the mix-and-match plus cost reduction mechanisms account for −11.43%, and the remainder is due to the value enhancement mechanism. The results indicate that the three mechanisms make comparable contributions to the market expansion effect of fast-fashion copycats on mega couture brands. In addition, the results indicate that both static and dynamic considerations affect users’ brand choices in their posts on social media.

Discussion

Firm Strategy

Thus far, because we lack firm-side data, we have focused on the partial-equilibrium case in which firms do not adjust prices, and fast-fashion firms do not strategically react to the counterfactual policy by changing their styles. Next, we speculate about the most likely reactions of firms to the counterfactual policy and the consequent impact on high-end brands.

Under the counterfactual policy, fast-fashion firms would have to design their own styles, and the styles probably would be less distinctive and less compatible with styles from high-end brands than before the policy change. Therefore, fast-fashion firms might decrease prices to attract consumers, creating an even starker difference in prices with the high-end brands and potentially hurting the business of the latter. Alternatively, fast-fashion brands may invest more in developing styles that are both distinctive and compatible with high-end brands, thereby deviating from the original premise of fast fashion and perhaps evolving into designer brands.

Meanwhile, high-end brands may adapt by starting an affordable product line that provides styles that resemble the high-end product line but cost less (i.e., umbrella branding). This strategy may erode the parent brand's value, however, and the net effect is not clear without additional information and further empirical study. Alternatively, high-end brands might simply lower their prices to attract more users. In this case, a mega couture brand's profit would be lower than before because it would have to charge a lower price to achieve the same demand. Moreover, under stringent copyright protection, the high-end brands may have less incentive to innovate new styles as they no longer have a constant need to differentiate their styles from those of fast-fashion brands. This may lead to less product variety in the market and lower consumer welfare. These possible firm-side reactions are also consistent with Qian's (2008) findings that under less government protection against counterfeits, brands differentiate their products through more innovation, vertical integration, and high-price signals.

Overall, we can see that if fast-fashion copycats were prohibited, neither fast-fashion brands nor mega couture brands could easily combat the loss.

Generalizability

As we acknowledged up front, the average consumer may differ from the average user on our fashion-specific social media platform. Strictly speaking, the findings in this research apply only to social media users; additional data and studies are necessary to determine the extent to which our context generalizes to the whole market and population. However, we can increase the generalizability of our results from the current data set by focusing on those who are not professional fashion bloggers, as professional bloggers’ preferences may differ from the average consumer's. We exclude the professional bloggers (17.5% of the users in our sample) and check the robustness of the results in the analysis of counterfactual policy B (which prohibits moderately similar designs). The results are reported in Column 3 of Table 12; the effects of the policy on posting, the number of fast-fashion items posted, and the number of mega couture items posted are negative but with smaller magnitudes than in Column 2 (full sample). We reason that professional bloggers derive more utility from popularity, and the lack of fast-fashion copycats makes it harder for them to build popularity. Therefore, the value enhancement and cost reduction mechanisms are weaker for users who are not professional bloggers.

Note that among the three mechanisms in the market expansion effect, the cost reduction mechanism applies only to social media users who get financial benefits from posting fashion looks, whereas the mix-and-match and value enhancement mechanisms apply to offline consumers as well as social media users. It seems reasonable to imagine that many offline consumers also like to mix and match items from different brand levels, and although they cannot derive utility from social media likes, they receive feedback from peers in the real world, and they may wish to build their popularity or attract attention just as social media users seek likes.

Managerial and Policy Implications

Our results have implications for both managers and policy makers. Managerially, our results show that, in the context of social media, the impact of prohibiting copycats depends on the threshold of similarity. If the ban covers moderately similar fast-fashion items, then high-end brands may suffer. This somewhat surprising result suggests that among moderately similar fast-fashion styles, the market expansion effect dominates the cannibalization effect, thereby benefiting high-end brands by increasing the number of high-end items featured in social media posts. This insight may alleviate some of the stress of high-end brand managers about the threat posed by fast-fashion copycats, though high-end brands could benefit from advocating for a policy to prohibit extremely similar designs (while letting the moderately similar copycats exist for the benefit of fast-fashion brands and high-end brands alike). In addition, based on our investigation into the mechanisms, high-end brands could enable consumers to mix and match items from different price tiers by launching affordable versions of their own high-end styles, capturing the unserved demand themselves rather than leaving it to fast-fashion brands. For example, Armani launched Emporio Armani as a diffusion brand that is much cheaper than the other Armani lines and targets less wealthy customers. Similarly, See by Chloé is an affordable line launched by the high-end brand Chloé.

For policy makers, we provide insights into the potential consequences of alternative copyright policies for fashion designs. Although we do not precisely calculate total welfare, we may draw some qualitative implications from the counterfactual results. Our findings suggest that a harsh ban like policy B would decrease demand for all brand levels, which would negatively affect all the firm-side players. Moreover, high-end brands might be less innovative as they would face less constant pressure to differentiate their styles from those of fast-fashion brands (e.g., Qian, Gong, and Chen 2015), and this may lead to lower consumer welfare (as consumers would enjoy less product variety). That said, under the prohibition of copying fashion designs, fast-fashion brands may develop their own styles that differ from those of high-end brands. Then, consumers would not necessarily see a decrease in product variety, but they might need to pay higher prices, given that fast-fashion brands would incur higher costs from the effort of designing unique styles. Put together, our results suggest that total welfare may be lower under a harsh prohibition of fast-fashion copycats. In contrast, if the copycat prohibition is mild (i.e., banning only extremely similar designs as in policy A), then high-end brands would benefit while fast-fashion firms may be worse off, at least as measured by the number of items featured in social media posts. The welfare implications are less clear in this scenario.

Contributions and Limitations

Fast-fashion brands often are accused of copying the designs of high-end fashion brands, thereby reducing the distinctiveness of the high-end brands and eroding their brand equity. However, no prior systematic empirical study has attempted to investigate whether and how fast-fashion copycats affect high-end brand equity, likely because large-scale data on individual consumers’ choices of brands and styles are unavailable, and it is difficult to quantify fashion styles in a scalable way. In this research, we obtain user-generated data from a large fashion-specific social media platform, apply state-of-the-art deep learning methods to conduct image analytics, and use structural modeling to investigate users’ brand and style choices. Social media users influence fashion trends and other consumer choices, so understanding their decision processes sheds light on the choices of fashion consumers at large.

Our results show that styles and brands are quite substitutable for most social media users, who often can derive the same utility by substituting a high-end brand with a good style from a fast-fashion brand. This is the basis of the concern of high-end brands: they might be losing customers to fast-fashion brands that provide comparable styles. We also find that most users derive positive utility from popularity (operationalized as receiving peer “likes”), but they vary widely in how much they value popularity. Additionally, higher popularity can increase the utility of posting a fashion look and can reduce the associated cost (as popularity may bring sponsorship opportunities).

In counterfactual analyses, we test the effects of two copyright policies that ban copycat styles. Our results may inform the product strategies of high-end fashion companies and may relieve stress about the threat from fast-fashion copycats. We find that the outcomes depend on the threshold of similarity used in the policy. High-end brands might benefit from advocating for a copyright policy that bans only extremely similar designs, which seems to be effective at alleviating the cannibalization effect. Meanwhile, a harsher copyright policy that prohibits even moderately similar designs might hurt high-end brands by suppressing the market expansion effect, which dominates the cannibalization effect among moderately similar copycats. We find three mechanisms that contribute to the market expansion effect. First, fast-fashion brands have more restricted style offerings under the harsh policy, so it is more difficult and expensive to create satisfactory ensembles by mixing and matching items from fast-fashion brands and high-end brands. The cost will be prohibitively expensive for some, leading to a decrease in the posting of high-end items. Second, the lack of good styles from fast-fashion brands makes it harder for users to build their popularity, which limits the utility they can expect to derive from their posts (including those featuring high-end brands). Third, with fewer options available for building popularity, users have a lower chance of obtaining sponsorships, which would have made high-end brands more affordable. High-end brands might take advantage of these mechanisms, even in the absence of a copyright policy, by offering more frugal product lines that are compatible with their high-end lines, thereby meeting the needs of consumers who otherwise would turn to fast fashion.

Our research has limitations. First, we do not have sales information, so our findings reflect only the posting behaviors of social media users, though posts may be correlated with demand. Additional data and studies are necessary to determine whether the cannibalization effect or the market expansion effect dominates sales. That said, the novel mechanisms we propose in this research can guide future research explorations above and beyond the context of social media. Second, we treat compatibility and distinctiveness as objective traits, but consumers probably are heterogeneous in their evaluations of styles. Readers should interpret our results based on the objective style measures we use. Future research can conduct more targeted analyses with granular information about how individual consumers evaluate styles; for instance, researchers could collect evaluations in lab experiments or surveys and then match people's evaluations of styles based on these observables. Lastly, we do not include accessories because computer vision techniques are not yet precise enough for small items (e.g., jewelry, hat, glasses, shoes), so our findings speak only to clothing and do not necessarily generalize to other product categories. That said, once the detection algorithm becomes accurate enough, it would be straightforward to incorporate accessories into our image processing framework, and the same structural model could be used for analyses.

Supplemental Material

sj-pdf-1-mrj-10.1177_00222437231164403 - Supplemental material for How Do Fast-Fashion Copycats Affect the Popularity of Premium Brands? Evidence from Social Media

Supplemental material, sj-pdf-1-mrj-10.1177_00222437231164403 for How Do Fast-Fashion Copycats Affect the Popularity of Premium Brands? Evidence from Social Media by Zijun (June) Shi, Xiao Liu, Dokyun Lee and Kannan Srinivasan in Journal of Marketing Research