Social influence during public crises: Weekly dynamics and adaptive patterns of conformity to the collective following the COVID-19 outbreak

Significance

Conformity to the collective significantly increased after the first wave of the nationwide COVID-19 crisis erupted in China compared with prior weeks. These increases were determined through nonparticipatory observation of real-world online shopping transactions, which provided evidence with high ecological validity and temporal resolution for prepandemic theoretical predictions. This study, emphasizing dynamic perspectives, demonstrated the weekly synchronization between crisis development and conformity behavior and the association between this adaptive shift and better antipandemic outcomes during the early stages. It further explored the associations between the magnitude of the increase and regional indicators like rice farming, collectivism, cultural tightness, and behavioral constraints. These findings will inform broader discussions on social changes as potential results of the pandemic, e.g., more conservative attitudes.

Abstract

Regional collectivism has been observed to contribute to better coping with public crises such as the COVID-19 pandemic. This study poses a reverse question: Does the eruption of public crises increase people’s conformity to the collective? To answer this question, we analyzed real-world transactions on Taobao (the largest e-commerce platform in China), each with a purchase decision and a list of candidates considered before purchasing. Conformity to the collective was measured using two indicators: whether the decision-maker opted for the A) most-sold and B) best-rated options within the candidate option set. The results reveal that both conformity variables were significantly higher in the 10 wk subsequent to January 19, 2020 (when the nationwide COVID-19 crisis erupted in China), than in the 8 wk prior. These shifts were common across subpopulations, regions, and product categories and remained significant after strictly matching across weeks and after using a within-person, longitudinal sample. These shifts were more confidently attributed to the pandemic by further conducting difference-in-differences analyses to compare pandemic-affected regions with their unaffected, comparable counterparts using data from six subsequent regional waves in China. Furthermore, regions with larger increases in conformity during the early stage of the pandemic achieved better antipandemic outcomes. These findings provide real-world evidence for previous theories on behavioral immune systems, terror management, and compensatory control. Additionally, cross-regional comparisons of effect sizes offer exploratory insights into cultural psychology. In summary, these findings capture how human societies dynamically adjust their values to better adapt to unanticipated survival challenges.

Previous studies have demonstrated that regional collectivism—a cultural value associated with stronger norms for external behaviors (1)—contributes to better coping with public crises such as the COVID-19 pandemic (2, 3). This study poses a reverse question: Does the eruption of public crises increase individuals’ conformity to the collective? This topic is crucial as it may help identify how human societies dynamically adjust their values to better adapt to unanticipated survival challenges. Additionally, greater conformity can potentially explain irrational responses to public crises, such as panic hoarding and overtrusting disinformation. Conformity has long been a core area of study in social psychology, reflecting individuals’ susceptibility to social influences and reliance on observational learning (4, 5). If conformity values increase, our world may witness decreased desires for scientific and technological innovation (6), reduced needs for personal uniqueness (7), harsher moral judgments about counternormative behaviors (8), and greater identification with conservatism (9). Therefore, determining whether public crises like the COVID-19 pandemic increase conformity behavior is meaningful.

This study addressed this question using datasets that had several advantages. First, they provided high temporal resolution, demonstrating weekly synchronization between crisis development and conformity behavior, which distinguished this study from previous studies that have mainly focused on static patterns. Second, they had high ecological validity by utilizing objective indicators of everyday behaviors instead of self-reports, and without interfering with individuals being observed. Third, they entailed a high sample diversity, encompassing a wide range of different subpopulations in China. In addition to the primary question, this study offered more comprehensive, nuanced insights into the dynamic, adaptive patterns of conformity shifts by exploring how the shift differed across regions, what mechanisms might explain the shift, whether this adaptive shift of conformity (rather than static traits) was associated with antipandemic outcomes, and whether (and when) it reverted alongside the downgrading of the crises.

Regarding the primary question, three streams of previous theories and evidence provided bases for hypothesizing that individuals’ conformity to the collective increased in response to the COVID-19 outbreak. The first stream of literature is the behavioral immune system (8). In cultural psychology, studies have demonstrated that countries/regions with a higher prevalence of disease-causing pathogens in the early 1900s tend to establish solid and strict norms (10), fostering cultural atmospheres that prioritize conformity to the collective, indicated by enhanced effects in behavioral conformity experiments, stronger emphasis on obedience, and less tolerance for nonconformity (11). Furthermore, the historical prevalence of pathogens and infectious diseases is associated with collectivism (12, 13). In addition to historical experiences, even brief exposure to infectious disease information can increase participants’ conformity to majority opinions (14, 15). The COVID-19 pandemic acts as a natural experiment, exposing people to salient threats of infectious diseases, which may trigger their aversion to risks resulting from deviance, leading to a greater inclination toward conforming with the collective.

The second stream of literature is the terror management theory, based on existential psychology and psychodynamics (16, 17). According to this theory, conformity to the collective serves as a defensive means of coping with the anxiety and fear induced by existential threats (18). Experiments have shown that after mortality salience is experimentally increased, participants tend to conform more to others’ opinions, particularly regarding aesthetic judgments and social issues (18). A survey conducted in February 2020 in China found that respondents perceiving greater threats of death from the pandemic reported a stronger need for social belonging and greater engagement in informational conformity consumer behavior (e.g., observing other people’s purchases) (19). This tendency also reflects a “survival wisdom” accumulated through prolonged evolutionary processes: Social isolation can be highly detrimental in life-and-death situations, whereas social linkages may offer certain survival advantages (20), making individuals feel safer (21).

The third stream of literature includes a series of studies on compensatory control in the realms of clinical and social psychology. People generally desire control over their lives; however, this is not always possible, for example, during a pandemic. In response to the lack of control, people may seek simple, clear, and consistent interpretations of the world, even in areas unrelated to where they experience a lack of control, to regain a sense of control (22). Thus, they seek structure while avoiding choices associated with unpredictable risks (23). Conforming to collective norms and opinions may not be the ideal alternative; however, it is arguably predictable and safe, offering a sense of compensatory control. Conversely, diverging from the majority is riskier. In accordance with this notion, studies have found that participants experiencing lower personal control or higher environmental variability tend to opt for the more popular product (i.e., selecting items selected by a larger number of others) (24), prefer regions and organizations with tighter norms (23), attach greater importance to similarities between their coworkers and themselves (25), and judge and punish wrongdoers more harshly (26).

The COVID-19 pandemic represents a confluence of these three streams as it combines factors such as pathogen prevalence, mortality salience, and lack of control, leading to a reasonable prediction of an increase in conformity to the collective. These theories and their predictions were proposed prior to the pandemic and supported by certain laboratory and survey evidence; however, they still require real-world examinations. As Schaller et al. (8) commented, it would be inhumane to expect an opportunity to examine these theories’ predictive validity in the real world, yet the pandemic has heartbreakingly provided such an opportunity. Nonetheless, only two relevant studies (27, 28) were located, both of which adopted a cross-sectional, self-report design and examined static patterns. Thus, whether conformity increased after COVID-19 imposed pathogens, salient mortality, and a loss of control over our world remains unexplored.

We addressed this gap by constructing a real-world dataset that preserves objective indicators of conformity behaviors in the weeks before and after the eruption of China’s first wave of the COVID-19 pandemic, enabling a pre–post comparison. Moreover, we strictly matched observations across weeks on a wide range of attributes or constructed a within-person, longitudinal sample for a cleaner comparison. Furthermore, we conducted difference-in-differences (DID) analyses using data from subsequent regional waves in China to make more confident attributions to the pandemic.

Our dataset originates from Taobao, China’s largest e-commerce platform. Consumers are typically exposed to two signals when browsing product pages on Taobao that can potentially exert social influences: A) the cumulative quantity sold and B) the previous buyers’ ratings. These indicators inform the decision-maker about the popularity of an item among other consumers and how strongly it is recommended by them. Before making a final decision, consumers may consider several alternatives (operationally defined as intrasubcategory items), which differ in terms of quantity sold and ratings. Among these options, the decision-maker can conform to others’ choices and opinions by purchasing the most-sold and/or best-rated items. Against this background, each row of our dataset corresponds to a purchase decision and a list of candidate options that the consumer considered before the purchase. For each row, we created two dummy variables representing whether the eventually purchased item was the A) most-sold and B) best-rated option within the candidate option set. Importantly, our observations of conformity behaviors align with those in previous laboratory tasks. For example, Murray and Schaller (14) asked participants to choose between an option chosen by 3 people and one chosen by 25 people. Additionally, Renkema et al. (18) and Wu and Chang (15) asked participants to rate their preferences for drawings, each accompanied by a visible high, medium, or low rating from the general public. Tan et al. (24) asked participants to choose from five products or service providers, each varying in the number of reviews and average ratings received. Chen (29) also regarded consumers’ reactions to star ratings and sales volume as indicators of online herd behavior. Notably, none of these tasks require participants to consciously attribute their choices to information about others’ opinions. Similarly, the conformity tasks in this study require decision-makers to choose from options varying in terms of the number of people who have chosen them and the ratings assigned to them.

Results

Initial Model.

January 19, 2020, was regarded as the day when the COVID-19 pandemic erupted in China as a nationwide public health crisis. On that day, the first cases outside of Wuhan were confirmed and reported in Beijing and Shenzhen. The expert team assembled by the National Health Commission announced the potential for the new virus to spread between humans. The following day, President Xi, Premier Li, and the State Council of China initiated significant nationwide prevention and control actions, marking the beginning of China’s comprehensive battle against the pandemic. Within the first week after January 19, all provinces, municipalities, autonomous regions, and special administrative regions except Xizang reported confirmed cases, indicating a pervasive, simultaneous public health crisis nationwide. Against this background, we constructed Subdataset 1, comprising randomly selected 600,000 rows of decision-making records from the 8 wk prior to and 10 wk following January 19 (i.e., November 24, 2019, to March 28, 2020; additional rationales are provided in the Materials and Methods section).

Our analyses revealed a significant shift in consumers’ conformity to the collective after the pandemic erupted. The likelihood of consumers selecting the most-sold option was 28.511% during the 8 wk before January 19, increasing to 35.067% during the 10 wk following the outbreak. This increase was significant [log odds ratio (LOR) = 0.303, SE = 0.006, P and P_Bonferroni < 0.001, Cohen’s d = 0.167; SI Appendix, Table S1]. Similarly, the likelihood of consumers selecting the best-rated option rose from 70.693 to 81.804% significantly (LOR = 0.623, SE = 0.006, P and P_Bonferroni < 0.001, Cohen’s d = 0.343; SI Appendix, Table S1). Even after controlling for potential confounding variables at the individual, regional, item, shop, and row levels, the increases in both outcomes remained significant (elaborated in SI Appendix, Table S2).

Weekly trends are depicted in Fig. 1 (SI Appendix, Tables S3 and S4). Compared with Week 8 (i.e., the week before January 19), none of the preceding weeks showed a significant difference in the most-sold indicator after performing the Bonferroni correction. Although statistical significance was found for the best-rated indicator, no consistent temporal trend was apparent before Week 8. Importantly, immediately after the pandemic erupted in Week 9, both outcome indicators surged significantly and plateaued for about a month. Starting from Weeks 13 and 14, both indicators gradually decreased. This pattern corresponds to the State Council’s guidelines issued on February 17 (in Week 13), requiring local authorities to ensure an orderly return to work and normal life, and the downgrade of public health emergency response levels from February 21 to 24 in most regions. Therefore, these conformity variables mirror the real-world fluctuations in pandemic risk with high synchronization.

Fig. 1.

Weekly trends of conformity to the collective before and after the COVID-19 pandemic erupted in China. Odds represent the ratio of the likelihood of selecting the most-sold (or best-rated) options in candidate option sets to the likelihood of selecting any other options. Solid lines represent the estimated marginal means output by generalized linear models with week as the explained variable. Light-colored shaded areas above and below the solid lines denote the 95% CI. Detailed numbers can be found in SI Appendix, Table S4. Red (or purple) stars represent that the odds in that week were significantly higher (or lower) than those in Week 8 (reference group) after the Bonferroni correction, whereas gray squares denote nonsignificance (P_Bonferroni > 0.05); three stars: P_Bonferroni < 0.001 (SI Appendix, Table S3). Subdataset 1 was used.

Are Shifts in Conformity Driven by Lockdowns Instead of Crisis?

According to the weekly trends (Fig. 1), we argue that a plausible alternative explanation—that the increase in conformity was due to lockdowns instead of the crisis per se—cannot be supported. Lockdowns lagged behind the initial outbreak of the pandemic, as this was the first time China encountered such an unprecedented crisis. Throughout Week 9, only 11 regions within Hubei and 1 region outside of Hubei (Qinhuangdao) implemented strict anticontagion policies (i.e., lockdowns) (30), accounting for merely 3.827% of our subdataset’s observations. However, the conformity variables already showed significant increases in Week 9, with no further increases in Week 10 or 11 when more regions (13 and 103, respectively) implemented such policies.

Are Shifts in Conformity Permanent? (Exploratory).

To explore longer-term effects, we supplemented an unpreregistered analysis by extending the observation period by 10 wk, with the final date set to June 6 (Subdataset 2). The significant differences (relative to Week 8) persisted until Week 20 (April 5 to 11; using the Bonferroni correction; SI Appendix, Tables S5 and S6). These findings imply that the observed shifts were seemingly not permanent.

Are Shifts in Conformity Unique?

We evaluated whether the conformity trend was unique or whether similar patterns emerged for other consumer decision preferences. Using Subdataset 1, two typical preferences were quantified as dummy variables: selecting the A) largest-shop and B) lowest-price option within the candidate option set. The presence of famous brands (indexed by shop scale) can arguably reduce the likelihood of making poor decisions and sometimes satisfy the desire for prestige, whereas low prices tend to save money. While both constitute classic consumer decision-making logic, they do not imply conformity. As depicted in Fig. 2 (SI Appendix, Tables S7 and S8), neither the largest-shop nor the lowest-price indicators displayed a temporal trend similar to that of the conformity indicators. In brief, Fig. 2 implies that consumers relied less on shop scale and item price when making decisions in Week 10 (the second week following January 19).

Fig. 2.

Weekly trends of other consumer decision preferences before and after the COVID-19 pandemic erupted in China. Odds represent the ratio of the likelihood of selecting the largest-shop (or lowest-price) options in candidate option sets to the likelihood of selecting any other options. Solid lines represent the estimated marginal means output by generalized linear models with week as the explained variable. Light-colored shaded areas above and below the solid lines denote the 95% CI. Detailed numbers can be found in SI Appendix, Table S8. Red (or purple) stars represent that the odds in that week were significantly higher (or lower) than those in Week 8 (reference group) after the Bonferroni correction, whereas gray squares denote nonsignificance (P_Bonferroni > 0.05); one star: P_Bonferroni < 0.05; two stars: P_Bonferroni < 0.01; three stars: P_Bonferroni < 0.001 (SI Appendix, Table S7). Subdataset 1 was used.

Are Shifts in Conformity Confounded by Individual, Regional, or Product Category Differences?

The above analyses rely on a simple random sample, potentially posing a reasonable challenge that fundamentally different groups might have been selected for various periods. For example, consumers might have purchased products from different categories after the pandemic erupted, potentially resulting in differences between periods. We addressed this possibility by constructing two additional subdatasets. Subdataset 3 comprises 599,994 rows strictly matched on nine attributes (i.e., gender, age level, consumption capacity level, prefecture-level region, item category, item price level, item popularity level, shop scale level, and option set size level) across 18 wk. This careful matching enhances homogeneity and comparability across periods, allowing cross-week differences to be attributed to temporal factors rather than individual, regional, or product category differences with greater confidence. Subdataset 4 includes a within-person, longitudinal sample comprising 33,333 consumer accounts × 18 wk. Both subdatasets possess certain advantages and disadvantages. Subdataset 4 thoroughly rules out the potential confounding effects of individual differences; however, consumers who have continuously purchased for 18 wk may not be representative of the general population.

As depicted in Fig. 3 (SI Appendix, Tables S9–S14), the results of Subdataset 3 (the effect of period on most-sold: LOR = 0.481, SE = 0.006, P and P_Bonferroni < 0.001, Cohen’s d = 0.265; on best-rated: LOR = 0.687, SE = 0.006, P and P_Bonferroni < 0.001, Cohen’s d = 0.379) and Subdataset 4 (most-sold: LOR = 0.418, SE = 0.006, P and P_Bonferroni < 0.001, Cohen’s d = 0.231; best-rated: LOR = 0.723, SE = 0.006, P and P_Bonferroni < 0.001, Cohen’s d = 0.399) are consistent with those of the simple random sample, suggesting that even a homogeneous group or identical individuals may have experienced the proposed increases in conformity after the pandemic erupted.

Fig. 3.

Weekly trends of conformity to the collective before and after the COVID-19 pandemic erupted in China. Odds represent the ratio of the likelihood of selecting the most-sold (or best-rated) options in candidate option sets to the likelihood of selecting any other options. Solid lines represent the estimated marginal means output by generalized linear models with week as the explained variable. Light-colored shaded areas above and below the solid lines denote the 95% CI. Detailed numbers can be found in SI Appendix, Tables S11 and S14. Red (or purple) stars represent that the odds in that week were significantly higher (or lower) than those in Week 8 (reference group) after the Bonferroni correction, whereas gray squares denote nonsignificance (P_Bonferroni > 0.05); two stars: P_Bonferroni < 0.01; three stars: P_Bonferroni < 0.001 (SI Appendix, Tables S10 and S13). Subdatasets 3 and 4 were used.

Are Shifts in Conformity Generalizable across Subpopulations, Regions, and Product Categories?

We extracted a series of subsamples with specific characteristics from Subdataset 1 and tested the differences in conformity before versus on/after January 19 using each subsample. As presented in SI Appendix, Table S15, between-periods differences in the most-sold and best-rated indicators were significant for male and female consumers and for all levels of age, consumption capacity, regional economic development, regional medical resources, regional population density, regional daily sales, item price, item popularity, shop scale, and option set size. Additionally, between-periods differences were significant regardless of whether regional relational or group collectivism was high or low, whether or not the eventually purchased option was the largest-shop or the lowest-price option within the option set, whether or not the region reported confirmed COVID-19 cases throughout the 18 wk, and whether or not the region implemented strict anticontagion policies throughout the 18 wk. Notably, between-periods differences were significant for all 15 tested item categories (i.e., athletic gear, clothing, shoes, jewelry, makeup, skincare, hygiene products, home and living, major appliances, small appliances, cellphones, computers, pet supplies, food, and books), although three differences became nonsignificant following the Bonferroni correction (i.e., jewelry for both indicators and cellphones for the most-sold indicator). This finding rules out the possibility that the observed temporal trends were simply driven by COVID-19 prevention products or by unusual stockout due to panic buying, because such trends were also observed in categories unrelated to antipandemic usage and in categories with a low possibility of panic buying.

In Which Regions Were the Shifts in Conformity Larger? (Exploratory).

Although the results have demonstrated that the shifts in conformity were significant across types of regions, further comparing effect sizes between regions and exploring which types of regions experienced larger increases in conformity are meaningful. Toward this end, we conducted an exploratory, unpreregistered analysis. To obtain effect sizes comparable across regions, we strictly matched all records on the eight abovementioned attributes (excluding prefecture) across not only 18 wk but also 31 province-level regions, thus retaining 274,770 rows for each region (Subdataset 5). We calculated LOR by rerunning the “initial models” but within each province-level region. For all regions, between-periods differences in both the most-sold and best-rated indicators were significant, with LOR ranging from 0.484 to 0.734 (SI Appendix, Table S16). Subsequently, we calculated correlations between these LOR values and 76 province-level indicators, including economic, demographic, cultural, psychological, government, and geographical characteristics, as well as regional average beliefs (Fig. 4 and SI Appendix, Table S17). We focused on the magnitude and direction of the correlations without obtaining P-values due to the small sample size (≤31) and the exploratory nature of the analysis.

Fig. 4.

Exploration of cross-regional differences in effect sizes. Correlation refers to the correlation between a province-level variable (e.g., sex ratio) and effect size (i.e., log odds ratio) obtained from the regression model within provinces (listed in SI Appendix, Table S16). BC = behavioral constraint. Full results, variable operationalization, and detailed numbers can be found in SI Appendix, Table S17. Subdataset 5 was used.

According to the findings, in which regions did conformity increase by a larger extent after the pandemic erupted? First, the role of psychological distance is seemingly critical. The most strongly correlated indicator was the straight-line distance to Wuhan (i.e., the “epicenter” of the first wave). Its correlation with the most-sold LOR (r_MS) was −0.694, while that with the best-rated LOR (r_BR) was −0.459, which indicated that regions at shorter distances from Wuhan experienced larger increases in conformity. Additionally, regions with larger proportions of residents whose household registration areas were in Hubei (where Wuhan is located)—who assumably felt psychologically closer to the “epicenter”—experienced larger increases in the best-rated indicator (r_BR = 0.431). Second, regions with stricter behavioral constraints (i.e., respondents perceived behaviors, such as eating and swearing, as more inappropriate in situations such as buses and classrooms) (31) experienced larger increases in conformity (r_MS: 0.236 to 0.499; r_BR: 0.306 to 0.635). Consistent with this result, regions with lower average levels of openness to experience experienced larger increases in the best-rated indicator (r_BR = 0.550). However, a direct measure of cultural tightness (31) exhibited weak correlations with LOR (r_MS = −0.089; r_BR = 0.132). Third, the increase in conformity seemingly reflected compensation for relatively less cohesive regions instead of amplification for previously cohesive ones. Regions with lower levels of group collectivism (r_MS = −0.451; r_BR = −0.313), lower average levels of general trust (r_MS = −0.447; r_BR = −0.319), and smaller proportions of state-holding enterprises (which assumably emphasize collective interests more; r_MS = −0.482; r_BR = −0.411) experienced larger increases in conformity. Interestingly, larger proportions of rice farming—an economic characteristic theorized to foster interdependence (32)—were correlated with larger increases in conformity (r_MS = 0.458; r_BR = 0.426). Notably, all findings in this paragraph are exploratory descriptions instead of confirmatory tests of any priorly proposed hypotheses.

Are Shifts in Conformity Attributable to the COVID-19 Pandemic?

Despite differences in magnitude, heterogeneity and correlation analyses demonstrate that shifts in conformity were common across regions during the first wave of COVID-19. Conformity levels also significantly increased even in regions that reported zero COVID-19 cases throughout the focal 18 wk. This finding makes sense because perceived risks were widespread nationwide in January and February 2020, irrespective of local cases. However, this presents a critical concern: Due to the absence of a control group unaffected by the pandemic, it is difficult to affirm that the observed increases in conformity can be attributed to the pandemic rather than reflect an episode of naturally seasonal shifts, festival variations, or ongoing social changes. Fortunately, China provides valuable opportunities to conduct DID analyses, enabling the comparison between pandemic-affected regions and control groups (unaffected regions), because China successfully suppressed the pandemic and limited its spread to a small number of regions in the following years. Undeniably, the psychological connotations of subsequent waves differed from those of the first wave; nevertheless, we argue that analyses of subsequent waves can be connected and offer an indispensable complement to those of the first wave. The primary reason is that, unlike other countries, numerous regions in China maintained zero local cases on most days after the first wave. Out of the 356 prefecture-level regions, the monthly number of regions that reported one or more local cases ranged from 0 to 33 in 2020 (except for January and February) and 2021. In contrast, people living in the 300+ other regions were not threatened by proximal risks; thus, there was no need to be in a state of tension. When subsequent regional outbreaks suddenly disrupted this normal life, assuming that psychological reactions similar to (though not equal to) those during the first wave were activated is reasonable. Additionally, considerable societal efforts were devoted to preventing spillover to other regions. As a result, other regions, especially geographically distant ones, frequently maintained normalcy without being affected by remote outbreaks. This stark contrast provided a robust basis for DID analyses, and we established strict criteria to select waves that align with these conditions. If we observe increases in conformity variables in pandemic-affected regions while observing no similar change in comparable, unaffected counterparts, we can attribute the shifts to the pandemic with greater confidence.

Against this background, we first identified pandemic-affected regions based on the following criteria: A) The regions had to report more than 50 new local confirmed COVID-19 cases within a month, ensuring the risks were obviously perceivable. B) In the same month, the number of prefecture-level regions in the Chinese mainland reporting new local confirmed cases had to be fewer than 10, ensuring that the perceived risks were not widespread across the country. These subsequent waves were much shorter than the first wave; thus, we simply equalized the pre- and postperiods (i.e., 8 + 8 wk) instead of observing across 18 wk. For each selected region, we identified counterpart regions that met the following criteria: A) The counterpart regions should strictly match the affected region on five attributes (i.e., economic development level, medical resource level, relational and group collectivism level, and population density level), ensuring comparability. B) The counterpart regions and the affected region should not be in the same province or autonomous region, nor should they share a border, ensuring that the counterpart regions are not psychologically impacted by the pandemic risks in nearby regions. C) The counterpart regions should report zero new local confirmed COVID-19 cases throughout the focal 8 + 8 wk and in 1 mo preceding this period. Using these criteria, we found six pairs of affected regions and their counterparts. For each pair, we constructed a subdataset comprising 100,000 rows that strictly matched on the aforementioned eight attributes (except region) across 16 wk and across the affected region and their counterparts.

The first pair is Urumqi and its counterparts (Dongying, Jiayuguan, and Kunming). The DID analyses indicated a significant interaction between treatment (affected regions versus unaffected counterparts) and period (8 wk before versus on/after July 15, 2020, i.e., the date the first case of this wave was confirmed) for the most-sold indicator (LOR = 0.449, SE = 0.033, P and P_Bonferroni < 0.001, Cohen’s d = 0.248; SI Appendix, Table S18). The interaction was also significant for the best-rated indicator but became nonsignificant after the Bonferroni correction (LOR = 0.128, SE = 0.041, P = 0.002, P_Bonferroni = 0.506, Cohen’s d = 0.071). The results of event study are depicted in Fig. 5 (SI Appendix, Table S19). None of the interactions between treatment and week were significant before July 15, supporting the DID model assumption that the affected and counterpart regions followed parallel trends prior to the pandemic. In Week 9, the interactions became significant for both indicators; that is, the likelihood of selecting the most-sold and best-rated options in Urumqi (relative to counterpart regions) significantly increased when this wave of pandemic began in Urumqi. Similar patterns for the best-sold indicator were also observed in the other five pairs (Fig. 5 and SI Appendix, Tables S20–S29): A) Dalian versus Huaian and Shaoxing, July 22, 2020; B) Dehong versus Shangluo and Tongren, March 30, 2021; C) Guangzhou versus Beijing, Changsha, Chengdu, Ningbo, Wuhan, Xiamen, Zhengzhou, and Zhoushan, May 21, 2021; D) Xiamen versus Nanchang, Ningbo, Zhoushan, and Zhuhai, September 12, 2021; and E) Harbin versus Fushun, September 21, 2021. The increased conformity in these subsequent waves declined faster than in the initial model (i.e., the first wave in China). For the other indicator, the interactions between treatment and period on the best-rated indicator were significant only in earlier cases (Dalian, Dehong, and Guangzhou) and not in later cases (Xiamen and Harbin); moreover, they remained significant only for Guangzhou after the Bonferroni correction.

Fig. 5.

Results of event study. Odds ratio (solid lines; referring to the left axis) represents the interaction effect between each week and treatment (1 = pandemic-affected regions, 0 = unaffected counterparts). The light orange and blue areas above and below the solid lines indicate the 95% CI, calculated using the Wald interval. Week 8 is the reference group and, thus, has no CI. Red (or purple) stars represent that the odds ratio in that week was significantly higher (or lower) than 0 after Bonferroni correction, whereas gray squares denote nonsignificance (P_Bonferroni > 0.05); one star: P_Bonferroni < 0.05; two stars: P_Bonferroni < 0.01; three stars: P_Bonferroni < 0.001. Detailed numbers can be found in SI Appendix, Tables S19, S21, S23, S25, S27, and S29. The green areas (referring to the right axis) represent the number of new confirmed COVID-19 cases in the affected region each week. Subdatasets 6 to 11 were used.

What Mechanisms Can Explain Shifts in Conformity? (Exploratory).

After confirming the increase in conformity after the COVID-19 outbreak, the next aspect to examine is which mechanisms underlie this phenomenon. In the introduction section, the hypotheses were formulated using three streams of theories, namely, parasite prevalence, mortality salience, and loss of control. Although a thorough test may exceed the potential of the current datasets, we conducted unpreregistered analyses to explore which of these three theories can best explain the data. Toward this end, we used new confirmed cases and death cases per week at the prefecture level as proxies for parasite prevalence and mortality salience, respectively. Identifying a comprehensive quantification for loss of control is difficult; thus, we adopted partial reflection: the extent of the decrease in regional population mobility, including within-city movement, move-out, and move-in, compared with the prepandemic baseline. We connected these five variables to Subdatasets 1, 3, and 4 (i.e., the first wave in China) as parallel mediators, tested using a mediation model with prefectures as the cluster. Additionally, we replaced confirmed cases and death cases with cases per million to check robustness. The parallel mediation approach enabled the five mediators to function as control variables for one another, and no multicollinearity was detected, given that the variance inflation factor (VIF) ranged from 1.533 to 3.997.

Throughout these trials, the most consistent and largest indirect effect was through the decrease in within-city movement (ranging from 0.020 to 0.069, all P and P_Bonferroni < 0.001; SI Appendix, Tables S30 and S31). This variable indicated the extent to which people avoided or were warned against moving freely, which partially (although not comprehensively) reflected the disruption of people’s control over daily life. Thus, the significant mediating effect of this variable preliminarily pointed to the role of compensatory control in the increase in conformity. On the contrary, support for the roles of parasite prevalence (confirmed cases) and mortality salience (death cases) were inconsistent; however, these findings should not be viewed as a rejection of these theories because risks were widespread nationwide at that time and, thus, the nonsignificance may be attributed to the lack of regions where people did not feel threatened by parasite and mortality.

Did Regions with Larger Shifts in Conformity Achieve Better Antipandemic Outcomes? (Exploratory).

Unpreregistered analyses were further conducted to identify the role of increased conformity in combating COVID-19. Focusing on the first wave in China (10 wk after January 19, 2019; Subdataset 12), we tested a moderation model (prefecture level) in which the main effect was whether regions with larger increases in average conformity levels in a given week would see lower levels of COVID-19 severity in the following week. This effect may vary over time, so we entered “week” as a categorical moderator. Following Talhelm et al. (33), we adopted log-transformed COVID-19 cases per million as a quantification of antipandemic outcomes (i.e., dependent variable in the model). A mixed-effects model demonstrated that the main effect of the increase in the most-sold indicator was significant (b = −68.477, SE = 8.154, P and P_Bonferroni < 0.001, β = −1.285; SI Appendix, Table S32). Simple slopes analyses indicated that this effect was significant only for 5 wk after the outbreak and thereafter became nonsignificant (Fig. 6, Left and SI Appendix, Table S33). This finding implies that, during the early stage, if the regional average level of the most-sold indicator increased by a larger extent in a given week (relative to the prepandemic baseline) in a region than in other regions, then the COVID-19 prevalence of this region may be less severe in the subsequent week. For the other variable of conformity, the main effect indicated that the association between the increase in the best-rated indicator in the first week after the outbreak and second-week severity was nonsignificant (b = 16.961, SE = 18.871, P = 0.369, P_Bonferroni > 1, β = 0.063); however, in the third and fourth weeks, the effects became significant (third: b = −42.257, 95% CI = [−73.430, −11.083]; fourth: b = −59.497, 95% CI = [−102.882, −16.111]; Fig. 6, Right). Additionally, we controlled for several variables that might influence COVID-19 severity, but these adjustments did not change the overall patterns of results (SI Appendix, Table S34). These findings highlight the potential role of dynamic conformity tides in the fight against the COVID-19 pandemic, although they do not establish causality.

Fig. 6.

Exploration of associations between conformity shifts and pandemic outcomes. COVID-19 severity refers to log-transformed COVID-19 cases per million in the subsequent week. For example, severity in Week 10 is explained by “increase in conformity” in Week 9 [i.e., conformity in Week 9 ÷ average (conformity during Weeks 1 to 8) − 1]. Solid lines represent the estimated marginal means output by mixed-effects models controlling for the random effect of prefecture. Light-colored shaded areas above and below the solid lines denote the 95% CI. Detailed numbers can be found in SI Appendix, Table S33. Subdataset 12 was used.

Discussion

In conclusion, the present work confirmed the hypothesized increase in conformity to the collective when the COVID-19 pandemic erupted. After the nationwide public health crisis broke out in China on January 19, 2020, online consumers became more likely to opt for choices that the largest number of others had chosen and those that others had expressed the most positive opinions about among their considered alternatives. Using multiple sampling and statistical strategies, we ruled out various confounding factors and alternative explanations, enabling a more confident attribution to the public health crisis. A key highlight is our adoption of the DID approach to analyze the subsequent regional waves of the pandemic in China. The results revealed that across all six events, the likelihood of selecting the most-sold options in pandemic-affected regions (relative to unaffected, comparable counterparts) increased during 2 to 4 wk following the onset of each wave. These findings provide real-world evidence for predictions based on previous theories on the behavioral immune system (11, 14), terror management (18), and compensatory control (22) with high ecological validity and temporal resolution, which demonstrates that prepandemic theories validly predicted conformity tides following the outbreak of a real pandemic.

The practical implications of such conformity tides are rich but bittersweet. On the one hand, people may benefit by conforming to the authorities in their anti-infection efforts, primarily when authorities provide beneficial guidelines (e.g., wearing masks) (8). At the societal level, battling public crises like the COVID-19 pandemic typically requires increased social coordination, which could be more easily achieved when most people prioritize alignment with others over personal unique pursuits. Studies have indicated that people in more collectivist and less neoliberal regions, as well as people with a higher level of national identification, were more proactive in carrying out actions to help society fight the pandemic, even if it meant sacrificing personal comfort and convenience (3, 34, 35); however, these structural characteristics were primarily regarded as a static trait of a region in previous studies. This study captures the dynamic adjustment and adaptation of society to sudden survival challenges by revealing the temporal synchronization between crisis development and conformity behaviors, as well as the association between conformity adjustment and antipandemic outcome. At the individual level, conformity behaviors offer a sense of social connection, reducing anxiety and increasing feelings of security (21, 24) as people feel they are not facing crises alone, although this sense of regaining control may merely be compensatory and does not effectively address the areas where people lose control (22). On the other hand, pandemic-induced conformity tendencies may function as underlying mechanisms for collective irrationality, such as panic hoarding and overtrusting disinformation (36). Moreover, studies have indicated relationships of higher conformity values with weaker pursuits of scientific innovation (6) and personal uniqueness (7).

These bright and dark sides of potential consequences help enhance the current understanding of social changes—at least short-term ones—after the COVID-19 outbreak. One notable potential change is a trend toward conservative attitudes, as evidenced by previous short-term studies. For example, after the outbreak of the pandemic, researchers found that respondents perceived nontraditional women and gay people as more threatening (37), expressed stronger beliefs in traditional gender stereotypes (38), and were less supportive of gender equality (39). Anti-Asian sentiment (40) and hate crimes (41) in the United States increased with the emergence of COVID-19. Additionally, after reading a report on the pandemic outbreak (versus climate change), participants were more likely to intend to vote for conservative candidates such as Donald Trump (42). Increased conformity to the collective offers a lens through which to interpret these traces of conservative attitudes.

In terms of long-term changes, however, our results reveal that the COVID-19 crisis did not result in a permanent or irreversible increase in conformity; instead, the degree of conformity gradually declined with the decrease in infection risks. Contextualizing these findings within the specific social context is essential: China successfully zeroed local cases of COVID-19 and restored normalcy within a relatively short period. In contrast, whether increased conformity would persist and become entrenched if the COVID-19 risk—and its associated disruption of personal control—were never eliminated in other societies remains an open question. Speculating that increased conformity and its consequences (e.g., conservative attitudes, weak innovation pursuits, and intolerance of nontraditional identities) in such societies may endure and become embedded as part of the regional culture is reasonable, given the strong synchronization between real-world risks and conformity levels revealed by our analyses. This speculation is inspired by previous evidence that demonstrates that regions with a history of higher levels of pathogen prevalence have retained stricter norms, higher conformity, and greater collectivism even today (10–13). Nevertheless, this speculation warrants an empirical examination.

Importantly, this study contributes to prior theories by not only confirming the existence of shifts in conformity but also revealing the positive meanings of this dynamic adaptation. From an evolutionary perspective, motives to avoid diseases are adaptive regulatory systems that detect the threat of infection, generate affective, and cognitive responses and accordingly guide behavioral decisions to mitigate infection risks (8). Previous studies mainly focused on whether and how threats elicit responses, but the present study further illustrates that threat-induced responses may be helpful in addressing threats per se. Specifically, regions with larger increases in conformity achieved better antipandemic outcomes during the early stage of the pandemic. As the crisis progressed, other sources of antipandemic factors, such as public administration and medical intervention, may have gained prominence, which potentially explains the gradually diminished significance of the effect of conformity shifts. However, in the first stage, these shifts served as a timely, spontaneous response system. Thus, conformity shifts, as part of the behavioral immune system, are rewarding, problem-oriented, and evolutionarily adaptive, which is similar to the physical immune system.

Given this positive role, another question emerges: What are the characteristics of the regions that adjusted their conformity by a larger extent in response to the pandemic (thereby benefiting more from this system)? Our exploratory analyses offer several insights into this issue and enrich the current understanding of cross-cultural differences. First, Talhelm et al. (33) demonstrated that regions that were mainly reliant on rice farming achieved better antipandemic outcomes, but they attributed this relationship to static regional traits such as cultural tightness and relational mobility. As a meaningful complement, the present study reveals a dynamic mechanism: Rice-growing regions increased conformity by a larger extent to cope with the pandemic outbreak. These regions may have historically accumulated a schema of social coordination due to the higher demands for irrigation and cooperative labor exchange (32). This schema may have shaped people’s responses to public crises. Second, although previous studies found that collectivist cultures used masks more (3) and achieved better antipandemic outcomes (2), the present study reveals that regions with higher levels of group collectivism experienced smaller increases in conformity. A potential explanation is that a collectivist society inherently maintains high levels of social coordination and does not need a temporary, compensatory increase in conformity. In contrast, an individualist society may respond differently. For instance, Huang et al. (2) found an association between individualism and a stronger fear of death during the pandemic. According to terror management theory, fear of death may induce conformity because belonging to a group provides individuals with psychological assurance (18). Integrating these studies might help explain why individualist regions showed a greater increase in conformity. The role of collectivism–individualism in antipandemic psychology may be more complex than originally expected. Third, a contextualized, behavior-oriented measure of behavioral constraints (rating the appropriateness of a series of specific behaviors according to various scenarios) is more relevant to conformity shifts compared with a general, perception-oriented measure of cultural tightness (31), although the former is regarded as a behavioral criterion for the latter. Conformity to the collective increased by a larger extent in response to the pandemic in regions with stricter constraints on daily behavior prior to the pandemic.

Admittedly, the results of the DID analyses on the best-rated indicator are inconsistent, as the effects were smaller in size, shorter in duration, and only significant in some cases. Several post hoc explanations are as follows: A) The subsequent waves were indeed much less severe and shorter in duration than the first wave in China. B) People repeatedly exposed to different waves of the pandemic may have developed coping strategies and/or pandemic fatigue (43), responding to pandemic threats with diminished intensity. C) Unlike the most-sold indicator (e.g., M = 0.315 in Subdataset 1), the best-rated indicator has a high average level (M = 0.757), suggesting a potential risk of ceiling effects. This disparity may be a culturally specific phenomenon. For example, researchers found that, for interdependent consumers, cues that convey the choices of other people were less persuasive than cues that convey others’ attitudes, but the two cues did not differ for consumers with independent self-construal (44). The present data also demonstrated that Chinese online consumers typically place substantial importance on others’ ratings (attitudes), which left little room for pandemic-induced rises. D) Certain previous studies have suggested that consumers’ reactions to the number of people selecting a particular option may be a better indicator of conformity than their reactions to an average rating (24).

Although we have shown that increased conformity following the outbreak of the pandemic can be generalized across a wide range of subpopulations, regions, and product categories, three generalizability limitations should be noted: A) Whether conformity also increases in response to other public crises aside from the pandemic remains an open question. Researchers in pathogen prevalence studies have typically answered “no” to this question, arguing that the effects of infectious disease threats on conformity are unique (14). However, studies on mortality salience and compensatory control have shown increased conformity after participants expressed their thoughts about death (18) or read charts indicating extreme weather (24), implying that the conformity shifts observed in this study have the potential to generalize to other public crises unrelated to pathogens. B) The finding’s generalizability to contexts other than China remains untested. Our results preliminarily illustrate that conformity also significantly increased in relatively individualist regions within China; however, data are lacking on more individualist countries. C) We only tested conformity behaviors in online shopping settings and hence cannot determine whether similar effects would emerge in other conformity settings. Nevertheless, we argue that our nonparticipatory observation of objective behaviors can avoid social desirability concerns and potential discrepancies between self-reported scores and actual deeds. Furthermore, it can capture conformity behaviors in more self-driven, consequential tasks compared to previous laboratory simulations. These three aforementioned questions remain unaddressed in the present work and require further research in future studies.

Materials and Methods

Overall Inclusion Criteria.

All subdatasets throughout the study were based on Taobao transactions, with each row corresponding to one consumer account’s purchase of one item. We ensured comparability and explainability by only obtaining transactions from 15 selected mainstream categories: athletic gear, clothing, shoes, jewelry, makeup, skincare, hygiene products, home and living, major appliances, small appliances, cellphones, computers, pet supplies, food, and books. We did not obtain transactions from consumer accounts younger than 18 y old in 2019. For each row (i.e., each transaction), we generated a list of candidate options, referring to the intrasubcategory items browsed, added to carts, and/or added to the favorites by the consumer account 24 h before the focal purchase. The Taobao system determined subcategories; examples include T-shirts, one-piece dresses, liquid essence, eyeliner, and puffed food. If the candidate option set size equaled one (i.e., the consumer account directly bought the item without considering other alternatives), the transaction was dropped because a comparison and selection process was not involved.

Targeted Regions, Targeted Periods, and Subdataset Generation.

For the first wave of the COVID-19 pandemic, we targeted 356 prefecture-level regions in China; thus, transactions outside these regions were not considered (see the SI Appendix, Text A for the rationale on why several regions were not included in the analysis). The first case outside Wuhan was confirmed and reported on January 19, 2020; thus, we considered this date as the start of the public crisis. The following day, the national government initiated the nationwide comprehensive fight against the pandemic. Although as early as December 31, 2019, the Wuhan government publicly reported 27 cases of “pneumonia of unknown cause,” the potential risks were not immediately known across the country. Therefore, we set 8 wk before and 10 wk after January 19 as the targeted period in our initial model (i.e., November 24, 2019–March 28, 2020) and set Week 8 as the reference group (i.e., January 12 to 18, 2020). The ending week was chosen because it was 1 wk after the date the Chinese mainland had zero new domestic cases for the first time (i.e., March 18, 2020), marking the end of this public crisis.

The full dataset comprising all transactions fulfilling requirements would be vast, and the security department of Alibaba Group did not permit downloading the original data onto personal computers. Thus, according to the security rules and software affordability, we determined 600,000 rows as the maximum sampling size of any subdataset. To construct Subdataset 1 (a simple random sample), we used DataWorks (a cloud computing tool of Alibaba) to order all transactions fulfilling requirements within the targeted period (i.e., the 18 wk) by random numbers generated by the rand() function and subsequently retain the first 600,000 rows. This subdataset comprises 597,050 unique consumer accounts, with demographic descriptions presented in SI Appendix, Table S35. This subdataset covers all 356 prefecture-level regions, showing a high correlation between the regional headcounts of Subdataset 1 and the actual regional population in 2019 (r = 0.846, P < 0.001). Subdataset 1’s descriptive statistics and correlation analysis results are presented in SI Appendix, Tables S36 and S37.

To explore longer-term effects, we constructed Subdataset 2 (unpreregistered) by extending the observation period by 10 wk. We randomly selected 600,000 rows from transactions within the extended period (i.e., 28 wk; SI Appendix, Tables S35 and S38).

To construct Subdataset 3 (a strictly matched sample), we first partitioned all transactions fulfilling requirements within the targeted period into cells by 10 attributes (i.e., week, gender, age level, consumption capacity level, prefecture-level region, item category, item price level, item popularity level, shop scale level, and option set size level). Rows with missing values for any attribute were dropped. Within each cell, all rows were numbered according to random orders. We generated 18 independent tables for the 18 wk. We used the “inner join” syntax to match the “Week 1” table with the “Week 2” table, retaining only those rows that exactly matched all nine attributes (except week) and row numbers. We subsequently matched the resultant table with the “Week 3” table and continued this process until the “Week 18” table. Cells that only appeared in some weeks were dropped because they were not comparable across weeks and, thus, might introduce confounding effects. In contrast, cells found in all 18 wk were retained; of these, we randomly selected 33,333 rows, corresponding to 599,994 transactions. This subdataset comprises 594,786 unique consumer accounts, with demographic descriptions in SI Appendix, Table S35. SI Appendix, Table S39 presents the results of Subdataset 3’s descriptive statistics and correlation analyses.

To construct Subdataset 4 (a within-person, longitudinal sample), we first identified all consumer accounts with at least one transaction in each of the 18 wk. We then randomly selected 33,333 consumers, whose demographic descriptions are presented in SI Appendix, Table S35. For each consumer, we randomly selected one transaction each week to construct the subdataset, with the descriptive statistics and correlation analysis results presented in SI Appendix, Table S40.

To explore cross-regional differences, we constructed Subdataset 5 (unpreregistered) using the procedure used for Subdataset 3. The difference was that we strictly matched all transactions not only across weeks but also across province-level regions to ensure comparability. Initially, we aimed to similarly select 599,994 transactions randomly for each of the 31 province-level regions; however, only 274,770 rows for each region remained after matching, which were all included in the analyses (SI Appendix, Tables S35 and S41).

In addition to the first pandemic wave in China, we also targeted the subsequent regional events to enable DID analyses. The procedures and criteria to determine the targeted regions and their counterparts have been reported in Results section. The data on newly confirmed COVID-19 cases were obtained from the official websites of provincial, autonomous-regional, and municipal people’s governments, health commissions, and centers for disease prevention and control. For each event, we focused on 8 wk before and 8 wk starting from the date the first case of that wave was confirmed in the affected region as the targeted period. The matching procedures were similar to those used to construct Subdataset 3 with a slight difference: The data were matched on eight attributes (except region) not only across all 16 wk but also across the affected region and its counterpart(s). Deviating from the preregistration, we selected 100,000 rows for each of the six events (see the SI Appendix, Text B for the rationale; SI Appendix, Tables S35 and S42–S47).

Variables and Measurements.

For each row of our dataset, we generated a list of candidate options considered by the consumer before purchasing. We obtained four aspects of background information for each candidate option: A) the cumulative quantity of the item sold in the 180 d before the focal purchase; B) the rate of nonpositive remarks that the item received in the 180 d before the focal purchase; C) the shop’s gross merchandise volume in the 180 d before the focal purchase; and D) price. If the purchased item was the most-sold option in the candidate option set, we coded 1 for the most-sold variable. Even if the eventually purchased item and abandoned option(s) tied for the first place, we still coded 1. In contrast, if at least one abandoned option was sold more than the eventually purchased item, we coded 0, indicating that the consumer did not choose the most-sold option in this round of decision-making. Similarly, we created the best-rated, largest-shop, and lowest-price variables. Details about other variables, along with the methods used to transform continuous variables into categorical ones, are provided in SI Appendix, Text C.

Statistical Analyses.

The initial model was a generalized linear regression with a “binomial” family, using the “period” dummy variable as the explanatory variable and the “most-sold” (or “best-rated”) dummy variable as the explained variable. To address specific research questions, we modified the model by replacing the explanatory and/or explained variables, additionally controlling fixed and/or random effects, using different subdatasets, or adopting different functions. All these details can be found in SI Appendix, Text D.

The DID analyses were another stream of analysis tasks. Classical DID analyses require controlling the individuals’ fixed effects in the regression model. Because it is reasonable to assume that few consumers continuously purchased for 16 wk and, even if such consumers did exist, they might not represent a broader population, the present analyses did not restrict data to the 16-wk repeated measures of the same individuals. Instead, we only required data homogeneity on eight attributes across weeks and across the affected region and its counterpart(s). Against this background, we replaced the “individual” in classical DID analyses with a new column titled “group_id,” which concatenated gender, age level, consumption capacity level, item category, item price level, item popularity level, shop scale level, option set size level, affected versus unaffected region, and row number in cells. In other words, such a combination was regarded as an “individual,” the temporal change of which was our focus. Following the DID approach, we performed a generalized linear regression with three terms: the interaction between period (pre versus post) and treatment (affected region versus unaffected counterparts), the fixed effects of week, and the fixed effects of “group_id.” Including these two fixed effects can control for time-invariant factors specific to each “group_id” and weekly fluctuations shocking all “group_ids.” Therefore, the interaction effect denotes the change in conformity variables after the pandemic broke out in the affected region (relative to unaffected counterparts). Subsequently, we conducted an event study by replacing the dummy variable “period” with the categorical variable “week,” setting Week 8 as the reference group, to test the parallel-trends assumption and depict the weekly trends.

Due to the multiple comparisons, P-values were adjusted using the Bonferroni correction (P_Bonferroni). These adjusted P-values were calculated by multiplying the original P-values by the total number of tests conducted in the study (i.e., 284).

Ethics Approval.

The present study was completed under Protocol #2023-12-01, which was approved by the Research Ethics Committee of the School of Psychological and Cognitive Sciences, Peking University. The researchers based informed consent on the user agreement with Taobao, which includes a statement that Taobao and its partners may use deidentified or anonymized data in scientific research, given that data security and legitimate purposes are ensured. Thus, informed consent was indirectly obtained because this study conducted anonymous archival analyses without recruiting human participants or using identifiable information.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

The hypotheses, design, and analysis of the study were preregistered at https://osf.io/48pqn/?view_only=f1fa0c2930564de39eafa544c355390f (45). SI Appendix, Table S48 summarizes the deviation from the preregistration and unpreregistered analyses. Data are not publicly available due to Alibaba Group’s rules for protecting consumer privacy and data security. Researchers interested in accessing the data for validating/reproducing this study may contact the first author to initiate a formal request. Such requests are highly likely to be approved, though the author cannot guarantee approval, because Alibaba Group (not the author) retains exclusive authority over data usage. According to its agreement with users, the company and its partner use deidentified or anonymized data in scientific research only if data security and legitimate purposes are ensured. Moreover, storing and processing such large-scale datasets requires the company to incur significant cloud computing and financial costs. Access is therefore subject to Alibaba Group’s internal review and approval process, which is outside the author’s control. However, the author promises to proactively assist with such requests and, if access is denied by the company, will communicate the company’s concerns to requesters and provide suggestions for addressing them.