Informational value of visual nudges during crises: Improving public health outcomes through social media engagement amid COVID‐19

It shall be noted that the specific details behind idiosyncratic content‐sharing strategies adopted by each university are different and are largely unobserved to us. For example, the only facet of the strategy that is observed to us is the proportion of the images containing masks, but the contextual diversity of those images, their editorial quality, posting guidelines, or internal content‐sharing policies remain unobserved to us. In other words, each university's content‐sharing strategy implies their own unique “best” way of communicating the desired message. Therefore, to account for potential bias associated with any differences in universities’ individual approaches to social media engagement, we use CEM (Iacus et al., 2012). Specifically, to reduce the confounding impact and produce an appropriate covariate balance between the groups with a variable proportion of the shared mask images, we group universities by size, type of ownership, and locale of the institution in terms of rurality or urbanicity. To control for the nudging effect of past IG images (as the main statistically significant finding of our analysis), we first use

I G _ M a s k i , t − 7 → t $IG\_Mas{k}_{i,t - 7 \to t}$

and then for each university, compute the median level

M i = median [ I G _ M a s k i , t − 7 → t ; t = 1 , … , 22 ] $M_i = {\rm{\ median}}[ {IG\_Mas{k}_{i,t - 7 \to t};t = 1, \ldots ,22} ]$

. ²⁴ Then, we compute the threshold classifier as the median over all universities

Mask threshold = median [ M i ; i = 1 , … , 104 ] . ${\rm{Mask\ threshold\ }} = {\rm{\ median}}[ {M_i;i = 1, \ldots ,104} ]{\rm{.}}$

We note that the sample size decreased from 117 to 104 universities, as for the current model setup, we removed observations with unbalanced panels (i.e., we excluded universities that did not have a complete 22 weeks of observed data due to variations in their start of the semester). ²⁵ Thus, we group the universities as follows:

Few mask images ≔ [ i = 1 , … , 104 ; such that M i < Mask threshold ] ${\rm{Few\ mask\ images}}\coloneq [i = 1, \ldots ,104;{\rm{\ such\ that\ }}{M}_i < {\rm{Mask\ threshold}}]$

, and

Many mask images ≔ [ i = 1 , … , 104 ; such that M i ≥ ${\rm{Many\ mask\ images}}\coloneq [ i = 1, \ldots ,104;{\rm{\ such\ that\ }}{M}_i \ge$

Mask threshold].

Last, for the outcome

IR i = Σ t = 1 , … , 22 Δ i t ${\rm{IR}}_i = {{{\Sigma}}}_{t = 1, \ldots ,22}\ {{{\Delta}}}_{it}\$

(where

Δ i t ${{{\Delta}}}_{it}$

is the positivity rate of newly tested individuals of university i at the end of week t), we compute the sample average treatment effect on the treated (SATT) by IG mask images in the infectivity rate per group. Estimation results provide suggestive evidence of the generally negative effect of the proportion of shared mask images,

S A T T I G = − 0.134 , p < 0.05 $SAT{T}_{IG} = - 0.134, p < 0.05$

. Overall, this result is qualitatively similar to our main results presented in Table 2. Additionally, Figure 3 below shows that the CEM stratification resulted in a total of 10 groups, most of which have a negative treatment effect of mask images.

OPEN IN VIEWER

FIGURE 3

Coarsened exact matching (CEM) analysis results.

Epidemiological model

Although we have made significant effort to account for observed confounding effects in our econometric model specification, a critical factor that remains unobserved in our specification is the true COVID‐19 infectivity on the university campuses. True infectivity is the key factor that ultimately drives the observed positivity rates and at the same time is likely to be correlated with managerial decisions to share more or less mask‐related content. Omission of this latent variable in our modeling approach can lead to biased estimates and misleading inferences. Therefore, to ensure correct identification of the previously identified statistically significant IG effect, we need a model specification that considers the disease dynamics originating at different levels of the population and allows estimation of true infectivity from the data. To this end, we adapt the widely accepted compartmental epidemiological SIR model (Gnanvi et al., 2021; Kermack & McKendrick, 1927).

S i t $S_{it}$

,

I i t $I_{it}$

, and

R i t $R_{it}$

denote the proportion of susceptible, infected, and removed (recovered or dead) individuals, respectively. We model the disease propagation in discrete time, where t stands for time, measured in weeks in our context; and i is a location denoting a university in the United States. Thus, we have the population balance condition

S i t + I i t + R i t = 1 $S_{it} + I_{it} + R_{it} = 1$

. The positivity rate of newly tested individuals of i at the end of week t is

Δ i t ${{{\Delta}}}_{it}$

. Finally, we let γ denote the rate at which infected individuals are removed from infection transmission. Given a 15‐day recovery period, ²⁶

γ = 1 15 = 0.066 $\gamma = \frac{1}{{15}} = 0.066$

is the probability of active recoveries on a given day with expected dynamics of the infected tested and recovered population as

γ I i t $\gamma I_{it}$

. The set of equations that represent the infection dynamics is stated in System (2):

2

Using the data available for each university, we benchmark four possible models to test Proposition 1. The corresponding models’ descriptions and associated dynamics are presented in Table 3.

OPEN IN VIEWER

TABLE 3

Benchmarked models and underlying dynamics.

Model	Description	Dynamics
M1	Base model (on‐campus): Considers only the internal infections that occur at each university. Each susceptible individual can meet with a proportion I i t $I_{it}$ of individuals that are infected. The parameter β i t M 1 $\beta _{it}^{M1}$ denotes the rate of contact and infection in university i at time t. Therefore, the new infections are:	Δ i t = S i t I i t β 1 i t M 1 $\ {{{\Delta}}}_{it} = S_{it}\ I_{it}\beta _{1it}^{M1}$
M2	Base model with mask: We enhance the previous model by controlling for the intensity of mask images signaled by university i at time t. I G _ M a s k ¯ i t = 1 − I G _ M a s k i , t − 7 → t − 1 ${\overline {IG\_Mask} }_{it} = 1 - IG\_Mas{k}_{i,t - 7 \to t - 1}$ :	Δ i t = S i t I i t β 1 i t M 2 + β 2 i t M 2 I G _ M a s k ¯ i t ${{{\Delta}}}_{it} = S_{it}\ I_{it} \left( {\beta _{1it}^{M2} + \beta _{2it}^{M2}IG\_{{\overline {Mask} }}_{it}} \right)$
M3	Base model with off‐campus effects: While institutions can ensure compliance with safety measures within their limits, the chances of members of the university population contracting the virus increase substantially when they leave campus and behave voluntarily. To account for such effects, we consider university i belonging to county j, and its corresponding indexes of internal mobility CMI j t ${\rm{CMI}}_{jt}\$ and social distancing CCI j t ${\rm{CCI}}_{jt}$ . Each susceptible individual of university i can meet with a proportion I j t $I_{jt}$ of individuals from county j that are infected:	Δ i t = S i t I i t β 1 i t M 3 + S i t I j t β 3 i t M 3 CMI j t + β 4 i t M 3 CCI j t ${{{\Delta}}}_{it} = S_{it}\ I_{it}\beta _{1it}^{M3} + S_{it}{I}_{jt} \left( {\beta _{3it}^{M3}{\rm{CMI}}_{jt} + \beta _{4it}^{M3}{\rm{CCI}}_{jt}} \right)$
M4	Base model with off‐campus effects and mask: Final model that combines mobility, social distancing, and masked images:	Δ i t = S i t I i t β 1 i t M 4 + β 2 i t M 4 I G _ M a s k ¯ i t + S i t I j t β 3 i t M 4 CMI j t + β 4 i t M 4 CCI j t ${{{\Delta}}}_{it} = S_{it}\ I_{it} \left( {\beta _{1it}^{M4} + \beta _{2it}^{M4}{{\overline {IG\_Mask} }}_{it}} \right) + S_{it}{I}_{jt} \left( {\beta _{3it}^{M4}{\rm{CMI}}_{jt} + \beta _{4it}^{M4}{\rm{CCI}}_{jt}} \right)$

Abbreviations: CCI, Cuebiq's Contact Index; CMI, Cuebiq's Mobility Index.

For each model, university

i = 1 , … , 104 $i = 1, \ldots ,104$

²⁷ and week

t = 8 , … , 21 $t = 8, \ldots ,21$

, we first used the available data of the previous 7 weeks (i.e.,

t − 7 , … , t − 1 $t - 7, \ldots ,t - 1$

) to regress the parameters of week t using a lasso regularized generalized linear model (Friedman et al., 2010). For Model M4, the regression objective function to minimize is:

The first summand corresponds to the square errors of the regression, and the second summand penalizes deviations of β’s across consecutive weeks. This penalty models the intuition that β’s should vary gradually over time. The last summand adds a lasso penalized regularization that reduces the tendency of the estimator to explain solely using one of the covariates. For the estimation process, we heuristically tuned the lasso meta‐parameters, obtaining the lowest out‐of‐sample error with

λ = 10 − 3 , v 1 = 0.7 , v 2 = 0.7 , v 3 = 1.3 , v 4 = 1.3 . $\lambda = {10}^{ - 3} , v_1 = 0.7, v_2 = 0.7, v_3 = 1.3, v_4 = 1.3.$

Next, we predict the new cases for out‐of‐sample week t,

Δ ̂ i t ${\hat{\Delta }}_{it}$

. Note that we use the inverse of mask intensity (i.e.,

I G _ M a s k ¯ i t = 1 − I G _ M a s k i , t − m a s k t w → t − 1 ${\overline {IG\_Mask} }_{it} = 1 - IG\_Mas{k}_{i,t - masktw \to t - 1}$

) such that values approaching 0 indicate a higher proportion of mask images shared, while 1 implies no mask images at all.

m a s k t w $masktw$

is a parameter to calibrate and denotes the number of past weeks used to construct

I G _ M a s k ¯ i t ${\overline {IG\_Mask} }_{it}$

.

To compare models, we computed the root mean square error (RMSE) as

RMSE = ( ∑ i = 1 104 ∑ t = 7 22 ( Δ i t − Δ ̂ i t ) 2 ) 104 × ( 22 − 6 ) ${\rm{RMSE\ }} = \sqrt {\frac{{( {\mathop \sum \nolimits_{i = 1}^{104} \mathop \sum \nolimits_{t = 7}^{22} {{({{{\Delta}}}_{it} - \ {{{{\hat{\Delta }}}}}_{it})}}^2} )}}{{104\ \times ( {22 - 6} )}}} \$

and plotted it on Figure 4 for

m a s k t w = 1 , … , 16 $masktw = 1, \ldots , 16$

. M4 (Base_Mask_cmi_cci) indicated the lowest RMSE across all four models tested. Additionally, we can see that the predictions are most accurate when the time window of accumulated mask effect ranges between 3 and 6 weeks. This observation is qualitatively consistent with the temporal dynamics of the effect identified using econometric model results (Table 2 and Figure 2).

OPEN IN VIEWER

FIGURE 4

Root mean square error (RMSE) comparison of epidemiologic models M1–M4.

Last, we compare the estimated magnitude of the effect of visual nudges between epidemiologic and econometric models (in the case of both models, we observe the negative direction of the effect). That is, we compare the average of

β 2 i t M 4 $\beta _{2it}^{M4}$

estimated using epidemiologic model M4 (= 0.0096) to the average of statistically significant

β ( p < 0.05 ) $\beta \ (p < 0.05)$

estimates of Equation (1) (=−0.190) presented in Table 2. Given the difference in the units of the outcome variables, we back‐transform β as exp(−0.190)/100 = 0.0082. This result provides suggestive evidence of the general consistency in the observed magnitudes of the mask effect across econometric and epidemiological models.

Temporal analysis of estimated coefficients

It shall be noted that in our models that account for the mask effect (models M2 and M4), we use an additive approach to estimate the base infectivity. That is, we take

β 1 i t M 1 $\beta _{1it}^{M1}$

and construct it as a summation of

β 1 i t M 2 $\beta _{1it}^{M2}$

(base infectivity) and

β 2 i t M 2 $\beta _{2it}^{M2}$

that represents the response of university i at time t to changes in

I G _ M a s k ¯ i t ${\overline {IG\_Mask} }_{it}$

. As a result, when

I G _ M a s k ¯ i t ${\overline {IG\_Mask} }_{it}$

approaches 1, the base infectivity increases, while when

I G _ M a s k ¯ i t ${\overline {IG\_Mask} }_{it}$

approaches 0, the base infectivity decreases. While Model 2 accounts for on‐campus infectivity only, in Model 4, we include

β 3 i t M 4 $\beta _{3it}^{M4}$

and

β 4 i t M 4 $\beta _{4it}^{M4}$

that represent off‐campus effects associated with mobility and social distancing, respectively. Figure 5 presents the changes in infectivity associated with its different constituents.

OPEN IN VIEWER

FIGURE 5

25%–75% quantiles of beta coefficients over schools per week.

First, the upper‐left graph represents the base infectivity

β 1 i t M 4 $\beta _{1it}^{M4}$

, which as expected remains largely constant over time. Next, the upper‐right graph shows that the infectivity component driven by the intensity of shared mask images,

β 2 i t M 4 $\beta _{2it}^{M4}$

, also remains constant during the active part of the semester. However, it starts to gradually decline around the time of Thanksgiving break, after which the universities in our sample went to an online mode of instruction and most students left campus. This observation could also imply that the intensity of mask images has a less prominent effect as the pandemic grows more mature and the campus population becomes more accustomed and used to visual nudges, perceiving them as less relevant. The observed nature of the declining mask effect suggests a potential existence of a structural break, that is, an unexpected change over time in the parameter (

β 2 i t M 4 $\beta _{2it}^{M4}$

) of our model. Therefore, to detect and formally test for the presence of a structural break, we use a heteroskedasticity‐ and autocorrelation‐consistent estimator implemented in Stata's—xtbreak—module. Our test results reject the null hypothesis of no breaks

s u p W τ = 82.45 , p < 0.001 $sup\ W_\tau = 82.45, p < 0.001$

and suggest an estimated break on November 29, 2020, a week corresponding to the Thanksgiving holidays. This result confirms our expectations about the potential reason behind the reduced effect of mask images driven by the fact that at least a portion of the university population left campus and the remaining part of the semester was held in on‐line mode only. This finding also reinforces our argument behind the temporal dynamics of the informational value of visual nudges. Last, both mobility and social distancing effects, that is,

β 3 i t M 4 $\beta _{3it}^{M4}$

and

β 4 i t M 4 $\beta _{4it}^{M4}$

, reflect the relative consistency of underlying infectivity over time with minor variations attributable to seasonal population traveling and interaction.

To conclude, using the SIR model, we were able to alleviate the omitted variable bias by taking into account the COVID‐19 infectivity originating at different levels of the population. To this end, our comprehensive results provide evidence of the relative consistency in the magnitude, direction, and temporal dynamics of the IG mask effect between epidemiologic and econometric models.

DISCUSSION

Our work aims to show how innovative data analytics research can help promote public health. While nudge theory is a well‐established theoretical framework in public health research, we present a novel mechanism for applying nudge theory in decentralized, informal communication channels directly by community leaders such as individual organizations, instead of the top‐down approach practiced by government entities. This new approach to nudge theory, leveraging information and communication technologies, makes several theoretical and practical contributions to the literature and to health practitioners.

First, our findings contribute to prior literature by elucidating the informational value of visual nudges to promote safe behaviors in the context of mass health crises such as the global COVID‐19 pandemic. Nudge theory's effectiveness in the study of public health phenomena has been demonstrated in the areas of patient acceptance of AI‐augmented healthcare delivery systems (Dai & Tayur, 2022), dietary behavior (Arno & Thomas, 2016), cardiometabolic health (Lakerveld et al., 2018), and diabetes management (Kwan et al., 2020), among others. With the advent of the COVID‐19 pandemic, scientists have turned to nudge theory once again due to its decidedly noncoercive yet compelling effect on shaping public behavior when official policies fall short (Einfeld, 2019). Debnath and Bardhan (2020), for instance, show how the Indian government used nudge‐like text messaging strategies to promote compliance with public guidelines such as mask‐wearing and social distancing. Similarly, Pierre (2020) attributes the feasible handling of the pandemic in Sweden, a country that kept its economy and social structures largely open yet kept the number of positive cases under control, to the clever use of nudging strategies by the national public health infrastructure; in particular, complementing “hard” coordination, which includes policy levers such as budget allocations and official guidelines, with “soft” coordination nudges, or the daily routine exchange of information, has been credited with much of the nation's pandemic control success. Our study goes a step further by demonstrating the added value of using nudge techniques not just at the official government level but also in a diffused fashion by decentralized economic and social actors such as individual organizations. Specifically, our results indicate that in times of national and international crises such as the COVID pandemic, communicating information about safe behaviors through central channels like the CDC may not be sufficient to ensure the adoption of such behaviors by the public. The power distance between the public and such central agencies may be too great to sustain consistent compliance, but when these vital messages are disseminated through more proximal channels such as one's university, and in a more evocative way, through visual nudges, there is substantial informational value as evidenced by the significance of our results.

Second, we contribute to the nudge theory by introducing the concept of temporal dynamics in the application of visual nudges, which enables us to consider the time it might take from initial infection and incubation period to symptom onset and subsequent testing result to model the effect of visual nudges dynamically. The dynamic impact of nudges, and specifically the effect of prior nudges on future behavior, which goes beyond the limited framework of short‐term decisions to capture long‐term conditioning, has been studied in the realm of economic theory and altruistic behavior in the context of charity (Loomba et al., 2021). Through a series of behavioral experiments and empirical analyses, the authors find that exposure to nudges in the past does indeed inform future behavior, thus demonstrating the learned nature of nudging influence and its internalization potential (Loomba et al., 2021). The results of our study reinforce this notion by showing that signals received ahead of time, from 2 to 5 weeks as our results suggest, have a considerable effect on behavioral outcomes at time t. However, the effect starts to dampen considerably if signaled more than 6 weeks ahead of time. This finding highlights the challenge of “timeliness of information spread during a disease outbreak” to coordinate a speedy intervention (Gour et al., 2022) and further demonstrates the value of combing standard epidemiological models with heterogenous data‐driven modeling approaches (Evgeniou et al., 2023).

Third, our findings have significant implications for practitioners. To facilitate comprehension of our contributions to policymakers and public health officials, we conduct a counterfactual analysis. To do so, we use an epidemiological model (M4) to forecast the number of new cases per 1000 university population under three counterfactuals of mask images: (1) the observed intensity of images displayed,

M a s k ¯ i t ${\overline {Mask} }_{it}$

; (2) a high intensity of mask images,

0.2 × M a s k ¯ i t $0.2 \times \ {\overline {Mask} }_{it}$

; and (3) a low intensity of mask images,

min [ 1 , 1.8 × M a s k ¯ i t ] $\min [ {1,1.8 \times \ {{\overline {Mask} }}_{it}} ]$

. Figure 6 presents the results of our counterfactual analysis for schools with a lower and higher average of cumulative masked images.

OPEN IN VIEWER

FIGURE 6

Counterfactual analysis.

We can see that for both scenarios, the predicted number of new cases is lower when universities use a higher intensity of signaling images containing masks. The direction of the effect and its magnitude (ranging from as low as 3% to as high as 15%) are generally consistent with those observed in the econometric and epidemiologic models. Furthermore, we find evidence of the temporal heterogeneity; that is, the mask effect is considerable during the active phase of the semester but declines during holidays and breaks. These findings can be of interest to policymakers and public health officials already using forms of visual nudges in their communication strategies by presenting a viable approach for maximizing the effect of their current and future campaigns. We also find that the effectiveness of visual nudges unfolds in a cumulative fashion, so practitioners may benefit from the intense sharing of content associated with mask‐wearing over a prolonged period of 3 to 5 weeks. Additionally, our findings provide insights for those actors in the information environment that have not yet adapted social media engagement in their crisis management practices. For example, we show such actors that their participation in the crisis information effort is likely to make a difference and also provide guidance on how to initiate such engagement by means of visual nudges most effectively.

Fourth, the results of our analysis revealed that the effect of visual nudges is not uniform across different social media platforms. The “boundary condition” of the visual nudge effect could arise from the following conditions. First, as our data show, there are considerable irregularity in posting visual content including that containing masks on different platforms. This observation is especially notable in the time preceding and around the beginning of the fall 2020 semester, when most of the mask‐related content was consistently shared on IG but not on other platforms. For example, while the average number of weekly photos was five on IG, six on TW, and three on FB, the corresponding share of masked images was 21%, 16%, and 11%, respectively. Second, when examining the visual nudge effect, one should account for the platform differences in terms of the modality of communication. That is, while IG emphasizes visual image sharing, TW and FB emphasize mostly text‐based sharing that can be complemented by visuals. Therefore, it is feasible to expect content shared through IG to activate university stakeholders’ visual information processing and interpretation to a greater extent compared to other platforms given IG's superior aesthetic and visual culture of imagery. Third, as with most social media engagement strategies, it is crucial to consider the compositional differences of the stakeholder audiences that follow the corresponding universities’ social media accounts. To this end, comparing FB, TW, and IG, the largest age group on FB between 2020 and 2022 is that of users between the ages of 25 and 34 (31.5%); IG reports a similar volume of audiences from that age group (31.2%), and a close second is the group between the ages of 18 and 24 (31%); and the largest age group on TW (42%) comprises users between the ages of 18 and 29. ²⁸ Although seemingly similar, these statistics show that unlike the other two platforms, IG maintains a firm hold on Gen Z and millennial audiences who jointly constitute two‐thirds of its users. This is a key differentiating characteristic in the context of university stakeholders since the most populous stakeholder group is that of current students whose age range overlaps with the majority of IG's following.

Finally, there are other contexts at the macro‐ and micro‐levels where visual nudges can be used for public health applications. For instance, our findings can be used in the context of public COVID‐19 vaccination attitudes, especially amid reports of rising reluctance to receive the vaccine among certain social groups and the worrisome spread of COVID misinformation (Loomba et al., 2021). Numerous government officials and public figures such as the Dalai Lama, the US President and Vice President, and all nine Supreme Court justices announced their respective vaccinations, often through visual social media posts, ²⁹ and research shows that this type of nudge is often preferable to traditional regulation levers, which may result in costly, ineffective, and invasive campaigning (Hansen et al., 2016). The effect of the university's social media nudging on COVID‐19 infection rates found by our study could, at least in theory, be related to localized empirical evidence from health communication studies in the COVID‐19 context. For example, a case study conducted by the city of San Jose, California, in collaboration with the Knight Foundation found that the use of “trusted messengers” (or “micro‐influencers”) at the local community level, rather than global celebrities or spokespeople, was associated with increased COIVD‐19 vaccination rates. ³⁰ Unlike the high‐profile social influence campaigns public figures may engage in, the type of visual nudge explored in this paper represents a particular member of the family of classic nudges (discussed in Supporting Information S2) known as a soft nudge (Thaler & Sunstein, 2009). In the context of the dual process theory, which distinguishes between the automatic and reflective cognitive systems employed by decision‐makers, classic nudges differ from other forms of behavioral interventions such as social advertising in that they seek to engage the automatic system responsible for quick and intuitive decisions (Bovens, 2009). In this automatic system framework, soft nudges seek to alter choices without constraining the decision‐maker's opportunity sets; they do not provide material or monetary incentives for changing behavior, nor do they threaten negative consequences if the suggested behavior is not adopted, but rather they help individuals consider aspects of the decision‐making process they would not otherwise have considered (Ehrig et al., 2015). Thus, soft visual nudges constitute images simply depicting desirable public health behaviors such as mask‐wearing, without explicitly stating the value or, alternatively, the cost of (not) adopting this behavior, through text or other communicative strategies. These characteristics make soft visual nudges effective without posing an ethical threat to the communities and individuals exposed to them and are therefore safe and noninvasive. Such nudges can be used in addition to larger‐scale social welfare interventions aimed at changing social norms. At the same time, our findings indicate that in certain cases, soft visual nudges have the potential to cover ground in individual decision‐making left unaccounted for by large‐scale social norm campaigns because of their noninvasive character restricted to invoking the automatic cognitive system only.

CONCLUSION

Using a theoretically informed framework rooted in nudge theory in combination with novel deep learning techniques, we developed a robust econometric model seeking to illuminate the role of organizational social media strategy in mitigating the spread of COVID‐19. We found that using social media posts as a form of visual nudging to promote safe behaviors has a marked effect on COVID‐19 positivity rates, even after controlling for a variety of organizational characteristics and disease dynamics at different levels. We further show that the effect of nudges may not be immediately detectable, as it may take time to be processed and internalized. Specifically, our results suggest that the impact of visual nudges peaks between 4 to 6 weeks after initial exposure. Additionally, we demonstrated the marginal utility of nudging strategies depending on the level of virus spread across organizations, with the most heavily impacted organizations benefiting the most from the use of visual nudges. We believe that our model can be fruitfully adopted not just by medical practitioners but also by large‐scale organizations more broadly in ways that creatively use a channel of communication already in place to overcome the challenges posed by COVID‐19.

Our study is not without limitations. First, we acknowledge that in practice, it is not feasible to definitively say how many students, faculty members, or staff members follow a university's social media account since not all of them will do so. Second, while we address self‐selection bias originating at the university level, such bias can also be present at the individual level in instances when university stakeholders may choose to follow its social media and engage with it or not. Finally, it shall be noted that the inherent volatility of user preferences and engagement leads to constant shifts in platform popularity such that, although the platforms examined in this study have undoubtedly been the preferred online media communication channels for universities and their stakeholders during the analyzed timeframe, new social media platforms (e.g., TikTok, BeReal, VSCO) with a different set of affordances appear on the market and could change the observed dynamics of the nudges.

Footnotes

ACKNOWLEDGMENTS

This project has been funded by the Jump ARCHES endowment through the Health Care Engineering Systems Center.

1

https://www.statista.com/statistics/272014/global‐social‐networks‐ranked‐by‐number‐of‐users/.

2

https://www.pewresearch.org/internet/2021/04/07/social‐media‐use‐in‐2021/.

3

https://www.statista.com/statistics/272014/global‐social‐networks‐ranked‐by‐number‐of‐users/.

4

https://www.pewresearch.org/internet/2018/05/31/teens‐social‐media‐technology‐2018/.

5

https://www.statista.com/statistics/246221/share‐of‐us‐internet‐users‐who‐use‐facebook‐by‐age‐group/.

6

https://www.statista.com/statistics/246199/share‐of‐us‐internet‐users‐who‐use‐instagram‐by‐age‐group/.

7

https://www.pewresearch.org/internet/2018/03/01/social‐media‐use‐in‐2018/.

8

In case one of the variables was not provided but the total positivity rate was reported, we used the reported positivity rate to calculate the missing parameter. We focused on the data that represent aggregated student, faculty, and staff measurements. If combined totals were not available, we collected student‐level data (< 8% of entities). Finally, we primarily focused on proactive surveillance/asymptomatic testing.

9

In most cases, entities updated their data on Sunday; however, about 20% did so on Monday or Tuesday. The results of our data collection initiative were publicly available via a dashboard published online from September 7, 2020, to June 1, 2022.

10

https://www.cdc.gov/museum/timeline/covid19.html.

11

https://instalooter.readthedocs.io/en/latest/.

12

https://github.com/JustAnotherArchivist/snscrape.

13

A verified badge is a check that appears next to an IG account's name in searches and on the profile. It means that IG has confirmed that an account is the authentic presence of the public figure, celebrity, or brand it represents. IG verification prevents impersonation. This is why verified accounts are exclusive for celebrities, public figures, or global brands. Being verified means that a company is a “big deal” because it is important enough to be at risk of impersonation. Other social media platforms such as TW or LinkedIn use similar features. University account verification is voluntary and largely based on the account managers’ strategic choices behind social media engagement practices.

14

When it comes to COVID‐19 testing policies, in general, universities (and other entities) adopt a testing strategy that implies (i) testing everyone, (ii) testing selected groups, or (iii) a combination of both. In the case of selected group testing, the groups to be tested are selected based on individual entities’ COVID‐19 response strategies and can include but are not limited to undergraduate students (due to the higher risk associated with their behavior), members of sports teams (due to the higher risk associated with potential exposure to others when travelling), staff (due to lower vaccination rates), unvaccinated population, and so forth.

15

See, for example, reports regarding disciplinary actions taken in response to violations of COVID‐19 regulations on college campuses such as: Dennon, Anne (October 2020) “College Students Suspended, Fined for Breaking COVID‐19 Rules” on BestColleges.com (https://www.bestcolleges.com/blog/college‐students‐breaking‐covid‐19‐rules‐suspended‐fined/), and Albert, Victoria (August 2020) “36 Purdue Students Suspended after Off‐campus Party” on CBS News (https://www.cbsnews.com/news/purdue‐university‐36‐students‐suspended‐over‐co‐op‐party‐at‐circle‐pines/).

16

https://github.com/nytimes/covid‐19‐data/blob/master/mask‐use/README.md.

17

https://www.dynata.com/.

18

https://www.cuebiq.com/visitation‐insights‐covid19.

19

https://github.com/tonmcg/US_County_Level_Election_Results_08‐20.

20

https://github.com/meteostat/meteostat‐python.

21

https://gist.github.com/philngo/d3e251040569dba67942.

22

https://www.cdc.gov/coronavirus/2019‐ncov/symptoms‐testing/symptoms.html.

23

https://www.pfizer.com/news/articles/14‐Day‐Quarantine‐Incubation‐Period.

24

To compute the median, we arranged all of the scores in numerical order and then calculated the average between the two middle values given the even number of total observations.

25

We used the median to impute the missing values and ensure consistency of the estimates; no bias was observed using all available observations. Instead, the observed effects were more pronounced.

26

https://www.cdc.gov/coronavirus/2019‐ncov/hcp/duration‐isolation.html.

27

Similar to the case with CEM model estimation, the sample size has reduced because of the missing values that preclude the current model estimation setup. We used the median to impute the missing values and ensure consistency of the estimates; no bias was observed using all available observations. Instead, the observed effects were more pronounced.

28

https://www.statista.com/markets/424/topic/540/social‐media‐user‐generated‐content/#overview.

29

https://people.com/health/celebrities‐and‐politicians‐get‐covid‐19‐vaccine‐photos/.

30

https://datasmart.ash.harvard.edu/power‐%E2%80%98micro‐influencers%E2%80%99‐fight‐against‐covid‐19.

ORCID

Anton Ivanov

Zhasmina Tacheva

Sebastian Souyris

References

1.

Abaluck J. Kwong L. H. Styczynski A. Haque A. Kabir M. A. Bates‐Jefferys E. Crawford E. Benjamin‐Chung J. Raihan S. Rahman S. Benhachmi S. Bintee N. Z. Winch P. J. Hossain M. Reza H. M. Jaber A. A. Momen S. G. Rahman A. Banti F. L. , … Mobarak A. M. (2022). Impact of community masking on COVID‐19: A cluster‐randomized trial in Bangladesh. Science, 375(6577), eabi9069. https://doi.org/10.1126/science.abi9069

Go to Reference

2.

Arno A. Thomas S. (2016). The efficacy of nudge theory strategies in influencing adult dietary behaviour: A systematic review and meta‐analysis. BMC Public Health, 16(1), 1–11. https://doi.org/10.1186/s12889‐016‐3272‐x

Go to Reference

3.

Bhuiyan M. M. Horning M. Lee S. W. Mitra T. (2021). Nudgecred: supporting news credibility assessment on social media through nudges. Proceedings of the ACM on Human‐Computer Interaction, 5(CSCW2), (pp. 1–30).

+ Show References

4.

Bovens L. (2009). The ethics of nudge. In Grüne‐Yanoff T. Hansson S. O. (Eds.), Preference change (pp. 207–219). Springer.

Go to Reference

5.

Camargo K. Jr. Grant R. (2015). Public health, science, and policy debate: Being right is not enough. American Journal of Public Health, 105(2), 232–235. https://doi.org/10.2105/AJPH.2014.302241

Go to Reference

6.

Charry K. Tessitore T. (2021). I tweet, they follow, you eat: Number of followers as nudge on social media to eat more healthily. Social Science & Medicine, 269, 113595.

+ Show References

7.

Chou W. ‐Y. S. Gaysynsky A. Vanderpool R. C. (2021). The COVID‐19 misinfodemic: Moving beyond fact‐checking. Health Education & Behavior, 48(1), 9–13.

Go to Reference

8.

Cui R. Gallino S. Moreno A. Zhang D. J. (2018). The operational value of social media information. Production and Operations Management, 27(10), 1749–1769. https://doi.org/10.1111/poms.12707

Go to Reference

9.

Dai T. Tayur S. (2022). Designing AI‐augmented healthcare delivery systems for physician buy‐in and patient acceptance. Production and Operations Management, 31(12), 4443–4451. https://doi.org/10.1111/poms.13850

+ Show References

10.

Dayaratna K. Tyrrell P. Vanderplas A. (2020). A comparative analysis of policy approaches to COVID‐19 around the world, with recommendations for US lawmakers. The Heritage Foundation, 234, 1–57.

Go to Reference

11.

Debnath R. Bardhan R. (2020). India nudges to contain COVID‐19 pandemic: A reactive public policy analysis using machine‐learning based topic modelling. PloS One, 15(9), e0238972. https://doi.org/10.1371/journal.pone.0238972

Go to Reference

12.

Ehrig T. Katsikopoulos K. V. Kuorikoski J. Poyhonen S. Sunder S. (2015). Limitations of behaviorally informed policy under social interaction. Available at SSRN 2685912.

Go to Reference

13.

Einfeld C. (2019). Nudge and evidence based policy: Fertile ground. Evidence & Policy: A Journal of Research, Debate and Practice, 15(4), 509–524.

Go to Reference

14.

Evgeniou T. Fekom M. Ovchinnikov A. Porcher R. Pouchol C. Vayatis N. (2023). Pandemic lockdown, isolation, and exit policies based on machine learning predictions. Production and Operations Management. Advance online publication. 32(5), 1307–1322 https://doi.org/10.1111/poms.13726

Go to Reference

15.

Friedman J. Hastie T. Tibshirani R. (2010). Regularization paths for generalized linear models via coordinate descent. Journal of Statistical Software, 33(1), 1–22. https://doi.org/10.18637/jss.v033.i01

Go to Reference

16.

Ganslmeier M. Furceri D. Ostry J. D. (2021). The impact of weather on COVID‐19 pandemic. Scientific Reports, 11(1), 1–7. https://doi.org/10.1038/s41598‐021‐01189‐3

Go to Reference

17.

Gnanvi J. E. Salako K. V. Kotanmi G. B. Kakaï R. G. (2021). On the reliability of predictions on COVID‐19 dynamics: A systematic and critical review of modelling techniques. Infectious Disease Modelling, 6, 258–272. https://doi.org/10.1016/j.idm.2020.12.008

Go to Reference

18.

Gong X. Lane K. M. D. (2020). Institutional Twitter usage among US geography departments. The Professional Geographer, 72(2), 219–237. https://doi.org/10.1080/00330124.2019.1653770

+ Show References

19.

Gour A. Aggarwal S. Kumar S. (2022). Lending ears to unheard voices: An empirical analysis of user‐generated content on social media. Production and Operations Management, 31(6), 2457–2476. https://doi.org/10.1111/poms.13732

+ Show References

20.

Gurbaxani B. M. Hill A. N. Paul P. Prasad P. V. Slayton R. B. (2021). Evaluation of different types of face masks to limit the spread of SARS‐CoV‐2: A modeling study. Scientific Reports, 12(1), 8630. https://doi.org/10.1038/s41598‐022‐11934‐x

Go to Reference

21.

Hanks A. S. Just D. R. Wansink B. (2013). Smarter lunchrooms can address new school lunchroom guidelines and childhood obesity. The Journal of Pediatrics, 162(4), 867–869. https://doi.org/10.1016/j.jpeds.2012.12.031

Go to Reference

22.

Hansen P. G. Skov L. R. Skov K. L. (2016). Making healthy choices easier: Regulation versus nudging. Annual Review of Public Health, 37, 237–251. https://doi.org/10.1146/annurev‐publhealth‐032315‐021537

Go to Reference

23.

Huh Y. E. Vosgerau J. Morewedge C. K. (2014). Social defaults: Observed choices become choice defaults. Journal of Consumer Research, 41(3), 746–760. https://doi.org/10.1086/677315

Go to Reference

24.

Iacus S. M. King G. Porro G. (2012). Causal inference without balance checking: Coarsened exact matching. Political Analysis, 20(1), 1–24. https://doi.org/10.1093/pan/mpr013

+ Show References

25.

Jung M. Cho D. Shin E. (2021). Repairing a cracked mirror: The heterogeneous effect of personalized digital nudges driven by misperception. Production and Operations Management, 30(8), 2586–2607. https://doi.org/10.1111/poms.13396

Go to Reference

26.

Kahane L. H. (2021). Politicizing the mask: Political, economic and demographic factors affecting mask wearing behavior in the USA. Eastern Economic Journal, 47(2), 163–183. https://doi.org/10.1057/s41302‐020‐00186‐0

Go to Reference

27.

Kahneman D. (2011). Thinking, fast and slow. Macmillan.

+ Show References

28.

Kermack W. O. McKendrick A. G. (1927). A contribution to the mathematical theory of epidemics. Proceedings of the Royal Society of London. Series A, Containing Papers of a Mathematical and Physical Character, 115(772), 700–721.

+ Show References

29.

Kraak V. Englund T. Misyak S. Serrano E. (2017). A novel marketing mix and choice architecture framework to nudge restaurant customers toward healthy food environments to reduce obesity in the United States. Obesity Reviews, 18(8), 852–868. https://doi.org/10.1111/obr.12553

Go to Reference

30.

Krawiec J. Piaskowska O. Piesiewicz P. Białaszek W. (2021). Tools for public health policy: Nudges and boosts as active support of the law in special situations such as the COVID‐19 pandemic. Global Health, 17(1), 132. https://doi.org/10.1186/s12992‐021‐00782‐5

Go to Reference

31.

Kwan Y. H. Cheng T. Y. Yoon S. Ho L. Y. ‐C. Huang C. Chew E. H. Thumboo J. Ostbye T. Low L. L. (2020). A systematic review of nudge theories and strategies used to influence adult health behaviour and outcome in diabetes management. Diabetes & Metabolism, 46(6), 450–460.

Go to Reference

32.

Lakerveld J. Mackenbach J. D. De Boer F. Brandhorst B. Broerse J. E. deBruijn G. ‐J. Feunekes G. Gillebaart M. Harbers M. Hoenink J. (2018). Improving cardiometabolic health through nudging dietary behaviours and physical activity in low SES adults: Design of the Supreme Nudge project. BMC Public Health, 18(1), 1–9. https://doi.org/10.1186/s12889‐018‐5839‐1

Go to Reference

33.

Lauer S. A. Grantz K. H. Bi Q. Jones F. K. Zheng Q. Meredith H. R. Azman A. S. Reich N. G. Lessler J. (2020). The incubation period of coronavirus disease 2019 (COVID‐19) from publicly reported confirmed cases: Estimation and application. Annals of Internal Medicine, 172(9), 577–582. https://doi.org/10.7326/M20‐0504

Go to Reference

34.

Lee S. Y. Qiu L. Whinston A. (2018). Sentiment manipulation in online platforms: An analysis of movie tweets. Production and Operations Management, 27(3), 393–416. https://doi.org/10.1111/poms.12805

Go to Reference

35.

Loomba S. deFigueiredo A. Piatek S. J. deGraaf K. Larson H. J. (2021). Measuring the impact of COVID‐19 vaccine misinformation on vaccination intent in the UK and USA. Nature Human Behaviour, 5, 337–348.

+ Show References

36.

Luttrell R. McGrath K. Medjesky C. A. (2017). Purpose, value, and unintended consequences of social media in higher education. In Powers K. Schloss P. J. (Eds.), Organization and administration in higher education (pp. 241–258). Routledge.

Go to Reference

37.

McKee R. (2013). Ethical issues in using social media for health and health care research. Health Policy, 110(2‐3), 298–301. https://doi.org/10.1016/j.healthpol.2013.02.006

Go to Reference

38.

Mirbabaie M. Ehnis C. Stieglitz S. Bunker D. Rose T. (2021). Digital nudging in social media disaster communication. Information Systems Frontiers, 23, 1097–1113. https://doi.org/10.1007/s10796‐020‐10062‐z

+ Show References

39.

Misawa D. Fukuyoshi J. Sengoku S. (2020). Cancer prevention using machine learning, nudge theory and social impact bond. International Journal of Environmental Research and Public Health, 17(3), 790. https://doi.org/10.3390/ijerph17030790

Go to Reference

40.

Moya C. (2020). Dynamics of behavior change in the COVID world. American Journal of Human Biology, 32(5), e23485. https://doi.org/10.1002/ajhb.23485

Go to Reference

41.

Mukherjee U. K. Bose S. Ivanov A. Souyris S. Seshadri S. Sridhar P. Watkins R. Xu Y. (2021). Evaluation of reopening strategies for educational institutions during COVID‐19 through agent based simulation. Scientific Reports, 11(1), 1–24.

+ Show References

42.

Nekmat E. (2020). Nudge effect of fact‐check alerts: Source influence and media skepticism on sharing of news misinformation in social media. Social Media + Society, 6(1).

+ Show References

43.

Ng K. C. Ke P. F. So M. K. Tam K. Y. (2023). Augmenting fake content detection in online platforms: A domain adaptive transfer learning via adversarial training approach. Production and Operations Management. Advance online publication. https://doi.org/10.1111/poms.13959

Go to Reference

44.

Pennycook G. McPhetres J. Zhang Y. Lu J. G. Rand D. G. (2020). Fighting COVID‐19 misinformation on social media: Experimental evidence for a scalable accuracy‐nudge intervention. Psychological Science, 31(7), 770–780. https://doi.org/10.1177/0956797620939054

Go to Reference

45.

Pierre J. (2020). Nudges against pandemics: Sweden's COVID‐19 containment strategy in perspective. Policy and Society, 39(3), 478–493.

Go to Reference

46.

Purohit A. K. Barclay L. Holzer A. (2020). Designing for digital detox: Making social media less addictive with digital nudges. Extended Abstracts of the 2020 CHI Conference on Human Factors in Computing Systems , Honolulu, HI (pp. 1–9).

+ Show References

47.

Purohit A. K. Holzer A. (2021). Unhooked by design: Scrolling mindfully on social media by automating digital nudges. Twenty‐Seventh Americas Conference on Information Systems, Montreal, Canada (pp. 1–10).

+ Show References

48.

Rainford P. Tinkler J. (2011). Designing for nudge effects: How behaviour management can ease public sector problems. LSE Public Policy Group.

Go to Reference

49.

Roozenbeek J. Freeman A. L. J. van derLinden S. (2021). How accurate are accuracy‐nudge interventions? A preregistered direct replication of Pennycook et al. (2020). Psychological Science, 32(7), 1169–1178. https://doi.org/10.1177/09567976211024535

Go to Reference

50.

Saifee D. H. Zheng Z. Bardhan I. R. Lahiri A. (2020). Are online reviews of physicians reliable indicators of clinical outcomes? A focus on chronic disease management. Information Systems Research, 31(4), 1282–1300. https://doi.org/10.1287/isre.2020.0945

Go to Reference

51.

Salathé M. Freifeld C. C. Mekaru S. R. Tomasulo A. F. Brownstein J. S. (2013). Influenza A (H7N9) and the importance of digital epidemiology. The New England Journal of Medicine, 369(5), 401. https://doi.org/10.1056/NEJMp1307752

Go to Reference

52.

Sasaki S. Kurokawa H. Ohtake F. (2020). Short‐term responses to nudge‐based messages for preventing the spread of COVID‐19 infection: Intention, behavior, and life satisfaction. Graduate School of Economics and Osaka School of International Public Policy (OSIPP) Osaka University Discussion Papers In Economics And Business, 20, 1–31.

+ Show References

53.

Serences J. T. Yantis S. (2006). Selective visual attention and perceptual coherence. Trends in Cognitive Sciences, 10(1), 38–45. https://doi.org/10.1016/j.tics.2005.11.008

Go to Reference

54.

Shane‐Simpson C. Manago A. Gaggi N. Gillespie‐Lynch K. (2018). Why do college students prefer Facebook, Twitter, or Instagram? Site affordances, tensions between privacy and self‐expression, and implications for social capital. Computers in Human Behavior, 86, 276–288. https://doi.org/10.1016/j.chb.2018.04.041

Go to Reference

55.

Srikoo I. R. P. Chung D. (2021). Descriptive analysis of social media adoption and utilization in AACSB schools. Issues in Information Systems, 22(4), 126–142.

+ Show References

56.

Louis C. St Zorlu G. (2012). Can Twitter predict disease outbreaks? BMJ, 344, e2353. https://doi.org/10.1136/bmj.e2353

Go to Reference

57.

Tang C. S. (2022). Innovative technology and operations for alleviating poverty through women's economic empowerment. Production and Operations Management, 31(1), 32–45. https://doi.org/10.1111/poms.13349

Go to Reference

58.

Thaler R. H. Sunstein C. R. (2009). Nudge: Improving decisions about health, wealth, and happiness. Penguin.

+ Show References

59.

Van Gestel L. Adriaanse M. De Ridder D. (2020). Do nudges make use of automatic processing? Unraveling the effects of a default nudge under type 1 and type 2 processing. Comprehensive Results in Social Psychology, 5(1‐3), 4–24.

+ Show References

60.

Wang Y. Leon P. G. Scott K. Chen X. Acquisti A. Cranor L. F. (2013). Privacy nudges for social media: An exploratory Facebook study. Proceedings of the 22nd International Conference on World Wide Web , Rio de Janeiro, Brazil (pp. 763–770).

+ Show References

61.

Wooldridge J. M. (2012). Introductory econometrics: A modern approach. Cengage Learning.

+ Show References

62.

Yan L. Pedraza‐Martinez A. J. (2019). Social media for disaster management: Operational value of the social conversation. Production and Operations Management, 28(10), 2514–2532. https://doi.org/10.1111/poms.13064

Go to Reference

63.

Young S. D. (2015). A “big data” approach to HIV epidemiology and prevention. Preventive Medicine, 70, 17–18. https://doi.org/10.1016/j.ypmed.2014.11.002

Go to Reference

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.36 MB