On spatial variation in the detectability and density of social media user protest supporters

Masías, Víctor Hugo et al. Article — Manuscript Version (Preprint) On spatial variation in the detectability and density of social media user protest supporters Telematics and Informatics Provided in Cooperation with: WZB Berlin Social Science Center Suggested Citation: Masías, Víctor Hugo et al. (2021) : On spatial variation in the detectability and density of social media user protest supporters, Telematics and Informatics, ISSN 1879-324X, Elsevier, Amsterdam, Vol. 65, pp. --, https://doi.org/10.1016/j.tele.2021.101730 This Version is available at: https://hdl.handle.net/10419/248284 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/4.0/ ON SPATIAL VARIATION IN THE DETECTABILITY AND DENSITY OF SOCIAL MEDIA USER PROTEST SUPPORTERS A PREPRINT Víctor H. Masías ∗1 , Fernando Crespo 2 , Pilar Navarro R. 3 , Razan Masood 1 , Nicole C. Krämer 4 , and H. Ulrich Hoppe 1 1Department of Computer Science and Applied Cognitive Science, University of Duisburg–Essen, Germany 2Facultad Tecnológica, Universidad de Santiago de Chile, Estación Central, Santiago, Chile 3E.T.S. of Computer and Telecommunication Engineering, University of Granada, Spain 4Social Psychology: Media and Communication, University of Duisburg–Essen, Germany ABSTRACT Although much has been published regarding street protests on social media, few works have attempted to characterize social media users’ spatial behavior in such events. The research reported here uses spatial capture-recapture methods to determine the influence of the built environment, physical proximity to protest location, and collective posting rhythm on variations in users’ spatial detectability and density during a protest in Mexico City. The best-obtained model, together with explaining the spatial density of users, shows that there is high variability in the detectability of social media user protest supporters and that the collective posting rhythm and the day of observation are significant explanatory factors. The implication is that studies of collective spatial behavior would benefit by focussing on users’ activity centres and their urban environment, rather than their physical proximity to the protest location, the latter being unable to adequately explain spatial variations in users’ detectability and density during the protest event. Keywords Protest Event ·Social Media ·Circadian Rhythms ·Spatio–Temporal Behavior ·Urban Environment 1 Introduction PROTEST EVENTS are a social phenomenon that contributes to processes of change in all political systems, whether democratic, non-democratic, or some hybrid of the two. In the last years, massive demonstration events held in the USA have generated renewed international attention from the scientific community on the phenomenon of street protest (Fisher et al.,2019). They typically take place in major urban centres and are inspired by a wide range of different motives. Recently, it has been found that there is a causal relationship between the moralization processes that occur in social media and the behavior of individuals in an offline environment (Mooijman et al.,2018). Little is known, however, about the spatial behavior of social media users who support street protests and how the urban environment of a city influences it. This paper focuses on examining in greater detail the spatial behavior of social media user protest supporters at the city level. Previous works have explored protest behavior in online environments and its correlations with the spatial dimension of offline protest (Chen and Pirolli,2012;Traag et al.,2017;Mooijman et al.,2018). Part of the contemporary research which studies protests in natural settings argues that the physical proximity to the protest location is an important explanatory of the spatial behavior of supporters during protest events. For example, it has been found that the place of residence, that is, if the social media user is frequently observed in the center on the periphery of the city is correlated with the level of support expressed in social media to a protest event (Chen and Pirolli, 2012;Barberá et al.,2015). Also, studies based on socio-physical models have proposed that physical distance to the protest location acts as an impedance to attendance (Traag et al.,2017). Other studies have proposed that ∗Corresponding author: [email protected]. arXiv:2103.06063v1 [cs.SI] 10 Mar 2021 On spatial variation in the detectability and density of social media user protest supporters the relationship between environmental characteristics and the distribution of individuals across space is more complex and dynamic, prompting researchers to investigate the relationship using simulation models (Davies et al.,2013;Lemos et al.,2016;Pires and Crooks,2017;Bacaksizlar,2019), whose results are, therefore, of unknown ecological validity. In this context, it is relatively unexplored whether existing elements of the urban environment influence the spatial behavior of social media user protest supporters. For this reason, we propose to conduct an observational interdisciplinary study that applies an ecological approach, known as spatial capture-recapture, to test a series of alternative hypotheses that theoretically have a better explanatory capacity in the context of social media research. Within this scenario, our research question is whether existing structures in the urban environment, such as street or subway networks, or whether the rate of social media posting in a given geographic area have significant ability to explain the detectability (i.e. the probability of detecting a user in a given place at a given occasion) and spatial density (i.e. the numbers of users divided by a spatial area) of the social media users supporting the protest. Thus, the objectives of this study can be summarized as follows: • Determine the factors that explain the variation in spatial detectability of social media user protest supporters. • Determine the factors that explain the variation in spatial density of social media user protest supporters. • Determine whether physical proximity to a protest location contributes to explaining the detectability and density of social media user protest supporters. • Compare different models and evaluate whether social media post rhythms in a given geographical region, transport network structures, the proximity of the users to the protest location, neighborhoodlevel socio-demographic variables, and the day of observation, contributes to explaining the detectability and density of social media user protest supporters. To determine these subjects, we will use multiple types of information that are generally difficult to use together under the same methodological framework. Our focus will be on protests at the city level to explain relationships between the spatial dimension and social media user behavior. More specifically, we examine how protests events are reflected in the spatial behavior of social media users who support a protest. 2 Conceptual-Analytical Framework Although the relationship between environmental elements and social media users’ spatial behavior has only recently attracted the attention of researchers, earlier work on the field of animal ecology had already thrown light on the impact of urbanization on species richness and diversity (McKinney,2008). Chronobiology research on cattle wandering the streets of India’s cities has found that their activity patterns (i.e. lying down, standing, walking, foraging) are correlated with environmental factors (Sahu et al.,2019). In the case of human beings, there is a complex interaction between land use and human activities that take place in the urban environment. Other studies have proposed that socio-ecological systems maintain reciprocal interactions between biophysical and socioeconomic structures (Arnaiz-Schmitz et al.,2018). In short, those studies suggest that there is a significant relationship between the environment and organisms’ spatial behavior. In this context, research in the existing literature suggests that the spatial behavior of individuals is influenced by factors related to both the individuals themselves and the geographic area they live in. In particular, numerous works have documented the way basic processes such as daily physical activity, cognitive performance, and locomotion are regulated by circadian rhythms (Valdez,2018). In evolutionary terms, it has been suggested that chrono-physiological processes control and facilitate the organization of spatial and temporal behavior in living beings. As an example, animal behavior studies indicate that wallabies follow circadian patterns in their search for food and shelter to avoid dangers in their environment (Fischer et al.,2019). Recently, in the field of social media research, it has been shown that daily posting rates of users can be used as a proxy for the endogenous circadian clock of individuals. For example, the daily use of social media has been used to examine how the rate of collective posting varies seasonally and geographically, and also to further demonstrate that social events and pressures disrupt patterns of social media activity, as occurs in the so-called “Twitter social jet lag” phenomenon (Leypunskiy et al.,2018). In another research more consistent with chronobiology studies, empirical evidence has been reported on the existence of the phenomenon of social synchronization in Facebook Messenger (Diwan et al.,2020). Generally speaking, research suggests that the behavior of posting social media content can be used as a proxy measure for the endogenous circadian clock of users of social networking sites (Murnane et al.,2015;Swain and Pati,2019). In light of the above, it seems plausible to consider that the daily 2 On spatial variation in the detectability and density of social media user protest supporters posting rates of social media content in a given geographical region can be a relevant factor that can help explain variations in the spatial behavior of social media user protest supporters. A parallel line of research has focused on investigating how the transport network configurations influence individuals’ spatial behavior. This research has found, for example, that street centrality is positively correlated with different types of land use (Rui and Ban,2014) and can be a good predictor of pedestrian flow (Bielik et al.,2018). Other authors have shown that street network configurations can explain serious outdoor violence (Summers and Johnson,2016). These findings imply that street centrality may affect individual spatial behavior, which in turn means it is reasonable to expect that the centrality of such spatial structures would also have the potential to explain variations, across a geographical region, of the spatial behavior of social media user protest supporters. Finally, neighborhood-level socio-demographic characteristics have been cited as having an impact on the spatial distribution of individuals over time and space. For example, a comparative study found that a set of geographical characteristics of the built environment consistently accounted for the greater part of the variability in population distribution in low- and medium-income countries (Nieves et al.,2017). Other studies have determined that a range of urban factors explain the density of mobile telephone activity within cities (Nadai et al.,2016;Liu et al., 2020). In general terms, it seems plausible that both neighborhood-level socio-demographic characteristics and the built environment have the potential to explain variations of the spatial behavior of social media user protest supporters. With the foregoing in mind and to enhance our understanding of the spatial behavior of social media user protest supporters, we designed an observational study based on the spatial capture-recapture approach (Royle et al., 2013) (see ‘Spatial capture-recapture analysis’ in the Methods section). More specifically, our approach attempted to determine whether and to what extent the physical proximity of the social media users (hereafter “SMUs” or simply “users”) to the protest location, socio-demographic characteristics, street network configurations, and the daily rhythms of social media posting in a given geographical region, and the day of observation, contribute in accounting for the variations in the detectability and density of SMUs at the city level. The event used as a real case study is the 40th annual Mexico City LGBT pride parade held on June 23, 2018. The organizers of the event proposed a march that has a distance of ≈ 4km. The main meeting place was planned to be in the Angel of Independence 2 at 10:00 a.m. From there, different groups marched to reach the so-called Zócalo 3 , where a concert was held during the afternoon, for concluding the whole event. This case is of particular interest because the parade is one of the largest mass gatherings held in the Mexican capital (Bosia et al.,2019). And unlike other types of collective behavior such as riots, the events organized by the Mexican LGBT movement are primarily of a collaborative, non-violent nature (Beer and Cruz-Aceves,2018).4 We used geotagged tweets to study the spatial behavior of users who supported this protest event. The geotagged social media data for the study were collected through the Twitter API, and the tweets were filtered based on geolocation information and constrained to Mexico City. To operationalize which SMUs supported the march, we coded 10,000 tweets by hand and used them to train a logistic regression classifier that then identified supportive tweets in our whole geotagged social media data sample (see ‘Identification of users supporting the protest event’ in the Methods section). To organize the data for the analysis of SMU’s spatial behavior during the march, we imposed a spatial grid over Mexico City in which each grid cell, hexagonal and covering 1.18km 2 , represented a trap where SMUs could be captured or recaptured in time and space (see Figure 1). For each such SMU capture or recapture, we indexed in a 3–dimensional array who ( i ), where ( j ), and when ( k ). Thus, yi,j,k = 1 denoted an individual user who was captured in a given cell on a given occasion, and yi,j,k = 0 indicated that the individual was not captured in that cell on that occasion. To identify each SMU ( i ), we use the unique identifier already assigned to each user by the Twitter API (the User ID field) and which was anonymized to maintain the user’s privacy. A unique identifier was also assigned to each grid cell (j) and each of the 24 hours in a day (k). Besides, three sets of proxies were used to serve as metrics of the urban environment factors and the SMU’s attributes hypothesized to influence the spatial detectability and density of SMU protest supporters. The first set of proxy measures corresponds to the concept of the physical proximity of the SMUs to the protest location (see 2 This is a frequent starting point for various demonstrations and not only for this case study (see, https://en. wikipedia.org/wiki/Angel_of_Independence). 3 This city square has an area of 57,600m 2 and is located in the downtown area of Mexico City (see, https://en. wikipedia.org/wiki/Zócalo). 4 This social movement has achieved in Mexico several successes in the political sphere such as the legalization of same-sex civil unions and LGBT adoptions, demonstrating its effectiveness in bringing about changes in human rights law. These changes have not only increased the acceptance of LGBT persons in Mexican society as a whole but have also improved their perspectives for embarking upon and developing careers in the sciences (Walker,2014). 3 On spatial variation in the detectability and density of social media user protest supporters (1) (0) (1) (0) (0) (1) (0) (1) (1) (0) (1) (0) (0) (1) (0) (1) (0) (1) (0) (1) (1) (0) (0) (0) (0) (1) (0) (1) (1) (0) (1) (1) (1) (0) (1) (0) (0) (1) (1) (1) (1) (0) (1) (0) (0) (1) (0) (0) Occasions (k=1,..., K) Traps (j=1,..., J) Users (i=1,..., I) A social media user was captured posting something today, here and now t−1t t + 1 sessions Protest event day Figure 1: Indexing social media users in a 3–dimensional array. ‘Physical proximity to the protest location’ in the Methods section). The second set consisted of a single proxy, which was the Twitter post rate of SMUs in the Mexican capital, which was measured using the Tweetogram metric (Leypunskiy et al.,2018) that expresses the normalized activity rate of users over 24 hours (see ‘Measuring the tweeting rhythm’ in the Methods section). The third set measured was the centrality of the Mexico City street and metro network. The network data were obtained from OpenStreetMap using the automated approach developed by Boeing (2017,2020). The metro network was based on the official information published by the Mexican metro system. The specific proxies used for quantifying street intersection centrality were the degree, betweenness, and closeness measures for weighted networks (Opsahl et al.,2010) (see “Measuring street and metro networks centralities’ in the Methods section). The fourth set represented socio-demographic and socioeconomic characteristics based on an anonymized data sample from the 2010 Mexican census of 58,064 blocks in Mexico City. One of these proxies measured population density while the other was a relative wealth index based on an aggregate measure of household assets, both calculated at the city block level (see ‘Demographic and economic indices’ in the Methods section). Our database consisted of i= 1,216 SMUs who supported the march and who were captured by j= 158 traps over 3 consecutive observation days (i.e. sessions ) encompassing the day of the march and one day on either side. The testing and evaluation of SMU’s detectability and density variations in the sample were done using spatial capture-recapture methods (Royle et al.,2013), which is an approach developed in the field of population and landscape ecology. This approach has definite advantages for identifying patterns in hard-to-find samples given that, even if the latter are small and incomplete, it can provide estimates of their detectability and density over a given geographical area. The method makes joint estimates of a pair of models, one for SMU’s detectability and the other for SMU’s spatial density (see Table 1). The detections decay as a function of the distance between the trap and the detected SMU’s activity centre d(s) , the latter considered as a latent variable (Royle et al.,2017). Thus, for the detection model, we use the so-called half–normal detection function. It has two parameters: baseline detection probability p0 , which determines the maximum detection probability at the SMU’s activity centre, and spatial decay σ , which controls how rapidly detection probabilities decrease with distance from it (see ‘Spatial capture-recapture analysis’ in the Methods section). These parameters are theoretically significant for understanding SMU’s spatial behavior during protest events given that d(s) allows us to estimate the variation if any, in SMU’s spatial density, p0 allows us to estimate the 4 On spatial variation in the detectability and density of social media user protest supporters Model Parameter Formula Baseline encounter model p0logit(p0) = α0+PkαkCovk Spatial decay model σlog(σ) = γ0+PkγkCovk Spatial point process model d(s) log(d(s))=β0+PkβkCovk Table 1: Parameters of the SMU’s detectability and density variation models. Covk are the different covariates affecting the parameters and αk,γk,βkare the covariate coefficients. maximum detectability of SMUs during a protest event, and with σ we can estimate any variation in the use of a space before, during and after the protest day. The importance of estimating these parameters lies in the fact that previous studies, especially those in sociology, have tended to focus on estimating the local number (i.e. at the protest location) of persons attending a demonstration (Rotman and Shalev,2020) without also considering the changing detectability, use of space, and density. This is relevant because it is usually not possible to make a direct estimation of user spatial density with geotagged social media data, and especially because the sampling mechanism of social media APIs is generally unknown. Also, the user’s location can only be observed when posting, so his or her activity centre is not directly observable, and the estimation of space use needs to be addressed differently. Also, given the understanding of SMU’s spatial behavior developed here, observed changes in density, detectability, or spatial decay on the protest day could be seen as an expression of the collective engagement. By the same token, a lack of significant change in the group’s spatial parameters could be said to indicate that there had been no change in its spatial behavior and level of engagement during the protest event. The approach proposed here is thus very powerful in its ability to identify behavior patterns using observational samples of SMUs and spatial covariates. Several candidate model configurations were used to test what factors explain the detectability and density of the observed SMUs. First, a null model was created, where the density of social media user, the baseline probability of detection, and the scale parameter is kept constant over the city. Second, a series of alternative models were created. For the baseline encounter model ( p0 ), the measures tested were demographic and economic indices of the neighborhood, the centrality of the street and metro network, the Tweetogram measure (which measures the rate of posting in the city for each hour), the day of observation (i.e. session ), and the physical proximity between the endpoint of the protest march and the trap where the SMU was captured or recaptured posting social media content. For the spatial decay model ( σ ) we tested whether observation day (i.e. session ) contributed to explaining variability in the use of space. Additionally, for the density model ( d ( s )), we include the demographic and economic indices of the neighborhood, the centrality of the street and metro network, the day of observation (i.e. session), and the physical proximity between the endpoint of the march and SMU’s possible activity centres (i.e. represented by a regular grid of points). Maximum likelihood was used to jointly estimate the model parameters and evaluate which candidate model best fit the data. Specifically, a likelihood analysis of various models were conducted using the R package OSCR (Sutherland et al.,2019), a type of generalized linear mixed model. As can be seen, this methodological design is quite flexible given that maximum likelihood allows us to compare multiple competing models and explanatory spatial and temporal factors. 3 Methods 3.1 Spatial capture-recapture analysis The following account of the spatial capture-recapture method is based on the work of Sutherland et al. (2019) which was adapted to our research context. In conceptual terms, an SMU is assumed to have an activity centre ( si = [ si,X , si,Y ]) that can be regarded as a spatial coordinate. However, an SMU’s location can only be known if it is captured in a trap, so si is considered to be a latent rather than an observable variable. SMUs are also presumed to live within a geographic area represented by a state space ( S ). In our methodological setting, the null model specifies that each SMU’s activity centre is distributed uniformly in space: si∼Uniform(S)(1) 5 On spatial variation in the detectability and density of social media user protest supporters To determine the activity centre values we use maximum likelihood estimation with spatial capture-recapture (SCR) models. This approach is based on the marginal likelihood that the unknown variable s is removed by averaging (or marginalizing) over the possible values of s . Thus, we begin by identifying the conditional-on-s likelihood. The user encounter model yijk for individual iat trap jon occasion k, conditional upon si, is yijk|si∼Bernoulli(p(x, si;θ)) (2) where p(x, s) = p0exp(−kx−sk2/(2σ2)) (3) is the detectability or probability of detection at trap x , which depends on s and θ= (p0, σ) (see Table 1for more details). The joint distribution of the data for individual i is the product of J×K such terms (i.e., contributions from each of Jtraps and Koccasions): [yi|si, θ] = J Y j=1 K Y k=1 Bernoulli(p(xj, si;θ)) (4) This assumes that an encounter of individual i in each trap and on each occasion is independent of encounters in every other trap and occasion, conditional upon si. To compute the marginal likelihood, consider a regular grid of G points denoted su that form cells of equal area and are indexed by u = 1,2, ..., G . In our design, we use a grid of 479 points. The marginal probability mass function (pmf) of yiis then approximated by [yi|θ] = 1 G G X u=1 [yi|su, θ](5) The joint likelihood for the data from n observed individuals, assuming independence of encounters among and between individuals, is the product of n such terms and a contribution of the n0=N−n uncaptured individuals. Each of these all-zero encounter histories (i.e., the capture histories of individuals not encountered) will have the same marginal pmf contribution in the likelihood given above, here denoted by π0 . Of the n observed individuals, there are N n = N! n!n0! ways to choose a sample of size n . A combinatorial term is therefore required, the joint likelihood then being given by L(θ, n0|y) = N! n!n0!(n Y i=1 [yi|θ])πn0 0(6) Our data are from distinct and more or less independent populations, referred to as sessions , that correspond to the three days in June considered in the study. The multi-session model integrates data from these different sessions . If Ng is the population size of group g , there are (Ng! ng!(Ng−ng)! ) ways to choose a sample. The multisession model assumes that Ng∼Poisson(λg) where the Ng are mutually independent random variables. We obtain the marginal likelihood as a function of λg by independently marginalizing over this distribution of Ngfor the data from each group: L(λg, θ) = ∞ X g=1 L(θ, n0g|y)Poisson(Ng;λg)(7) Explicit models can be formulated for λg: log(λg) = β0+β1Covg Spatial covariates can be obtained by raster data or by inverse distance-weighted interpolation, as in this study, which also contributes to maintaining offline the privacy of the studied geographical areas. For further details on the assumptions of spatial capture-recapture modelling, see Royle et al. (2013); Sutherland et al. (2019). 6 On spatial variation in the detectability and density of social media user protest supporters 3.2 Identification of users supporting the protest event The identification of SMUs who supported the protest was accomplished in two stages. The first stage was to manually label the data using qualitative coding in a small sample of social media data. The second stage utilized a machine learning approach to search in a larger sample for protest-related tweets. 3.2.1 Qualitative coding of data The data coding procedure consisted of five steps that were applied to the text of each tweet, as follows. Step 1: Determine whether it includes hashtags relating to the march. Step 2: Determine whether it contains emoticons related to this social movement. Step 3: Determine whether it has any other content supporting the protest (e.g., photos, videos, maps, or any other media content in the URL). Step 4: Read the entire text to determine whether it suggests support for the protest. Step 5: Code each tweet based on the results of the previous steps. This last step was carried out by two raters. If a tweet contained references to the protest, it was coded YES , otherwise, it was coded NO . 5 The outcome of this procedure was that, of the 10,000 coded tweets, 796 users supporting the march. The value of Cohen’s kappa coefficient for the two raters’ coding results ( κ= 0,84) indicated a high level of agreement between them. This initial codification was used for the next section. 3.2.2 Machine learning classification Once the 10,000-item database was codified, the classification of the entire sample using machine learning could be performed. The tweet texts to be classified were first lemmatized and cleaned by removing stop words, punctuation, URLs, and mentions. Hashtags were kept and emoticons were replaced with code words given that they played an important role in distinguishing related tweets. Logistic regression was used to classify the tweets and a grid search was used to refine the model parameters. We then applied a term-frequency transformer and performed a feature selection using the chi-square test to choose the most significant 500 terms. Because the dataset is unbalanced, we also employed combined under-sampling and over-sampling techniques using SMOTE+Tomek as suggested in Santos et al. (2018). The performance of the classifiers (see Table 2) was evaluated by 5-fold cross-validation on the test dataset measured in term of ROC-AUC, as well as the Macro-F1 score (M-F1) and the Matthews correlation coefficient (MCC), the last two often used to measure performance with unbalanced databases. The obtained model was then used to make predictions regarding the non-coded data in the sample. Logistic regression obtained 2,605 protest-related tweets and 114,848 unrelated ones. After filtering the database, the number of users identified as supporters on each of the three days covered by the data (i.e. 22, 23, and 24 of June) was 526, 846, and 539 respectively. Therefore, the use of machine learning made it possible to identify additional users supporting the protest march. Classifier M-F1 ROC-AUC MCC Logistic regression 0.95 0.98 0.91 Table 2: Summary of classification performance. 3.3 Measuring the tweeting rhythm To measure SMU’s tweeting activity in the city we used the measure (A(t)) (Leypunskiy et al.,2018), defined as: A(t) = 1 NX i fi(t) Pτfi(τ)(8) where N is the number of observed users normalized to a unit weight, and fi is the number of tweets posted by user i during a time bin t , in this case, 1 hour. The time series were then averaged across the 72 hours (i.e., the three observation days, see Figure 2). 3.4 Measuring street and metro networks centralities A graph of the Mexico City street network was generated that contained 112,188 nodes representing intersections and 164,586 edges representing streets. The mean and standard deviation of the edge lengths were 88.91 meters 5Additional details of the coding can be found in Masias et al. (2019). 7 On spatial variation in the detectability and density of social media user protest supporters 0.00 0.05 0.10 0.15 Jun 22 Jun 23 Jun 24 Jun 25 Date Tweetogram (A(t)) Figure 2: Normalized posting rate of users in Mexico City during three days of observation. On June 23, the day of the planned protest event, the maximum value of Tweetogram was reached at 3:00 p.m. and SD = 128.12, respectively. To measure street centrality as an explanatory factor of SMU’s density in the city, we used the centrality metrics αω -weighted degree, αω -weighted betweenness, and αω -weighted closeness (Opsahl et al.,2010). These indices allowed us to measure the centrality of street intersections, and metro stations (i.e. nodes) depending on the “number” or the “strength” of the links, or both. A tuning parameter α was incorporated to include additional information in the centrality measures. The values used for this parameter were: α= 0, which leaves the centrality measures in their traditional form; α= 0.5, which weights the measures by the number of edges and the latter’s weights (the lengths of the streets in meters, and the distance between pairs of connected metro stations); and α=1, which weights the measures only by the street lengths.6 3.5 Demographic and economic indices The neighborhood-level socio-demographic indicators were created with anonymized data from the 2010 Mexican census. 7 Household and dwelling attributes were used to create indices at the census block level, which were then normalized by census-block area. 3.5.1 Population density per census block This indicator described the population density of each census block in terms of its population density per dwelling and household. The information was aggregated and normalized by census-block area. A principal components analysis (PCA) was carried out to detect independent features characterizing the population density per census block. We used the first two main components (i.e. population density PC1 and population density PC2 ) that captured 64.3% and 32.6% percent of the total variation to create a feature vector that characterized the population density characteristics of each census block (n=58,064) of Mexico City. 3.5.2 Relative wealth per census block This index measured the relative wealth of a census block. It was constructed from census data on two categories of assets. The first category was household items such as information and communications technology devices (radios, televisions, computers, landlines, and mobile telephones, etc.), refrigerators, washing machines, and motor vehicles. The second category included dwelling characteristics such as internet access, electricity, running water, toilets, and connection to the city sewage system. The asset counts were aggregated by census block and 6 Further information on how these definitions of the weighted centrality measures compare with the original, non-weighted definitions may be found in Masías et al. (2016). Network data is publicly available from OpenStreetMap through OSMNX Package. See, Boeing (2017,2020). 7 The anonymized census data of Mexico City was made publicly available by Diego del Valle ( https://blog. diegovalle.net/2013/06/shapefiles-of-mexico-agebs-manzanas-etc.html). 8 On spatial variation in the detectability and density of social media user protest supporters C. Arnaiz-Schmitz, M.F. Schmitz, C. Herrero-Jáuregui, J. Gutiérrez-Angonese, F.D. Pineda, and C. Montes. Identifying socio-ecological networks in rural-urban gradients: Diagnosis of a changing cultural landscape. Science of The Total Environment, 612:625–635, jan 2018. Pablo Valdez. Homeostatic and circadian regulation of cognitive performance. Biological Rhythm Research, 50(1): 85–93, jul 2018. Manuela Fischer, Julian Di Stefano, Pierre Gras, Stephanie Kramer-Schadt, Duncan R. Sutherland, Graeme Coulson, and Milena Stillfried. Circadian rhythms enable efficient resource selection in a human-modified landscape. Ecology and Evolution, 9(13):7509–7527, jun 2019. Eugene Leypunskiy, Emre Kıcıman, Mili Shah, Olivia J. Walch, Andrey Rzhetsky, Aaron R. Dinner, and Michael J. Rust. Geographically resolved rhythms in Twitter use reveal social pressures on daily activity patterns. Current Biology, 28(23):3763–3775.e5, dec 2018. Ananya Diwan, Rakesh Kumar Swain, Sarojini Minz, Arti Parganiha, and Atanu Kumar Pati. Ultradian, circadian, and circaseptan rhythms in the patterns of usage of Facebook messenger. Biological Rhythm Research, pages 1–9, apr 2020. Elizabeth L Murnane, Saeed Abdullah, Mark Matthews, Tanzeem Choudhury, and Geri Gay. Social (media) jet lag: How usage of social technology can modulate and reflect circadian rhythms. In Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing - UbiComp ' 15, pages 843–854. ACM Press, 2015. Rakesh Kumar Swain and Atanu Kumar Pati. Habitual daily ‘Good Morning’ message senders reveal the status of their own circadian clock. Biological Rhythm Research, pages 1–12, jan 2019. Yikang Rui and Yifang Ban. Exploring the relationship between street centrality and land use in Stockholm. International Journal of Geographical Information Science, 28(7):1425–1438, mar 2014. M. Bielik, R. König, S. Schneider, and T. Varoudis. Measuring the impact of street network configuration on the accessibility to people and walking attractors. Networks and Spatial Economics, 18(3):657–676, sep 2018. Lucia Summers and Shane D. Johnson. Does the configuration of the street network influence where outdoor serious violence takes place? Using space syntax to test crime pattern theory. Journal of Quantitative Criminology, 33(2):397–420, jun 2016. Jeremiah J. Nieves, Forrest R. Stevens, Andrea E. Gaughan, Catherine Linard, Alessandro Sorichetta, Graeme Hornby, Nirav N. Patel, and Andrew J. Tatem. Examining the correlates and drivers of human population distributions across low- and middle-income countries. Journal of The Royal Society Interface, 14(137):20170401, dec 2017. Marco De Nadai, Jacopo Staiano, Roberto Larcher, Nicu Sebe, Daniele Quercia, and Bruno Lepri. The death and life of great italian cities. In Proceedings of the 25th International Conference on World Wide Web - WWW ' 16, pages 413–423. ACM Press, 2016. Yaolin Liu, Feiguo Fang, and Ying Jing. How urban land use influences commuting flows in Wuhan, Central China: A mobile phone signaling data perspective. Sustainable Cities and Society, 53:101914, feb 2020. J Andrew Royle, Richard B Chandler, Rahel Sollmann, and Beth Gardner. Spatial capture-recapture. Academic Press, 2013. Michael J. Bosia, Sandra M. McEvoy, and Momin Rahman, editors. The Oxford Handbook of Global LGBT and Sexual Diversity Politics. Oxford University Press, jan 2019. Caroline Beer and Victor D. Cruz-Aceves. Extending rights to marginalized minorities: Same-sex relationship recognition in Mexico and the United States. State Politics & Policy Quarterly, 18(1):3–26, jan 2018. Cameron Walker. Equality: Standing out. Nature, 505(7482):249–251, jan 2014. Geoff Boeing. OSMnx: New methods for acquiring, constructing, analyzing, and visualizing complex street networks. Computers, Environment and Urban Systems, 65:126–139, sep 2017. Geoff Boeing. The right tools for the job: The case for spatial science tool-building. Transactions in GIS, 24(5): 1299–1314, sep 2020. Tore Opsahl, Filip Agneessens, and John Skvoretz. Node centrality in weighted networks: Generalizing degree and shortest paths. Social Networks, 32(3):245–251, jul 2010. J. Andrew Royle, Angela K. Fuller, and Christopher Sutherland. Unifying population and landscape ecology with spatial capture-recapture. Ecography, 41(3):444–456, aug 2017. 15 On spatial variation in the detectability and density of social media user protest supporters Assaf Rotman and Michael Shalev. Using location data from mobile phones to study participation in mass protests. Sociological Methods & Research, page 004912412091492, apr 2020. Chris Sutherland, J. Andrew Royle, and Daniel W. Linden. oSCR: a spatial capture–recapture R package for inference about spatial ecological processes. Ecography, 42(9):1459–1469, jul 2019. Victor H. Masias, Tobias Hecking, Fernando Crespo, and H. Ulrich Hoppe. Detecting social media users based on pedestrian networks and neighborhood attributes: an observational study. Applied Network Science, 4(1), oct 2019. Miriam Seoane Santos, Jastin Pompeu Soares, Pedro Henrigues Abreu, Helder Araujo, and Joao Santos. Crossvalidation for imbalanced datasets: Avoiding overoptimistic and overfitting approaches [research frontier]. IEEE Computational Intelligence Magazine, 13(4):59–76, nov 2018. Víctor Hugo Masías, Paula Baldwin, Sigifredo Laengle, Augusto Vargas, and Fernando A. Crespo. Exploring the prominence of Romeo and Juliet’s characters using weighted centrality measures. Digital Scholarship in the Humanities, page fqw029, aug 2016. Chris Sutherland, David J. Muñoz, David A.W. Miller, and Evan H. Campbell Grant. Spatial capture–recapture: A promising method for analyzing data collected using artificial cover objects. Herpetologica, 72(1):6–12, mar 2016. David J. Muñoz, David A. W. Miller, Chris Sutherland, and Evan H. Campbell Grant. Using spatial capture–recapture to elucidate population processes and space-use in herpetological studies. Journal of Herpetology, 50(4):570–581, dec 2016. Jennifer Oser and Shelley Boulianne. Reinforcement effects between digital media use and political participation: A meta-analysis of repeated-wave panel data. Public Opinion Quarterly, 84(S1):355–365, 2020. Sophia J Wallace, Chris Zepeda-Millán, and Michael Jones-Correa. Spatial and temporal proximity: Examining the effects of protests on political attitudes. American Journal of Political Science, 58(2):433–448, 2014. Jeronimo Cortina. From a distance: Geographic proximity, partisanship, and public attitudes toward the U.S.–Mexico border wall. Political Research Quarterly, 73(3):740–754, jun 2019. Michael Biggs. Size matters. Sociological Methods & Research, 47(3):351–383, feb 2016. Anton Sobolev, M Keith Chen, Jungseock Joo, Zachary C Steinert-Threlkeld, et al. News and geolocated social media accurately measure protest size variation. American Political Science Review, 114(4):1343–1351, jun 2020. Karl-Dieter Opp. Theories of Political Protest and Social Movements. Routledge, apr 2009. Han Zhang. Physical exposures to political protests impact civic engagement: Evidence from 13 quasi-experiments with chinese social media. SSRN Electronic Journal, 2016. Han Zhang, Shawndra Hill, and David Rothschild. Geolocated twitter panels to study the impact of events. In 2016 AAAI Spring Symposium Series. AAAI press, Palo Alto., 2016. Alireza Karduni and Eric Sauda. Anatomy of a protest: Spatial information, social media, and urban space. Social Media and Society, 6(1):205630511989732, jan 2020. Yingjie Hu and Ruo-Qian Wang. Understanding the removal of precise geotagging in tweets. Nature Human Behaviour, 4(12):1219–1221, sep 2020. 16