Background: A growing amount of health research uses social media data. Those critical of social media research often cite that it may be unrepresentative of the population; however, the suitability of social media data in digital epidemiology is more nuanced. Identifying the demographics of social media users can help establish representativeness.
Objective: This study aims to identify the different approaches or combination of approaches to extract race or ethnicity from social media and report on the challenges of using these methods.
Methods: We present a scoping review to identify methods used to extract the race or ethnicity of Twitter users from Twitter data sets. We searched 17 electronic databases from the date of inception to May 15, 2021, and carried out reference checking and hand searching to identify relevant studies. Sifting of each record was performed independently by at least two researchers, with any disagreement discussed. Studies were required to extract the race or ethnicity of Twitter users using either manual or computational methods or a combination of both.
Results: Of the 1249 records sifted, we identified 67 (5.36%) that met our inclusion criteria. Most studies (51/67, 76%) have focused on US-based users and English language tweets (52/67, 78%). A range of data was used, including Twitter profile metadata, such as names, pictures, information from bios (including self-declarations), or location or content of the tweets. A range of methodologies was used, including manual inference, linkage to census data, commercial software, language or dialect recognition, or machine learning or natural language processing. However, not all studies have evaluated these methods. Those that evaluated these methods found accuracy to vary from 45% to 93% with significantly lower accuracy in identifying categories of people of color. The inference of race or ethnicity raises important ethical questions, which can be exacerbated by the data and methods used. The comparative accuracies of the different methods are also largely unknown.
Conclusions: There is no standard accepted approach or current guidelines for extracting or inferring the race or ethnicity of Twitter users. Social media researchers must carefully interpret race or ethnicity and not overpromise what can be achieved, as even manual screening is a subjective, imperfect method. Future research should establish the accuracy of methods to inform evidence-based best practice guidelines for social media researchers and be guided by concerns of equity and social justice.
Research Using Twitter Data
Twitter data are increasingly being used as a surveillance and data collection tool in health research. When millions of users post on Twitter, it translates into a vast amount of publicly accessible, timely data about a variety of attitudes, behaviors, and preferences in a given population. Although these data were not originally intended as a repository of individual information, Twitter data have been retrofitted in infodemiology to investigate population-level health trends [- ]. Researchers often use Twitter data in consort with other sources to test the relationship between web-based discourse and offline health behavior, public opinion, and disease incidence.
The appeal of Twitter data is clear. Twitter is one of the largest public-facing social media platforms, with an ethnically diverse user base [, ] of more than 68 million US Twitter users, with Black users accounting for 26% of that base [ ]. This diverse user base gives researchers access to people they may have difficulty reaching using more traditional approaches [ ]. However, promising insights that can be derived from Twitter data are often limited by what is missing, specifically the basic sociodemographic information of each Twitter user. The demographic attributes of users are often required in health research for subpopulation analyses, to explore differences, and to identify inequity. Without evidence of the distal and proximal factors that lead to racial and ethnic health disparities, it is impossible to address and correct these drivers. Insights from social media data can be used to inform service provision as well as to develop targeted health messaging by understanding public perspectives from diverse populations.
Extracting Demographics From Twitter
However, to use social media and digital health research to address disparities, we need to know not only what is said on Twitter but also who is saying what . Although others have discussed extracting or estimating features, such as location, age, gender, language, occupation, and class, no comprehensive review of the methods used to extract race or ethnicity has been conducted [ ]. Extracting the race and ethnicity of Twitter users is particularly important for identifying trends, experiences, and attitudes of racially and ethnically diverse populations [ ]. As race is a social construction and not a genetic categorization [ , ], the practice of defining race and ethnicity in health research has been an ongoing, evolving challenge. Traditional research has the advantage of identifying the person in the study and allowing them to systematically identify their racial and ethnic identities. In digital health research [ , ], determining a user’s race or ethnicity by extracting data from a user’s Twitter profile, metadata, or tweets is a process that is inevitably challenging, complex, and not without ethical questions.
Furthermore, although Twitter is used for international research, an international comparative study of methods to determine race or ethnicity is difficult, practically impossible, given that societies use different standardized categories that describe their own populations . A common approach in the United States is based on the US Census Bureau practice to allow participants to identify with as many as 5-6 large racial groupings (Black, White, Asian, Pacific Islander, Native, and other), while separately choosing one ethnicity (Hispanic) [ ]. However, race and ethnicity variables continue to be misused in the study design or when drawing conclusions. For example, race or ethnicity is often incorrectly treated as a predictor of poor health rather than as a proxy for the impact of being a particular race or ethnicity has on that person’s experience with the health system [ ]. Simply put, health disparities are driven by racism, not race [ - ]. Although race or ethnicity affiliation is an important factor in understanding diverse populations, digital research must tread lightly and thoughtfully both the collection and assignment of race or ethnicity.
The lack of basic sociodemographic data on Twitter users has led researchers to apply a variety of approaches to better understand the characteristics of the people behind each tweet. The breadth of the landscape of approaches to extracting race or ethnicity is currently unknown. Our overall aim was to summarize and assess the range of computational and manual methods used in research based on Twitter data to determine the race or ethnicity of Twitter users.
We conducted a comprehensive scoping review of extraction methods and offered recommendations and cautions related to these approaches . We selected Twitter, as it is currently the most commonly used social media platform in health care research, and it has some unique intrinsic characteristics that drive the methods used for mining it. Thus, we felt that the methods, type of data, and social media platforms used are related in such a way that comparing methods for different social media would add too many variables and would not be truly comparing like with like. A detailed protocol was designed for the methods to be used in our scoping review, but we were unable to register scoping reviews on PROSPERO. We report our methods according to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) scoping review statement [ ].
We devised strict inclusion criteria for our review based on the Population, Intervention, Comparators, Outcomes, and Study design format. Although this was not a review of effectiveness, we felt that the Population, Intervention, Comparators, Outcomes, and Study design question breakdown  was still the most appropriate one available for our question format [ ]. The inclusion criteria are described in the following sections.
We included only data sets of Twitter users. Studies were eligible for inclusion if they collected information to extract or infer race or ethnicity directly from the users’ tweets, their profile details (such as the users’ photo or avatar, their name, location, and biography [bio]), or their followers. We excluded studies that extracted race or ethnicity from social media platforms other than Twitter, from unspecified social media platforms, or those that used multiple social media platforms that included Twitter, but the data relating to Twitter were not presented separately.
Studies were included where the methods to extract or infer the race or ethnicity data of Twitter users were stated. Articles that used machine learning (ML), natural language processing (NLP), human-in-the-loop, or other computationally assisted methods to predict race or ethnicity of users were included, as were manual or noncomputational methods, including photo recognition or linking to census data. We excluded studies for which we were unable to determine the methods used or for which we extracted data solely on other demographic characteristics, such as age, gender, or geographic location.
The use of a comparison of the methods used was not required. A method could be compared with another (such as a gold standard), or no comparison could be undertaken.
The extraction or inference of the race or ethnicity of Twitter users was the primary or secondary outcome of the study. As this was a scoping review in which we aimed to demonstrate the full landscape of the literature, no particular measurement of the performance of the method used was required in our included studies.
Any type of research study design was considered relevant. Discussion papers, commentaries, and letters were excluded.
No restrictions on date, language, or publication type were applied to the inclusion criteria. However, no potentially relevant studies were identified in any non-English language, and the period by default was since 2006, the year of the inception of Twitter.
A database search strategy was derived by combining three facets: facet 1 consisted of free-text terms related to Twitter (Twitter OR Tweet* OR Tweeting OR Retweet* OR Tweep*); facet 2 consisted of terms for race or ethnicity; and facet 3 consisted of terms for methods of prediction, such as ML, NLP, and artificial intelligence–related terms (Table S1 in[ , , , , , , - ]). All ethnology-related subject terms were adapted for different database taxonomies and syntax, with standard methods for predicting subject terms in MEDLINE and other database indexing. The methods of predicting term facets were expanded using a comprehensive list of specific text analysis tools and software names extracted from the study by Hinds and Joinson [ ], which included a comprehensive list of automated ML processes used in predicting demographic markers in social media. Additional terms have been added from a related study [ ].
A wide range of bibliographic and gray literature databases were selected to search for topics on computer science, health, and social sciences. The databases () were last searched on May 15, 2021, with no date or other filter applied.
Reference checking of all included studies and any related systematic reviews identified by the searches were conducted. We browsed the Journal of Medical Internet Research, as this is a key journal in this field, and hand searched 2 relevant conferences, the International Conference on Weblogs and Social Media and Association for Computational Linguistics proceedings.
Citations were exported to a shared Endnote library, and duplicates were removed. The deduplicated records were then imported into Rayyan to facilitate independent blinded screening by the authors. Using the inclusion criteria, at least two screeners (SG, RS, KO, or RJ) from the research team independently screened each record, with disputes on inclusion discussed and a consensus decision reached.
Only the first 50 records from ACL and the first 100 records from a Google Scholar search were screened during two searches (March 11, 2020, and May 24, 2021) as these records are displayed in order of relevance, and it was felt that after this number no relevant studies were being identified [, , - , ].
|Database||Total results, n|
|ACL Anthology||Screened first 50 records from 2 searches|
|ACM Digital Library||150|
|Conference Proceedings Citation Index—Science||84|
|Conference Proceedings Citation Index—Social Science||7|
|Emerging Sources Citation Index||41|
|Google Scholar||Screened first 100 records from 2 searches|
|Library and Information Science Abstracts||120|
|ProQuest dissertations and theses—United Kingdom and Ireland||195|
|Science Citation Index||56|
|Social Science Citation Index||111|
For each included study, we extracted the following data on an excel spreadsheet:
year of publication, study country and language, race or ethnicity categories extracted (such as for race—Black, White, or Asian or for ethnicity—Hispanic or European), and paper type (journal, conference, or thesis). We also extracted details on extraction methods (such as classification models or software used), features and predictors used in extraction (tweets, profiles, and pictures), number of Twitter users, number of tweets or images used, performance measures to evaluate methods used (validation), and results of any evaluation (such as accuracy). All performance measure metrics were reported as stated in the included studies. All the extracted data were checked by 2 reviewers.
There was no formally approved quality assessment tool for this type of study. As this was a scoping review, we did not carry out any formal assessment. However, we assessed any validation performed and whether the methods were reproducible.
We have summarized the stated performance of the papers that included validation. However, we could not compare approaches using the stated performance, as the performance measures and validation approaches varied considerably. In addition, there is no recognized gold standard data set for comparison.
A total of 1735 records were entered into an Endnote library (Clarivate), and duplicates were removed, leaving 1249 (72%) records for sifting (). A total of 1080 records were excluded based on the title and abstract screening alone. A total of 169 references were deemed potentially relevant by one of the independent sifters (RS, GG, RJ, SG, and KO). The full text of these articles was screened independently, and 67 studies [ , , - , ] met our inclusion criteria and 102 references were excluded [ , , - ]. The main reason for exclusion was that although the abstract indicated that demographic data were collected, it did not include race or ethnicity (most commonly, other demographic attributes such as gender, age, or location were collected). Other reasons for exclusion were that the researchers collected demographic data through surveys or questionnaires administered via Twitter (but not from data posted on Twitter) or that the researchers used a social media platform other than Twitter.
Characteristics of the Included Studies
Most of the studies (51/67, 76%) stated or implied that they were based solely or predominantly in the United States and were limited to English language bios or tweets. A total of 6 studies were multinational [, , , , , ]; 1 was UK based (also in English) [ ], another was based in Qatar [ ], and 12% (8/67) of studies extracted data from tweets in multiple languages [ , , , , , , , ] (Table S2 in ).
The most common race examined was White (58/67, 87%), followed by Black or African American (56/67, 84%), Asian (45/67, 67%), and the most common ethnicity examined was Hispanic/Latino (43/67, 64%).
Some studies (12/67, 18%) treated race as a binary classification, such as African American or not or African American or White, whereas others created a multiclass classifier of 3 (15/67, 22%) or 4 classes (33/67, 49%) or a combination of classes. A total of 6 studies identified >4 classes; however, these often included ethnicity or nationality classifiers as well as race [, , , , , ]. Wang and Chi [ ] was a conference paper which did not report the race types extracted.
The data objects from Twitter used to extract race or ethnicity varied, with the use of profile pictures or Twitter users’ names being the most common. Others have also used tweets in the users’ timeline, information from Twitter bios, or Twitter users’ locations. Most studies (39/67, 58%) used more than one data object from Twitter data. In addition, the data sets within the studies varied in size between 392 and 168,000,000, with those using manual methods having smaller data sets ranging from just 392  to 4900 [ ].
Unfortunately, although performance has been measured in 67% (45/67) of studies (this was inconsistently measured). The metrics used to report results were particularly varied for studies using ML or NLP and included the F1 score (which combines precision and recall), accuracy, area under the curve, or mean average precision. lists the methods, features, and reported performance of the top model from each study.
|Study||Classifier||MLa model||Features||Results reported|
|Accuracy||F1 score||Area under curve|
|Pennacchiotti and Popescu, 2011 ||Binary||GBDTb||Images, text, topics, and sentiment||N/Ac||0.66||N/A|
|Pennacchiotti and Popescu, 2011 ||Binary||GBDT||Images, text, topics, sentiment, and network||N/A||0.70||N/A|
|Bergsma et al, 2013 ||Binary||SVMd||Names and name clusters||0.85||N/A||N/A|
|Ardehaly and Culotta, 2017 ||Binary||DLLPe||Text and images||N/A||0.95 (image); 0.92 (text)||N/A|
|Volkova and Backrach, 2018 ||Binary||LRf||Text, sentiment, and emotion||N/A||N/A||0.97|
|Wood-Doughtry et al, 2018 ||Binary||CNNg||Name||0.73||0.72||N/A|
|Saravanan, 2017 ||Ternary||CNN||Text||NRh||NR||NR|
|Ardehaly and Culotta, 2017 ||Ternary||DLLP||Text and images||N/A||0.84 (image); 0.83 (text)||N/A|
|Gunarathne et al, 2019 ||Ternary||CNN||Text||N/A||0.88||N/A|
|Wood-Doughtry et al, 2018 ||Ternary||CNN||Name||0.62||0.43||N/A|
|Culotta et al, 2016 ||Quaternary||Regression||Network and text||N/A||0.86||N/A|
|Chen et al, 2015 ||Quaternary||SVM||n-grams, topics, self-declarations, and image||0.79||0.79||0.72|
|Markson, 2017 ||Quaternary||CNN||Synonym expansion and topics||0.76||N/A||N/A|
|Wang et al, 2016 ||Quaternary||CNN||Images||0.84||N/A||N/A|
|Xu et al, 2016 ||Quaternary||SVM||Synonym expansion and topics||0.76||N/A||N/A|
|Ardehaly and Culotta, 2015 ||Quaternary||Multinomial logistic regression||Census, name, network, and tweet language||0.83||N/A||N/A|
|Ardehaly, 2014 ||Quaternary||LR||Census and image tweets||0.82||0.81||N/A|
|Barbera, 2016 ||Quaternary||LR with ENi||Tweets, emojis, and network||0.81||N/A||N/A|
|Wood-Doughty 2020 ||Quaternary||CNN||Name, profile metadata, and text||0.83||0.46||N/A|
|Preotiuc-Pietro and Ungar, 2018 ||Quaternary||LR with EN||Text, topics, sentiment, part-of-speech tagging, name, perceived race labels, and ensemble||N/A||N/A||0.88 (African American), 0.78 (Latino), 0.83 (Asian), and 0.83 (White)|
|Mueller et al, 2021 ||Quaternary||CNN||Text and accounts followed||N/A||0.25 (Asian), 0.63 (African American or Black), 0.28 (Hispanic), and 0.90 (White)||N/A|
|Bergsma et al, 2013 ||Multinomial (>4)||SVM||Name and name clusters||0.81||N/A||N/A|
|Nguyen et al, 2018 ||Multinomial (>4)||Neural network||Images||0.53||N/A||N/A|
aML: machine learning.
bGBDT: gradient-boosted decision tree.
cN/A: not applicable.
dSVM: support vector machine.
eDLLP: deep learning from label proportions.
fLR: logistic regression.
gCNN: convolutional neural network.
hNR: not reported.
iEN: elastic net.
A total of 12 studies used manual techniques to classify Twitter users into race or ethnicity categories [, , , - , , , - ]. These studies generally combined qualitative interpretations of recent tweets, information in user bios making an affirmation of racial or ethnic identity, or photographs or images in the user timeline or profile.
In most cases, tweets were first identified by text matching based on terms of interest in the research topic, such as having a baby with a birth defect , commenting on a controversial topic [ , ], or using potentially gang- or drug-related language [ ]. Researchers then identified the tweet authors and, in most cases, assigned race or ethnicity through hand coding based on profile and timeline content. Some studies coded primarily based on self-identifying statements of race used in a tweet or in users’ bios, such as people stating that they are a Black American [ , , , ] or hashtags [ ] (such as #BlackScientist). Others coded exclusively based on the research team’s attribution of racial identity through the examination of profile photographs [ , ] or avatar [ ]. Some authors coded primarily with self-declarations, with secondary indicators, such as profile pictures, language, usernames, or other content [ , , , , ]. In most cases, it appears reasonable to infer that coding was performed by the study authors or members of their research teams, with the exception of those using the crowdsourcing marketplace, Amazon Mechanical Turk [ , ].
The agreement among coders was sometimes measured, but validity and accuracy measurements were not generally included. A study , however, documented 78% reliability for coding race compared with census demographics, with Black and White users being coded accurately 90% of the time and Hispanic or Asian users being accurately coded between 45% and 60% of the time. The high accuracy of Black users was based on the higher likelihood of Black users to self-identify.
Another approach to predict race or ethnicity is to use demographic information from the national census and census-like data and transfer it to the social media cohort. The US-based studies largely used census-based race and ethnicity categories: Asian and Pacific Islander, Black or African American, Latino or Hispanic, Native American, and White. A UK-based study included the categories British and Irish, West European, East European, Greek or Turkish, Southeast Asian, other Asian, African and Caribbean, Jewish, Chinese, and other minorities .
We identified 14 studies [, , , , , , , , , , - , ] that used census geographic data, census surname classification, or a combination of both. A total of 6 studies incorporated geographic census data [ , , , , , ]. For example, Blodgett et al [ ] created a simple probabilistic model to infer a user’s ethnicity by matching geotagged tweets with census block information. They averaged the demographic values of all tweets by the user and assumed this to be a rough proxy for the user’s demographics. Stewart [ ] collected tweets tagged with geolocation information (longitude and latitude). The ZIP code of the user was derived from this geolocation information and matched with the demographic information found in the ZIP Code Tabulation Area defined by the Census Bureau. This information was used to find a correlation between ethnicity and African American vernacular English syntax [ ].
Other studies have used the census-derived name classification system to determine race or ethnicity based on user names. We identified 12 studies that predicted user race or ethnicity using surnames [, , , , , , , - , , ]. Surnames were used to assign race or ethnicity using either a US census-based name classification system or, less commonly, an author in-house generated classification system. Of these 12 studies, 7 (58%) relied solely on the user’s last names [ , , , , , , ]. Of those that reported validating the system, validation methods of this name-based system alone were not reported, but 4 (33%) of the 12 studies reported an accuracy between 71.8% and 81.25% [ , , , ]. Of note, a study reported vastly different accuracies in predicting whiteness versus blackness (94% predicting White users vs 33% predicting African American or Black users) [ ]. The remaining 2 studies augmented name-based predictions with aggregate demographic data from the American Community Survey or equivalent surveys. For example, statistical and text mining methods have been used to extract surnames from Twitter profiles, combining this information with census block information based on geolocated tweets to assess the probability of the user’s race or ethnicity [ ]. However, these studies did not report validation or accuracy.
Ad Hoc ML or NLP
A total of 24 papers [- , , , , , , , - , , , - , - , ] used ML or NLP to automatically classify users based on their race or ethnicity. ML and NLP methods were used to process the data made available by Twitter users, such as profile images, tweets, and location of residence. These studies almost invariably consisted of larger cohorts, with considerable variation in the specific methods used.
Supervised ML models (in which some annotated data were used to train the system) were used in 12 (50%) of the 24 studies. The models used include support vector machine [, , ], gradient-boosted decision trees [ , ], and regression models [ , , , , ].
Semisupervised (where a large set of unannotated data is also used for training the system, in addition to annotated data) or fully unsupervised models using neural networks or regression were used for classification in 10 (42%) of the 24 studies [, , , , , , , - ].
A total of 2 studies used an ensemble of previously published race or ethnicity classifiers by processing the data through 4 extant models and using a majority rule approach to classify users based on the output of each classifier [, ].
ML models use features or data inputs to predict desired outputs. Features derived from textual information in the user’s profile description, such as name or location, have been used in some studies [, , , , , , , , , ]. Other studies included features related to images, including but not exclusively profile images [ , , , ], and facial features in those images [ ]. Some studies have used linguistic features to classify a user’s race or ethnicity [ , , , , , , , , , , , - , ]. Specific linguistic features used in the models include n-grams [ , , , - ], topic modeling [ , , ], sentiment and emotion [ ], and self-reports [ , , ]. Information about a user’s followers or network of friends was included as a feature in some studies under the assumption that members of these networks have similar traits [ , , , , ].
Labeled data sets are used to train and test supervised and semisupervised ML models and to validate the output of unsupervised learning methods. Some of the studies used previously created data sets that contained demographic information, such as the MORPH longitudinal face database of images , a database of mugshots [ ], or manually annotated data from previous studies [ , ]. Others created ground truth data sets from surveys [ ] or by semiautomatic means, such as matching Twitter users to voter registrations [ ], using extracted self-identification from user profiles or tweets [ , , ], or using celebrities with known ethnicities [ ]. Manual annotation of Twitter users was also used based on profile metadata [ , , , ], self-declarations in the timeline [ , ], or user images [ , ]. summarizes the best performing ML approach, features used, and the reported results for each study that used automatic classification methods. In the table, the classifier is the number of race or ethnicity classification groups, ML model is the top performing algorithm reported, and features are the variables used in the predictions.
Data from Twitter are inherently imbalanced in terms of race and ethnicity. In ML, it is important to attempt to mitigate the effects of the imbalance, as the models have difficulty learning from a few examples and will tend to classify to the majority class and ignore the minority class. Few studies (12/67, 18%) have directly addressed this imbalance. Some opted to make the task binary, focusing only on their group of interest versus all others [, , ] or only on the majority classes [ , ]. Others choose modified performance metrics that account for imbalance when reporting their results [ , , ]. A group, which was classified based on images, supplemented their training set from an additional data source for the minority classes [ , ]. Only 2 studies have experimented with comparator models trained on balanced data sets. In a study by Wood-Doughty et al [ ], the majority class was undersampled in their training sets and [ ] the minority classes were oversampled. In both cases, the overall performance of the models decreased in accuracy from 0.83 to 0.41 (on their best performing unbalanced model) and 0.84 to 0.68. [ ], as the performance boost from the models, the superior performance on the majority class was eradicated.
A total of 17 studies [, , - , , , , , , , , , , ] used off-the-shelf software packages to derive race or ethnicity. Moreover, 10 studies [ , , , , , , , , , ] used Face++ [ ], 5 studies [ , - , ] used Demographics Pro [ ], and 2 studies used Onomap [ ] software to determine ethnicity [ , ]. Face++ is a validated ML face detection service that analyzes features with confidence levels for inferred race attributes. Specifically, it uses deep learning to identify whether profile pictures contain a single face and then the race of the face (limited to Asian, Black, and White) and does not infer ethnicity (eg, Hispanic) [ ]. Demographics Pro estimates the demographic characteristics based on Twitter behavior or use using NLP, entity identification, image analyses, and network theory [ ]. Onomap is a software tool used for classifying names [ ]. A total of 3 studies that used Face++ used the same baseline data set [ , , ], and one used a partial subset of the same data set [ ].
In total, 2 studies that used Face++ [, ] did not measure its performance. Another study [ ] stated that Face++ could identify race with 99% confidence or higher for 9% of total users. In addition, 2 studies [ , ] used Face++ along with other methods. One of these studies used Face++ in conjunction with demographics, using a given name or full name from a database that contains US census data for demographics. This study simply measured the percentage of Twitter users for which race data could be extracted (46% college students and 92% role models) but did not measure the performance of Face++ [ ]. Another study [ ] built a classifier model on top of using Face++ and recorded an accuracy of 83.8% when compared with users who stated their nationality.
A total of 4 studies [, , , ] (with the same data set in full or in part) used the average confidence level reported by Face++ for race which was 85.97 (SD 0.024%), 85.99 (SD 0.03%), 86.12 (SD 0.032%), respectively, with a CI of 95%. When one of these studies [ ] carried out its own accuracy assessment, they found an accuracy score of 79% for race when compared with 100 manually annotated pictures. Huang et al [ ] also carried out an accuracy assessment and found that Face++ achieved an averaged accuracy score of 88.4% for race when compared with 250 manually annotated pictures.
A total of 5 studies [, - , ] used Demographics Pro, and although they reported on Demographics Pro success in general, they did not directly report any metrics of its success. The 2 studies using Onomap provided no validation of the software [ , ].
In light of our results, we have compiled our recommendations for best practice, which are summarized inand further examined in the Discussion section.
As there are no currently published guidelines or even best practice guidance, it is no surprise that researchers have used a variety of methods for estimating the race or ethnicity of Twitter users. We identified four categories for the methods used: manual screening, census-based prediction, ad hoc ML or NLP, and off-the-shelf software. All these methods exhibit particular strengths, as well as inherent biases and limitations.
Comparing the validity of methods for the purpose of deriving race or ethnicity is difficult as classification models differ not only in approach but also in the definition of the classification of race or ethnicity itself [, , ]. There is also a distinct lack of evaluation or validation of the methods used. Those that measured the performance of the methods used found accuracy to vary from 45% to 93%, with significantly lower accuracy in identifying categories of people of color.
This review sheds little light on the performance of commercial software packages. Previous empirical comparisons of facial recognition application programming interfaces have found that Face++ achieves 93% accuracy  and works comparatively better for men with lighter skins [ ]. The studies included in our review suggested a lower accuracy. However, data on accuracy were not forthcoming in any of the included studies using Demographics Pro [ ]. Even when performance is assessed, the methodology used may be biased if there are issues with the gold standard used to train the model.
In addition to the 4 overarching methods used, the studies varied in terms of the features used to determine or define race or ethnicity. Furthermore, the reliability of the features used to determine or define race or ethnicity for this purpose is questionable. Specifically, the use of Twitter users’ profile pictures, names, and locations, the use of unvalidated linguistic features attributed to racial groups (such as slang words, African American vernacular English, Spanglish, or Multicultural London English), and the use of training data that are prone to perpetuate biases (eg, police booking photos or mug shots) were all of particular concern.
Issues Related to the Methods Used
Approaches that include or rely solely on profile pictures to determine race or ethnicity can introduce bias. First, not all users have a photograph as their profile picture, nor is it easy to determine whether the picture used is that of the user. A study on the feasibility of using Face++ found that only 30.8% of Twitter users had a detectable single face in their profile. A manual review of automatically detected faces determined that 80% could potentially be of the user (ie, not a celebrity) . Human annotation may introduce additional bias, and studies have found systematic biases in the classification of people into racial or ethnic groups based on photographs [ , ]. Furthermore, humans tend to perceive their own race more readily than others [ , ]. Thus, race or ethnicity in the annotation team has an impact on the accuracy of their race or ethnicity labels, potentially skewing the sample labels toward the race or ethnicity of the annotators [ , ]. Given ML and NLP methods are trained on these data sets, the human biases transfer to automated methods, leading to poorly supervised ML and training, which has been shown to result in discrimination by the algorithm [ - ]. These concerns did not appear to be interrogated by the study designers. Without exception, they present categorization of persons into race or ethnicity, assuming that a subjective reading of facial features or idiomatic speech is the gold standard both for coding of race or ethnicity and for training and evaluation of automated methods.
Other methods, such as using geography or names as indicators of race, may also be unreliable. One could argue that the demographic profile for a geographic region is a better representation of race or ethnicity in the demographic environment than an individual’s race or ethnicity. Problems in using postcodes or locations to decipher individual social determinants are well documented . The use of census data from an area that is too large may skew the results. Among the studies reviewed, some used census block data, which are granular, whereas others extrapolated from larger areas, such as city- or county-level data. For example, Saravanan [ ] inferred the demographics of users in a city as a certain ethnic group based on a city with a large population of that group; however, no fine-grained analysis was performed either for the city chosen or for geolocation of the Twitter user. Thus, the validity of their assumption that a user in Los Angeles County is of Mexican descent [ ] is questionable. As these data were then used to create a race or ethnicity dictionary of terms used by that group to train their model, the questionable assumption further taints downstream applications and results. The models also do not consider the differences between the demographics of Twitter users and the general demographics of the population.
In addition, census demographic data that uses names are also questionable because of name-taking in marriage and indiscernible names.
The practice of using a Twitter user’s self-reported race or ethnicity would provide a label with high confidence but restrict the amount of usable data and introduce a margin of error depending on the method used to extract such self-reports. For example, in a sample of 14 million users, >0.1% matched precise regular expressions created to detect self-reported race or ethnic identity . Another study used mentions of keywords related to race or ethnicity in a user’s bio; however, limited validation was conducted to ensure that the mention was actually related to the user’s race or ethnicity [ , ]. This lack of information gathered from the profile information leads to sampling bias in the training of the models [ ].
Some models trained on manually annotated data did not have high interannotator agreement; for example, Chen et al  crowdsourced annotation agreement measured at 0.45. This can be interpreted as weak agreement, with the percentage of reliable data being 15% to 35% [ ]. Training a model on such weakly labeled data produces uncertain results.
It is not possible to assume the accuracy of black box proprietary tools and algorithms. The only race or ethnicity measure that seems empirically reliable is self-report, but this has considerable limitations. Thus, faulty methods continue to underpin digital health research, and researchers are likely to become increasingly dependent on them. The gold standard data required to know the demographic characteristics of the Twitter user is difficult to ascertain.
Strengths and Limitations
We did not limit our database searches and other methods by study design; however, we were unable to identify any previous reviews on the subject. To the best of our knowledge, this is the first review of methods used to extract race or ethnicity from social media. We identified studies from a range of disciplines and sources and categorized and summarized the methods used. However, we were unable to obtain information on the methodologies used by private-sector companies that created software for this purpose. Marketing and targeted advertising are common on social media and are likely to use race as a part of their algorithms to derive target users.
We did not limit our included papers to those in which the extraction of race or ethnicity was the primary focus. Although this can be conceived as a strength, it also meant that reporting of the methods used was often poor. The accurate recreation of the data lost was hampered by not knowing how decisions were made in the original studies, including what demographic definitions of race or ethnicity were used, or how accuracy was determined. This limited the assessment of the included studies. Few studies have validated the methods or conducted an error analysis to assess how often race is misapplied and those that did, rarely used the most appropriate gold standard. This makes it difficult to directly compare the results of the different approaches.
Future studies should investigate their methodological approaches to estimate race or ethnicity, offering careful interpretations that acknowledge the significant limits of these approaches and their impact on the interpretation of the results. This may include reporting the results as a range that communicates the inherent uncertainty of the classification model. Social media data may best be used in combination with other information. In addition, we must always be mindful that race is a proxy measure for the much larger impact of being a particular race or ethnicity in a society. As a result, the variability associated with race and ethnicity might reveal more about the effects of racism and social stratification than about individual user attributes. To conduct this study ethically and rigorously, we recommend several practices that can help reduce bias and increase reproducibility.
We recommend acknowledging the researchers’ bias that can influence the conceptualization of the implementation of the study. Incorporating this reflexivity, as is common in qualitative research, allows for the identification of potential blind spots that weaken the research. One way to address homogenous research teams is through the inclusion of experts in race or ethnicity or in those communities being examined. These biases can also be reduced by including members of the study population in the research process as experts and advisers . Although big data from social media can be collected without ever connecting with the people who contributed the data, it does not eliminate the ethical need for researchers to include representative perspectives in research processes. Examples of patient-engaged research and patient-centered outcomes research, community-based participatory research, and citizen science (public participation in scientific research) within the health and social sciences amply demonstrate the instrumental value and ethical obligation of intentional efforts to involve nonscientist partners in cocreation of research [ ]. The quality of data science can be improved by seriously heeding the imperative, Nothing about us without us [ ]. Documenting and establishing the diverse competence attributes of a research team should become a standard. Emphasizing the importance of diverse teams within the research process will contribute to social and racial justice in ways other than improving the reliability of research.
In terms of the retrieved data, the most reliable (though imperfect) method for ascertaining race was when users self-identified their racial affiliation. Further research on overcoming the limitations of availability and sample size may be warranted. Indeed, a hybrid model with automated methods and manual extraction may be preferred. For example, automation methods could be developed to identify potential self-declarations in a user profile or timeline, which can then be manually interpreted.
Finally, we call for greater reporting of the validation by our colleagues. Without error analysis, computational techniques would not be able to detect bias. Further research is needed to establish whether any bias is systematic or random, that is, whether inaccuracies favor one direction or another.
We identified major concerns that affect the reliability of the methods and bias the results. There are also ethical concerns throughout the process, particularly regarding the inference of race or ethnicity, as opposed to the extraction of self-identity. However, the potential usefulness of social media research requires thoughtful consideration of the best ways to estimate demographic characteristics such as race and ethnicity . This is particularly important, given the increased access to Twitter data [ , ].
Therefore, we propose several approaches to improve the extraction of race or ethnicity from social media, including representative research teams and a mixture of manual and computational methods, as well as future research on methods to reduce bias.
This work was supported by the National Institutes of Health (NIH) National Library of Medicine under grant NIH-NLM 1R01 (principal investigator: GG, with coapplicants KO and SG) and NIH National Institute of Drug Abuse grant R21 DA049572-02 to RS. NIH National Library of Medicine funded this research but was not involved in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; or decision to submit the manuscript for publication.
The included studies are available on the web, and the extracted data are presented in Table S2 in. A preprint of this paper is also available: Golder S, Stevens R, O’Connor K, James R, Gonzalez-Hernandez G. 2021. Who Is Tweeting? A Scoping Review of Methods to Establish Race and Ethnicity from Twitter Datasets. SocArXiv. February 14. doi:10.31235/osf.io/wru5q.
SG, RS, KO, RJ, and GG contributed equally to the study. RS and GG proposed the topic and the main idea. SG and RJ were responsible for literature search. SG, RS, KO, RJ, and GG were responsible for study selection and data extraction. SG drafted the manuscript. SG, RS, KO, RJ, and GG commented on and revised the manuscript. SG provided the final version of this manuscript. All authors contributed to the final draft of the manuscript.
Conflicts of Interest
Search strategies and characteristics of included studies.DOCX File , 59 KB
- Golder S, Norman G, Loke YK. Systematic review on the prevalence, frequency and comparative value of adverse events data in social media. Br J Clin Pharmacol 2015 Oct;80(4):878-888. [CrossRef] [Medline]
- Sarker A, Ginn R, Nikfarjam A, O'Connor K, Smith K, Jayaraman S, et al. Utilizing social media data for pharmacovigilance: a review. J Biomed Inform 2015 Apr;54:202-212 [FREE Full text] [CrossRef] [Medline]
- Bhattacharya M, Snyder S, Malin M, Truffa MM, Marinic S, Engelmann R, et al. Using social media data in routine pharmacovigilance: a pilot study to identify safety signals and patient perspectives. Pharm Med 2017 Apr 17;31(3):167-174. [CrossRef]
- Convertino I, Ferraro S, Blandizzi C, Tuccori M. The usefulness of listening social media for pharmacovigilance purposes: a systematic review. Expert Opin Drug Saf 2018 Nov;17(11):1081-1093. [CrossRef] [Medline]
- Golder S, Smith K, O'Connor K, Gross R, Hennessy S, Gonzalez-Hernandez G. A comparative view of reported adverse effects of statins in social media, regulatory data, drug information databases and systematic reviews. Drug Saf 2021 Feb 01;44(2):167-179 [FREE Full text] [CrossRef] [Medline]
- Bychkov D, Young S. Social media as a tool to monitor adherence to HIV antiretroviral therapy. J Clin Transl Res 2018 Dec 17;3(Suppl 3):407-410 [FREE Full text] [Medline]
- Kalf RR, Makady A, Ten HR, Meijboom K, Goettsch WG, IMI-GetReal Workpackage 1. Use of social media in the assessment of relative effectiveness: explorative review with examples from oncology. JMIR Cancer 2018 Jun 08;4(1):e11 [FREE Full text] [CrossRef] [Medline]
- Golder S, O'Connor K, Hennessy S, Gross R, Gonzalez-Hernandez G. Assessment of beliefs and attitudes about statins posted on Twitter: a qualitative study. JAMA Netw Open 2020 Jun 01;3(6):e208953 [FREE Full text] [CrossRef] [Medline]
- Golder S, Bach M, O'Connor K, Gross R, Hennessy S, Gonzalez Hernandez G. Public perspectives on anti-diabetic drugs: exploratory analysis of Twitter posts. JMIR Diabetes 2021 Jan 26;6(1):e24681 [FREE Full text] [CrossRef] [Medline]
- Hswen Y, Naslund JA, Brownstein JS, Hawkins JB. Monitoring online discussions about suicide among Twitter users with schizophrenia: exploratory study. JMIR Ment Health 2018 Dec 13;5(4):e11483 [FREE Full text] [CrossRef] [Medline]
- Howie L, Hirsch B, Locklear T, Abernethy AP. Assessing the value of patient-generated data to comparative effectiveness research. Health Aff (Millwood) 2014 Jul;33(7):1220-1228. [CrossRef] [Medline]
- Cavazos-Rehg PA, Krauss MJ, Costello SJ, Kaiser N, Cahn ES, Fitzsimmons-Craft EE, et al. "I just want to be skinny.": a content analysis of tweets expressing eating disorder symptoms. PLoS One 2019;14(1):e0207506 [FREE Full text] [CrossRef] [Medline]
- Ahmed W, Bath PA, Sbaffi L, Demartini G. Novel insights into views towards H1N1 during the 2009 Pandemic: a thematic analysis of Twitter data. Health Info Libr J 2019 Mar;36(1):60-72 [FREE Full text] [CrossRef] [Medline]
- Cook N, Mullins A, Gautam R, Medi S, Prince C, Tyagi N, et al. Evaluating patient experiences in dry eye disease through social media listening research. Ophthalmol Ther 2019 Sep;8(3):407-420 [FREE Full text] [CrossRef] [Medline]
- Roccetti M, Salomoni P, Prandi C, Marfia G, Mirri S. On the interpretation of the effects of the Infliximab treatment on Crohn’s disease patients from Facebook posts: a human vs. machine comparison. Netw Model Anal Health Inform Bioinforma 2017 Jun 26;6(1):10.1007/s13721-017-0152-y. [CrossRef]
- Madden ML, Cortesi S, Gasser U, Duggan M, Smith A, Beaton M. Teens, social media, and privacy. Pew Internet & American Life Project. 2013. URL: http://www.pewinternet.org/2013/05/21/teens-social-media-and-privacy/ [accessed 2022-04-19]
- Chou WS, Hunt YM, Beckjord EB, Moser RP, Hesse BW. Social media use in the United States: implications for health communication. J Med Internet Res 2009;11(4):e48 [FREE Full text] [CrossRef] [Medline]
- Social media use in 2018. Pew Research Center. URL: https://www.pewresearch.org/internet/2018/03/01/social-media-use -in-2018/ [accessed 2022-04-19]
- Bowleg L, Teti M, Malebranche DJ, Tschann JM. "It's an Uphill Battle Everyday": intersectionality, low-income black heterosexual men, and implications for hiv prevention research and interventions. Psychol Men Masc 2013 Jan 1;14(1):25-34 [FREE Full text] [CrossRef] [Medline]
- Sloan L. Social Science 'Lite'? Deriving demographic proxies from Twitter. In: The SAGE Handbook of Social Media Research Methods. Thousand Oaks, California: SAGE Publications; 2016.
- McCormick TH, Lee H, Cesare N, Shojaie A, Spiro ES. Using Twitter for demographic and social science research: tools for data collection and processing. Sociol Methods Res 2017 Aug;46(3):390-421 [FREE Full text] [CrossRef] [Medline]
- Smedley A, Smedley BD. Race as biology is fiction, racism as a social problem is real: anthropological and historical perspectives on the social construction of race. Am Psychol 2005;60(1):16-26. [CrossRef]
- Yudell M, Roberts D, DeSalle R, Tishkoff S. NIH must confront the use of race in science. Science 2020 Sep 10;369(6509):1313-1314. [CrossRef]
- Davenport L. The fluidity of racial classifications. Annu Rev Polit Sci 2020 May 11;23(1):221-240. [CrossRef]
- Resident population and net change. U.S. Census Bureau. URL: https://www.census.gov/quickfacts/fact/note/US/RHI625219 [accessed 2022-04-19]
- Zuberi T. Thicker Than Blood How Racial Statistics Lie. Minneapolis: University of Minnesota Press; 2001.
- Hardeman RR, Karbeah J. Examining racism in health services research: a disciplinary self‐critique. Health Serv Res 2020 Sep 25;55(S2):777-780. [CrossRef]
- Jenkins W, Schoenbach V, Rowley D, Ford C. 2. Overcoming the impact of racism on the health of communities: what we have learned and what we have not. In: Racism: Science & Tools for the Public Health Professional. Washington, D.C: American Public Health Association; 2019.
- Jones CP. Toward the science and practice of anti-racism: launching a national campaign against racism. Ethn Dis 2018 Aug 08;28(Supp 1):231. [CrossRef]
- Tricco AC, Lillie E, Zarin W, O'Brien KK, Colquhoun H, Levac D, et al. PRISMA Extension for Scoping Reviews (PRISMA-ScR): checklist and explanation. Ann Intern Med 2018 Oct 02;169(7):467-473. [CrossRef] [Medline]
- Updated guidance for trusted systematic reviews: a new edition of the Cochrane Handbook for Systematic Reviews of Interventions. Cochrane Database of Systematic Reviews. URL: https://www.cochranelibrary.com/cdsr/doi/10.1002/14651858.ED000142/full [accessed 2022-04-21]
- An J, Weber I. # greysanatomy vs # yankees: demographics and hashtag use on Twitter. arXiv 2016 [FREE Full text] [CrossRef]
- Lightly supervised machine learning for classifying online social data. ProQuest. 2017. URL: https://www.proquest.com/openview/25ccbef5caa83249e9dc363bdb196827/1?pq-origsite=gscholar&cbl=18750 [accessed 2022-04-21]
- Ardehaly E, Culotta A. Inferring latent attributes of Twitter users with label regularization. In: Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2015 Presented at: 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies; 2015; Denver, Colorado. [CrossRef]
- Ardehaly E, Culotta A. Co-training for demographic classification using deep learning from label proportions. In: Proceedings of the IEEE International Conference on Data Mining Workshops, ICDMW. 2017 Presented at: IEEE International Conference on Data Mining Workshops, ICDMW; Nov 18-21, 2017; Sorrento, Italy URL: https://doi.org/10.1109/ICDMW.2017.144 [CrossRef]
- Auguste D, Polman J, Miller S. A data science approach to STEM (science, technology, engineering and math) identity research for African American communities. ProQuest. URL: https://www.proquest.com/openview/ee3643a744b7c01262 cb9a917611f812/1.pdf?pq-origsite=gscholar&cbl=18750&diss=y [accessed 2022-04-21]
- Barbera P. Less is more? How demographic sample weights can improve public opinion estimates based on Twitter data. Work Paper NYU. 2017. URL: http://pablobarberacom/static/less-is-morepdf [accessed 2022-04-19]
- Bergsma S, Dredze M, Van Durme B, Wilson T, Yarowsky D. Broadly improving user classification via communication-based name and location clustering on Twitter. In: Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2013 Presented at: 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies; Jun, 2013; Atlanta, Georgia.
- Blodgett S, Wei J, O'Connor B. Twitter Universal Dependency Parsing for African-American and Mainstream American English. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2018 Presented at: 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers); Jul 15 - 20, 2018; Melbourne, Australia. [CrossRef]
- Borradaile G, Burkhardt B, LeClerc A. Whose tweets are surveilled for the police: an audit of a social-media monitoring tool via log files. In: Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency. 2020 Presented at: FAT* '20: Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency; Jan 27 - 30, 2020; Barcelona Spain. [CrossRef]
- Cavazos-Rehg P, Krauss M, Grucza R, Bierut L. Characterizing the followers and tweets of a marijuana-focused Twitter handle. J Med Internet Res 2014;16(6):e157 [FREE Full text] [CrossRef] [Medline]
- Cavazos-Rehg PA, Krauss M, Fisher SL, Salyer P, Grucza RA, Bierut LJ. Twitter chatter about marijuana. J Adolesc Health 2015 Feb;56(2):139-145. [CrossRef] [Medline]
- Cavazos-Rehg PA, Zewdie K, Krauss MJ, Sowles SJ. "No high like a brownie high": a content analysis of edible marijuana Tweets. Am J Health Promot 2018 May;32(4):880-886. [CrossRef] [Medline]
- United we tweet?: a quantitative analysis of racial differences in twitter use. ResearchWorks Archive. URL: https://digital.lib.washington.edu/researchworks/handle/1773/40971 [accessed 2022-04-21]
- Chakraborty A, Messiaso J, Benevenutoo F, Ghosh S, Ganguly N, Gummadi K. Who makes trends? Understanding demographic biases in crowdsourced recommendations. In: Proceedings of the 11th AAAI International Conference on Web and Social Media (ICWSM ). 2017 Presented at: 11th AAAI International Conference on Web and Social Media (ICWSM ); May 15-18, 2017; Montreal, Quebec, Canada.
- Chen X, Wang Y, Agichtein E, Wang F. A comparative study of demographic attribute inference in twitter. Proc Int AAAI Conf Web Social Media 2021;9(1):590-593 [FREE Full text]
- Culotta A, Ravi NK, Cutler J. Predicting Twitter user demographics using distant supervision from website traffic data. J Artificial Intell Res 2016 Feb 19;55:389-408. [CrossRef]
- De Choudhury M. Tie formation on Twitter: homophily and structure of egocentric networks. In: Proceedings of the 2011 IEEE Third International Conference on Privacy, Security, Risk and Trust and 2011 IEEE Third International Conference on Social Computing. 2011 Presented at: 2011 IEEE Third International Conference on Privacy, Security, Risk and Trust and 2011 IEEE Third International Conference on Social Computing; Oct 9-11, 2011; Boston, MA. [CrossRef]
- Firmansyah F, Jones J. Did the black panther movie make blacks blacker? Examining black racial identity on twitter before and after the black panther movie release. Lecture Notes Comput Sci 2019;11864:66-78. [CrossRef]
- Golder S, Chiuve S, Weissenbacher D, Klein A, O'Connor K, Bland M, et al. Pharmacoepidemiologic evaluation of birth defects from health-related postings in social media during pregnancy. Drug Saf 2019 Mar;42(3):389-400 [FREE Full text] [CrossRef] [Medline]
- González Y, Cutter S. Leveraging geotagged social media to monitor spatial behavior during population movements triggered by hurricanes. Scholar Common. URL: https://scholarcommons.sc.edu/etd/5367/ [accessed 2022-04-21]
- Haffner M. A spatial analysis of non-English Twitter activity in Houston, TX. Transact GIS 2018 Apr 11;22(4):913-929. [CrossRef]
- He L, Murphy L, Luo J. Using social media to promote STEM education: matching college students with role models. In: Machine Learning and Knowledge Discovery in Databases. Cham: Springer; 2016.
- Hswen Y, Hawkins JB, Sewalk K, Tuli G, Williams DR, Viswanath K, et al. Racial and ethnic disparities in patient experiences in the United States: 4-year content analysis of Twitter. J Med Internet Res 2020 Aug 21;22(8):e17048 [FREE Full text] [CrossRef] [Medline]
- Huang W, Weber I, Vieweg S. Inferring nationalities of Twitter users and studying inter-national linking. In: Proceedings of the 25th ACM conference on hypertext and social media. 2014 Presented at: HT '14: Proceedings of the 25th ACM conference on hypertext and social media; Sep 1-4, 2014; New York, NY, United States. [CrossRef]
- Huang X, Xing L, Dernoncourt F, Paul M. Multilingual Twitter corpus and baselines for evaluating demographic bias in hate speech recognition. In: 12th Language Resources and Evaluation Conference, European Language Resources Association. 2020 Presented at: Proceedings of the 12th Language Resources and Evaluation Conference, European Language Resources Association; May 11-16, 2020; Marseille, France.
- Karlsen AS, Scott KD. Making sense of Starbucks’ anti-bias training and the arrests of two African American men: a thematic analysis of Whites’ Facebook and Twitter comments. Discourse Context Media 2019 Dec;32:100332. [CrossRef]
- Kteily NS, Rocklage MD, McClanahan K, Ho AK. Political ideology shapes the amplification of the accomplishments of disadvantaged vs. advantaged group members. Proc Natl Acad Sci U S A 2019 Jan 29;116(5):1559-1568 [FREE Full text] [CrossRef] [Medline]
- Longley PA, Adnan M. Geo-temporal Twitter demographics. Int J Geographical Inf Sci 2015 Sep 24;30(2):369-389. [CrossRef]
- Luo F, Cao G, Mulligan K, Li X. Explore spatiotemporal and demographic characteristics of human mobility via Twitter: a case study of Chicago. Applied Geography 2016 May;70:11-25. [CrossRef]
- Markson C. Detecting user demographics in twitter to inform health trends in social media. New Jersey Institute of Technology. URL: https://digitalcommons.njit.edu/dissertations/36/ [accessed 2022-04-21]
- Messias J, Vikatos P, Benevenuto F. White, man, and highly followed: gender and race inequalities in Twitter. In: IEEE/WIC/ACM International Conference on Web Intelligence (WI'17). 2017 Presented at: Proceedings of the IEEE/WIC/ACM International Conference on Web Intelligence (WI'17); Aug 23-26, 2017; Leipzig, Germany. [CrossRef]
- Mislove A, Lehmann S, Ahn Y, Onnela JP, Rosenquist J. Understanding the Demographics of Twitter Users. In: Proceedings of the International AAAI Conference on Web and Social Media. 2011 Presented at: Fifth International AAAI Conference on Weblogs and Social Media; 2011; Barcelona Spain.
- Mohammady E. Using county demographics to infer attributes of Twitter users. In: Proceedings of the ACL Joint Workshop on Social Dynamics and Personal Attributes in Social Media. 2014 Presented at: ACL Joint Workshop on Social Dynamics and Personal Attributes in Social Media; Jun 27, 2014; Baltimore, Maryland. [CrossRef]
- Murthy D, Gross A, Pensavalle A. Urban social media demographics: an exploration of Twitter use in major American cities. J Comput Mediat Commun 2015 Nov 19;21(1):33-49. [CrossRef]
- Nguyen V, Tran M, Luo J. Are French really that different? Recognizing Europeans from faces using data-driven learning. In: Proceedings of the 2018 24th International Conference on Pattern Recognition (ICPR). 2018 Presented at: 24th International Conference on Pattern Recognition (ICPR); Aug 20-24, 2018; Beijing, China. [CrossRef]
- Pennacchiotti M, Popescu AM. Democrats, republicans and starbucks afficionados: user classification in twitter. In: Proceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining. 2011 Presented at: KDD '11: Proceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining; Aug 21- 24, 2011; KDD '11: San Diego, CA, USA. [CrossRef]
- Pennacchiotti M, Popescu AM. A machine learning approach to twitter user classification. In: Proc International AAAI Conference Web Social Media. 2011 Jul Presented at: Proceedings of the Fifth International Conference on Weblogs and Social Media; Jul 17-21, 2011; Barcelona, Catalonia, Spain.
- Reis J, Kwak H, An J, Messias J, Benevenuto F. Demographics of news sharing in the U.S. Twittersphere. In: Proceedings of the 28th ACM Conference on Hypertext and Social Media. 2017 Presented at: HT '17: Proceedings of the 28th ACM Conference on Hypertext and Social Media; Jul, 2017; Prague, Czech Republic. [CrossRef]
- Sadah SA, Shahbazi M, Wiley MT, Hristidis V. A study of the demographics of web-based health-related social media users. J Med Internet Res 2015;17(8):e194 [FREE Full text] [CrossRef] [Medline]
- Sadah SA, Shahbazi M, Wiley MT, Hristidis V. Demographic-based content analysis of web-based health-related social media. J Med Internet Res 2016 Jun 13;18(6):e148 [FREE Full text] [CrossRef] [Medline]
- Saravanan M. Determining Ethnicity of Immigrants using Twitter Data. In: Proceedings of the 4th Multidisciplinary International Social Networks Conference. 2017 Presented at: MISNC '17: Proceedings of the 4th Multidisciplinary International Social Networks Conference : Association for Computing Machinery; Jul 17 - 19, 2017; Bangkok, Thailand. [CrossRef]
- Sowles SJ, Krauss MJ, Connolly S, Cavazos-Rehg PA. A content analysis of vaping advertisements on Twitter, November 2014. Prev Chronic Dis 2016 Sep 29;13:E139 [FREE Full text] [CrossRef] [Medline]
- Stewart I. Now we stronger than ever: African-American English syntax in twitter. In: Student Research Workshop at the 14th Conference of the European Chapter of the Association for Computational Linguistics. 2014 Presented at: Proceedings of the Student Research Workshop at the 14th Conference of the European Chapter of the Association for Computational Linguistics; 2014; Gothenburg, Sweden. [CrossRef]
- Vikatos P, Messias J, Manoel M, Benevenuto F. Linguistic diversities of demographic groups in twitter. In: Proceedings of the 28th ACM Conference on Hypertext and Social Media. 2017 Presented at: HT '17: Proceedings of the 28th ACM Conference on Hypertext and Social Media; Jul 4-7, 2017; Prague, Czech Republic. [CrossRef]
- Volkova S, Backrach Y. Inferring perceived demographics from user emotional tone and user-environment emotional contrast. In: 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2018 Presented at: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers); Aug, 2018; Berlin, Germany. [CrossRef]
- Wang W, Chi G. Who are you? Estimating demographics of twitter users. PAA. URL: http:///C:/Users/user/Downloads/PAA _2017_Twitter_edited%20(1).pdf [accessed 2022-04-21]
- Wang Y, Li Y, Luo J. Deciphering the 2016 U.S. Presidential Campaign in the Twitter sphere: a comparison of the Trumpists and Clintonists. In: Proceedings of the 10th International AAAI Conference on Web and Social Media. 2016 Presented at: 10th International AAAI Conference on Web and Social Media; May 17-20, 2016; Cologne, Germany.
- Wood-Doughty Z, Andrews N, Marvin R, Dredze M. Predicting Twitter User Demographics from Names Alone. In: Proceedings of the Second Workshop on Computational Modeling of People's Opinions, Personality, and Emotions in Social Media, Association for Computational Linguistics. 2018 Presented at: Proceedings of the Second Workshop on Computational Modeling of People's Opinions, Personality, and Emotions in Social Media, Association for Computational Linguistics; Jun 6, 2018; New Orleans, Louisiana. [CrossRef]
- Wood-Doughty Z, Smith M, Broniatowski D, Dredze M. How Does Twitter User Behavior Vary Across Demographic Groups? In: Proceedings of the Second Workshop on Natural Language Processing and Computational Social Science, Association for Computational Linguistics. 2017 Presented at: Second Workshop on Natural Language Processing and Computational Social Science, Association for Computational Linguistics; Aug 3, 2017; Vancouver, Canada. [CrossRef]
- Wood-Doughty Z, Xu P, Liu X, Dredze M. Using noisy self-reports to predict twitter user demographics. In: Proceedings of the Ninth International Workshop on Natural Language Processing for Social Media, Association for Computational Linguistics. 2021 Presented at: Ninth International Workshop on Natural Language Processing for Social Media, Association for Computational Linguistics; 2021; Online. [CrossRef]
- Xu S, Markson C, Costello KL, Xing CY, Demissie K, Llanos AA. Leveraging social media to promote public health knowledge: example of cancer awareness via Twitter. JMIR Public Health Surveill 2016;2(1):e17 [FREE Full text] [CrossRef] [Medline]
- Ye J, Han S, Hu Y, Coskun B, Liu M, Qin H. Nationality classification using name embeddings. In: Proceedings of the 2017 ACM on Conference on Information and Knowledge Management. 2017 Presented at: CIKM '17: Proceedings of the 2017 ACM on Conference on Information and Knowledge Management; Nov, 2017; Singapore p. 1897-1906. [CrossRef]
- Yin J, Chi G, Hook J. Evaluating the representativeness in the geographic distribution of twitter user population. In: Proceedings of the 12th Workshop on Geographic Information Retrieval. 2018 Presented at: GIR'18: Proceedings of the 12th Workshop on Geographic Information Retrieval; Nov 6, 2018; WA, USA. [CrossRef]
- Automated analysis of user-generated content on the web. ProQuest. 2021. URL: https://www.proquest.com/openview/d035cbe40f3a459a9aa347273139233f/1?pq-origsite=gscholar&cbl=18750&diss=y [accessed 2022-04-21]
- Adnan M, Longley PA, Khan SM. Social dynamics of Twitter usage in London, Paris, and New York City. First Monday 2014;19(5). [CrossRef]
- Coleman LS. “we’re a part of this city, too”: an examination of the politics of representation of D.C. Native via #dcnativesday. Social Media Soc 2021 Jan 21;7(1):205630512098444. [CrossRef]
- Saha K, Yousuf A, Hickman L, Gupta P, Tay L, De Choudhury M. A social media study on demographic differences in perceived job satisfaction. Proc ACM Human Comput Interaction (HCI) 2021 Apr 13;5(CSCW1):1-29. [CrossRef]
- Hong T, Wu J, Wijaya D, Xuan Z, Fetterman J. JUUL the heartbreaker: Twitter analysis of cardiovascularhealth perceptions of vaping. Tobacco Induced Diseases 2021 Jan 8;19(January):1-6. [CrossRef]
- Jiang J, Vosoughi S. Not judging a user by their cover: understanding harm in multi-modal processing within social media research. In: Proceedings of the 2nd International Workshop on Fairness, Accountability, Transparency and Ethics in Multimedia: ACM. 2020 Presented at: 2nd International Workshop on Fairness, Accountability, Transparency and Ethics in Multimedia: ACM; Oct 12, 2020; Seattle WA USA. [CrossRef]
- Mueller A, Wood-Doughty Z, Amir S, Dredze M, Nobles AL. Demographic representation and collective storytelling in the me too twitter hashtag activism movement. Proc ACM Human Comput Interaction 2020;5(CSCW1):1-28. [CrossRef]
- Aguirre C, Harrigian K, Dredze M. Gender and racial fairness in depression research using social media. In: Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics. 2021 Presented at: 16th Conference of the European Chapter of the Association for Computational Linguistics; Apr 19-23, 2021; Online. [CrossRef]
- Aguirre C, Dredze M. Qualitative analysis of depression models by demographics. In: Proceedings of the Seventh Workshop on Computational Linguistics and Clinical Psychology: Improving Access. 2021 Presented at: Seventh Workshop on Computational Linguistics and Clinical Psychology: Improving Access; 2021; Online. [CrossRef]
- Gunarathne P, Rui H, Seidmann A. Racial Discrimination in Social Media Customer Service: Evidence from a Popular Microblogging Platform. Manoa: University of Hawaii; 2019.
- Ye J, Skiena S. The Secret Lives of Names? Name Embeddings from Social Media. In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2019 Presented at: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining; Aug 4 - 8, 2019; Anchorage, AK. [CrossRef]
- Preotiuc-Pietro D, Ungar L. User-level race and ethnicity predictors from twitter text. In: Proceedings of the The 27th International Conference on Computational Linguistics; Association for Computational Linguistics. 2018 Presented at: The 27th International Conference on Computational Linguistics; Association for Computational Linguistics; Aug 20-26, 2018; Santa Fe, New Mexico, USA URL: https://wwwaclweborg/anthology/C18-1130
- Hinds J, Joinson AN. What demographic attributes do our digital footprints reveal? A systematic review. PLoS One 2018;13(11):e0207112 [FREE Full text] [CrossRef] [Medline]
- Abubakar U, Bashir SA, Abdullahi MB, Adebayo OS. Comparative study of various machine learning algorithms for tweet classification. J Comput Sci 2019;6(4):12-24. [CrossRef]
- Ardehaly E, Culotta A, Raghavan V, Aluru S, Karypis G, Miele L. Mining the demographics of political sentiment from twitter using learning from label proportions. In: Proceedings of the IEEE International Conference on Data Mining Workshops, ICDMW. 2017 Presented at: IEEE International Conference on Data Mining Workshops, ICDMW; Nov 18-21, 2017; New Orleans, Louisiana, USA. [CrossRef]
- An J, Ciampaglia GL, Grinberg N, Joseph K, Mantzarlis A, Maus G, et al. Reports of the workshops held at the 2017 international AAAI conference on web and social media. AI Magazine 2017 Dec 28;38(4):93-98. [CrossRef]
- Anindya I. Understanding and mitigating privacy risks raised by record linkage. The University of Texas at Dallas. 2020. URL: https://utd-ir.tdl.org/handle/10735.1/9373 [accessed 2022-04-21]
- Bardier C. Detecting electronic cigarette user disparity behaviors: an infovelliance study on twitter. ProQuest. URL: https://www.proquest.com/openview/dbcad596abc1e82eb6718d504134ec17/1?pq-origsite=gscholar&cbl=18750&diss=y [accessed 2022-04-21]
- Basterra L, Worthington T, Rogol J, Brown D. Socio-Temporal Trends in Urban Cultural Subpopulations through Social Media. New York: IEEE; 2017.
- Beretta V, Maccagnola D, Cribbin T, Messina E. An interactive method for inferring demographic attributes in twitter. In: Proceedings of the 26th ACM Conference on Hypertext & Social Media. 2015 Presented at: HT '15: Proceedings of the 26th ACM Conference on Hypertext & Social Media; Sep 1-4, 2015; Guzelyurt Northern Cyprus. [CrossRef]
- Bergsma S, Van Durme B. Using conceptual class attributes to characterize social media users. In: Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2013 Presented at: Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers); Aug 4-9, 2013; Sofia, Bulgaria.
- Bi B, Shokouhi M, Kosinski M, Graepel T. Inferring the demographics of search user: social data meets search queries. In: Proceedings of the International World Wide Web Conference Committee (IW3C2). 2013 Presented at: International World Wide Web Conference Committee (IW3C2); May 13-17, 2013; Rio de Janeiro Brazil. [CrossRef]
- Blevins T, Kwiatkowski R, Macbeth J, McKeown K, Patton D, Rambow O. Automatically processing tweets from gang-involved youth: towards detecting loss and aggression. In: COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers. 2016 Presented at: COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers; Dec 11-16, 2016; Osaka, Japan.
- Blodgett S, Wei J, O'Connor B. A dataset and classifier for recognizing social media english. In: Proceedings of the 3rd Workshop on Noisy User-generated Text: Association for Computational Linguistics. 2017 Presented at: 3rd Workshop on Noisy User-generated Text: Association for Computational Linguistics; Sep, 2017; Copenhagen, Denmark p. 56-61. [CrossRef]
- Bokányi E, Kondor D, Dobos L, Sebők T, Stéger J, Csabai I, et al. Race, religion and the city: twitter word frequency patterns reveal dominant demographic dimensions in the United States. Palgrave Commun 2016 Apr 26;2(1):1-9. [CrossRef]
- Racial identities on social media: projecting racial identities on Facebook, Instagram, and Twitter. Minnesota State University. URL: https://cornerstone.lib.mnsu.edu/etds/781/ [accessed 2022-04-21]
- Burnap P, Colombo G, Amery R, Hodorog A, Scourfield J. Multi-class machine classification of suicide-related communication on Twitter. Online Soc Netw Media 2017 Aug;2:32-44 [FREE Full text] [CrossRef] [Medline]
- Cesare N, Grant C, Nguyen Q, Lee H, Nsoesie E. How well can machine learning predict demographics of social media users? arXiv 2017:1-24.
- Chan MS, Winneg K, Hawkins L, Farhadloo M, Jamieson KH, Albarracín D. Legacy and social media respectively influence risk perceptions and protective behaviors during emerging health threats: a multi-wave analysis of communications on Zika virus cases. Soc Sci Med 2018 Sep;212:50-59 [FREE Full text] [CrossRef] [Medline]
- Chenworth M, Perrone J, Love JS, Greller HA, Sarker A, Chai PR. Buprenorphine initiation in the emergency department: a thematic content analysis of a #firesidetox Tweetchat. J Med Toxicol 2020 Jul;16(3):262-268. [CrossRef] [Medline]
- Cheong M, Lee V. Integrating web-based intelligence retrieval and decision-making from the twitter trends knowledge base. In: Proceedings of the 2nd ACM workshop on Social web search and mining. 2009 Presented at: SWSM '09: Proceedings of the 2nd ACM workshop on Social web search and mining; Nov 2, 2009; Hong Kong, China p. 1-8. [CrossRef]
- Chi G, Giles L, Kifer D, Van Hook J, Yin J. Predicting twitter user demographics as a first step in big data for population research: developing unsupervised, scalable methods using real-time, large-scale twitter data. In: Proceedings of the 2017 International Population Conference. 2017 Presented at: 2017 International Population Conference; Oct 29 -Nov 3, 2017; Cape Town, South Africa.
- Claude F, Konow R, Ladra S. Fast compressed-based strategies for author profiling of social media texts. In: Proceedings of the 4th Spanish Conference on Information Retrieval. 2016 Presented at: CERI '16: Proceedings of the 4th Spanish Conference on Information Retrieval; Jun 14-16, 2016; Granada Spain. [CrossRef]
- Compton R, Lee C, Lu T, De Silva L, Macy M. Detecting future social unrest in unprocessed Twitter data: “emerging phenomena and big data”. In: Proceedings of the 2013 IEEE International Conference on Intelligence and Security Informatics. 2013 Presented at: 2013 IEEE International Conference on Intelligence and Security Informatics; Jun 4-7, 2013; Seattle, WA, USA. [CrossRef]
- Augmenting household travel survey and travel behavior analysis using large-scale social media data and smartphone GPS data. ProQuest. URL: https://www.proquest.com/openview/30e2c6f084eb32378522fe9929604037/1?pq-origsite=gscholar &cbl=18750&diss=y [accessed 2022-04-21]
- Dai H, Hao J. Mining social media data for opinion polarities about electronic cigarettes. Tob Control 2017 Mar;26(2):175-180. [CrossRef] [Medline]
- Daughton AR, Paul MJ. Identifying protective health behaviors on Twitter: observational study of travel advisories and Zika virus. J Med Internet Res 2019 May 13;21(5):e13090 [FREE Full text] [CrossRef] [Medline]
- DeJohn AD, Schulz EE, Pearson AL, Lachmar EM, Wittenborn AK. Identifying and understanding communities using twitter to connect about depression: cross-sectional study. JMIR Ment Health 2018 Nov 05;5(4):e61 [FREE Full text] [CrossRef] [Medline]
- Diaz F, Gamon M, Hofman JM, Kıcıman E, Rothschild D. Online and social media data as an imperfect continuous panel survey. PLoS One 2016;11(1):e0145406 [FREE Full text] [CrossRef] [Medline]
- Using social media to evaluate public acceptance of infrastructure projects. Digital Repository at University of Maryland. 2018. URL: https://drum.lib.umd.edu/handle/1903/20848 [accessed 2022-04-21]
- Eisenstein J. Phonological factors in social media writing. In: Proceedings of the Workshop on Language in Social Media (LASM 2013). 2013 Presented at: Proceedings of the Workshop on Language Analysis in Social Media; Jun 13, 2013; Atlanta, Georgia URL: https://aclanthology.org/W13-1102.pdf
- A case study on Black Twitter's reactions to the framing of blacks in Dove's 2017 Facebook advertisement. Digital Communs @ University of South Florida. URL: https://digitalcommons.usf.edu/etd/8446/ [accessed 2022-04-21]
- Filho R, Almeida J, Pappa G. Twitter Population Sample Bias and its impact on predictive outcomes: a case study on elections. In: Proceedings of the 2015 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining 2015. 2015 Presented at: IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM); Aug 25-28, 2015; Paris France. [CrossRef]
- Filippova K. User demographics and language in an implicit social network. In: Proceedings of the Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning. 2012 Presented at: Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning; Jul 12-14, 2012; Jeju Island, Korea URL: https://aclanthology.org/D12-1135/
- A world of one-way and two-way streetsxploring the nuances of fan-athlete interaction on Twitter. Indiana University. URL: https://scholarworks.uark.edu/cgi/viewcontent.cgi?article=3185&context=etd [accessed 2022-04-21]
- Georgiou T, Abbadi A, Yan X. Privacy cyborg: towards protecting the privacy of social media users. In: Proceedings of the 2017 IEEE 33rd International Conference on Data Engineering (ICDE). 2017 Presented at: 2017 IEEE 33rd International Conference on Data Engineering (ICDE); Apr 19-22, 2017; San Diego, CA, USA. [CrossRef]
- Ghazouani D, Lancieri L, Ounelli H, Jebari C. Assessing socioeconomic status of Twitter users: a survey. In: Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP ). 2019 Presented at: International Conference on Recent Advances in Natural Language Processing (RANLP ); Sep 2-4, 2019; Varna, Bulgaria. [CrossRef]
- Gibbons J, Malouf R, Spitzberg B, Martinez L, Appleyard B, Thompson C, et al. Twitter-based measures of neighborhood sentiment as predictors of residential population health. PLoS One 2019;14(7):e0219550 [FREE Full text] [CrossRef] [Medline]
- Gilchrist-Herring N. An analysis of attitudes towards transgender individuals utilizing social media usage, ethnicity, gender, age range, and level of education. ProQuest. 2020. URL: https://www.proquest.com/openview/1275cca853a493adc44dad6ad 0f3d1ab/1?pq-origsite=gscholar&cbl=2026366&diss=yq [accessed 2022-04-21]
- Giorgi S, Yaden DB, Eichstaedt JC, Ashford RD, Buffone AEK, Schwartz HA, et al. Cultural differences in Tweeting about drinking across the US. Int J Environ Res Public Health 2020 Feb 11;17(4):1125 [FREE Full text] [CrossRef] [Medline]
- Profiling social media users with selective self-disclosure behavior. Singapore Management University (Singapore). URL: https://ink.library.smu.edu.sg/etd_coll_all/1/#:~:text=Among%20the%20social%20media%20user,is%20called%20selective %20self%2Ddisclosure [accessed 2022-04-21]
- Towards secure and privacy-preserving online social networking services. University of California, Berkeley. URL: https://escholarship.org/uc/item/1b14t6kq [accessed 2022-04-21]
- Gundecha P, Ranganath S, Feng Z, Liu H. A tool for collecting provenance data in social media. In: Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining. 2013 Presented at: KDD '13: Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining; Aug 11-14, 2013; Chicago Illinois USA. [CrossRef]
- Guo G, Zhu F, Chen E, Liu Q, Wu L, Guan C. From footprint to evidence: an exploratory study of mining social data for credit scoring. ACM Trans Web 2016 Dec 27;10(4):1-38. [CrossRef]
- Gupta H, Lam T, Pettigrew S, Tait RJ. The association between exposure to social media alcohol marketing and youth alcohol use behaviors in India and Australia. BMC Public Health 2018 Jun 13;18(1):726 [FREE Full text] [CrossRef] [Medline]
- Haffner M. A place-based analysis of #BlackLivesMatter and counter-protest content on Twitter. GeoJournal 2018 Sep 1;84(5):1257-1280. [CrossRef]
- Ikeda K, Hattori G, Ono C, Asoh H, Higashino T. Twitter user profiling based on text and community mining for market analysis. Knowledge Based Syst 2013 Oct;51:35-47. [CrossRef]
- Ireland ME, Chen Q, Schwartz HA, Ungar LH, Albarracin D. Action Tweets linked to reduced county-level HIV prevalence in the United States: online messages and structural determinants. AIDS Behav 2016 Jun;20(6):1256-1264 [FREE Full text] [CrossRef] [Medline]
- Jha D, Singh R. SMARTS: the social media-based addiction recovery and intervention targeting server. Bioinformatics 2019 Oct 24;36(6):1970-1972. [CrossRef] [Medline]
- Jimenez S, Dueñas G, Gelbukh A, Rodriguez-Diaz C, Mancera S. Automatic detection of regional words for pan-hispanic spanish on twitter. In: Proceedings of the Ibero-American Conference on Artificial Intelligence. 2018 Presented at: Ibero-American Conference on Artificial Intelligence; Nov 13-14, 2018; Trujillo (Perú). [CrossRef]
- Jones T. Toward a description of African American vernacular English dialect regions using "Black twitter''. American Speech 2015;90:2015-2040. [CrossRef]
- Jørgensen A, Hovy D, Søgaard A. Learning a POS tagger for AAVE-like language. In: Proceedings of the 2016 conference of the North American chapter of the association for computational linguistics: human language technologies. 2016 Presented at: 2016 conference of the North American chapter of the association for computational linguistics: human language technologies; Jun 12-17, 2016; San Diego, California. [CrossRef]
- Kang Y, Zeng X, Zhang Z, Wang Y, Fei T. Who are happier? Spatio-temporal analysis of worldwide human emotion based on geo-crowdsourcing faces. In: Proceedings of the Ubiquitous Positioning Indoor Navigation and Location Based Service (UPINLBS). 2018 Presented at: Ubiquitous Positioning Indoor Navigation and Location Based Service (UPINLBS); Mar 22-23, 2018; Wuhan, China. [CrossRef]
- Kent JD, Capello HT. Spatial patterns and demographic indicators of effective social media content during theHorsethief Canyon fire of 2012. Cartography Geographic Inf Sci 2013 Mar;40(2):78-89. [CrossRef]
- Debiasing 2016 Twitter election analysis via Multi-Level Regression and Poststratification (MRP). University of Illinois. 2019. URL: https://indigo.uic.edu/articles/thesis/Debiasing_2016_Twitter_Election_Analysis_via_Multi-Level_Regression _and_Poststratification_MRP_/10904234 [accessed 2022-04-21]
- Kostakos P, Pandya A, Kyriakouli O, Oussalah M. Inferring demographic data of marginalized users in twitter with computer vision APIs. In: Proceedings of the 2018 European Intelligence and Security Informatics Conference (EISIC). 2018 Presented at: 2018 European Intelligence and Security Informatics Conference (EISIC); Oct 24-25, 2018; Karlskrona, Sweden. [CrossRef]
- Kotzé E, Senekal B. Employing sentiment analysis for gauging perceptions of minorities in multicultural societies: an analysis of Twitter feeds on the Afrikaner community of Orania in South Africa. J Transdisciplinary Res Southern Africa 2018 Nov 15;14(1):11. [CrossRef]
- Kumar D, Ukkusuri SV. Enhancing demographic coverage of hurricane evacuation behavior modeling using social media. J Comput Sci 2020 Sep;45:101184. [CrossRef]
- Lachlan KA, Spence PR, Lin X. Expressions of risk awareness and concern through Twitter: on the utility of using the medium as an indication of audience needs. Comput Human Behav 2014 Jun;35:554-559. [CrossRef]
- Lama Y, Chen T, Dredze M, Jamison A, Quinn SC, Broniatowski DA. Discordance between human Papillomavirus Twitter images and disparities in human Papillomavirus risk and disease in the United States: mixed-methods analysis. J Med Internet Res 2018 Sep 14;20(9):e10244 [FREE Full text] [CrossRef] [Medline]
- Virtual Homespace (Re)constructing the Body and Identity Through Social Media. Binghamton: State University of New York; 2016.
- Lee-Won RJ, White TN, Potocki B. The Black catalyst to tweet: the role of discrimination experience, group identification, and racial agency in Black Americans’ instrumental use of Twitter. Inf Commun Soc 2017 Mar 23;21(8):1097-1115. [CrossRef]
- Li J, Ritter A, Hovy E. Weakly supervised user profile extraction from twitter. In: Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2014 Presented at: Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers); Jun 22-27, 2014; Baltimore, Maryland p. 165-174. [CrossRef]
- Lienemann BA, Unger JB, Cruz TB, Chu K. Methods for coding tobacco-related Twitter data: a systematic review. J Med Internet Res 2017 Mar 31;19(3):e91 [FREE Full text] [CrossRef] [Medline]
- Lin Y. Assessing sentiment segregation in urban communities. In: Proceedings of the International Conference on Social Computing. 2014 Presented at: SocialCom '14: Proceedings of the International Conference on Social Computing; Aug 4-7, 2014; Beijing China. [CrossRef]
- As seen on Twitter: African-American rhetorical traditions gone viral. Michigan University. 2012. URL: https://docgo.net/as-seen-on-twitter-african-american-rhetorical-traditions-gone-viral [accessed 2022-04-21]
- Human activity recognition: a data-driven approach. UC Irvine. URL: https://escholarship.org/uc/item/4w98w1zd [accessed 2022-04-21]
- Lwowski B, Rios A. The risk of racial bias while tracking influenza-related content on social media using machine learning. J Am Med Inform Assoc 2021 Mar 18;28(4):839-849 [FREE Full text] [CrossRef] [Medline]
- Magdy A, Ghanem T, Musleh M, Mokbel M. Understanding language diversity in local twitter communities. In: Proceedings of the 27th ACM Conference on Hypertext and Social Media. 2016 Presented at: 27th ACM Conference on Hypertext and Social Media; Jul 10 - 13, 2016; Halifax Nova Scotia Canada. [CrossRef]
- Maheshwari T, Reganti A, Chakraborty T, Das A. Socio-ethnic ingredients of social network communities. In: Proceedings of the Companion of the ACM Conference on Computer Supported Cooperative Work and Social Computing. 2017 Presented at: CSCW '17 Companion: Companion of the ACM Conference on Computer Supported Cooperative Work and Social Computing; Feb 25- Mar 1, 2017; Portland, Oregon, USA. [CrossRef]
- Meng H, Kath S, Li D, Nguyen QC. National substance use patterns on Twitter. PLoS One 2017;12(11):e0187691 [FREE Full text] [CrossRef] [Medline]
- Montasser OK. Predicting Demographics of High-Resolution Geographies with Geotagged Tweets. In: Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence AAAI Press. 2017 Presented at: Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence AAAI Press; Feb 4 - 9, 2017; Vancouver, Canada p. 1460-1466.
- Nguyen TT, Adams N, Huang D, Glymour MM, Allen AM, Nguyen QC. The association between state-level racial attitudes assessed from Twitter data and adverse birth outcomes: observational study. JMIR Public Health Surveill 2020 Jul 06;6(3):e17103 [FREE Full text] [CrossRef] [Medline]
- Mulders D, Bodt C, Bjell J, Pentl A, Verleysen M, Montjoye Y. Improving individual predictions using social networks associativity. In: Proceedings of the 12th International Workshop on Self-Organizing Maps and Learning Vector Quantization, Clustering and Data Visualization (WSOM). 2017 Presented at: 12th International Workshop on Self-Organizing Maps and Learning Vector Quantization, Clustering and Data Visualization (WSOM); Jun 28-30, 2017; Nancy, France p. 1-8. [CrossRef]
- Nelson J, Quinn S, Swedberg B, Chu W, MacEachren A. Geovisual analytics approach to exploring public political discourse on Twitter. Isprs Int J Geo Inf 2015 Mar 05;4(1):337-366. [CrossRef]
- Nguyen QC, Kath S, Meng H, Li D, Smith KR, VanDerslice JA, et al. Leveraging geotagged Twitter data to examine neighborhood happiness, diet, and physical activity. Appl Geogr 2016 Aug;73:77-88 [FREE Full text] [CrossRef] [Medline]
- Nguyen QC, Li D, Meng H, Kath S, Nsoesie E, Li F, et al. Building a national neighborhood dataset from geotagged Twitter data for indicators of happiness, diet, and physical activity. JMIR Public Health Surveill 2016 Oct 17;2(2):e158 [FREE Full text] [CrossRef] [Medline]
- Novak AN, Johnson K, Pontes M. LatinoTwitter: discourses of Latino civic engagement in social media. First Monday 2016 Jul 24;21(8). [CrossRef]
- Odlum M, Cho H, Broadwell P, Davis N, Patrao M, Schauer D, et al. Application of topic modeling to Tweets as the foundation for health disparity research for COVID-19. Stud Health Technol Inform 2020 Jun 26;272:24-27 [FREE Full text] [CrossRef] [Medline]
- Oktay H, Firat A, Ertem Z. Demographic breakdown of Twitter users: an analysis based on names. In: Proceedings of the Academy of Science and Engineering (ASE). 2014 Presented at: Academy of Science and Engineering (ASE); Jan 11, 2014; Birmingham.
- Orsolini L, Papanti GD, Francesconi G, Schifano F. Mind navigators of chemicals' experimenters? A web-based description of e-psychonauts. Cyberpsychol Behav Soc Netw 2015 May;18(5):296-300. [CrossRef] [Medline]
- Pick J, Sarkar A, Rosales J. Social media use in American counties: geography and determinants. Isprs Int J Geo Inf 2019 Sep 19;8(9):424. [CrossRef]
- Developing computational approaches to investigate health inequalities. University of Washington. 2017. URL: https://soc.washington.edu/research/graduate/developing-computational-approaches-investigate-health-inequalities [accessed 2022-04-21]
- Priante A, Hiemstra D, Saeed A, van den Broek T, Ehrenhard M, Need A. #WhoAmI in 160 characters? Classifying social identities based on twitter profile descriptions. In: Proceedings of the First Workshop on NLP and Computational Social Science. 2016 Presented at: Proceedings of the First Workshop on NLP and Computational Social Science: Association for Computational Linguistics; Nov 5, 2016; Austin, Texas. [CrossRef]
- Riederer C, Zimmeck S, Phanord C, Chaintreau A, Bellovin S. I don't have a photograph, but you can have my footprints: revealing the demographics of location data. In: Proceedings of the ACM on Conference on Online Social Networks. 2015 Presented at: OSN '15: Proceedings of the ACM on Conference on Online Social Networks; Nov 2-3, 2015; Palo Alto California USA. [CrossRef]
- Roberts MJ, Perera M, Lawrentschuk N, Romanic D, Papa N, Bolton D. Globalization of continuing professional development by journal clubs via microblogging: a systematic review. J Med Internet Res 2015;17(4):e103 [FREE Full text] [CrossRef] [Medline]
- Roy S, Ghosh P. A comparative study on distancing, mask and vaccine adoption rates from global Twitter trends. Healthcare 2021 Apr 21;9(5):488. [CrossRef]
- Rummo PE, Cassidy O, Wells I, Coffino JA, Bragg MA. Examining the relationship between youth-targeted food marketing expenditures and the demographics of social media followers. Int J Environ Res Public Health 2020 Mar 03;17(5):1631 [FREE Full text] [CrossRef] [Medline]
- Runge K. "Social" science, spider goats and American science audiences: investigating the effects of interpersonal networks on perceptions of emerging technologies. ProQuest. 2017. URL: https://www.proquest.com/openview/85b51025bda09f9a6941 fd9b6e7cc054/1?pq-origsite=gscholar&cbl=18750&diss=y [accessed 2022-04-21]
- Sijtsma B, Qvarfordt P, Chen F. Tweetviz: visualizing tweets for business intelligence. In: Proceedings of the 39th International ACM SIGIR conference on Research and Development in Information Retrieval. 2016 Presented at: 39th International ACM SIGIR conference on Research and Development in Information Retrieval; Jul 17 - 21, 2016; Pisa Italy. [CrossRef]
- Singh M, Singh A, Bansal D, Sofat S. An analytical model for identifying suspected users on Twitter. Cybernetics Syst 2019 Apr 02;50(4):383-404. [CrossRef]
- Tomeny TS, Vargo CJ, El-Toukhy S. Geographic and demographic correlates of autism-related anti-vaccine beliefs on Twitter, 2009-15. Soc Sci Med 2017 Dec;191:168-175 [FREE Full text] [CrossRef] [Medline]
- Tulloch J. An appraisal of health datasets to enhance the surveillance of Lyme disease in the United Kingdom. ProQuest. 2019. URL: https://www.proquest.com/openview/72acae0b592c9f8564d2204eaa84f5d1/1?pq-origsite=gscholar&cbl=44156 [accessed 2022-04-21]
- Vydiswaran VG, Romero DM, Zhao X, Yu D, Gomez-Lopez I, Lu JX, et al. Uncovering the relationship between food-related discussion on Twitter and neighborhood characteristics. J Am Med Inform Assoc 2020 Feb 01;27(2):254-264 [FREE Full text] [CrossRef] [Medline]
- Wang Y, Feng Y, Luo J, Zhang X. Voting with feet: who are leaving Hillary Clinton and Donald Trump. In: Proceedings of the IEEE International Symposium on Multimedia (ISM). 2016 Presented at: IEEE International Symposium on Multimedia (ISM); Dec 11-13, 2016; San Jose, CA. [CrossRef]
- Weeg C, Schwartz HA, Hill S, Merchant RM, Arango C, Ungar L. Using Twitter to measure public discussion of diseases: a case study. JMIR Public Health Surveill 2015;1(1):e6 [FREE Full text] [CrossRef] [Medline]
- Wright M, Adams T. #KnowBetterDoBetter: an examination of Twitter impact on disaster literacy. ProQuest. 2019. URL: https://www.proquest.com/openview/1d20d09437e5921d1b996c6657c29011/1?pq-origsite=gscholar&cbl=18750&diss=y [accessed 2022-04-21]
- Yazdavar AH, Mahdavinejad MS, Bajaj G, Romine W, Sheth A, Monadjemi AH, et al. Multimodal mental health analysis in social media. PLoS One 2020;15(4):e0226248 [FREE Full text] [CrossRef] [Medline]
- Ying Q, Chiu D, Venkatramanan S, Zhang X. Profiling OSN users based on temporal posting patterns. In: Proceedings of The Web Conference. 2018 Presented at: WWW '18: Proceedings of The Web Conference; Apr 23-27, 2018; Lyon, France. [CrossRef]
- Yuan F, Li M, Zhai W, Qi B, Liu R. Social media based demographics analysis for understanding disaster response disparity. In: Proceedings of the Construction Research Congress 2020: Computer Applications. 2020 Presented at: Construction Research Congress 2020: Computer Applications; Mar 8–10, 2020; Tempe, Arizona. [CrossRef]
- Zhang Z, Bors G. “Less is more” : mining useful features from Twitter user profiles for Twitter user classification in the public health domain. Online Inf Rev 2019;44(1):213-237 [FREE Full text]
- Zhao P, Jia J, An Y, Liang J, Xie L, Luo J. Analyzing and predicting emoji usages in social media. In: Proceedings of the The Web Conference 2018. 2018 Presented at: WWW '18: Companion Proceedings of the The Web Conference 2018; Apr 23 - 27, 2018; Lyon France. [CrossRef]
- Zhong Y, Yuan N, Zhong W, Zhang F, Xie X. You are where you go: inferring demographic attributes from location check-ins. In: Proceedings of the Eighth ACM International Conference on Web Search and Data Mining. 2015 Presented at: WSDM '15: Proceedings of the Eighth ACM International Conference on Web Search and Data Mining; Feb 2-6, 2015; Shanghai China. [CrossRef]
- Jiang Y, Li Z, Ye X. Understanding demographic and socioeconomic biases of geotagged Twitter users at the county level. Cartography Geographic Inf Sci 2018 Feb 09;46(3):228-242. [CrossRef]
- Face++ homepage. Face++. URL: https://www.faceplusplus.com/ [accessed 2022-04-21]
- Powerful audience demographics. DemographicsPro. URL: https://www.demographicspro.com/ [accessed 2022-04-21]
- Onomap is changing. Onomap. URL: https://www.onomap.org/ [accessed 2022-04-21]
- Academic research: preparing for the academic research application: learn everything there is to know about applying for the academic research product track. Twitter Inc. URL: https://developer.twitter.com/en/solutions/academic-research/application-info [accessed 2022-04-19]
- Twitter grants academics full access to public data, but not for suspended accounts. Reuters. URL: https://www.usnews.com/news/technology/articles/2021-01-26/twitter-grants-academics-full-access-to-public-data-but-not-for-suspended-accounts [accessed 2022-04-19]
- Jung S, An J, Kwak H, Salminen J, Jansen B. Assessing the accuracy of four popular face recognition tools for inferring gender, age, and race. In: Proceedings of the Twelfth International AAAI Conference on Web and Social Media. 2018 Presented at: Twelfth International AAAI Conference on Web and Social Media; Jun 25-28, 2018; Palo Alto, California, USA.
- Buolamwini J, Gebru T. Gender shades: intersectional accuracy disparities in commercial gender classification. In: 1st Conference on Fairness, Accountability and Transparency. 2018 Presented at: Proceedings of the 1st Conference on Fairness, Accountability and Transparency; 2018; New York.
- Jung S, An J, Kwak H, Salminen J, Jansen B. Inferring social media users? Demographics from profile pictures: a face++ analysis on twitter users. In: Proceedings of The 17th International Conference on Electronic Business. 2017 Presented at: Proceedings of The 17th International Conference on Electronic Business (ICEB); 2017; Dubai, UAE.
- Jointly de-biasing face recognition and demographic attribute estimationiasing Face Recognition and Demographic Attribute Estimation. In: Computer Vision – ECCV 2020. Cham: Springer; 2020.
- Fu S, He H, Hou Z. Learning race from face: a survey. IEEE Trans Pattern Anal Mach Intell 2014 Dec 1;36(12):2483-2509. [CrossRef]
- Goldinger SD, He Y, Papesh MH. Deficits in cross-race face learning: insights from eye movements and pupillometry. J Experimental Psychol Learn Memory Cognition 2009;35(5):1105-1122. [CrossRef]
- Meissner CA, Brigham JC. Thirty years of investigating the own-race bias in memory for faces: a meta-analytic review. Psychol Public Policy Law 2001;7(1):3-35. [CrossRef]
- Jofre A, Berardi V, Brennan K, Cornejo A, Bennett C, Harlan J. Crowdsourcing image extraction and annotation: software development and case study. Digital Humanities Q 2020;14(2) [FREE Full text]
- King RD, Johnson BD. A punishing look: skin tone and afrocentric features in the halls of justice. Am J Sociol 2016 Jul;122(1):90-124. [CrossRef]
- Cavazos JG, Phillips PJ, Castillo CD, O'Toole AJ. Accuracy comparison across face recognition algorithms: where are we on measuring race bias? IEEE Trans Biom Behav Identity Sci 2021 Jan;3(1):101-111. [CrossRef]
- Buolamwini J, Gebru T. Gender shades: intersectional accuracy disparities in commercial gender classification. In: Proceedings of the 1st Conference on Fairness, Accountability and Transparency. 2018 Presented at: 1st Conference on Fairness, Accountability and Transparency; Feb 23-24, 2018; New York.
- Torralba A, Efros A. Unbiased look at dataset bias. In: Proceedings of CVPR 2011. 2011 Presented at: CVPR 2011; Jun 20-25, 2011; Colorado Springs, CO, USA. [CrossRef]
- Moscrop A, Ziebland S, Bloch G, Iraola JR. If social determinants of health are so important, shouldn’t we ask patients about them? BMJ 2020 Nov 24:m4150. [CrossRef]
- McHugh ML. Interrater reliability: the kappa statistic. Biochem Med (Zagreb) 2012;22(3):276-282 [FREE Full text] [Medline]
- Developer terms: more about restricted uses of the Twitter APIs. Developer Platform. URL: https://developer.twitter.com/en/developer-terms/more-on-restricted-use-cases [accessed 2022-03-04]
- Alsaied T, Allen KY, Anderson JB, Anixt JS, Brown DW, Cetta F, et al. The Fontan outcomes network: first steps towards building a lifespan registry for individuals with Fontan circulation in the United States. Cardiol Young 2020 Jul 08;30(8):1070-1075. [CrossRef]
|ML: machine learning|
|NIH: National Institutes of Health|
|NLP: natural language processing|
|PRISMA: Preferred Reporting Items for Systematic Reviews and Meta-Analyses|
Edited by T Leung; submitted 17.12.21; peer-reviewed by JW Kwon, D Storman, A Joseph; comments to author 23.02.22; revised version received 08.03.22; accepted 23.03.22; published 29.04.22Copyright
©Su Golder, Robin Stevens, Karen O'Connor, Richard James, Graciela Gonzalez-Hernandez. Originally published in the Journal of Medical Internet Research (https://www.jmir.org), 29.04.2022.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on https://www.jmir.org/, as well as this copyright and license information must be included.