香京julia种子在线播放

To cope with the huge amount of data, researchers have turned to artificial intelligence (AI) algorithms (Gupta et al., 2016; Wahl et al., 2018) to mine for insights. A key benefit of AI algorithms is their ability to computationally crunch large number of factors or variables in both interactive linear and non-linear ways to detect patterns in data (Kreatsoulas and Subramanian, 2018). While researchers have recognized that there are areas in the problem of data deluge could be effectively dealt with by AI (Maddox et al., 2019), such as identifying complex patterns in clinical settings in the area of cancer tumors detection (Bi et al., 2019), its efficacy is limited in other areas, especially in understanding motivations behind why people adhere to or reject certain health behaviors (e.g., vaccination). A blind and unassuming faith in health big data, or AI to make sense of it, would contribute to the problem of data hubris.

The issue of overstating causal relationships from big data is particularly important for health communication researchers. Although health communication is a broad and multidisciplinary field, a central goal of health communication research is to understand how the process of communication influences health outcomes (Kreps, 2001; Beck et al., 2004; Schiavo, 2013). This means that an underlying and unifying goal of health communication research is in identifying causal communication-related mechanisms that have an effect on individual's health-related outcomes and behaviors. In other words, causal inferences are fundamental in health communication research.

Undergirding various causal claims in health communication are theories that specify how and why health-related behaviors and health outcomes are enacted (e.g., Janz and Becker, 1984; Bandura, 2004; Ryan and Deci, 2007; Goldenberg and Arndt, 2008; Rimal, 2008; Fishbein and Ajzen, 2010). Solid theoretical grounding allows for robust predictions of health behaviors, leading to more confident empirically based recommendations for policymakers and healthcare practitioners to effect real change. While big data offer an opportunity to examine health communication phenomena on an unprecedented scale, using them without theoretical considerations to guide analyses can be highly problematic. This is particularly pertinent to health communication research, as one issue we face in our field is the lack of substantial theoretical development (Beck et al., 2004; Sandberg et al., 2017). Inductive approaches to big data analysis can exacerbate this.

Specifically, purely data-driven big data analyses have the potential to generate large amounts of spurious findings. Spurious correlations refer to variables being associated with each other, either entirely from chance, or due to confounding factors. The enormous amount of data available in big data analyses mean that spurious correlations are more likely to be encountered (Calude and Longo, 2017). While not exactly digital traces, researchers who scanned a dead salmon in an fMRI machine while exposing it to photos of humans in a variety of social situations, found that 16 out of 8,064 brain regions in the dead salmon had a statistically significant response to the images (Bennett et al., 2009). The sheer amount of data in an fMRI scan meant that impossible findings can also be flagged out as significant. As big data datasets could be exponentially larger than an fMRI scan, spurious correlations are even more likely to be identified through various analyses. Instead of producing robust knowledge, data-driven approaches could potentially lead to a deluge of unimportant associations being flagged out. In other words, patterns derived from big data analysis are far from reliable, and causal claims arising from such findings can be nothing more than hubris. For example, the increase in Google search frequency of disease-related words during disease outbreaks is more likely due to extensive media coverage, rather than actual disease outbreak (Cervellin et al., 2017). After all, social media and search engines have the propensity to amplify risks (Strekalova, 2017; Strekalova and Krieger, 2017). Adopting the social amplification of risk framework (SARF), Strekalova and Krieger (2017) highlighted that health risks could be heightened via social media or online platforms through the mechanisms of (a) user engagement: where content of health risks are freely generated and diffused among social networks, (b) media richness (e.g., vivid descriptions or photographs depicting health risks), and (c) signal sharing: online communities adopting certain words, phrases, and pictures to create shared socially constructed meaning (e.g., posting selfies of people with mustache during November as part of the Movember campaign which raises awareness for prostate cancer) and experiences of health risks (Jacobson and Mascaro, 2016). In a study examining people's response to Centers for Disease Control and Prevention's (CDC) Facebook page, Strekalova (2017) found that even though the CDC had fewer posts of Ebola as compared to posts on health promotion posts, the former received more attention from Facebook users.

A second problematic issue is the assumption that the enormous number of observations in big data analysis leads to generalizable findings. In contrast, big data have significant issues pertaining to sampling bias and representativeness. Specifically, existing big data studies in health research tend to utilize one platform, such as Twitter, Facebook, or EHRs (Hargittai, 2015). These different platforms have sample bias problems, with race, gender, and class being factors that determine their use (Tufekci, 2014; Hargittai, 2015). For example, only 40% of patients in the US have their information in EHRs, while gender, race, and income predict Facebook, LinkedIn, and Twitter use (Kaplan et al., 2014; Hargittai, 2015). This means that findings and conclusions drawn from single-platform big data studies are limited to the type of user that is characteristic to each platform. To add to this, supposedly random samples drawn from Twitter's Application Programming Interfaces (APIs) are sometimes skewed by the presence of sample cheaters, corporate spammers, and frequency bots (Pfeffer et al., 2018).

Relatedly, self-selection bias is particularly prevalent in social media studies that utilize hashtags. By selecting only hashtagged tweets or posts to answer certain research questions, it excludes other related content. This self-selected sample of content might lead to systematic errors which can severely impact the validity of findings (Tufekci, 2014). In sentiment analysis, hashtagged content might reflect more polarizing opinions, as in the case of the #himtoo hashtag during Brett Kavanaugh's supreme court confirmation, or the #vaccinescauseautism and #vaccineskill hashtags among the antivax population. It is possible that users who utilize hashtags have a stronger opinion on certain issues, while content produced by users who are neutral are ignored by studies that draw their data from hashtags.

The problem of data opacity in data acquisition is not unique to social media platforms. Researchers utilizing mobile phones for health communication and public health also face the problem of data opacity. Like social media companies, researchers often do not have access to large-scale mobility data from smartphones, and would rely on data provided by telecommunication companies which often operate within strict regulations (Wesolowski et al., 2016). While researchers may be able to access large scale mobility data form telecommunication companies from open challenges such as Data for Refugees Turkey—an initiative by Turk Telekom (2018) to release large anonymized mobile datasets to researchers for utilizing big data solutions to improve living conditions for Syrian refugees—researchers ultimately would not have full details as to how the data was prepared, and if certain details were left out.

Data opacity at the data collection level poses practical challenges for health researchers who rely on them for their research. Recently, Facebook partnered with Harvard University and the Social Science Research Council to forge an industry-academic partnership, and pledged to support academic scholars to use a subset of Facebook data to study the impact of social media on democracy (Social Science One, 2018). The collaboration would involve Facebook providing URLs to researchers who have been granted ethical approval from their respective Institutional Review Boards (IRB) for their research. Yet, due to Facebook being unable to provide what was promised in the original request for proposals in 2018 (King and Persily, 2019), researchers would need to work with reduced versions of the data until the full release. While this is not in the health context, it shows the extent of reliance that health researchers have on companies who play a gate-keeping role in health big data.

In addition to data acquisition, data opacity occurs at the level of analysis as well. For instance, for health researchers analyzing social media data, one very common algorithm to deploy is topic modeling. Topic modeling is a text mining approach for identifying “themes” computationally in a large corpus of unstructured texts (Richardson et al., 2014). One of the most widely used algorithms for topic modeling is Latent Dirichlet Allocation (LDA), which is a popular technique used by researchers to examine what people are talking about online pertaining to health (Bian et al., 2017). In using LDA, researchers need to specify a priori how many latent topics there are. For deciding on the final number of topics there are in a given text, researchers typically turn to computational measures such as perplexity scores or the likes—a measure of goodness of fit, where lower scores meant that the number of topics fit the LDA model (Blei et al., 2003; Fung et al., 2017). The strict reliance on computational power ignores nuanced meaning in languages that may skew the results, or how certain geo-political environmental contexts may augment the nature of the topics.

A second example of data opacity in the context of data analysis is in the use of geospatial analytics on location or physiological data obtained from smartphones or wearables. While some would argue that mobile sensing in public health has a higher degree of objectivity in terms of the data it could capture (Chaix, 2018)—GPS location would inform researchers where individuals have been—yet, there are multiple factors at the analysis stage that dilute this objectivity. For instance, in analyzing proximity to key features (e.g., hospitals, services), a common technique is buffer analysis, which allows researchers to create artificial boundaries around the features of interests and count how many data points (e.g., mobility patterns) are within the artificial boundaries (ArcMap, 2019). For instance, research in understanding opioid overdose has used buffer analysis to identify the total number of discarded needles in in a given area (Bearnot et al., 2018). The selection of such boundaries for buffer analysis is inherently subjective, and the assumptions of such selection would need to be supported theoretically.

While existing research in health communication research utilize surveys and experiments to measure abstract concepts such as attitudes and beliefs, it might be possible to view social media content as operationalizations of abstract concepts as well. Specifically, social media posts can be operationalizations of sentiment, or attitude, toward certain health behaviors. For example, in the context of alcohol consumption, production of social media content—tweets or posts—can be conceptualized as an underlying positive or negative attitude toward drinking, while the consumption of social media content (what people are exposed to) regarding drinking can be conceptualized as social norms (Cavazos-Rehg et al., 2015). In combination with other measures, these can be used as measures within existing theoretical frameworks such as the theory of normative social behavior or the theory of planned behavior (Ajzen, 1991; Rimal and Lapinski, 2015).

Despite this, formal explication of how existing theoretical constructs such as attitude and beliefs can be measured through big data is rare. In one study, researchers looked at how both visual and textual posts about alcohol on Facebook predicts binge drinking, in addition to attitude and injunctive norms (D'Angelo et al., 2014). Instead of viewing the social media posts as a distinct and separate construct to attitude in a predictive model, we propose that health communication researchers should aim to explicate what these specific data are reflecting. Could they be reflective of an individuals' attitude toward drinking, or something else? Health communication researchers must begin clarifying what specific pieces of big data mean.

Beyond the process of concept explication, big data also offer opportunities for the testing of existing theories that was previously impossible. For example, traditional ways of testing priming effects utilize laboratory experiments where researchers manipulate stimuli of interest (e.g., Alhabash et al., 2016). Instead of testing priming effects in a lab, big data tools allow us to examine these effects in a naturalistic setting.

Take for example screenomics, a novel approach in measuring digital media use through the capturing of screenshots periodically (e.g., screenshot every 5 s), before being encrypted and sent to university servers for processing, in which textual and image data are extracted using computer vision technology and natural language processing (Reeves et al., 2019). Instead of manipulating Facebook alcohol ads to investigate priming effects intention to consume alcohol, as Alhabash et al. (2016) had done, screenomics paired with ecological momentary assessment (EMA) measuring alcohol consumption intentions can provide us with a unique way of investigating priming effects naturalistically. Furthermore, with the sheer amount of data generated, in addition to examining priming effects across participants, it is possible to examine priming effects within an individual (e.g., whether exposure to alcohol images and posts primes a specific individual's intention to consume alcohol) by using a person's daily screenshot data and EMA responses as a single data point. This leads to the potential testing of health communication theories through both an idiographic and nomothetic approach. Such naturalistic ways of collecting media and behavioral data, addresses one of the biggest criticisms of traditional experimental media effects research—the artificiality of the environment in which stimuli are manipulated, and the artificiality of the stimuli themselves (Livingstone, 1996).

First, as health communication scholars increasingly need to work with big data from diverse sources, there is a potential to connect the dots by drawing upon diverse datasets on how social determinants exacerbate health disparities (Lee and Viswanath, 2020). There are two key developments that would enable health data sense-making for the context of health disparities. First, the incorporation of factors such as social determinants and geographical variables into existing EHRs (Zhang et al., 2017). The second is the move toward utilizing digital and traditional media and social norms in nudging individuals to take responsibilities of their own health through active health monitoring (Patel et al., 2015; Hentschel et al., 2016; Ho et al., 2016).

The potential to link social determinants, demographic, and geographical variables with how people are utilizing digital devices for health and subsequent health outcomes would shed light on how different forms of communication inequalities—differences among social groups in generation and use of information that contribute to differential health outcomes—serve as intermediary mechanisms that moderate the effects of social determinants on health disparities (Viswanath et al., 2012; Lee and Viswanath, 2020). The ability to synthesize insights from datasets would contribute to macro health communication theorizing (e.g., ecological health model, social cognitive model), and enable scholars to specifically tease out how external and individual determinants factors collectively influence health outcomes and behaviors (Bandura, 2001; Richard et al., 2011; Lee et al., 2016; Lee and Viswanath, 2020).

The second way new health communication theories could emerge is in the area of quantifying the interactivity of media consumption and interpersonal networks on health communication behaviors. Screenomics, for example, has the potential to allow health communication researchers to explicate health information seeking beyond traditional survey measures, and specifically examine the process of health information seeking, and how different groups seek health information (Reeves et al., 2019). In terms of quantifying the interactivity of media and network effects, researchers could test their a-priori hypotheses on how they expect information flow and interpersonal networks to influence health outcomes through the use of dynamic or agent-based modeling (Zhang et al., 2015; Song and Boomgaarden, 2017), and compare how reinforcing spirals (Slater, 2007) of media effects and interpersonal influence on health behaviors differ in a simulated environment and in real environment.

The third area which there is a potential for emergence of new health communication theories is in the area of explicating different dimensions of locations within health communication theories and examine the variability in health outcomes at a granular level. Traditionally, health communication theories such as the structural influence model of communication postulates that social determinants such as geographical locations in terms of neighborhood and urbanicity (Kontos and Viswanath, 2011) are underlying macro-level factors influencing health outcomes. Yet, how scholars conceptualize locations are often based on artificial administrative boundaries (e.g., census tracts or zip codes), which would be problematic as people are highly mobile and not constrained to a single location.

The advancement in collection of location data, together with physiological and behavioral data through mobile health apps and wearables, would enable researchers to do a deep dive into how specific locations matter in health outcomes. For instance, by drawing upon geolocations, galvanic skin response, skin temperature, and heart rate variability, researchers could identify areas quantify moments and movements of stress in an urban environment, such as cyclists' emotionala states when traveling in certain paths (Kyriakou et al., 2019). In other words, coupled with geospatial analytical tools such as emerging hotspot analysis, researchers could accurately identify potential hotspots where traffic accidents are likely to happen because of the stress-clustering. In addition, by utilizing geolocation tracking from smartphones, researchers would be conceptualize and concretize mobility signatures, and connect how real-time exposure to point-of-sale tobacco marketing, differential tobacco product pricing, location of tobacco retail outlets, influence quit attempts (Kirchner et al., 2012).