Structural Ageism in Big Data Approaches

Digital systems create logs of all activities. Terabytes of data generated by real users in digital environments help the monitoring, or tracking, of everyday activities, leading to the so-called big data era (boyd & Crawford, 2012). The availability of these massive and complex data sets, or big data, is changing the paradigms of research and innovation in diverse disciplines (Kitchin, 2014), among them computer science, the economy, bio-informatics, sociology and political science (boyd & Crawford, 2012). Big data approaches, which use digital logs to learn from user behaviours, are nowadays pervasive (Schäfer & Van Es, 2017). They can perform systematic analyses of continuous data flows based on statistical instruments (Böhmer et al., 2011; Kosinski et al., 2013). In this sense, big data are used to understand digital practices through statistical models with the intention of predicting the behaviour of the population (e.g., Oulasvirta et al., 2012; Ørmen & Thorhauge, 2015). Algorithms are at the core of intelligent systems and serve different purposes, for example predicting personality, detecting crime or helping developers in tailoring products to users’ interests.

However, big data approaches come with questions, not only regarding privacy and ethical protocols (Jensen, 2013; Smith et al., 2012) but also in methodological and epistemological terms (boyd & Crawford, 2012; Schäfer & Van Es, 2017). Digital tracking is meant to provide more objective data (Karikoski & Soikkeli 2013; Möller et al., 2013). As Karikoski and Soikkeli explained, objective data provide a measurement of activities that is exempt from the declarative subjectivities that are common in interviews or surveys (Karikoski & Soikkeli, 2013). Classical sampling techniques can be overcome, because the data volumes are large enough to allow an understanding of trends and correlations at the population level (Ferreira et al., 2012). However, the size of a database does not necessarily grant representativeness at the population level (Ferreira et al., 2012).

Moreover, data are a socio-technical phenomenon: the “crumbs” of a complex system of interrelations between data, systems and society (Letouzé, 2015). Tracked logs respond to the technical processes of the systems, which are not necessarily aligned to the needs of specific analyses and should be understood as a by-product of digital systems. Therefore, an interpretation of the available data is required that introduces subjectivity, which researchers often do not even recognize (boyd & Crawford, 2012). Thus, from the social sciences point of view, critical positions are arising towards the current techno-optimism that accompanies big data, which may suffer from limitations brought by the underlying assumptions, values and biases (boyd & Crawford, 2012).

Age, old age and aging are socially constructed concepts. The concept of a “senior person” often has negative connotations (Garattini & Prendergast, 2015), among others because of the power structures that exclude most older people from decision making (Calasanti & King, 2015). Otherwise, we are living in an aged and aging (Eurostat, 2017) networked society (Castells, 2009), in which an increasing percentage of older people adopt and use digital technologies (Eurostat, 2018). However, older people are a minority in digital media due to lower levels of adoption (ibid.) and, once adopted, lower levels of use than younger generations (Rosales & Fernández-Ardèvol, 2016a).

Digitally mediated communication has been studied extensively in children and youths, but it has been developed little in the case of the elderly (Mihailidis, 2014). Adolescents and youths have been the main reference generations in ICT studies because they establish trends that later become generalized to the rest of the population (Castells et al., 2006). Otherwise, the older population has systematically been excluded from ICT studies (e.g., Eurostat, 2018; Roca Salvatella, 2016), preventing the development of diverse models of ageing in the digital world. In turn, older people are rarely involved in the design of new technologies, some exceptions include projects developed by Rogers and colleagues (2014) and Righi and colleagues (2018). The use of digital devices and the goals of mediated communication change over time (Ling et al., 2012), because individual values and interests change (Neugarten, 1996). Aging in the broad sense conditions communicative habits and the choice of digital channels (Selwyn et al., 2003). However, when effectively incorporated into empirical analyses, their specificity is often not considered, which makes older people invisible (Hendricks, 2005) by ignoring their particular uses, interests and values.

Older people are often portrayed as less avid users, lacking interest or unable to use digital technologies properly: a thesis that would sustain the statistics because, from the age of 55 years onwards, ownership and use fall below the average (Eurostat, 2018). However, older people constitute a diverse user group (Sawchuk & Crow, 2011). An increasing number of studies have reported that, when older people become users, they can be active users of digital systems (Choudrie & Vyas, 2014; Rosales & Fernández-Ardèvol, 2016a), though it is necessary to distinguish quantity and quality because the centrality of digital media does not depend on the intensity of use. Although older people can (or tend to) be less active users, their use of ICT is relevant and meaningful (Rosales & Fernández-Ardèvol, 2016a).

Big data approaches rely on intelligent systems for decision making by considering individuals’ general characteristics, such as where they live, their social network or their credit score, which can lead to race and socio-economic discrimination (O’Neil, 2016). All intelligent systems use a data set for learning the patterns; however, if the learning data set uses collective data and the data are biased against a particular community, the system will reproduce such a stereotype. Instead of evaluating individuals, they are “unfairly treating people on the basis of their belonging to a specific group” (Hajian & Domingo-Ferrer, 2013: 1445). Discrimination also operates by making some collectives invisible (Hendricks, 2005), ignoring their particular practices of use, interests and values in the design of new technologies. Invisibilization reinforces inequality and injustice (Stocchetti, 2018), as poor, marginalized and vulnerable collectives may suffer the negative consequences from weapons of math destruction because data models do not consider their interests (O’Neil, 2016).

Another particular issue creating exclusion is the design of algorithms, based on the beliefs and values of researchers (Uricchio, 2017) that are not supported by empirical evidence. However, while discrimination in big data approaches has attracted attention regarding, sex, race, socio-economic background and sexual orientation (Eubanks, 2018; Hajian & Domingo-Ferrer, 2013; O’Neil, 2016), it has generated less interest regarding age.

Predictive systems rely on learning data sets to learn common patterns of use. Intelligent systems compare new users with the patterns of the learning data set to make predictions. However, in some cases, the characteristics of the learning data set are not provided or not discussed. Thus, it becomes impossible to identify biases, which might entail the exclusion of minorities. Particularly, the learning data set characteristics should be compared with the studied population, as the recruitment can favour the inclusion of some specific profiles. In such cases, predictions would be less accurate for those collectives that are not properly represented in the learning data set.

Three examples illustrate this case. First, Kosinski and colleagues (2013) research on Facebook and Bi and colleagues (2013) study of search queries relied on the same learning data set based on information from 58,000 volunteers. The authors did not explicitly provide the demographics of this learning data set. Thus, it is not possible to evaluate its possible biases. It is likely that older people are a minority in the learning data set, because they are a minority in the digital world, which will shape the predictions of both systems. Second, Oktay and colleagues (2012) predictions relied on the first name of US-based Twitter users. Their learning data set was the US public statistics on naming, at birth, particularly “the frequency of more than 150K different baby names for each year starting from 1881” (Oktay et al., 2012: 2). Migratory movements were not considered in these particular statistics – and some age cohorts might be more affected by immigration from different countries than others, which would shape the learning data set’s representativeness. Finally, Liu and Yang (2012) predicted demographics on last. fm based on music. Their learning data set was a public data set of last.fm users (Liu & Yang, 2012); however, the possible biases of the learning data set were not discussed.

On other occasions, the bias is more explicit and hides the nuances of older age. This is the case when the research defines age groups that are too broad (e.g., those over 40 years old) or establishes upper age limits based on technical restrictions. Research on Twitter conducted in 2012 limited the predictions until 80 years old, because otherwise they “do not seem to contain genuine guesses” (Nguyen et al., 2014: 4). Another study, conducted in 2010, focused on individuals born between 1977 and 1988, because they “use social networking, blogs, and instant messaging more than their elders” (Rosenthal & McKeown, 2011: 769).

By building on homophilic ideas, some developers have assumed that an algorithm that suits some users well should suit all users well, particularly minority users. This is often associated with the limitations of the algorithms in making accurate predictions for some groups. Algorithms often use the average age of a user’s network to predict his or her age (e.g., Culotta et al., 2016; Perozzi & Skiena, 2015). However, this equation does not apply to older users, as the average age of their network on social media sites is often lower than their age – because very often older people use social media to follow their children and grandchildren, while most of their friends are not part of their digital network. In line with this, authors such as Culotta and colleagues (2016) recognized that their age predictions worked better for younger cohorts.

Although many projects have argued that they predict the age of participants using text provided by users, the algorithm only manages to distinguish between younger cohorts and adults (Liao et al., 2014; Nguyen et al., 2013; Peersman et al., 2011), partly due to the limitation of the algorithm in making better predictions and partly due to all the nuances of the language used by adults or the limited use of language among young people.

Most ageist practices in intelligent systems design are related to the data set limitations in terms of the representativeness of the studied population and particularly to recruitment procedures that tend to exclude older people.

Smartphone log studies often offer participants a particular service to engage users and allow them to track their logs through an installed app (for instance Böhmer et al., 2011). In such cases, the mobile application should provide an added value, an appealing reason for the users to install it and contribute voluntarily to the study. However, the interest in the service or the research goal is not necessarily neutral and the resulting sample will not be random but incidental. The (learning) sample of the study would therefore be biased towards those individuals who perceive the advantages of the services linked to the research project. For instance, Securacy allows users to assess the security of their network (Jones et al., 2015) and Appazaar supports users in finding new apps (Böhmer et al., 2011) while collecting data for its studies. However, the users of these two apps might already have been concerned with network security or might be early adopters. The participants could be non-average users, and the “sample may not be representative of what all smartphone owners would do” (Ferreira et al., 2012: 12).

Another recruitment approach relies on snowball sampling procedures. Such is the case of Böhmer and colleagues (2011), who advertised their call on popular blogs and social network sites. These media do not necessarily reach a representative sample of the population. For instance, a study promoted a project on Facebook and Twitter “to reach a diverse population sample” (Jones et al., 2015: 1199); however, it made no reflection on the comparatively high popularity of those media among younger cohorts (Greenwood et al., 2016).

Maybe a more relevant issue is the fact that most revised studies did not present the demographics of their samples, which prevents an accurate evaluation of biases. While some articles did not discuss such a lack of information (Ferdous et al., 2015; Wagner et al., 2013), others showed a clear position. For instance, some preferred not to collect demographic data, such as gender or age, to grant the privacy and security of data (Böhmer et al., 2011). Others, instead, argued that the extra steps necessary to collect personal data – such as the acceptance of informed consent – would reduce the willingness to participate in the research (Jones et al., 2015).

Tracking systems are not universal. They can monitor a limited number of smartphones depending on their operating system (OS) version and model. In addition, the technical condition of the smartphone shapes the efficiency of the tracker. These two dimensions exemplify the non-neutrality of the measurement tool and serve to illustrate how tools shape data collection.

Android and iOS dominate the smartphone operating system market, and trackers that focus on them would allow access to most smartphone users. However, different OS versions are in operation at the same time. Notably, the spectrum of Android OS versions is broader than that of iOS, with several handset models of multiple brands relying on different Android OS versions. Any tracking system should ensure its compatibility with all the OSs, brands and models available on the market. However, most of the revised tracking systems argue that they track Android and/or IOS devices but provide no further detail (Holz et al., 2015; Jones et al., 2015; Lee et al., 2014). Few articles have described the limits of the measurement tool, as Wagner and colleagues (2013) did by explaining that the system tracks 1277 different types of devices. Design limitations in the tracking system might exclude some OS versions, which would result in the exclusion of minority handset models from the sample, such as very new or very old models.

Digital logs are the raw data of what are seen as non-intrusive methods for data collection (Kiukkonen et al., 2010; Xu et al., 2016). Researchers have often argued that tracking systems provide everyday life data in everyday settings with little or no disturbance and require no effort from the participants (Holz et al., 2015; Lee et al., 2014; Shin et al., 2012). However, tracking systems affect the battery, memory and processor of the tracked device (Wagner et al., 2013). In this vein, some studies have expressed concern about the battery drain of monitoring systems (Rahmati et al., 2012; Srinivasan et al., 2014; Yan et al., 2012). Concretely, tracking systems have a considerable impact on smartphones with limited RAM and weak batteries. The capacities of the smartphone, therefore, could act as a limiting technical dimension preventing individuals from participating in a tracking study. However, none of the previous studies have reported the available capacities of the tracked smartphones.

Thus, the capacity of the tracking systems and the technical conditions of the smart-phones prevent the inclusion of the oldest smartphones in tracking studies. This affects older people more, as they tend to have the oldest devices, renew them less frequently (Jacobson et al., 2017) and often use second-hand devices inherited from their relatives (Fernández-Ardèvol & Ivan, 2013; Oreglia & Kaye, 2012).