Introduction

People’s perceptions about health risks are impacted in part by who they see portrayed as at risk, including media portrayals [1]. Specifically, theories in communication and social psychology posit that the perception of prevalence (commonly called the descriptive norm) governs how individuals assess their own risk and adopt prophylactic behaviors [2,3][4,5]. Accordingly if an individual perceives there is low risk among persons they know or see, they likely infer they too are at low risk. In the case of media and disease, these theories assume that when viewers engage with media they ask themselves “do I look like the people who are portrayed as at risk?” [6,7].

It is also well established in the literature on persuasion and health behaviors that people that are perceived as similar to oneself are more influential than people who are perceived as not similar [8]. For instance, through constant social comparison processes, individuals adjust their risk prevalence estimate by changing the reference group to persons who share their traits. Consequently, messages for altering risk reduction behaviors will be more persuasive if delivered by similar others. When the portrayal of those at risk in media is aligned with those of the risk group it is likely to modify perception of risk from those who are indeed most at risk.

These theories garner substantial attention in the evaluation of traditional media (e.g., print or television) [9]. Despite some work on social media [10–12], most analyses are heavily focused on written content in part because early social media was primarily text-based and the most well known analytic strategies rely on text analysis [12–14]. The omission of image analyses focused on social normative theories is critically important as social media is now the dominant source of media consumption among the public [15] and the fastest growing social media channels focus on user generated images and/or video such as Instagram and YouTube. [16].

Furthermore, health disparities are well documented among minority populations, and significantly degrade health outcomes [17]. Perceived lack of representation in visual health communications may further these disparities. This is especially concerning since for many health conditions minority populations are at higher-risk.

Methods

We assembled a dataset of public Instagram posts from January 2017 and July 2018 using InstaLooter [35]. Posts with the hashtag “#HIV” were collected to capture images that posters self-labeled as primarily related to HIV. Hashtag(s) serve as anchoring devices to label the content of Instagram posts and make it searchable for Instagram users. The posts were then restricted to those authored in American English (i.e., posts with spellings common to the vernacular in the United States) using automated language identification [36] applied to the caption of the post to capture posts that are most likely intended for a US-based audience.

The primary outcomes focused on assessing the demographic attributes of persons included in images with #HIV. We focused on images that were shared and not the account holder’s profile image, to measure how specific images about #HIV include persons reflective of the at-risk population. We relied on automated image recognition by Clarifai [37] to evaluate images to discover faces and then estimate its associated demographic profile (age, gender, and race or ethnicity). If more than one face was present in the image, demographics were estimated for each detectable face; if a face was not detectable the person would not be counted (e.g., a back profile of someone looking into the distance). These techniques rely on machine learning, specifically convolutional neural networks (CNN), that are optimized to mirror human judgment, similar to if a researcher reviewed each of the images gathered from Instagram. The CNN ingests an image and returns a probability distribution of the demographic profile of each face. We assigned the demographic characteristic with the highest probability (e.g., we assigned ‘female’ for gender if the model returns a corresponding probability of 70% for female and 30% for male). These methods typically achieve 85% accuracy in face recognition, 88% accuracy in gender recognition, 79% accuracy in racial/ethnic recognition, and 93% accuracy for age estimation when age is grouped [10,38–40]. The automated image recognition analyses can be replicated using the documentation provided by Clarifai [37].

The demographic profiles in the images were normalized to the total number of detectable faces, thereby reflecting the prevalence of each demographic group (age, gender, and race/ethnicity) within #HIV conversations. The resulting representation of demographics in the images in the dataset were described using general descriptive statistics including means, standard deviations, frequencies, and percentages, and compared to the most recent HIV incidence data from the US CDC [29]. Because we analyzed the full dataset of #HIV instagram posts we did not use confidence intervals. In addition to #HIV, we carried out the aforementioned steps on the subset of #HIV posts related to PrEP by restricting the analyses to posts that also contained the keywords “PrEP” or “Truvada”, regardless of capitalization and inclusive of partial words (e.g., “preplife,” “#prep,” “truprep”), anywhere in the caption.

This study was deemed exempt from ethics board review by the University of California San Diego institutional review board because all data was publicly available and the researchers did not interact with the users. However, given the sensitivity of topics discussed and vulnerable groups participating in #HIV conversations, we decided to not include example images after several peer reviewers brought forth concerns about sharing the blurred example images that were included in our initial submission.

Results

During the 18 month study period, we discovered 26,766 Instagram posts containing #HIV authored in American English. These included 10,036 (37.5%) images with a detectable human face showing a total of 18,227 faces (mean = 1.8, standard deviation [SD] = 1.7). Disparities between the demographic profile of faces in #HIV posts were contrasted with the national HIV at-risk demographic profile according to the estimates in the 2017 US CDC HIV Surveillance Report [29] (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Demographic profiles of #HIV instagram posts vs. at-risk population.

https://doi.org/10.1371/journal.pone.0231155.t001

The distribution of faces in #HIV images in terms of age was skewed higher than the distribution of newly diagnosed HIV-positive individuals. For example, the majority (47%) of faces represented in #HIV images were 35–39 years old; however, most new HIV diagnoses occur among people that are 25–29 years old.

The distribution of women in #HIV images was skewed higher than the distribution of newly diagnosed HIV-positive individuals. For example, 41% of faces in #HIV images were female, which was more than twice that of new HIV diagnoses (19.2%). Conversely, men were underrepresented in #HIV images, with 58.7% of the faces being male compared to most new HIV diagnoses (80.8%).

The distribution of faces in #HIV images in terms of race and ethnicity skewed towards people who were white, and among minority populations, overrepresented Asians and underrepresented blacks and Hispanics. For example, the majority (42.6%) of faces represented in #HIV images were identified as white; however, most new HIV diagnoses in the US occur among black people (43.6%). Only 30.8% of the faces in #HIV images were identified as black. Similarly, only 13.4% of the faces represented in #HIV images were identified as Hispanic compared to new HIV diagnoses (24.7%). Conversely, Asians were overrepresented representing 12.9% of the faces in #HIV images compared to new HIV diagnoses (2.5%).

Among the subset of #HIV images that also included terms indicative of discussing PrEP, the demographic misrepresentation attenuated only slightly. For instance, the subset of PrEP related facial images also skewed towards people who are older (52.2% were 35–39 years old) and female (30.7%), and less towards those who were black (33.2%) or Hispanic (16.2%).

Discussion

Our case study found that thousands of Instagram images are contributing to the #HIV discussion; however, the demographic profile of faces in these images does not reflect the demographic profile of the HIV at-risk community. Low risk persons (e.g., white females) are consistently overrepresented and persons with high risk (e.g., black and Hispanic men) are underrepresented. This may be furthering existing health disparities affecting these groups by biasing their perceptions of disease risk. Acting to mitigate this disparity can theoretically make HIV prevention and control advocacy on Instagram more effective. Just as important, our case study is designed to raise questions about why these findings exist for HIV and the larger question about how the field can begin adopting strategies to study norm setting in health-related images on social media.

Limitations

Not everyone at risk uses social media; however, Instagram’s demographics align with the target audience of reaching young persons who are at greater risk of acquiring HIV [29]. For example, Instagram is the most popular among young adults (18–29 years old) and people who identify as minorities, especially black or Hispanic [16]. Since we only used public images, we did not capture how HIV prevention and control messages are being broadcast in private networks. Second, our study did not examine underlying factors that could be contributing to the skewed demographics. Although our work identified skewed demographics and suggested a way to fix it, we acknowledge that future work should focus on factors, such as existing social and cultural norms, that could be contributing to this disparity to address the root cause. Third, we relied on usage of English language as a proxy for location. Unlike Twitter, Instagram does not geo-tag posts. Last, at present the technologies around automated image analyses misrepresent multiethnic and non-binary gendered persons. However, we use these technologies to set an early baseline for how to consider demographic norms and biases in health-related social media. In the future as methods for automated image analyses improve it may be possible to even study potential multiethnic or non-binary faces. As algorithmic identification of these identities become available it may help improve our understanding of norm setting on social media or perpetuate stereotypes that might undermine these communities.

Conclusions

Our case study of demographic profiles on Instagram among #HIV posts charts the course for future research to track and respond to disparities between at-risk population trends and their presentation on social media. There is tremendous potential as automated image recognition methods improve and this underappreciated issue comes to light to engender changes that can impact population-level HIV prevention and control. These changes can be further amplified by replicating our strategy for other health issues and on other social media platforms with visual content.