Abstract
Understanding the impact of digital media on news inequalities is critical for democracy. The literature on incidental exposure challenges the idea that major platforms shrink information gaps, and research has turned to the identification of variables that explain how those gaps widen or persist. We use Latent Class Analysis to operationalize the metaphor of “attracting the news” and investigate incidental exposure as both an individual trait and temporal state. We link the top stories on Facebook during an election cycle with incidental exposure measures to explore news attraction, incidental exposure, and news engagement. We find some evidence incidental exposure may close information gaps, but that it does not close engagement gaps. Results contribute to theory about how digital media shape inequalities in news consumption and engagement.
Geolocation Information
This study was conducted in and analyzes data collected in the United States.
Disclosure Statement
The author declares no conflicts of interest.
Data Availability
The dataset supporting this analysis is available at doi: 10.17632/8v9594h3jm.1.
Notes
1 The weights do not inflate standard errors. We compared the first and third models in Table 1 to models without weights. The unweighted estimates are similar to the weighted estimates.
2 We compared the first and third models in Table 1 to models using listwise deletion. The substantive interpretations of model estimates are similar.
3 Previous studies suggest that survey respondents typically underestimate their exposure, resulting in point estimates lower than the actual population parameters (González-Bailón and Xenos Citation2023). However, this error in measurement may not bias causal inferences if all variables are impacted in the same way (King et al. Citation2021).
4 As a robustness check, we also tested a “high-effort” dependent variable consisting only of those engagement behaviors requiring relatively high effort to perform. However, an exploratory factor analysis (EFA) found no empirical difference between the high- and low-effort items, and we therefore excluded the variable from the analysis.
5 We compared the first and third models in Table 1 with ordinary least squares (OLS) models. In both cases, the MLM model fits the data better than the OLS model, displaying lower information criteria and statistically significant log-likelihood tests (for the first model, χ2 = 5.16, p = .023; for the second model, χ2 = 29.17, p < .001).