This paper examines language patterns in US television news coverage of police killings. We first document that the media employ semantic structures that obfuscate responsibility—such as passive voice, nominalizations, and intransitive verbs—more frequently for police killings than for civilian killings. Using field variation and an online experiment, we demonstrate that these language differences matter. In the field, we find that people who happened to have taken a survey just after more obfuscatory coverage of a police killing are more likely to support police funding. In our online experiment, participants are less likely to hold police officers morally responsible and demand penalties when exposed to obfuscatory language, especially when the victim is unarmed. Returning to the news data, we find higher use of obfuscatory language when victims are unarmed, when video footage is available, or when the suspect is not fleeing—in other words, situations when obfuscation matters most. Turning to the causes of this differential obfuscation, our evidence is not consistent with either demand-side drivers or supply-side factors associated with TV station ownership and political leaning. Instead, our results point to original narratives crafted by police departments as a more likely driver of obfuscation. Our study emphasizes the importance of considering semantic language structures in understanding how media shapes perceptions, extending beyond coverage quantity and slant.

The global increase in refugee flows and anti-migrant politics has made it increasingly urgent to understand how migration translates into anti-migrant sentiment. We study the mass exodus of Venezuelans across Latin America, which coincided with an unprecedented decrease in migrant sentiment in the countries which received the most Venezuelans. However, we find no evidence that this decrease occurred in the regions within-country that received the most migrants. We do this using multiple migrant sentiment outcomes including survey measures and social media posts, multiple levels of geographic variation across seven Latin American countries, and an instrumental variable strategy. We find little evidence for heterogeneity along a range of characteristics related to labor market competition, public good scarcity, or crime. If anything, local migration increases migrant sentiment among those most directly exposed to these pressures. We also find that local migration induces meaningful, repeated contact between migrants and natives. The results are consistent with anti-migrant sentiment being driven by national-level narratives divorced from local experiences with migrants.

This paper provides evidence of the causal effect of church attendance on petty crime by using quasi-random variation in the number of Sundays when it precipitated at the specific time of most religious services. Using a novel strategy, I find a narrow time window when most individuals attend church. Based on a panel between 1980 and 2016, I find that one more Sunday with precipitation at the time of church increases yearly drug-related, alcohol-related, and white-collar crimes. I do not find an effect for violent or property crimes. These effects are driven by more religious counties. Previous evidence showing negative effects of church attendance on the demand for alcohol and drugs is consistent with a demand-driven interpretation of the results presented.

Standardized test scores aggregate item (question) responses into a single scalar, collapsing distinct skills into an undifferentiated measure of proficiency. Which of these component skills matter most for long-run economic outcomes is a question that aggregate scores cannot answer. We develop a framework that looks both inside the score — re-weighting items by their predictive power for a chosen outcome — and inside the item — using the digitized text of each question to identify what skills drive the variation in item-level returns. We apply this framework to over 3,500 items linked to approximately 1 billion student–item records and adult earnings from Texas administrative data. Achievement scales that weight items by estimated economic ‘prices’ yield white-minority gaps roughly 45% larger than conventional scales and substantially reorder individual student rankings. To interpret these prices, we show that item text carries economically relevant information beyond standard psychometric characteristics, and we develop a novel text-based mapping of items to over 600 Common Core State Standards. The mapping reveals that procedural, spatial, and automation-exposed mathematics skills have the highest estimated returns, while basic reading comprehension dominates more fine-grained reading skills. To our knowledge, this provides the first standards-based evidence on which K–12 curricular skills predict long-run labor-market outcomes.

We estimate how exposure to religious services affects U.S. voting. Novel sermon corpora show a sharp spike in political content on the Sunday before presidential elections. Exploiting quasi-random rainfall during typical service hours before elections — Precipitation at Time of Church (PTC) — and controlling for election day and weekly precipitation, a one–standard deviation increase in PTC lowers county two-party Republican vote share by 0.6 percentage points. The effect is driven by reduced Republican turnout, with near zero effects on Democratic turnout, indicating demobilization rather than persuasion. Individual-level CES estimates confirm that effects concentrate among regular church-attending Christians — particularly White Evangelicals — and are absent for non-churchgoers, consistent with church-based mobilization. Placebo tests further support the exclusion restriction: effects are absent for non-churchgoers who face the same weather, and rain outside the service window — including on Saturday and Sunday afternoon — has no detectable effect on voting.

I estimate how routine crime coverage capitalizes into housing prices by separating the realized amenity from the information conveyed about it. I assemble a novel dataset linking nearly all U.S. gun homicides (2014-2018) to 1.5 million station-days of local news and to housing transactions, which allows me to benchmark crimes the media could cover against those it does. Identification leverages predictable crowd-out in airtime from local sports schedules. Coverage of a nearby homicide lowers sale prices by 8.1-8.5 percent. Additional evidence points to demand-side responses: buyers react to news reports while short-run housing supply remains unchanged. Having established this information channel, I document systematic coverage disparities across race: incidents with non-white victims or suspects are more likely to be covered than white-white cases, even conditional on extensive incident and neighborhood characteristics. Using the price estimates, I quantify the implied valuation effects of these coverage gaps.