Racial Framing in Quarterback Commentary: The Research

Four separate studies, spanning 1996 to 2019 and two entirely different research methods, have each independently found the same specific pattern in how quarterbacks are described by race: athletic ability attributed disproportionately to Black players, intelligence and mental preparation attributed disproportionately to White players. This is a review of what those studies actually measured, what they found, and how the field’s method for measuring it changed from manual transcript coding to large-scale computational text analysis.

The Starting Point: Rada (1996)

James Rada’s 1996 study, published in the Howard Journal of Communications, was among the first to apply systematic content analysis to how network television broadcasters described players during professional football coverage. Using a coding scheme applied to broadcast transcripts, the study found that announcers described Black players in terms weighted toward physical and athletic qualities more often than White players, while intellectual and character-based descriptors — composure, leadership, football intelligence — were disproportionately applied to White players. Rada characterized the pattern as a largely unconscious feature of broadcast language, not an accusation against any individual commentator, a framing every subsequent study in this line of research has maintained.

A Partial Replication: Billings (2004)

Andrew Billings revisited the same question eight years later with a comparable design: a systematic coding of quarterback descriptors across a large sample of college and professional broadcast commentary, using an 18-item coding scheme applied to thousands of individual descriptors. The result was not a simple confirmation of Rada’s findings — it was a more specific and, on balance, more encouraging one. Billings found that most of the categories where Rada had documented a racial gap showed no significant difference by the early 2000s: composure, commitment, courage and experience were described similarly regardless of a quarterback’s race. One pattern, however, held: Black quarterbacks were still disproportionately described as succeeding because of athletic ability, and White quarterbacks were still disproportionately described as failing because they lacked it. The finding mattered specifically because it wasn’t a blanket claim that nothing had changed — it isolated exactly which part of the earlier pattern had persisted and which parts had not, which is a harder and more useful result than either “bias is gone” or “bias is unchanged” would have been.

Extending the Question to Draft Coverage: Mercurio and Filak (2010)

Eugenio Mercurio and Vincent Filak took the same underlying question into a different kind of text: pre-draft scouting coverage, not live broadcast commentary. Their 2010 study, also published in the Howard Journal of Communications, coded 4,745 individual attributions describing Black and White NFL quarterback prospects across a ten-year span of draft coverage. The pattern echoed what Billings had found in broadcast commentary: Black prospects were more frequently described in terms of physical tools and less often in terms of mental readiness, while White prospects were more often described as mentally prepared and less prone to in-game mental errors, even when the physical form the coverage took — written scouting analysis, not live commentary — was completely different from either Rada’s or Billings’ source material. That consistency across two very different genres of football media was itself part of what the study treated as significant.

From Manual Coding to Computational Text Analysis: Merullo et al. (2019)

All three of the studies above depended on human coders: reading transcripts, applying a coding scheme, and checking that independent coders agreed closely enough for the results to be treated as reliable. That approach is labor-intensive by design, which limits how much broadcast text any single study can realistically cover — Billings’ 162 hours and Mercurio and Filak’s decade of draft coverage were both, by the standards of the field, large undertakings, and both predate the kind of large, freely available play-by-play data that eventually reshaped the rest of football analytics.

A 2019 paper by Jack Merullo and coauthors, presented at the EMNLP-IJCNLP conference, took a different approach entirely. The researchers built a dataset called FOOTBALL: 1,455 broadcast transcripts spanning roughly six decades of American football coverage, automatically annotated to identify around 250,000 individual player mentions and linked to racial metadata for each player. The study did not manually code each mention into a category. Instead, it used a statistical language model to measure which words were disproportionately associated with mentions of white versus non-white players, while explicitly controlling for a major confound the earlier studies had to manage through sampling design — player position, since certain positions have historically had different racial compositions and different commentary conventions attached to them independent of any individual player’s race.

The computational result lined up with the three manual-coding studies that preceded it: after controlling for position, commentators disproportionately used athleticism- and physicality-related language for non-white players and intelligence- and character-related language for white players. The significance of this isn’t that a fourth study found a new bias — it’s that a completely different measurement method, applied to a dataset roughly two orders of magnitude larger than what manual coding could realistically process, converged on the same specific asymmetry that Rada had first documented in 1996.

How Coding Reliability Works in Manual Content Analysis

The three manually coded studies share a methodological structure worth spelling out, since it’s what separates a systematic content analysis from an individual’s subjective read of a broadcast. Each used a predefined coding scheme — a fixed list of descriptor categories, decided before the coding began, not categories invented on the fly to fit whatever was found. Multiple independent coders then applied that same scheme to the same transcripts, and the studies report a measure of agreement between those coders (inter-coder reliability) as a check that the categories were being applied consistently, not reflecting one person’s individual judgment. A descriptor only counted toward a study’s reported pattern if independent coders, working from the same fixed scheme, classified it the same way. That structure is what allows a claim like Billings’ “athletic explanation persisted while other categories converged” to be read as a measured result, not an impression — the categories were fixed in advance, and the counts came from more than one person’s reading of the same material.

How This Research Connects to the Broader Positional-Stacking Literature

The quarterback-specific studies above sit inside a wider body of sports-sociology research on what’s generally called racial stacking (or positional segregation) — the pattern, documented across multiple sports, of players being disproportionately represented in certain positions along racial lines, with positions broadly characterized as “central” or “thinking” positions (quarterback among them) historically overrepresented by White players, and positions characterized as relying more heavily on raw athleticism overrepresented by Black players. The broadcast-commentary and draft-coverage research summarized above doesn’t measure stacking directly — it measures language, not roster composition — but the specific descriptive pattern it identifies (athletic framing versus intelligence framing) is the same conceptual pairing that positional-stacking research uses to describe how positions themselves have historically been characterized. The two literatures ask different questions with different methods, but they converge on the same underlying distinction between how physical ability and mental preparation get attributed along racial lines in football specifically — a distinction that also shows up, framed very differently, in research on how teams themselves evaluate quarterback decision-making as a situational skill, not a purely physical one.

What the Pattern Across Four Studies Actually Shows

Read together, these four studies don’t describe a single unchanging bias so much as a specific, narrow, and unusually well-replicated one — which is itself worth pausing on. Research on media bias is often criticized, fairly, for resting on a single sample or a single research team’s coding decisions. This particular finding does not have that vulnerability: it was measured by three separate research teams, using two fundamentally different methods (manual content coding and automated statistical text analysis), across two different genres of football media (live broadcast commentary and written draft coverage), spanning more than two decades, and it held up every time. The broader claim that broadcasters describe Black and White athletes differently across every dimension has not held up consistently — Billings’ own data showed real convergence in several categories over time. What has held up, independently, across manual coding in the 1990s, manual coding in the 2000s applied to a different medium, and automated analysis of a much larger corpus in the 2010s, is one specific pairing: athletic explanation for success and failure attached more to Black quarterbacks, mental and preparation-based explanation attached more to White quarterbacks. None of the researchers involved frame this as a claim about any individual commentator’s intent. It’s a pattern detectable only in aggregate, across large numbers of broadcasts and years — which is precisely why it took four studies, three different research teams, and eventually a purpose-built quarter-million-mention dataset to establish it with any confidence.