Airline Accidents and Media Bias:
New York Times 1978-1994

Note: This is an update of a study that was was originally published in 1996. The original study categorized the broad categories of events involving airline passenger fatalities that received coverage, and the magnitude of that coverage. The study described below covers the same time frame as the original study, 1978-1994, but uses an updated database of accident information, as well as a combination of linear and logistic regression techniques provide insights into the coverage the newspaper gave to events that involved at least one death to a passenger (excluding eventers where the only passengers killed were sabateurs, hijackers, or stowaways.

• What independent variables were most significant with respect to predicting if a particular fatal airline event would receive coverage in the New York Times.
• Whether a model based on a logistic regression based on those independent variables can serve as a good predictor of whether an event would receive coverage.
• What kind of regression model could be used to estimate the number of articles that a particular accident my generate.

A common perception by the aviation safety community is that the news media gives extra attention to some kinds of fatal airline events. This study intended to do three things to show if the data supported this common perception:

1. Identify categories of events that generated a disproportionate amount of new coverage,
2. Develop a regression model that would identify the characteristices associated with an event that would generate one or more New York Times articles,
3. Develop a model to predict the number of articles that an event would generate.

Summary: An analysis of events from 1978 to 1994 and New York Times articles about those events showed that 59.5% (289 of the 486) generated at least one article, and that there were a total of 2,730 articles. Just over half (145 of 289) of all the covered events generated only article. While only 9.3% (27) generated 10 or more articles, these events accounted for 72.1% of of all the articles written. A closer review revealed the following subcategories with a disproporitonal share of the articles:

• US events (occurring in the US or on US-registered aircraft) - 19.1% of all events, 28% of the covered events, and 57.6% of all articles
• Events that caused deaths due to deliberate actions (sabotage, hijacking, and military action) - 7.4% of all events, 10.7% of the covered events and 50.1% of all articles
• US events that were due to deliberate actions - 1.2% of all events, 2.1% of the covered events, and 22.1% of all articles
• Events involvng jet transports - 37.9% of all events, 63.7% of the covered events, and 90.8% of all articles

Predicting coverage with two regression models
Two regression models were created to identify the independent variables that could be used to predict the likelihood that a particular event would receive coverage. In the first model, the dependent variable was a binary varible indicating if an event led to at least one article, and 12 independent variables were analyzed to see which were useful in an appropriate regression model, and four of them were statistically significant .

Using a randomized training set comprised of 75% of the entire data set in a generalized linear model, four of the 12 independent variables were used to predict New York Times coverage on the test set with 76% accuracy on the test set, compared to a baseline value of 59.5% (the percentage of events in the test set that had at least one article).

The independent variables used in first model were:

1. Fatal.Pax - The number of passengers members killed.
2. Jet.Transport - A binary variable indicating if the event involved a jet transport.
3. deliberate - A binary variable indicating if the event was due to some deliberate action such as hijacking, sabotage, or military action.
4. USA.Event - A binary variable indicating if the event either occurred on or above US territory or involved a US-registered aircraft.

In the second regression model, which assumed a Poisson distribution of the number of articles, the number of artilces was the dependent variable. The entire fatal event data set was used, and seven of the 12 independent variables, including the four from the first model, were used to predict the number of articles.

The independent variables used in second model were:

1. Fatal.Pax - The number of passengers members killed.
2. Fatal.Other - The number of people killed who were neither passengers or crew onboard the aircraft.
3. Total.pax - The total number of passengers onboard the aircraft.
4. Jet.Transport - A binary variable indicating if the event involved a jet transport.
5. Scheduled - A binary variable indicating if the event involved a scheduled flight.
6. deliberate - A binary variable indicating if the event was due to some deliberate action such as hijacking, sabotage, or military action.
7. USA.Event - A binary variable indicating if the event either occurred on or above US territory or involved a US-registered aircraft.

Detailed information about the analysis
New York Times Index news coverage data
Updated fatal event data
Fatal event data from the original study
R code used for the analysis
Completed analysis (HTML)
Completed analysis (on RPubs)