This appendix provides methodological details for the statistical model introduced in Chapter 4. It also details the data-generation process for latent Dirichlet allocation (Blei et al. 2003) used in Chapter 2 to examine the articles that cite Fenno (1978).

STATISTICAL MODEL AND ESTIMATION

Suppose that each senator i during each year in which he or she is in Congress t (t = 2005, 2006, 2007) can allocate attention to K (k = 1, …, K) issues in his or her press releases. Represent the attention senator i allocates to issue k during legislative session t as πitk· πitk is the expected proportion of press releases from senator i that is about the kth issue. Collect the attention a senator allocates to all K issues during a session into the vector πit = (πit1, …, πit K).

The attention a senator allocates to the issues probabilistically determines the topic of each press release. Suppose that each press release j (j = 1,…, Dit)from senator i in the tth session has only one topic, which I represent with τitj:if τitjk = 1, then the press release is about the kth topic, and the other K – 1elements of τitj are equal to zero. It follows that τitj is a draw from a multinomial distribution:

τijtπit ∼ multinomial(1, πit)

Given the topic of each press release, I draw its content. As a result of the preprocessing steps, I represent the jth press release from senator i in the tth legislative session as a count vector yijt, where typical element yijtz counts the number of times the zth word appears in press release yijt.