3.2 Identifying and Validating Frames by Issue Area
3.2.1 Identifying Frames
As a reminder, the data set includes speeches in the Congressional Record from the104th(1994/1995) through the112th(2011/2012) Congresses on policy. Additionally, general issue areas are coded according to the Comparative Agendas Project major topic codes.1 Once classified by issue area, I break down the speeches within each issue area by paragraph. Then I fit a dynamic topic model to identify the latent topics present. For the dynamic topic model, paragraphs are treated as documents and congresses are treated as time units.
A topic model identifies latent concepts or topics in text by identifying which terms or words cluster together. Additionally, it estimates the probability that a document belongs to the identified clusters. Latent topics are the clusters identified by the model. There are different types of topic models that one can use to do this. I specifically fit a dynamic topic model that uses non-negative matrix factorization (Greene and Cross, 2017, 2015) rather than a topic model fit using latent Dirichlet allocation (LDA) (Blei, Ng and Jordan, 2003; Grimmer, 2010) or a structural topic model (STM) (Roberts et al., 2014). I do so because the language used may change even though the frame or general topic of discussion used remains constant. This problem is called linguistic drift. Dynamic topic models allow for and model this possible fluctuation by fitting separate topic models for each time period (here for each Congress) and then smoothing across them. Either of the other models does not incorporate this smoothing process.
For each issue area, I fit a separate dynamic topic model, which allows for different numbers of latent topics being identified for each issue area. Between 15 and 43 policy relevant latent topics are identified for each issue area.2 For the policy relevant latent topics, the latent topics are then assigned to what frame each uses. As a result, multiple latent topics can be assigned to a single frame, and not every frame is used to discuss every issue area. Each paragraph is then assigned to one latent topic, and the frame assigned to that latent topic is attributed to it.
Put another way, within the policy-relevant latent topics for each issue area, each is assigned to a general frame. The general frames transcend issue area, which means the same set of seventeen general frames are used across the issue areas. There are a number of benefits to the development of
1
For a longer discussion of how the Record is acquired and then coded for policy topic, please see the discussion in Chapter 2 and the appendix at the conclusion of the dissertation.
2
a general scheme that transcends issue area, just as the Comparative Agendas Project provides issue area codes that transcends venue and agenda. The two primary benefits to this approach are: (1) it provides a coding scheme that allows for the direct comparison of the use of frames across different issue areas; and (2) it provides a near complete coding scheme with which almost all documents and paragraphs can be classified. However, one drawback is that individual, particular frames are obscured by the broader categories. These are presented in table 3.1, which provides the name of the frame and a brief description.
I make two modest changes to the list of frames used by the Media Frames Corpus. First, three additional frames are added: “Linked to another issue,” “Vulnerable populations,” and “Placing blame.” The first of these, “Linked to another issue,” is included to indicate discussion in a given paragraph of a different issue area than the one the majority of the speech addresses. As a reminder, issue area labels are assigned at the speech level, while frames are identified at the paragraph level. In Congressional discussion, this might occur for a number of reasons, such as a member, coalition, or party seeking to tie two or more issues together, two or more issues already being tied together in a single piece of legislation, or error in the classification of the speeches into the CAP major topic codes. The second two, “Vulnerable populations” and “Placing blame,” indicate specific subpopulations that members may want to highlight as vulnerable or at fault, such as highlight the vulnerable position of small family businesses or as blaming Fannie Mae for the recession. These are drawn from the framework developed by Schneider and Ingram (1993) and used by Boushey (2016). Finally, the “public opinion” and “quality of life” frames in the original scheme have been collapsed and re-named “the public.” This was done to capture the broader discussion that may occur concerning what the American people want or how Congress and its members works for the American people.
Using the twenty most identifying words for each latent topic within each issue area, each latent topic is categorized into one of the 17 frame categories. This means that multiple latent topics within a single issue area are associated with each frame. To determine how reliable this classification process is, two independent coders identify whether or not a latent topic is on policy and then what frame it is associated with. The inter-coder agreement is at least 90% across the issue areas. This process means that multiple latent topics within in a single issue area may be classified as using the same frame and that some frames may not be used in every issue area. Once all of the latent topics
Table 3.1: Adjusted Categories of the Media Frames Corpus
Frame Description
Economic Discussion and consideration of the costs, benefits, and eco-
nomic considerations and trade-offs.
Capacity & Resources Discussion and consideration of the availability of physical, human or financial resources, and capacity of current systems to carry out policy.
Morality Discussion and considerations of the religious or ethical impli-
cations and virtue (or lack) of what is being considered. Fairness & Equality Whether or not the current law, proposed law, current circum-
stances, or potential circumstances is fair. Considerations of who a bill or policy works for or against.
Legality & Constitutionality Protection of rights and freedoms of individuals, corporations, and the government.
Policy Prescription & Evaluation Discussion and consideration of the goals of the proposed policy, the specific policies aimed at addressing problems, and specific evaluation of them.
Implications of Laws Discussion and consideration of the effectiveness of a law, its enforcement, and implications of passage.
Security & Defense Discussion and consideration of threats to the safety of the nation, community, or individual.
Health & Safety Discussion and consideration of health care, sanitation, and public safety.
Vulnerable Populations** Discussion and consideration of social groups that are viewed as particularly vulnerable.
Placing Blame** Discussion and consideration of who is to blame for the current problem.
The Public* Discussion and consideration of what the public wants and
public opinion.
Cultural Identity Discussion and consideration traditions, customs, or values of a social group with a relation to the issue area.
Political Process Discussion and consideration of the process, politics, and politi- cians.
Size of Government Discussion and consideration of the size and scope of govern- ment.
External Regulation & Reputation Discussion and consideration of the international reputation, position, and foreign policy of the US.
Linked to Another Issue** Nested discussion of another issue area.
Other Miscellaneous and unidentified frames used.
Note: * indicates that “public opinion” and “quality of life” codes from the Card et al. (2015) have been collapsed. ** indicates new codes.
Table 3.2: Five Example Latent Topics and their Assigned Frames from Macroeconomics
Latent Topic Frame Top 20 Terms
D01 Policy Prescription &
Evaluation
tax, cuts, cut, tax cuts, tax cut, income, taxes, estate, relief, credit, pay, tax relief, estate tax, families, middle, tax credit, income tax, increase, code, class
D03 Political Process budget, resolution, budget resolution, balanced, spend-
ing, balanced budget, fiscal, balance, committee, con- gressional, congressional budget, balance budget, office, budget office, budget committee, congressional budget office, president budget, republican, chairman, priorities
D04 Morality work, hard work perseverance, work perseverance, perse-
verance essential, develop work, levels strive education, ethic guide rest lives, guide rest lives, levels strive, de- velop work ethic guide rest, dedication demonstrated, hard work perseverance essential, essential students lev- els, education develop work, work ethic guide, strive education, strive education develop, perseverance es- sential students levels, students levels strive education develop, work ethic guide rest,
D08 Size of Government government, federal, federal government, spending,
money, taxes, state, programs, spend, local, government spending, congress, private, revenue, spent, pay, dollars, taxpayers, washington, control,
D11 The Public people, american, american people, congress, country,
people want, speaker, america, believe, deserve, ameri- can people want, washington, american families, give, work, people country, people know, working, families, understand,
are assigned to a general frame, paragraphs are assigned a general frame based on which latent topic is assigned by the dynamic topic model.
To illuminate this process a little more, I return to the sample of the latent topics from macroeco- nomics. Table 3.2 presents the top twenty terms and assigned frame for five latent topics. For each, latent topic the top twenty words are drawn, and a frame is assigned.