Elsevier

Social Science Research

Volume 45, May 2014, Pages 73-88
Social Science Research

Modeling time-series count data: The unique challenges facing political communication studies

https://doi.org/10.1016/j.ssresearch.2013.12.008Get rights and content

Highlights

  • Presents benefits of Poisson autoregression (PAR) for temporal media count data.

  • Replicates three previously published studies in the analysis.

  • Compares PAR to transfer function, negative binomial, Koyck and log-normal models.

  • Describes how to illustrate dynamic count model results.

Abstract

This paper demonstrates the importance of proper model specification when analyzing time-series count data in political communication studies. It is common for scholars of media and politics to investigate counts of coverage of an issue as it evolves over time. Many scholars rightly consider the issues of time dependence and dynamic causality to be the most important when crafting a model. However, to ignore the count features of the outcome variable overlooks an important feature of the data. This is particularly the case when modeling data with a low number of counts. In this paper, we argue that the Poisson autoregressive model (Brandt and Williams, 2001) accurately meets the needs of many media studies. We replicate the analyses of Flemming et al., 1997, Peake and Eshbaugh-Soha, 2008, and Ura (2009) and demonstrate that models missing some of the assumptions of the Poisson autoregressive model often yield invalid inferences. We also demonstrate that the effect of any of these models can be illustrated dynamically with estimates of uncertainty through a simulation procedure. The paper concludes with implications of these findings for the practical researcher.

Introduction

Political communication scholars commonly study an outcome that is a count variable, such as the number of stories printed on a subject in a given time frame (Flemming et al., 1997, Fogarty, 2005, Peake and Eshbaugh-Soha, 2008, Rhee, 1996, Sellers, 2000, Ura, 2009). Since the unit of analysis is a day, week, or month of coverage, these data often are time dependent as well. Many studies of such data focus primarily on specifying a reasonable time series model as if these outcomes were normally distributed, ignoring the count aspect of these data. Certainly, if we were limited to addressing only one of these issues, time dependence is the more important, as it speaks to non-independence of observations and more importantly the dynamic form of causality in time dependent data. (In other words, an input can have a consequence on the outcome many time periods into the future.) However, substantial developments have been made in the creation of Poisson models for time dependent data (Brandt et al., 2000, Brandt and Williams, 2001, Blundell et al., 2002, Schwartz et al., 1996), implying that models of media outcomes can handle both time ordering and the proper distribution for the outcome.1

In this paper, we illustrate how to apply time-series count models to media outcomes by replicating several important studies in the field of political communication. In so doing, we accomplish two things: We show how several results change under this new specification, and we illustrate how to interpret quantities in a count model for time series data. We begin this paper by describing the background on research about media event counts. Second, we describe various approaches to modeling time series of event counts, including the Poisson autoregressive model and its properties. Next, we conduct three replication analyses, describe how model-sensitive results are in each, and illustrate the best way to interpret these results. The three replications revisit Flemming et al.’s (1997) analysis of news stories about church and state issues, Peake and Eshbaugh-Soha’s (2008) study of television coverage of energy policy, and Ura’s (2009) study of newspaper coverage of homosexuality. We conclude by discussing the implications of our results for the practical researcher.

Section snippets

Background on models of event counts in political communication

As with most social science research, the sophistication of political communication analysis has been increasing over the past few decades. This includes an interest in modeling media count data over time. Instead of relying on descriptive measures alone, scholars are increasingly trying to systematically explain variations in coverage of political phenomena over days, months, and years (Flemming et al., 1997, Fogarty, 2005, Peake and Eshbaugh-Soha, 2008, Ura, 2009). At the most basic level,

Approaches to modeling time-series count data

Within the literature on political communication, most of the empirical models take some approach that handles either the count aspect of the data or the time series aspect. To illustrate how inferences and forecasts can differ among plausible models, our paper refits several past models from this subfield using four models that are commonly used for analyzing count data in political communication research. The first such model is transfer function analysis (Box and Tiao, 1975, Box et al., 2008

Replication analyses

In this section, we reanalyze Flemming et al.’s (1997) study of church and state coverage, Peake and Eshbaugh-Soha’s (2008) study of energy policy coverage, and Ura’s (2009) study of national coverage of homosexuality. We choose these three studies because they all have made significant contributions to the study of political communication, helping to establish the substantial effect that presidential and Supreme Court actions have on the news media’s agenda. Further, these studies all use

Implications for the practical researcher

Political communication scholars over the past decade have enhanced our understanding of the press and politics by augmenting qualitative and cross-sectional studies with time-series studies of media coverage of political phenomena. By adding macro-level explanations of press attention to politics and government to micro-level explanations, the richness of political communication theories grows. In studies where time defines the unit of analysis, scholars have rightfully begun to incorporate

Acknowledgments

For helpful assistance, we thank Keith T. Poole, Richard L. Vining Jr., and several anonymous reviewers. For sharing code, we thank Patrick T. Brandt. For sharing data used in this paper, we thank B. Dan Wood, Joseph D. Ura, Jeffrey S. Peake, and Matthew Eshbaugh-Soha. For additional data shared, we thank Tobey Bolsen and Patrick Sellers. Complete replication data and code are posted on Monogan’s Dataverse: http://hdl.handle.net/1902.1/16677.

References (57)

  • R. Blundell et al.

    Individual effects and dynamics in count data models

    J. Econom.

    (2002)
  • M. Barakso et al.

    Winning coverage: news media portrayals of the women’s movement, 1969-2004

    Harvard Int. J. Press/Polit.

    (2006)
  • K. Barnhurst

    The makers of meaning: National Public Radio and the new long journalism, 1980-2000

    Polit. Commun.

    (2003)
  • M.A. Baum et al.

    Economic class and popular support for Franklin Roosevelt in war and peace

    Public Opin. Quart.

    (2001)
  • R. Best

    Situation or social problem: the influence of events on media coverage of homelessness

    Soc. Probl.

    (2010)
  • T. Bolsen

    The construction of news: energy crises, advocacy messages, and frames toward conservation

    Int. J. Press/Polit.

    (2011)
  • G. Box et al.

    Intervention analysis with applications to economic and environmental problems

    J. Am. Stat. Assoc.

    (1975)
  • G.E.P. Box et al.

    Time Series Analysis: Forecasting and Control

    (2008)
  • P.T. Brandt et al.

    A Bayesian Poisson vector autoregression model

    Polit. Anal.

    (2012)
  • P.T. Brandt et al.

    A linear Poisson autoregressive model: the Poisson AR(p) model

    Polit. Anal.

    (2001)
  • P.T. Brandt et al.

    Dynamic modeling for persistent event-count time series

    Am. J. Polit. Sci.

    (2000)
  • C. Brown et al.

    Daring to be heard: advertorials by organized interested on the Op-Ed page of the NYT, 1985–1998

    Polit. Commun.

    (2001)
  • C.G. Broyden

    The convergence of a class of double-rank minimization algorithms

    J. Inst. Math. Appl.

    (1970)
  • R.H. Byrd et al.

    A limited memory algorithm for bound constrained optimization

    SIAM J. Sci. Comput.

    (1995)
  • A.C. Cameron et al.

    Microeconometrics

    (2005)
  • A.C. Cameron et al.

    Regression Analysis of Count Data

    (2013)
  • J.J.F. Commandeur et al.

    An Introduction to State Space Time Series Analysis

    (2007)
  • W. Enders

    Applied Econometric Time Series

    (2009)
  • S. Farnsworth et al.

    The international two-step flow in foreign news: Canadian and U.S. television news coverage of U.S. affairs

    Int. J. Press/Polit.

    (2010)
  • R.B. Flemming et al.

    One voice among many: the Supreme Court’s influence on attentiveness to issues in the United States, 1947–1992

    Am. J. Polit. Sci.

    (1997)
  • R.B. Flemming et al.

    Policy attention in a system of separated powers: the dynamics of American agenda setting

    J. Polit.

    (1999)
  • R. Fletcher

    A new approach to variable metric algorithms

    Comput. J.

    (1970)
  • B.J. Fogarty

    Determining economic news coverage

    Int. J. Public Opin. Res.

    (2005)
  • C. Forbes et al.

    Statistical Distributions

    (2011)
  • D. Goldfarb

    A family of variable metric updates derived by variational means

    Math. Comput.

    (1970)
  • T. Groeling et al.

    Is network news coverage of the president biased?

    J. Polit.

    (1998)
  • G.K. Grunwald et al.

    Non-gaussian conditional linear AR(1) models

    Aust. New Zeal. J. Stat.

    (2000)
  • D.N. Gujarati et al.

    Basic Econometrics

    (2009)
  • View full text