Jump to ContentJump to Main Navigation
The New Economic PopulismHow States Respond to Economic Inequality$

William Franko and Christopher Witko

Print publication date: 2017

Print ISBN-13: 9780190671013

Published to Oxford Scholarship Online: November 2017

DOI: 10.1093/oso/9780190671013.001.0001

Show Summary Details
Page of

PRINTED FROM OXFORD SCHOLARSHIP ONLINE (www.oxfordscholarship.com). (c) Copyright Oxford University Press, 2018. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a monograph in OSO for personal use (for details see http://www.oxfordscholarship.com/page/privacy-policy). Subscriber: null; date: 18 August 2018

(p.177) Appendix A Measurement and Methodology

(p.177) Appendix A Measurement and Methodology

The New Economic Populism

William W. Franko

Christopher Witko

Oxford University Press

Measuring the Public’s Aggregate Awareness of Income Inequality over Time in the State

The most common approach to measuring state opinion in the absence of polls specifically designed to sample state populations (which are quite rare) is to use some form of disaggregation. Disaggregation generally involves combining many national surveys and then disaggregating responses by state. Since the common sample size of a national survey is around 1,000 respondents, the main drawback of this method is that a large number of surveys are required in order to produce accurate estimates of state opinion. Disaggregation is particularly problematic for this study since not nearly enough polls asked the rich-poor question to measure state opinion over time for every state. Recent advances in public opinion estimation, however, have given researchers an alternative to disaggregation when studying attitudes at the state-level. Multilevel regression and poststratification (MRP) is a measurement strategy that allows for the estimation of state opinion using typical national opinion polls. Research has shown that MRP provides accurate estimates of state and local opinion even when using a single national survey (Lax and Phillips 2009a; 2009b; 2012; Park, Gelman, and Bafumi 2006). This is the approach taken here to create a unique measure of state-level awareness of growing economic inequality (also see Franko 2017). Table A1 provides information on the collection of survey data we use to estimate inequality awareness, including the polling organizations that conducted the surveys and the years the surveys were conducted, the exact question wording, coding, and sources.1

Estimating opinion using MRP involves two steps. The first is to model individual responses to the survey question of interest—in this case, whether the individual agrees or disagrees with “the rich are getting richer (p.178) and the poor are getting poorer” statement—using multilevel regression. These models include basic demographic and geographic characteristics of the survey respondents. Similar to pervious work, this study uses the following characteristics to model awareness of inequality: race (black, white, or other), gender (female or male), age (18–29, 30–44, 45–64, or 65+), education (less than high school graduate, high school graduate, some college, or college graduate), state of residence (all 50 states and DC),2 the percentage of Republican Party identifiers in each state, the year of the survey (more on this below), and an indicator for each survey used to account for any potential differences across polls.

The results of the model are then used to predict the probability of agreeing that inequality is growing for every possible individual type (e.g., a white female who is 30–44 years of age with some college education living in Ohio), resulting in a total of 4,896 predicted values. These probabilities are then used in the second step of the estimation, which is poststratification. Poststratification is the process of weighting each individual type probability estimate by the actual proportion of each type in the population using data from the US census.3 This part of the procedure adjusts for any differences between the individuals surveyed in each state and the true state population.

The strategy for creating the over-time state opinion estimates developed here departs slightly from previous studies using MRP to measure attitudes over time. One potential strategy to account for change in opinion over time using MRP is to complete the steps described above for each year under analysis (see Enns and Koch 2013). This approach is most useful when the questions used to measure opinion are asked several times every year so that enough respondents from smaller states are used to have more precise over-time estimates. While the “rich are getting richer” question is asked regularly, it is only asked more than once in a given year on a few occasions. When the survey questions being examined are not asked multiple times on an annual basis, an alternative option is to increase state sample sizes by combining surveys across multiple years. Rather than estimating state opinion for every year, surveys are pooled over specified blocks of time (e.g., a three- or five-year window) to increase the amount of information used when modeling opinion (see Pacheco 2011).

This study expands on these methods by using a completely pooled approach to estimate state opinion over time. In other words, all available survey questions for all available years are included in a single model of (p.179) individual opinion (i.e., the first step of MRP). The procedure is relatively straightforward and only requires researchers to add a time component to their multilevel model. An indicator of time (in this case, the year the survey was conducted) is interacted with state of residence so that a random effect is allowed for every state-year combination.4 This allows for unique estimates of opinion for each state over time by using all available information in one model. The result is a series of aggregate state opinion from 1987 to 2012 indicating the percentage of the public agreeing with the “rich are getting richer” statement. This variable is used throughout the book to account for the public’s aggregate awareness of inequality. Additional details about the measure, along with a number of validity checks, can be found in Franko (2017). (p.180) (p.181)

Table A.1. Survey Questions Used to Estimate Aggregate State Awareness of Economic Inequality

Polling Organization & Question Wording

Year of Survey


ABC News: I’m going to read a few statements, for each, please tell me if you agree or disagree with it . . . The rich are getting richer and everyone else is getting poorer.


1 = agree; 0 = disagree

CBS News: These days, do you feel that the rich are getting richer and everyone else is getting poorer, or is that not the case?


1 = rich getting richer; 0 = not the case

CBS News / New York Times: These days, do you feel that the rich are getting richer and everyone else is getting poorer, or is that not the case?


1 = rich getting richer; 0 = not the case

Harris Poll: Now, we want to ask you about some things some people have told us they have felt from time to time. Do you tend to feel that . . . The rich get richer and the poor get poorer?

1991, 1996, 1997, 2002–2005, 2007–2009, 2011

1 = yes, feel this way; 0 = no, don’t feel this way

Marttila & Kiley: Now I am going to read you a series of statements that will help us understand how you feel about a number of things. Please tell me whether you completely agree, mostly agree, mostly disagree, or completely disagree with each statement I read . . . Today it’s really true that the rich just get richer while the poor get poorer.


1 = completely agree or mostly agree; 0 = mostly disagree or completely disagree

Pew Values Survey: I’m going to read you some more statements on a different topic. Please tell me how much you agree or disagree with each of these statements . . . Today it’s really true that the rich just get richer while the poor get poorer. Do you completely agree, mostly agree, mostly disagree, or completely disagree?

1987–1989, 1991, 1992, 1997, 1999, 2002, 2003, 2007, 2008 (Social Trends Survey), 2009, 2012

1 = completely agree or mostly agree; 0 = mostly disagree or completely disagree

Princeton Survey Research Associates: Here are some statements on a different topic. Please tell me how much you agree or disagree with each of these statements . . . Today it’s really true that the rich just get richer while the poor get poorer.

1991, 1992 (5)

1 = completely agree or mostly agree; 0 = mostly disagree or completely disagree

Note: The Harris Poll also asked the question in 1992 (2), 1993, 1994, 1999, and 2000, and the Pew Values Survey asked the question in 1990 and 1994. The question could not be used for these particular survey years, however, because state-of-residence identifiers are not included in the data.

Source: The Harris Poll surveys were accessed through the Odum Institute’s data archive (http://www.odum.unc.edu/odum/). The Pew Values Survey is available through the Pew Research Center’s website (http://www.people-press.org/values-questions/). All other surveys were accessed using the Roper Center’s iPOLL Databank (http://www.ropercenter.uconn.edu/data_access/ipoll/ipoll.html).

Table A.2. State Time-Series Variables List and Data Sources

Variable Name



Top 10% Income Share

Frank (2009)


Top 1% Income Share

Frank (2009)


Gini Coefficient

Frank (2009)


Poverty Rate

US Census


Union Membership

Union Membership and Coverage Database


Per Capita Income

US Census


Partisanship (Dem. – Rep.)

Enns and Koch (2013)


State Government Ideology

Berry et al. (2013)


Public Mood

Enns and Koch (2013)


% White

US Census


% Age 60+

US Census


Unemployment Rate

Bureau of Labor Statistics


Minimum Wage

Department of Labor


Initiative Qualification Index

Bowler and Donovan (2004), Table 1


State EITC Adoption

Tax Credits for Working Families


Estimating Statistical Models for Time-Series Cross-Sectional Data

While many of the relationships we assess in this book focus on the importance of variation across the states, we also have to consider the fact that economic inequality in the United States has grown dramatically in recent years and that the dynamics of this growth have played quite differently among the states (see chapter 3). This means that a number of (p.182) the statistical analyses we estimate must account for variation both across states and over time. In other words, the data we are modeling to assess some of our questions involve repeated measures of state characteristics over multiple years. This kind of data structure is often referred to time-series cross-sectional (or TSCS) data, and, therefore, a modeling approach for TSCS data is needed when we are analyzing this type of data. When modeling TSCS data researchers must be aware of issues related to repeated measures over time (e.g., nonstationarity and autocorrelation) as well as the clustered nature of the data (e.g., over-time measures grouped by state), and a number of strategies have been proposed to address these common methodological obstacles (Beck and Katz 1995; 1996; Wilson and Butler 2007). Similar to the approach used by numerous scholars, error correction models (ECMs) are used to estimate the models in the book that require the use TSCS data. The following equation, which includes only a single independent variable for clarity, is used to model variables that vary over time among the states:


The ECM is employed here since it is one of the most general time-series models and allows researchers to account for both long- and short-term effects over time (De Boef and Keele 2008; Kelly and Enns 2010). In the above equation each observation is a particular state j in a given time period t. The first difference of the dependent variable, Y, is regressed on a lagged version of the dependent variable and a lagged and differenced version of each explanatory variable. The effect of the independent variable on the dependent variable is represented by β‎1 and β‎2, with the former being an estimate of the short-term effect of the variable. The α‎1 estimate—also referred to as the error correction rate—together with the β‎2 coefficient provides the long-run effect of the variable. The total effect, or long-run multiplier, is estimated as (β‎2/ −α‎1). The main distinction to make between short- and long-term effects is that short-term effects occur immediately, while long-term effects are distributed over time. When the effect of a variable is distributed over time, the long-run multiplier provides an estimate of the total effect of the variable for all periods.

All of the ECMs in the book are estimated using linear regression analyses that allow the intercept for each state to vary randomly (represented by the uj term) as a way to account for any unexplained cross-sectional heterogeneity. Additionally, Hadri Lagrange multiplier stationarity tests were conducted for the main dependent variable used in the analyses. While a (p.183) number of tests indicate evidence of nonstationarity in the levels of our dependent variables, the differenced version of the variables—that is, the version used in all of the models estimated for the book—demonstrate no signs of nonstationarity. Although time-series models that include a lagged version of the dependent variable as a regressor (as all of our models do) tend to have very low, if any, serial correlation of the errors (Beck and Katz 2011), we attempt to limit any remaining serial correlation by including time indicators in the models to account for any trending in the first-differenced dependent variable. Specifically, a set of time variables estimating a second- or third-order polynomial are included in each analysis.


(1.) In addition to general web searches, the Roper Center’s iPOLL Databank (http://www.ropercenter.uconn.edu/data_access/ipoll/ipoll.html) and the Polling the Nations database (http://poll.orspub.com/) were used to find all instances of the “rich getting richer” question being asked.

(2.) Although all available respondents from all states are included in the individual models, many surveys did not interview residents of Alaska and Hawaii. This means that most years do not have opinion estimates for these states and they are not included in the analysis presented below.

(3.) The census’s Public Use Microdata Samples are used for the poststratification stage, which include the 1990 5% sample of population and housing, the 2000 5% sample of population and housing, and the five-year American Community Survey for years 2005–2011. Linear interpolation is used to estimate each population type for years between these surveys.

(4.) This is an extension of the initial MRP modeling strategy suggested by Lax and Phillips (2009b). Instead of allowing only a random effect for every state, a random effect is estimated for every state-year. The method extends the logic of partial pooling across states to partial pooling across states and time. It should be noted that simply adding a time component to the multilevel model without interacting the term with the state indicator (or some other geographic identifier) will not produce a dynamic opinion series.