Measuring the Partisan Behavior of U.S. Newspapers, 1880 to 1980

Abstract

In this paper, we study newspaper partisan behavior and content, which we measure using coverage of and commentary on partisan activities, institutions, and actors. We use this measure to describe the levels of relative partisan behavior during the period 1880 to 1900, and to describe changes over the period 1880 to 1980. We find that, on average, newspapers were initially highly partisan, but gradually became less partisan over time. Importantly, we find as much change after the 1910s as before, which contributes to the existing literature that focuses on changes in the nineteenth and early twentieth centuries. We also investigate words and phrases that had negative or positive partisan connotations in particular periods. Finally, we examine whether some of the common hypotheses offered in the literature can account for the changes. The initial findings suggest that these explanations can only account for part of the decline.

The press in the United States changed dramatically over the course of the nineteenth and twentieth centuries. One dimension in which it changed was partisanship. Initially, most newspapers were tied to a political party, and their content on political matters heavily favored that party. According to McGerr (1986, p. 14), “In 1850, 95 percent of the daily and weekly papers in America claimed loyalty to some party.” Data from Gentzkow, Shapiro, and Sinkinson (2011) show that by 1980, only 9 percent of daily newspapers in the United States identified themselves as Democratic or Republican. Conventional wisdom holds that non-partisan newspapers emerged in the United States in the nineteenth century. There is less agreement about why this happened or about how newspaper partisan behavior continued to evolve during the twentieth century.

Scholars have offered several explanations for the decline of partisan newspapers, which can be grouped into three broad classes. The first focuses on commercial forces, particularly the increasingly attractive market for advertising revenue, as well as technological changes that increased the profitability of higher circulation.¹ The second emphasizes the role of political forces, such as the amount of resources available to party organizations or the strength of partisan ties in the electorate.² The third argues that journalism became increasingly professionalized, leading publishers, editors, and reporters to increasingly value objectivity over time (Schudson 1978, 2001; Mindich 1998).

In terms of the empirical literature, most studies of the partisan press and its decline focus on the mid- to late-nineteenth century to the early twentieth century.³ For example, Hamilton (2004, p. 45) writes, “The most remarked upon change in daily newspapers in the period 1870–1900 was the emergence of the independent press.” Similarly, Baldasty (1992, p. 139) writes that in 1900, “Newspaper owners and editors were no longer primarily political activists obsessed with winning elections and filling their newspapers with political argument.” Gentzkow, Glaeser, and Goldin (2006, p. 190) write, “Sometime between 1870 and the early 1900s newspapers became demonstrably less connected to political parties” and focus their attention on the period 1870 to 1920. McGerr (1986, p. 118) argues that by the mid-1880s, “independent journalism was well established.” In contrast, Kaplan (2002, p. 16) argues that it was not until 1896 that “newspapers broke from parties and established their independence.”

In this paper, we focus on two main questions: What happened to newspaper partisanship during the twentieth century? Were the same economic and political forces continuing to affect newspaper partisanship throughout this period? To address these questions, we develop new measures of newspaper partisanship that cover the century between 1880 and 1980. Previous empirical studies have employed a variety of measures of partisan and ideological bias. However, these seem adequate only for studying short and/or specific time periods.⁴

Our key measures of newspaper partisanship are based on the coverage and commentary of partisan activities, institutions, and actors. We expect that partisan newspapers will devote especially large amounts of coverage to events involving their party and its members. We construct these measures via automated searches of online newspaper text. We use these measures to describe the levels of relative partisan behavior during the period 1880 to 1900, and to describe changes in partisan coverage over the period 1880 to 1980. In addition to our main measures, we also investigate measures that attempt to capture the “tone” rather than the intensity of coverage. These are based on “loaded” language, or news that is “good” or “bad” for a party. Furthermore, since there are well-documented differences in political and economic conditions across states and regions, we compare partisan coverage across newspapers within states and even within cities when possible (McGerr 1986; Kleppner 1987; Bensel 2000).

Our main finding is that newspaper coverage became less partisan between 1880 and 1980, and this decline was gradual and largely continuous. Consistent with much of the previous literature, we find noticeable changes through the 1910s. We also find that the decline in the partisanship of coverage continued after the 1910s, through 1980. Based on our measure, the decline between 1920 and 1980 was nearly as large as the decline between 1880 and 1920. To our knowledge, this pattern has not been previously documented in the existing literature. We also find no evidence that partisanship in newspaper coverage increased after 1980, at least through 2018.

Another finding is that our main measure is strongly correlated with explicit partisan positions taken on the editorial pages in the early years of the sample, 1880–1900. This suggests that the measures are capturing a similar dimension of newspaper partisanship. We also find high correlations between our measure, partisan positions on editorial pages, and newspapers’ self-reported partisan identifications, which have been used in previous studies. This demonstrates that self-identifications reflect significant variation in partisan content. Since self-identifications are easier to collect than content—and available for a larger set of newspapers—our findings support using this measure, at least for papers that identify with a party.⁵

In terms of heterogeneity, we find evidence suggesting that partisan differences in tone became less pronounced earlier than differences based on amounts of coverage. This is based on a limited set of terms and requires further investigation. We also find no clear differences in the partisan trends of southern vs. non-southern newspapers.⁶

While our paper is largely descriptive, the patterns we identify help assess some arguments and claims made in previous scholarship. For example, the fact that the decline was gradual and mostly continuous suggests that no single shock, such as the invention of the linotype, the appearance of radio, or the election of 1896, can account for the entire change unless its effect diffused quite gradually. Instead, if the decline in newspaper partisan coverage is driven by only one or two variables, then these variables probably also exhibit gradual trends. Taking an initial step, we examine several economic and political hypotheses regarding the transition from partisan to non-partisan newspapers.

First, many scholars argue that an increase in potential advertising revenue increases the pressure for newspapers to be more independent (Hamilton 2004; Besley and Prat 2006; Gentzkow, Glaeser, and Goldin 2006; Petrova 2011). We regress our newspaper coverage measure on population and wealth (for a short period), both of which are related to potential advertising revenues. Second, since radio and TV were relatively independent sources of news, newspapers might have responded to their entry by becoming less partisan (Besley and Prat 2006; Song 2020). We examine the share of households that owned radios and (later) televisions. We find that none of these variables account for the entire decline in partisan newspaper coverage—or even a substantively large part of it—on their own. Population and wealth only account for a small proportion of the observed change. We find no significant relationships between partisan coverage and the expansion of radio or television.

Regarding political explanations, we focus on two electoral measures. The first is vote share—compared to areas where voters lean heavily toward one party, there is pressure for newspapers to become less partisan in areas where voter partisanship is divided more equally (Hamilton 2004; Gentzkow and Shapiro 2010). The second is a measure of vote volatility—the standard deviation of the vote—which is related to the strength of voters’ partisan attachments and possibly also the strength of party organizations (McGerr 1986; Kaplan 2002). In areas where voters are more independent, newspapers have less incentive to be highly partisan. Again, we find that these variables account for only a small amount of the observed decline in partisan newspaper coverage.

Although the economic and political variables mentioned previously do not account for much of the decline in partisan newspaper coverage when analyzed one at a time, taken together they account for a noticeable share of the changes in within-newspaper behavior. The regression analyses may underestimate the impact of these variables for several reasons. First, following the literature (and due to data availability), our analyses are at the county level. This might approximate the market areas for some newspapers, but for many others, the relevant market area is probably smaller, for example, the municipality. A second concern is measurement error in the independent variables—for example, population and wealth are only proxies for newspapers’ potential advertising revenue. The dependent variable is also measured with error, which limits our ability to explain its variation. Finally, although we exploit the panel nature of the data—estimating models with newspaper-specific fixed effects—more work is needed to rule out potential biases due to omitted variables or endogeneity.

The ability to measure media biases is important because such biases may be consequential for democratic selection and accountability. In many formal models, biased reporting reduces voter information. Partisan media might affect election outcomes—for example, voters might choose lower-quality politicians if they are not given enough accurate information—or lead to increased polarization among citizens. Partisan news might also reduce electoral accountability, leading to more partisan behavior by politicians, more shirking, or increased catering to narrow, organized interest groups. Many theoretical papers have explored these arguments, and recent empirical studies have found evidence that biased news has persuasive and polarizing effects.⁷

The paper proceeds as follows. The next section describes our measures of partisan newspaper coverage and editorial stances. The subsequent two sections present our main descriptive findings for the 1880–1900 period and then the entire 1880–1980 period. Next, we focus on measures that capture differences in the tone of coverage. The penultimate section presents our preliminary analyses regarding possible explanations for variation in partisan coverage. The final section concludes.

DATA AND MEASURES

Our main measure of newspaper behavior is based on the amount of coverage of actors, institutions, or events associated with each party. Using this measure based on differential amounts of coverage, we focus on two time periods and two corresponding samples. The first covers 1880–1900, which, as we show, was a period of highly partisan behavior. The second covers 1880–1980. We also extend one analysis to 2018.

Our primary unit of observation is the newspaper-year. In both samples, we include a newspaper-year only if we can code both the partisanship of the newspaper’s editorial stance and the partisan slant of its content in that year. For the 1880–1900 sample, we use all available newspaper-years. For the 1880–1980 sample, we restrict attention to “long-lived” newspapers for which we can measure the relative partisan slant of their coverage for a sufficient number of years. We refer to this as the multi-decade sample.

We use information from the online archive Newspapers.com to code content; therefore, all of the newspapers in the sample appear in that archive.⁸ For the summary measure of partisan coverage, the 1880–1900 and 1880–1980 samples contain 1985 and 414 newspapers, respectively.

The number of newspapers in the United States increased dramatically between the Civil War and the 1910s. Similarly, the number of newspapers available in the Newspapers.com archive increases sharply over the same period. We began in 1880 as a compromise, trading off between the desire to cover a period as long as possible and to have a roughly comparable sample over time.⁹

We end our main analyses in 1980, primarily due to data limitations. The sample of available newspapers shrinks noticeably around 1980 and continues to decline almost every year afterward. The sample is further reduced because we need to define “Democratic” and “Republican” newspapers based on their historical editorial behavior or self-identification, as described next.

Classifying Newspaper Partisanship

We first classify newspaper partisanship based on its explicit editorial behavior. We call this Newspaper Party ID, or PID for short. We code each newspaper’s initial PID as Democratic or Republican, using only information from 1880 to 1920. This information comes from two sources.

The first source is the newspapers themselves. In the late nineteenth and early twentieth centuries, it was common for partisan newspapers to print the ticket of the party they supported directly under their banner/masthead.¹⁰ They did this for almost every issue in the weeks leading up to the election and stopped immediately after the election. We coded these for 2,386 newspapers found in the Newspapers.com archive.¹¹

The second source is newspapers’ partisan self-identification, as published in N.W. Ayer & Son’s American Newspaper Annual. For most daily papers, we found this in ICPSR 30261 United States Newspaper Panel, 1869–2004, and for non-dailies, we coded them ourselves.¹² This measure has been used previously by various scholars (Rutenbeck 1995; Hamilton 2004; Gentzkow, Glaeser, and Goldin 2006; Gentzkow, Shapiro, and Sinkinson 2011, 2014). Newspapers are coded as Democratic, Independent Democratic, Independent, Independent Republican, and Republican.¹³ For 3,057 newspaper-years between 1880 and 1920, we have both explicit editorial behavior and self-identification, and both are coded as either Democratic or Republican. In these cases, the two virtually always coincide. There are 1,127 cases where both are Democratic, 1,920 cases where both are Republican, and only 10 cases where the partisanship does not match.¹⁴,¹⁵

Since the two measures are so highly correlated for the period 1880 to 1920, we code PID using newspapers’ partisan self-identifications for the cases where we are unable to classify a newspaper based on its explicit editorial behavior. We do this only when the directory lists a newspaper as Democrat or Republican, not as Independent Democrat, Independent Republican, or something else. We code the partisanship of 730 newspapers based on this second source.

On rare occasions, newspapers changed their partisan affiliations from Democratic to Republican or Republican to Democratic. In these cases, we treat the newspaper as if it were two different newspapers—one before the switch and one after. Since we are only using information from 1880 to 1920 to code PID, if a newspaper changed partisanship from Democratic to Republican or vice versa after 1920, then we drop all observations for that newspaper after the change.

Sometimes newspapers consolidated, or one newspaper bought another. In the latter case, the purchased paper disappears from our sample in subsequent years, and the paper that made the purchase continues. In some consolidations, one paper was the “senior” partner and the other the “junior” partner.¹⁶ We treat these as if the senior partner bought the junior partner. When the differences were small or unclear, we treated the newly consolidated paper as a new newspaper, and both of the constituent papers disappeared from the sample years after the consolidation. Again, since we are using information from 1880 to 1920 to code PID, a new newspaper formed from a consolidation that occurred after 1920 never appears in our analyses.

Overall, we have 767 Democratic papers and 1,013 Republican papers. Note that many papers classified as Democratic or Republican in the early period switched their self-identification to Independent, Independent-Republican, or Independent-Democrat in later years. In these cases, we leave PID unchanged. We focus on the “initial” partisanship because we are investigating whether and when partisan newspapers changed the slant of their coverage, irrespective of when they started to self-identify as “Independent.”¹⁷

Measuring Partisanship in Newspaper Coverage

We measure newspaper behavior by looking at the amount of coverage of actors, institutions, or events associated with each party. In their survey on the measurement of media bias, Puglisi and Snyder (2015) distinguish between the “issue intensity” approach and approaches that attempt to capture “tone.”¹⁸ We use the term “topic intensity” rather than issue intensity, since it better captures the type of coverage that we measure. Many theoretical models of media bias assume that the primary choice newspapers make is over the amount of coverage to devote to specific events (Strömberg 2004; Besley and Prat 2006; Bernhardt, Krasa, and Polborn 2008; Chan and Suen 2009; Anderson and McLaren 2012). Gentzkow, Shapiro, and Stone (2015, p. 626) use the term “filtering” to describe topic intensity, and in their review of the literature, they write: “We are not aware of any systematic empirical evidence on the relative prevalence of these forms of bias. But it seems clear that filtering, in the form of both selection and summary, plays a large role in the way media bias occurs in practice.” Many scholars argue that this type of behavior is especially important because of its potential “agenda-setting” effects. As Cohen (1963, p. 13) famously wrote, the press “may not be successful much of the time in telling people what to think, but it is stunningly successful in telling its readers what to think about.”

One group of measures we use reflects the topic-intensity approach. More specifically, for each of the search terms shown next, we count the number of pages in each newspaper in each year in which the search term appeared one or more times—we refer to these as “hits.”¹⁹ For Republican terms, the search strings are: [Republican convention], [Republican primary], [Republican committee], and [Republican meeting OR Republican rally OR meeting of Republican].²⁰ For Democratic terms, we use the natural analogs. The phrases chosen were based on reading a large number of articles in many newspapers over many years. The phrases are common enough that we can detect meaningful differences across newspapers and years, rather than just small random fluctuations in the use of the terms. These terms also appear regularly throughout our period of study.²¹,²²

The main idea is that Democratic newspapers should devote more space to Democratic actors, institutions, or events because their readers are more interested in these topics. Republican newspapers should do the opposite. In some cases, the coverage has clear informational value. Republican readers will normally vote in Republican primaries, so it is natural for Republican papers to devote more space to Republican primaries and to provide information about the competing candidates’ names, issue positions, and background characteristics. In other cases, the coverage may be more for entertainment. Republican readers might simply want to read about the Republican Party’s activities and details about the lives of its candidates and leaders. Variation in coverage might also reflect the views of different publishers or editors. This idea is not new, of course—it is discussed in Kaplan (2002) and is applied in Gentzkow, Shapiro, and Sinkinson (2011).²³ Our contribution is to employ the idea in a more intensive and extensive manner.

We also examined coverage of pre-election forecasts and post-election wrap-ups. Partisan papers may emphasize their party’s success and the other party’s difficulties, just as candidates and parties do. The search string used for positive Republican coverage is [Republicans ahead OR Republicans lead OR Democrats behind OR Republican victory OR Republican landslide OR Republican triumph OR Republicans win OR Republicans won OR Republicans gain OR Democrat lost OR Democrats lose OR Democratic loss].²⁴ This search string captures both coverage amounts and tone, since success, even electoral success, is better than failure.

Finally, we examine a measure that even more clearly reflects (negative) tone in addition to relative amounts of partisan coverage. The string used for positive Republican coverage is [Democratic boss OR Democratic machine OR machine Democrat]. Throughout the period of study, both “boss” and “machine” had clearly negative connotations. For both the Forecasts/Wrap-ups coverage and the Boss/Machine measure, the Democratic strings are the natural analogs of the Republican strings.

For many analyses, we aggregate the six individual items—Committees, Meetings/Rallies, Conventions, Primaries, Forecasts/Wrap-ups, and Boss/Machine—into an overall measure of newspaper partisan behavior, which we call the Combined Index. This is simply an average of the partisan scores for the six underlying items.

Let R _ijt be the number of pages in newspaper i of state j in year t on which there was at least one hit on a Republican item.²⁵ Analogously, let D _ijt be the number of pages on which there was at least one hit on a Democratic item. Then define Republican Coverage Share, or RCS, as RCS _ijt = R _ijt/(R _ijt + D _ijt). We study various statistics based on RCS. Let RCS _Rt be the average of RCS taken over papers with Republican PID in year t. Analogously, let RCS _Dt be the average over papers with Democratic PID in year t. We calculate RCS for each of our six items. The Combined Index is the average RCS of the items.²⁶

The simplest measure we study is the difference between the average RCS for the two types of papers. We call this Partisan Gap:

(1) $$P{G_t} = RC{S_{Rt}} - RC{S_{Dt}}$$

The second measure focuses on the gaps within states. Let RCS _Rjt be the average RCS among Republican newspapers in state j in year t, let RCS _Djt be the average among Democratic newspapers, and let PG _jt = RCS _Rjt − RCS _Djt be the difference between the two within-state averages. The Within-State Partisan Gap is the average of PG _jt across all states in the sample:

(2) $$WSP{G_t} = \sum\nolimits_j {P{G_{jt}}/J}$$

where J is the number of states with at least one Republican newspaper and at least one Democratic newspaper, so we can compute PG _jt. We prefer this second measure because it helps separate partisan slant from coverage that may appear partisan but is actually based on relevance. Consider states in which Republicans have a large electoral advantage because of underlying partisan preferences among voters. In these states, Republican party activities are more deserving of attention and monitoring than Democratic activities because Republican candidates are more likely to win elections and hold office. Using WSPG subtracts out “bias” that is due to differences in state-specific party relevance, while the overall measure does not.²⁷

Misclassifying PID could make it appear that the newspapers are becoming less partisan, even when they are not. To take an extreme example, suppose that in year T half of all Republican papers became Democratic and half of all Democratic papers became Republican. Suppose also that, in all years, RCS = 1 for Republican papers and RCS = 0 for Democratic papers. Then PG before the switch would be 1 and PG after the switch would be 0 (since RCS _R = RCS _D = 0.5).

We address this issue in two ways. First, we attempt to find all cases where a newspaper clearly changed its partisan self-identification. We also checked whether newspapers changed their general election endorsement patterns, switching from one party to the other.²⁸ When we find cases after 1920 in which a newspaper with a Republican PID switched to Democratic/Independent Democratic or a newspaper with a Democratic PID switched to Republican/Independent Republican, we drop all observations for that newspaper after the switch.

Second, we examine the standard deviation of RCS by year. Let SD _jt be the standard deviation of RCS _ijt across all newspapers in state j in year t. We average SD _jt across all states in year t to create the average standard deviation, Within-State Standard Deviation _t, or WSSD _t. A decline in the standard deviation over time indicates that newspapers are becoming more similar to one another. Since WSSD does not use PID, any miscoding of PID classifications does not affect it.²⁹

Several details regarding the data and variables deserve mention. First, we combine odd-numbered years with the previous even-numbered year. Second, we only keep a newspaper-year for a search term if it has 20 or more hits. Online Appendix Figure A1 shows the average number of hits per newspaper by decade for each of the items. For the Combined Index we only require 10 or more hits for each item since the index is already an average of several items. Third, we only compute WSPG for years in which we can compute the gap between Democratic and Republican papers in at least 10 states.

We use a threshold of four hits for the Boss/Machine terms. These terms are highly discriminating, but they appear less often than the other items (see Online Appendix Figure A1). Interestingly, the usage of the Boss/Machine terms has remained relatively constant over our period of study. There are several possible reasons. First, according to some scholars, traditional, patronage-based party “machines” persisted at least into the 1960s (Stave 1970; Wolfinger 1972; Dorsett 1977; Trounstine 2008). Second, the terms “boss” and “machine” are used more broadly to describe powerful politicians in various circumstances, not only patronage-based organizations and leaders. Finally, journalists and others who want to criticize a party can reference prominent bosses or machines from other parts of the country or from earlier time periods, that are associated with that party.

Another potential concern is that optical character recognition (OCR) errors could affect our measures. As we discuss in Online Appendix E, however, these errors must be of a particular form in order to bias our measures in serious ways, and therefore it is unlikely that they account for much of the large changes we observe.

Finally, not all newspapers exist for the entire time period under study, and for some of those that do, Newspapers.com does not have complete coverage. We define the multi-decade sample as newspapers that, roughly speaking, have existed for at least 50 years.³⁰ More precisely, for each separate item and the Combined Index, let $$Y_{ij}^{min}$$ be the first year for which we can compute RCS _ijt for newspaper i in state j, and let $$Y_{ij}^{max}$$ be the last year for which we can compute RCS _ijt for that newspaper. Newspaper i is in the multi-decade sample if and only if $$Y_{ij}^{\max } - Y_{ij}^{\min }$$ is at least 50, and RCS _ijt is non-missing for at least half of the years between $$Y_{ij}^{min}$$ and $$Y_{ij}^{max}$$ . We allow newspapers to have gaps in their coverages because there are gaps or unreadable pages in the Newspapers.com archive.³¹

ERA OF PARTISAN PRESS, 1880–1900

As noted previously, historians and scholars of journalism describe U.S. newspapers as being highly partisan during the nineteenth century. We investigate this using our measures for the first two decades of our sample, 1880–1900. This also provides a benchmark from which we measure changes over the course of the twentieth century in the next section.

Table 1 presents the average RCS for Republican and Democratic newspapers, as well as the Partisan Gap (PG) between them. The table shows this for each of the six items as well as the Combined Index, for all available newspapers. Online Appendix Table B1 shows the analogous figures for papers in the multi-decade sample. For the replication files reproducing all tables and figures, see Hirano and Snyder (2024).

Table 1 NEWSPAPER PARTISAN CONTENT, 1880 TO 1900, ALL AVAILABLE NEWSPAPERS

Notes: Number of Observations is the number of newspaper-years used in calculating the RCS for either Democratic or Republican newspapers. Number of Papers is the number of newspapers that are used at least once.

Sources: Newspapers.com, ICPSR 30261, and N.W. Ayer & Son’s American Newspaper Annual (various years).

For all items in both samples, Republican newspapers appear to favor the Republican Party, at least when compared to Democratic newspapers. The overall PG is always positive, meaning that Republican newspapers have higher Republican Coverage Shares for these items compared to Democratic newspapers.³²,³³

Table 2 shows analogous statistics for WSPG. For almost all items, these gaps are smaller than those in Table 1. This is consistent with the argument noted previously that the overall gap might overstate partisan bias because it includes differences in the relative relevance of the two parties across states and localities. Nonetheless, for each item, WSPG exhibits the same general pattern as the corresponding PG. Online Appendix Table B2 again shows the analogous figures for papers in the multi-decade sample.

Table 2 NEWSPAPER WITHIN-STATE PARTISAN GAP, 1880 TO 1900, ALL AVAILABLE NEWSPAPERS

Notes: Number of Observations is the number of state-years used in calculating the average Within-State Partisan Gap. Number of Papers is the number of newspapers that are used at least once. Number of States is the number of states that are used at least once.

Sources: Newspapers.com, ICPSR 30261, and N.W. Ayer & Son’s American Newspaper Annual (various years).

Finally, Table 3 shows Within-City Partisan Gaps. These are analogous to WSPG, but at the city level rather than the state level.³⁴ We can calculate this gap for at least one city in 33 states. The patterns are similar to those in Table 2. Thus, evidently, we can account for most of the difference in newspaper coverage relating to differences in the relative relevance of the two parties using WSPG.

Table 3 NEWSPAPER WITHIN-CITY PARTISAN GAP, 1880 TO 1900

Notes: Number of Observations is the number of city-years used in calculating the average Within-City Partisan Gap. Number of Papers is the number of newspapers that are used at least once. Number of Cities is the number of cities that are used at least once.

Sources: Newspapers.com, ICPSR 30261, and N.W. Ayer & Son’s American Newspaper Annual (various years).

In all three tables, the partisan differences are substantively large. Consider, for example, the Combined Index and the sample of all available newspapers. The average RCS across Democratic papers is about 0.35 in Table 1. The WSPG from Table 2 is 0.26, implying that the average RCS in Republican papers would be 73 percent higher than the corresponding RCS in Democratic papers.

The patterns in Tables 1–3 are consistent with the conventional wisdom that newspapers were quite partisan during the late-nineteenth century. This increases our confidence in the measures. Also, since our measures are novel and based on more comprehensive data than previous studies, the findings provide new evidence for conventional wisdom. One implication is that many newspaper readers during this period consumed information filtered through highly partisan lenses.

TRENDS

We now study the long-term trends in newspaper partisan behavior between 1880 and 1980. In order to keep the set of newspapers roughly consistent over time, we restrict attention to the multi-decade sample—that is, newspapers for which we can measure RCS for at least 50 years and, except when using WSSD, assign a PID.³⁵

Figure 1 presents the results for the Combined Index. The upper-left panel shows PG, the upper-right panel shows WSPG, the lower-left panel shows WSSD, and the lower-right panel shows RCS for Democratic and Republican newspapers separately.

Figure 1 COMBINED INDEX, MULTI-DECADE SAMPLE

Sources: Newspapers.com, ICPSR 30261, and N.W. Ayer & Son’s American Newspaper Annual (various years).

Both upper panels display the same basic pattern of a steady and gradual decline. The WSSD panel also shows a long and slow decline. The panel in the lower-right shows that Republican and Democratic papers converged over time by roughly the same amount to a RCS of about 0.50.

The WSPG fell from an average of 0.26 in 1884–1890 to an average of 0.15 in 1914–1920, a decline of 0.11. The average continued to fall after 1920, reaching zero in 1974–1980, a decline of 0.15. Thus, based on our measure of newspaper content, partisan newspaper behavior had not disappeared by 1920. Instead, there was an even larger decline in WSPG between the late 1910s and late 1970s.

The patterns are consistent with the conventional wisdom that newspapers became increasingly independent of political parties around the turn of the twentieth century. However, many studies of the emergence and development of press independence ended their analyses in the early 1900s. For example, Hamilton (2004) focuses on the period up to 1900, and Kaplan (2002) focuses on the period until 1920. The figures show that substantial changes in partisan coverage occurred after 1900 and even after 1920.

The patterns are less clear regarding the claim in Kaplan (2002) that 1896 was a critical turning point in newspaper independence. The overall PG did not clearly begin to decline until the late 1890s, which is consistent with the argument. Also, WSPG appears to show a small but noticeable drop just after 1900 (between 1904 and 1906). Similarly, WSSD began its more apparent decline after 1900. When viewed in the context of the entire period 1880–1980, however, the changes around 1896 are not particularly large.³⁶

In the Online Appendix, we show PG, WSPG, and WSSD, as well as RCS by party, for each of the six items separately (Figures C8–C11). They exhibit roughly similar patterns—in particular, long, gradual declines in partisan behavior after 1900 or, in some cases, earlier.

We also present WSSD using all available newspapers, not only those in the multi-decade sample (Online Appendix Figure C7), and find the trend is similar to that in Figure 1. This mitigates concerns that the multi-decade sample might systematically differ from the full set of newspapers available on Newspapers.com. Finally, we examined other thresholds for including newspapers in the multi-decade sample—30, 40, 60, and 70 years. As shown in Online Appendix C.2, the patterns do not differ substantially depending on the threshold used.

In the analysis noted earlier, we restricted the sample to long-lived newspapers that existed for at least 50 years of our 100-year period of investigation. However, even in these figures, there is some entry and exit into the sample. In Online Appendix Figure C1, we examine the trends in PG, WSPG, and WSSD for overlapping 50-year windows, holding the sample of newspapers fixed within each window, and find patterns similar to those noted earlier. Thus, the decline shown in Figure 1 reflects changes in partisan behavior within newspapers, not merely changes over time in the newspapers included.

What about Post-1980?

The analyses mentioned previously stopped in 1980, primarily because of limited data availability and our decision to define PID using only information through 1920. As noted earlier, the sample of newspapers available in our data source shrunk noticeably just before 1980 and continued to decline almost every year afterward. The sample of newspaper years for which PID is non-missing is even smaller. Many of the newspapers that exist in the database post-1980 first appear after 1920; also, we “restart” newspapers as new objects after consolidations (except where one newspaper acquired the other or was the “senior” partner). This means that our preferred measure, WSPG, is based on relatively small samples for many states.

With that caveat in mind, it is nonetheless interesting to extend our figures to the present. In addition, WSSD does not rely on PID, so we have a reasonably large number of newspapers in the multi-decade sample to compute this measure through 2018. The RCS measure was non-missing for nearly 200 newspapers in the 2010s, as compared to more than 350 newspapers in the 1970s.³⁷

We again focus on the Combined Index. The left panel of Figure 2 shows WSPG and the right panel shows WSSD. Both panels show a similar pattern—that is, no significant change in either WSPG or WSSD between 1980 and 2018.³⁸

Figure 2 COMBINED INDEX, MULTI-DECADE SAMPLE, TO 2018

Sources: Newspapers.com, ICPSR 30261, and N.W. Ayer & Son’s American Newspaper Annual (various years).

The lack of change in newspaper behavior is interesting because it contrasts sharply with the increase in partisan polarization among political elites during this recent period. Polarization in roll-call voting increased in Congress and in most state legislatures (Shor and McCarty 2011; McCarty, Poole, and Rosenthal 2016). Other media might have been polarized during this period, with the growth of conservative talk radio, the introduction of cable news networks such as Fox News on the right and MSNBC on the left, and the expansion of the conservative Sinclair Broadcast Group (Ridout 2013). Moreover, two text-based measures find evidence of partisan polarization among political elites. Jensen et al. (2012) and Gentzkow, Shapiro, and Taddy (2019) examine the Congressional Record and find an increasing partisan divide in the speeches made by members of Congress since the 1980s.

Regional Variation: South versus Non-South

Historians have documented large economic, social, and cultural differences between the South and non-South during our period of study. Here we briefly analyze whether the trends in newspaper partisan behavior described earlier differ between the regions.

Focusing on the period 1880 to 1920, we calculate WSPG for six southern states and many non-southern states. The WSPG for the South and non-South are 0.25 and 0.23, respectively. The difference is not statistically significant. Thus, the partisan behavior appears to be similar in southern and non-southern newspapers for this period.

We cannot conduct a within-state analysis of the partisan gap for the South for our entire time period, because there are no southern Republican papers in the multi-decade sample (except one in Tennessee). We can, however, compare Democratic newspapers in the South and non-South. Figure 3 shows RCS for the Combined Index for southern newspapers (in black) and non-southern newspapers (in gray). For almost all years, Democratic newspapers in the South had lower values of RCS than Democratic newspapers elsewhere. However, the trends are similar in the two regions, suggesting that similar forces were at work outside and inside the South.

Figure 3 REPUBLICAN COVERAGE SHARE FOR DEMOCRATIC PAPERS IN THE SOUTH AND NON-SOUTH

Sources: Newspapers.com, ICPSR 30261, and N.W. Ayer & Son’s American Newspaper Annual (various years).

TONE-BASED PARTISAN DIFFERENCES

Does the decline in newspapers’ partisan behaviors mainly reflect a change in the amount of coverage given to each party’s activities, events, and elites, or was there also a change in the tone of the coverage? One of the items in the Combined Index, Boss/Machine, reflects a clear difference in the tone of the coverage of the political parties—“Republican machine” is not a favorable description of the Republican party. Forecasts/Wrap-ups also have a partisan tone when they reflect “cheerleading” for one party or the other. The remaining items have a more neutral tone.

Here, we examine two other types of news stories in which we can identify search strings that capture differences in tone. The first involves direct accusations of improper behavior, malfeasance, or incompetence by one of the parties. The second exploits the fact that Democrats and Republicans often use different language to frame issues and policies and, in some cases, used charged phrases that were widely adopted in the press. We do not include these in the Combined Index analyzed, either because they appear too infrequently in the newspapers compared to the items included in the Combined Index, or because they capture partisan tone only for specific years or decades.

Corruption, Ring, Scandal, and Waste

References to corruption and wastefulness in connection with one of the parties provide another possible way to measure tone, at least to the degree that the references are not merely reports about clear, specific acts of corruption or waste. In reading through numerous newspaper articles, we found that these accusations were commonly used to impugn the integrity of one of the parties, especially in the late nineteenth and early twentieth centuries. These items are similar in spirit to the Boss/Machine item used, but often involved more serious charges.

For references to Republican corruption, we used the following search strings: [Republican corruption OR corrupt Republican] and [Republican ring OR ring Republican]. For Republican waste, we used the following search string: [Republican extravagance OR extravagant Republican OR Republican waste OR wasteful Republican]. For scandals, we searched for the following string, which was more frequently used in the second half of the twentieth century: [Republican scandal OR scandal Republican]. For Democratic terms, we used natural analogs. Since the tone of these terms is negative, RCS is the number of pages with the terms referring to the Democratic Party divided by the total number of pages with these references applied to either party. Although these phrases appeared regularly, they were not used commonly enough to compute an accurate RCS measure at the newspaper-year level. Therefore, we aggregate across years by decade. We examine the period 1880–1980, and include 1980 in the decade of the 1970s.

Figure 4 shows the results. Again, we observe a steep decline in all three measures, PG, WSPG, and WSSD. Unlike Figure 1, the partisan differences appear to have bottomed out by the 1930s as opposed to 1980.

Figure 4 CORRUPTION, RING, SCANDAL, AND WASTE, MULTI-DECADE SAMPLE

Sources: Newspapers.com, ICPSR 30261, and N.W. Ayer & Son’s American Newspaper Annual (various years).

In addition to the changes in the relative frequencies shown in Figure 4, there were also changes in the overall usage of these highly charged phrases. In fact, it appears that all but one of them went almost completely out of fashion by 1980. The only phrases for which there was an increase in use over time were those involving the word scandal.

Tariff Issue

The tariff divided the Democrats and Republicans after the Civil War through the first decades of the twentieth century, with the Republican Party favoring high tariffs and the Democratic Party favoring low tariffs. We focus on the period 1888–1938, during which the tariff was a particularly partisan and salient issue.³⁹

We measure the partisan coverage of the tariff issue using two search strings that have a positive connotation for tariffs—[protective tariff OR tariff protection]—and four search strings that have a negative connotation—[high tariff OR monopoly tariff OR trust tariff OR tariff tax].⁴⁰ Because Republicans favored the protective tariff more than Democrats, RCS is the number of pages with terms that have a positive connotation divided by the total number of pages with either positive or negative terms. We include newspapers that existed in the archive for at least half of the years during this period.⁴¹

Figure 5 shows the results. The upper-left panel shows the PG, the upper-right panel shows the WSPG, the lower-left panel shows the WSSD, and the lower-right panel shows RCS for Democratic and Republican newspapers separately.

Figure 5 TARIFF, MULTI-DECADE SAMPLE Sources: Newspapers.com, ICPSR 30261, and N.W. Ayer & Son’s American Newspaper Annual (various years).

The patterns in Figure 5 are roughly similar to the patterns in Figure 1. Both of the upper panels show a steady and gradual decline in differential use of the positive and negative terms for the tariff by Democratic and Republican papers. The WSSD panel also exhibits a decline, although there are two outlier years. The lower-right panel shows that Republican and Democratic papers converged over time to a RCS of about 0.50 in 1938.

Other Specific Issues

Here we examine several other issues for which we identified simple phrases with a partisan slant. Unlike the tariff, these involved specific pieces of legislation or government actions that were newsworthy for only a short time. They are: (1) Federal Election Bill introduced in 1890; (2) Judicial Procedures Reform Bill of 1937; (3) President Harry Truman’s Healthcare Reform Plan of 1949; and (4) Bay of Pigs Invasion of 1961.

The Federal Election Bill was first introduced in 1890. Republicans supported it, and Democrats opposed it. According to proponents, the bill’s purpose was to ensure that voting rights were respected and to improve election administration, and it would have involved expanding the federal government’s role in congressional elections. Opponents referred to it as the “(Lodge) Force Bill.” We use the following search strings for the supporting and opposing sides, respectively: [election bill] and [force bill]. We search for these strings during the years 1890–1892.

The Judicial Procedures Reform Bill was proposed by Franklin D. Roosevelt and supported by Democrats in 1937. The bill contained several reforms, but the most controversial was the provision allowing the president to appoint new Supreme Court judges. Opponents, primarily Republicans (and some southern Democrats), referred to it as the “court packing plan.” We use the following search strings for the supporting and opposing sides, respectively: [court reform] and [court packing]. We search for these strings in 1937.

President Harry Truman strongly supported attempts to pass a national universal healthcare plan, especially in 1949, making it part of his Fair Deal. It was mainly supported by Democrats, while opponents referred to it as “socialized medicine.” We use the following search strings for the supporting side: [national health insurance OR national health program OR national health plan OR trumans health program OR trumans health plan OR medical insurance program OR medical insurance plan]. For the opposing side, we use: [socialized medicine]. We search for these strings in 1949 and 1950.

Finally, the Bay of Pigs Invasion was a failed military operation that attempted to stop the Cuban Revolution in 1961. Critics of the operation referred to it as a “fiasco,” “debacle,” or “disaster.” It was viewed as a major failure of John F. Kennedy’s administration, and Republicans were especially vocal in their criticism. We use the following search strings for the critics of the administration: [bay of pigs fiasco OR bay of pigs debacle OR bay of pigs disaster]. The search strings more favorable toward the Democratic administration are: [bay of pigs operation OR bay of pigs invasion]. We search for these strings during the years 1962–1964.

The results in Table 4 show that the partisan gap for the Election Bill was similar in magnitude to the gap of the Combined Index reported in Table 1 and the gap for the tariff issue for the period 1880–1900. The gap for the Court Reform Plan is lower but still fairly large, and in fact somewhat larger than the gap of the Combined Index in the 1930s, as shown in Figure 1. The gaps for the Health Insurance and Bay of Pigs items are both tiny.⁴²

Table 4 NEWSPAPER PARTISAN CONTENT ON SELECTED ISSUES, ALL AVAILABLE NEWSPAPERS

Notes: Number of Observations is the number of newspaper-years used in calculating the RCS for either Democratic or Republican newspapers.

Sources: Newspapers.com, ICPSR 30261, and N.W. Ayer & Son’s American Newspaper Annual (various years).

Taken together, the patterns documented in this section, together with the Online Appendix for Boss/Machine, suggest that stark partisan differences in tone fell out of usage even earlier than partisan differences in topic intensity. For each of the measures involving tone—Corruption/Ring/Scandal/Waste, the Tariff, the four specific issues, and Boss/Machine—the differences between Democratic and Republican newspapers were all small by 1950 or earlier. By contrast, recall that partisan differences in the Combined Index, which is dominated by measures that capture topic intensity, only became small and stable by about 1980.

POSSIBLE EXPLANATIONS

Why did newspaper partisan behavior decline during the twentieth century? While answering this question is beyond the scope of this paper, in this section we present some empirical findings that shed light on a few of the common explanations in the literature. While not conclusive, they provide some guidance for future research in this area. All analyses in this section focus on RCS of the Combined Index.

One argument is that as advertising markets became more lucrative, newspapers had an incentive to reduce their partisan slant to appeal to a broader cross-section of readers. It appears impossible to construct a comprehensive panel of information on newspaper advertising—including rates, line inches, and revenues—for the 100-year period we are studying. However, two of the main determinants of advertising market size are population and income. The number of consumers in a given market area is highly correlated with the population of that area, and advertising prices appear to be related to population as well—for example, Hamilton (2004) finds that in 1880, city population was positively correlated with advertising rates. Households with higher incomes can afford to buy more consumer goods. Since there was no consistently and regularly calculated income measure at the county level before 1950, we use the population in the county where a newspaper circulates as a rough proxy for the size of the paper’s potential advertising market.

In our multi-decade sample, there are modest but statistically significant correlations between newspaper RCS and the population of the counties where newspapers are based. Measuring population in logs, for Republican newspapers, the correlation is –0.30, and for Democratic papers, the correlation is 0.35.⁴³ Thus, in both cases, the correlations indicate that newspapers based in more populous counties are less partisan.

However, accounting for county population does not significantly affect the overall pattern of convergence in RCS of Democratic and Republican newspapers. To see this, consider the following analysis. We run two regressions of RCS on year-trends and newspaper fixed-effects, one that controls for the log of county population and one that does not. In the regressions, we use third-order polynomials for both the log of population and year-trend variables.⁴⁴ We then calculate the expected values of RCS, as predicted by the year-trend variables alone, for each of the two regressions. We do this separately for Democratic and Republican newspapers (as mentioned earlier, partisanship is defined using PID). The top left panel in Figure 6 shows the predicted point estimates. The two curves toward the top are for Republican newspapers, and the two lower curves are for Democratic newspapers. In both cases, the dashed curves are for the regressions with controls, and the solid curves are for those without controls. If changes in population accounted for a large amount of the changes in RCS, then there would be little variation left for the year-trend variables to capture, and the dashed curves in the figures would be flat lines near 0.5. As the top left panel of Figure 6 shows, however, the two curves for Republican newspapers are quite similar to one another, as are the curves for Democratic newspapers.

Figure 6 PREDICTED RCS OF COMBINED INDEX, WITH AND WITHOUT CONTROLS FOR LOG OF POPULATION, REPUBLICAN VOTE SHARE, STANDARD DEVIATION OF REPUBLICAN VOTE SHARE, OR LOG OF TOTAL HITS

Sources: Newspapers.com, ICPSR 30261, N.W. Ayer & Son’s American Newspaper Annual (various years), U.S. Census of Population (various years), ICPSR 1, ICPSR 13, and Hirano and Snyder (2019).

Measures of wealth at the county level exist for the period 1880 to 1912—the period of high partisan behavior—so we can compare population and wealth as measures of advertising market potential for this period. As shown in Online Appendix Figure G16, using wealth rather than population does not significantly affect the key estimates.

Another potential explanation is that newspapers exhibited less partisan behavior over time in response to the changing political preferences of their readers. Two of the most important characteristics of readers are the distribution and intensity of partisan attachments. In an area where a large fraction of the electorate favors one party, a newspaper has incentives to cater to the tastes of the advantaged party’s voters. By contrast, in an area where the electorate is divided evenly between the parties, a newspaper risks losing a large segment of its potential audience if it favors one party in its coverage. Similarly, in areas where voters do not have strong partisan attachments, a newspaper has less incentive to be highly partisan since readers may be interested in the coverage of both parties.

We begin with the distribution of partisan attachments. Let RVS _it be the average Republican party vote share in county i in year t.⁴⁵ Throughout the period of study, but especially in the early decades, there is a large and positive correlation between both PID and the newspaper RCS, and the partisanship of the county where the newspaper circulates. For example, focusing on the period 1880 and 1910, the correlation between RCS and RVS is 0.53, and the correlation between PID and RVS is 0.51. These correlations are consistent with the hypothesis that newspaper partisan behavior responds to reader demand, which previous researchers have also found (Gentzkow and Shapiro 2010).⁴⁶

However, systematic changes in Republican vote share do not account for a large amount of the convergence in RCS. We conduct an analysis analogous to that in the previous discussion, but with RVS in place of population. The upper right-hand panel of Figure 6 shows the resulting predicted values. The panel shows that the curves with and without the RVS control variables are similar.

This is perhaps not surprising, given the patterns we observe in RVS over time. In particular, the counties where Republican and Democratic newspapers circulated did not become steadily more similar in their partisan orientations. Consider the difference between the average RVS of counties where Republican newspapers circulated and the average RVS of counties where Democratic papers circulated. For the newspapers in our multi-decade sample, this difference did not decline monotonically over time. Instead, the difference in RVS increases from the 1880s to the 1910s, is flat from the 1910s to the 1950s, and falls from the 1950s through the 1970s.

We now turn to the intensity of partisan attachments. Measuring intensity is difficult, but one type of variable used in the literature is variability in the two-party vote share either across offices and/or over time.⁴⁷ Let SDVS _ct be the standard deviation of the Republican Party vote share in county c and year t.⁴⁸ Similar to the patterns for RVS, systematic changes in SDVS do not account for a large amount of the convergence in RCS. We conduct a regression analysis analogous to the previous two analyses and show the resulting predicted values in the lower left-hand panel of Figure 6. The panel shows that the curves with and without the SDVS control variables are almost identical to one another for both Republican and Democratic newspapers.

Overall, these findings suggest that the decline in newspaper partisan coverage was not driven exclusively by the broader changes in partisanship and voting that occurred during the period of our study. For local inter-party competition, this is not too surprising because this variable does not exhibit the same nationwide downward trend that we document for newspaper partisan coverage. Instead, when averaging across counties, this variable oscillates with no clear trend between 1900 and 1950 and only exhibits a steady downward trend from 1950 to 1980 (Hirano and Snyder 2019).⁴⁹

Between 1880 and 1980, the amount of coverage of the items in our Combined Index generally increased (see Online Appendix Figure A1). There are several reasons why higher levels of coverage on an issue might be associated with a less partisan slant in the coverage. Some events are so important that all news outlets should consider them newsworthy. Examples might include the parties’ national conventions and, within a state, the parties’ state conventions, as well as contested primaries for U.S. president, U.S. senator, or governor. To the degree that these events occur equally often for both parties, higher coverage will be correlated with values of RCS closer to 0.5. Journalistic norms might also have changed over time, leading newspapers to cover both parties more equally. If papers do this mainly by expanding coverage of “the other side” to avoid reducing coverage of the party they support, then higher amounts of coverage will again be correlated with values of RCS closer to 0.5.

To measure the total amount of space newspapers devote to each item, we find the number of pages with at least one Democratic hit and the number of pages with at least one Republican hit for that item, and sum them: T _ijt = R _ijt + D _ijt. We then calculate $${\bar T_{ijt}}$$ by averaging T _ijt across all six items in the Combined Index. Because $$\bar T$$ is right-skewed, we use log $$\left( {\bar T} \right)$$ in our analyses. The correlations between log $$\left( {\bar T} \right)$$ and RCS indicate that newspaper-years with higher $$\bar T$$ are a bit less partisan. For Republican newspapers, the correlation is −0.37, and for Democratic papers, the correlation is 0.34.

Accounting for this factor does not significantly affect the overall pattern of convergence in RCS of Democratic and Republican newspapers. The bottom left-hand panel of Figure 6 shows that the curves with and without log $$\left( {\bar T} \right)$$ as control variables are similar to one another, both for Republican and Democratic newspapers.

Figure 7 shows the predicted values from regressions that include all four variables together. Collectively, they appear to account for a non-trivial amount of the convergence between Republican and Democratic newspapers. Most of the change, however, remains unexplained.

Figure 7 PREDICTED RCS OF COMBINED INDEX, WITH AND WITHOUT CONTROLS FOR LOG OF POPULATION, REPUBLICAN VOTE SHARE, STANDARD DEVIATION OF REPUBLICAN VOTE SHARE, AND LOG OF TOTAL HITS

Sources: Newspapers.com, ICPSR 30261, N.W. Ayer & Son’s American Newspaper Annual (various years), U.S. Census of Population (various years), ICPSR 1, ICPSR 13, and Hirano and Snyder (2019).

Finally, we conducted two analyses to assess the potential effects of the introduction and growth of new competing media, one for radio and one for television. Models such as Besley and Prat (2006) suggest that the introduction of less partisan alternative sources of information could reduce the incentives for partisan behavior by newspapers. Song (2020) finds evidence suggesting that television had this effect.⁵⁰ Using data on radio ownership for 1930 and 1940, we calculate the growth in the fraction of households with radios in each county. The median grew sharply from 0.25 to 0.75. There is no significant relationship, however, between changes in radio ownership rates by county and the RCS of the Republican newspapers that circulate in the county. The same holds for Democratic newspapers. Similarly, using television ownership for 1950 and 1960, we calculate the growth in the fraction of households with televisions in each county—in this case, the median grew from 0.01 to 0.82. Again, we find no significant relationships between changes in television ownership rates by county and RCS of the newspapers that circulate in the county. See Online Appendix Table G5 for information about the data and the estimates.

Summarizing, even after controlling for the variables we have explored, a substantial amount of the decline in newspaper partisan behavior remains unexplained. This does not imply that the theoretical arguments underpinning these variables are invalid. First, some of the variables capture a non-trivial amount of the cross-sectional variation. Second, the variables we use are rough proxies for the underlying theoretical concepts, and therefore the estimates may understate the actual relationships due to measurement error. Better measures might account for a larger share of the variation over time.

Other Potential Explanations

Of course, we have not examined all of the factors that might be driving changes in newspaper partisanship. One class of explanation is market structure and competition. At least two factors increased the incentives for newspapers to expand in size and for newspaper markets to become more concentrated. One is technological change that increased the minimum efficient scale of newspapers. Another is declining transportation costs, which allowed newspapers to expand their geographic coverage more easily.

Another potentially important factor is the growth of newspaper chains. We considered the 11 largest and most well-known chains, as cataloged by the University of Illinois History, Philosophy, and Newspaper Library.⁵¹ Only a few newspapers in our sample belonged to a chain in some years but not in others, so the growth of chains cannot account for much of the observed change. Moreover, in a regression analysis including newspaper and year fixed effects, the point estimates on a chain membership indicator variable are small and not statistically significant.

Political factors, in addition to those studied earlier, might also have played a role. Two possibilities are intra-party divisions and institutional reforms. Important political movements, including populism, progressivism, and the New Deal, led to open intra-party conflicts. Newspapers might have responded by becoming less partisan to avoid alienating potential readers. Institutional reforms include changes in electoral laws such as direct primary and non-partisan elections, as well as other reforms that affected traditional party organizations, such as the adoption of civil service laws. Weaker party organizations had fewer resources and less ability to subsidize newspapers.

Another potential factor is a change in what constitutes newsworthy events. One example is the increasing role of the national government over the course of the twentieth century—newspapers covering many of the same national events might exhibit smaller partisan differences in their coverage. A related factor is the reliance on wire service content. If newspapers printed an increasing number of stories using content from wire services, they might have become more similar to one another. The amount of wire service content and whether to employ it in a partisan manner were still at the discretion of newspaper editors (see, e.g., Shaw 1967).⁵²

Finally, the norms of journalistic professionalism also changed over time, and “objectivity” became increasingly valued (Schudson 1978; Mindich 1998). One way for a newspaper to assert its objectivity is to be less partisan in its coverage. As noted previously, we find some evidence that tone-based partisan differences declined earlier than differences based on topic intensity. One possible explanation is that the development of journalistic professionalism may have discouraged obviously disparaging language.

CONCLUSION

In this paper, we develop a new measure of newspaper partisan behavior. We document: (1) a high degree of partisanship in the U.S. press between 1880 and 1900, consistent with conventional wisdom; and (2) a long, steady decline in partisan behavior in terms of both topic intensity and tone, beginning in 1900 or perhaps earlier and continuing all the way until 1980. We find no evidence of an increase in partisan polarization in coverage after 1980. There is some evidence that the partisan differences for the tone-based measures may have declined earlier than the topic intensity measures. We also provide some preliminary analyses of potential explanatory variables to account for socio-economic changes, partisan changes, and the introduction of radio and television, but we find that none of these variables individually accounts for a large fraction of the total change.

Clearly, there are many directions for future work. First, we would like to further investigate possible differences between the tone-based and topic intensity measures. In particular, we would like to include additional terms that capture tone over long periods of time, similar to the Boss/Machine and Forecasts/Wrap-ups items, or terms that capture tone for issues specific to particular years. This would increase our confidence that the patterns we identified earlier hold more generally.

Second, we have mostly focused on long-term trends rather than cross-sectional variation. However, there is a large amount of variation in RCS, even among papers with a Republican PID and among those that are Democratic. For example, during the period 1880–1900, non-daily newspapers were more partisan than dailies in their behavior. This is true even after controlling for population in the county where each newspaper circulates, as well as total coverage.⁵³ The RCS of Republican non-dailies is 0.031 higher than that of Republican dailies, while the RCS of Democratic non-dailies is 0.021 lower than that of Democratic dailies. These differences are potentially important because during this period there were many more non-dailies than dailies, and non-dailies accounted for more than 60 percent of total U.S. newspaper circulation.

Third, we have focused on newspapers with Democratic or Republican PIDs. Although less common around the turn of the twentieth century, there were also independent, non-partisan, and third-party newspapers. Interestingly, during the period 1880–1900, the average Combined Index RCS for these other types of papers is 0.54, which is close to the midpoint between the RCS for Democratic and Republican papers. The interquartile range is 0.47 to 0.60, which lies between the average RCS for Democratic and Republican newspapers of 0.35 and 0.66, respectively. Competition from these papers may have induced Democratic and Republican papers to moderate their behavior. This clearly deserves further investigation.

Fourth, one potential driver of a newspaper’s partisan behavior is the advertising potential of the newspaper’s market area. Following Hamilton (2004), we used county population as a rough proxy of the size of each newspaper’s advertising market. For a shorter time period, we also incorporated wealth. However, more must be done on this—for example, using income and other variables associated with consumer demand and measuring more accurately the geography of each paper’s (potential) market area. Since population and income tend to be trending variables, it is possible that advertising potential could account for a significant portion of the trending in our measures.

Finally, although the measures here appear to capture meaningful variation across newspapers and over time, they are limited. Online newspaper archives continue to grow, and natural language processing continues to improve. These advances should allow researchers to develop even more refined measures of partisan behavior, as well as other types of behavior, in the coming years.