All the Rioting That's Fit to Print:Selection EflFects in National NewspaperCoverage of Civil Disorders, 1968-1969

Daniel J. MyersUniversity of Notre Dame

Beth Schaefer CanigliaOklahoma Slate University

This study examined selection effects In newspaper reports about civil disorders in thelate 1960s, A comprehensive set of events recorded in newspapers across the UnitedStates was compared with the subsets of these events recorded in two nationalnewspapers often used to construct collective event data bases—the New York Times andthe IVashington Post. The results demonstrate that fewer than half of all disorders arecovered in these two newspapers combined, and thai those reported are selected on thebasis of event intensity, distance, event density, city population size, type of actor, andday ofthe week. To demonstrate the effects of these selection patterns on substantiveanalysis of civil disorder, the authors replicated earlier studies using all reported events,and then repeated the analyses using only the events reported in the Times and the Post.This procedure .showed some substantial differences in results. The implications of these

findings for event analyses and for substantive understandings of media selection arediscussed. ,

It is a truism that the mass media, both printand broadcast sources, have a powerful impact

on how we understanding the social world andhow we live our lives. The role ofthe media insocialization, for example, is well documented(Bandura and Walters 1963; Gitlin 1979), andtbe billions spent annually on advertising standas testament to the extent of media influence.News sources, in particular, do much to help usknow the world around us. In fact, most of ourknowledge about political and social conditions

Direct correspondence to Daniel J. Myers,Department of Sociology, Notre Dame, IN 46556(Myers.33@nd.edu). This research was supportedby grants SES 01-11217 and SBR 96-01409 from theNational Science Foundation, and by the Institutefor Scholarship in the Liberal Arts and the Francis M.Kobayashi Research Travel Program at the Universityof Notre Dame. The authors thank Michael Davem,Rhonda Stewart, Kimberly Berg. Keely Jones,Victoria Myers, Gregg Carter, Robin Gratz, FemeBaldwin, Charles Lamb, and Peter Lysy for assistancein accessing and collecting data, and Pam Oliver.Alexandra Coming, and Jerry Jacobs for commentson earlier drafts.

in our country and abroad comes either direct-ly or indirectly from the mass media. Of course,the media cannot inform us about everythinghappening in tbe world, and as reporters andnews editors select content, they emphasize cer-tain types of events while neglecting others.Tbese selection practices shape our collectivesense of reality—our perception of wbat is goingon in the social world.

Tbe influence of media does not, bowever,stop with the influence on the typical audiencemember. In recent decades, media records bavebecome core data in scholarly studies investi-gating tbe causes and outcomes of politicalevents. Likewise, government agencies andcommissions (e.g., the Kemer Commission intbe 1960s) have relied beavily on newspapers forenumeration and cbaracterization of events.Such studies further reify tbe version of reali-ty presented by the media and add an officialstamp of veracity reflecting tbe expertise ofthescbolars and authors.

In botb tbe public world and the scholarlyworld, some media outlets are considerablymore influential in this real ity-creating process.One of the most influential is the New YorkTimes, whicb has long held an elite position

AMERICAN SOCIOLOGICAL REVIHW. 2004, VOL. 69 (August 519-543)

52O AMERICAN SOCIOLOGICAL REVIiEW

among U.S. and world newspapers. An eventreported in the Times is not just news, after all;it is important news. The appearance of an eventin the Times certifies it as newsworthy ("fit toprint") and imbues it with cultural and politicalsignificance. This occurs in part because polit-ical and economic elites pay particular attentionto the Times. Thus, events that appear in theTimes play a larger role in the perceptions anddecision making of American leaders. But theeffects of the Times on understandings of polit-ical situations and social life do not stop withelites who read the paper firsthand. Secondaryspin-off effects give the newspaper even morepower over our collective reality. One such echooccurs when editors and reporters from othernews outlets (newspapers, magazine, radio, andtelevision) read the Times and draw materialfor constructing their own stories, sidebars,local follow-ups, and op-ed pieces. Events thatare not reported in the Times do not have thispotential pass-along effect, and therefore havetremendously diminished chances of becom-ing salient influences on public opinion andpolicy decisions.

Given the power of a news outlet such as theTimes, it is fitting to ask why some events arecovered whereas others are neglected. Whenwe as scholars and members of the publicdepend on single sources or a few key sourcesfor information, what are we missing? Considertwo events.

• In April of 1968, students at ColumbiaUniversity took over five campus buildingsto protest university construction on landcommonly used for recreation by residentsof Harlem. After several days, a physicalconfrontation developed when policeattempted to eject students from the build-ings. In the end, more than 1,000 studentswere arrested, more than 200 were injuredduring the clash, and property damage tothe extent of several thousand dollarsoccurred.

• Earlier, that same month in Cincinnati, ablack man defending his property againstsuspected looters accidentally shot a femaleacquaintance. A crowd gathered around thepolice officers sent to investigate, and rumorsspread that a white officer had shot a blackwoman. In the next few hours, violencespread quickly, and more than 50 firebomb-ings were reported. A National Guard unitwas quickly activated and a curfew imposed,but before the incident ended, 2 people had

been killed, 20 more had been seriouslyinjured, nearly 300 had been arrested andproperty damage exceeding $250,000 hadresulted.

By any conventional definition, these twocivil disorders were, in the words of the NewYork Times, "news that's fit to print." The char-acteristics of both events made them notewor-thy and likely to attract the interest of a widereadership. They were not, however, both cov-ered in the national newspaper of record, theNew York Times. The Times record of Americancivil disorders in 1968, then, is incomplete. Itis not surprising, of course, that the Times didnot cover every single civil violence event in theUnited States. The paper did not have theresources to find all of them, the space to printthem, or the audience interest to warrant suchcomprehensive coverage. But why was theColumbia University event reported by theTimes, but not the Cincinnati event? Was itbecause the Columbia event was local, mean-ing that our understanding of the riot wave isexcessively influenced by New York events? Orwas it because the actors at Columbia havehigher status, suggesting that the media rein-force political stratification? Furthermore, whatdo these kinds of selection processes mean forour substantive and social scientific under-standings of civil disorder and protest?

The current study offers insight into thesequestions by examining newspaper reports aboutcivil disorders in the late 1960s. In the first partof this analysis, we examine the question ofwhat is covered and what is missed by a promi-nent newspaper such as the New York Times,comparing the set of civil disorders reported inthe Times and the Washington Post with thosereported by all local newspapers in the UnitedStates. This comparison helps us understandwhy some events become important news tothe public and scholars by showing a number offactors that influence coverage, such as eventintensity, location, and the political significanceof the actors. In the second part of this study, weturn more directly to the scholarly use of news-paper data and consider how reliance on a sin-gle source such as the Times may affectsubstantive analyses of civil disorders. To do so,we examine models that predict disorder, first,using all events from all newspapers, and thencomparing the results with models using datafrom only the Times and the Post. We conclude

AUTHE RIOTING THATS FIT TO PRINT 521

by examining the methodological issues sur-rounding the use of such data before returningto the authors' larger concems about contentselection and presenting a broader model forunderstanding media selection that emphasizesthe role ofthe audience in the selection process.

BACKGROUND , | .

Questions about what is covered and what ismissed by newspapers are longstanding. In polit-ical science and sociology, tbese concernsemerged in full force mainly as a methodolog-ical question about the time scholars startedusing newspapers as data sources for politicalevents (see Mueller 1997b for detail).' As tbesekinds of studies have accumulated, early con-cems about the accuracy of newspaper data(e.g., Danzger 1975; Hazelwood and West 1974;Jackman and Boyd 1979; Snyder and Kelly1977; Taylor and Hudson 1972) have reemerged,and scholars have been taking a keen interest intbe biases tbat these data may be introducing(e.g., Barranco and Wisler 1999; Franzosi 1987;Hocke 1998; Koopmans 1998; McCarthy,McPhail, and Smith 1996; Mueller 1997a;1997b; Oliver and Maney 2000; Oliver andMyers 1999). Tbese recent studies make it plainthat tbe media are far from transparent con-duits of information about political events. Themedia use various filtering mecbanisms that

' Although those who study protest and collectiveviolence have made the most use of newspaper data(e.g.. Gurr 1968; Jenkins and Eckert 1986; Jenkinsand Perrow 1977; Kriesi etal. 1995; Lieberson andSilvermanI965;McAdam 1982;Olzak 1992; Shorterand Tilly 1974; Spilerman 1970; 1976; Minkoff1997), the practice is much more widespread.Analysts use newspaper data to study assassinations,intra- and interstate wars, coups and revolutions,state repression (Paige 1975; Rummel 1966; Taylorand Jodice 1983), elections (Banks 1997), manage-ment control of corporations {Burch 1972), long-term changes in public opinion (Galambos 1975),strikes (Shorter and Til ly 1974), publ ic signals aboutmonetary policy (Woolley 2000), and even earth-quakes (Taylor and Jodice 1983). Furthermore, news-paper records are the key elements in well-establishedand widely used data series such as Polity ///(Jaggarsand Gurr 1995) and the World Handbook of Politicaland Social Indicators (Russet 1964, Taylor andHudson 1972; Taylor and Jodice 1983).

produce substantial representation of someevents while neglecting others. A number ofstudies, for example, have demonstrated thesensible finding that large protest events aremore likely to be covered tban smaller ones(Barranco and Wisler 1999; McCartby, McPhail,and Smith 1996; Oliver and Maney 2000; Oliverand Myers 1999).

Besides tbe consistent effect of event size,otber factors also affect coverage. For example,McCartby. McPbail. and Smith (1996) showedtbat media attention cycles affect coverage. Associal issues move in and out of vogue, tbe cov-erage of related social movements ebbs andwanes. Barranco and Wisler (1999) found thatviolence connected with an event increases cov-erage. McCartby et al. (1998), Oliver and Myers(1999), and Mueller (!997a) found tbat thenewspaper's political position affected whichevents it reported, and Oliver and Maney (2000)and Oliver and Myers (1999) documented sea-sonal pattems in protest coverage and linkedselection pattems to reporting routines.

Despite sucb findings, analysts have remainedbopeful about newspaper data because, altboughimperfect, tbey often are tbe only reasonablycomplete and continuous sources available(Koopmans 1998). With the documentation ofsystematic coverage pattems, analysts have con-cluded that, if treated properly, event data arereasonably sound. For example, some baveargued that event intensity and size are the onlyimportant determinants of bias, and that if tbeseare controlled, analyses and the conclusionsdrawn from the data remain largely valid(Shanaban and Olzak 1999; Snyder and Kelly1977). Otbers have suggested tbat newspaperdata are useful as long as tbe bias that exists isconsistent over time. If so, according to theseanalysts- tbe peaks and valleys in tbe mediarecord reflect real cycles of activity, and tbus,many analyses are valid (Franzosi 1987;Koopmans 1998; Olzak 1992).2

DMITING NEWSPAPER SOURCES: THE NEW

YORK TIMES

Tbe recent surge of attention to newspaper eventrecords bas focused on comparing police records

^ Even consistent bias may produce problems inmultivariate analysis (Hug and Wisler 1998).

522 AMERICAN SOCIOLOGICAL REVIEW

of protest (including permit applications andincident reports) with what has been reported innewspapers (Barranco and Wisler 1999;McCarthy, McPhail, and Smith 1996; Oliverand Maney 2000; Oliver and Myers 1999).Through such comparisons, these studies havedetermined what characteristics of events causethem to appear in the media record. However,besides the general media selection processexamined in these studies, there is another pieceof the selection process that has received lessattention.

Because analysts of political behavior typi-cally do not have the resources necessary tocollect data from multiple newspapers (Jackmanand Boyd 1979; Koopmans 1998), most news-paper-based event data are tabulated from onlya small number of newspapers—often only one.Thus, selection bias in these data developsthrough two steps. First, within the universe ofevents relevant to a particular study, there is asubset reported by some newspaper. This is theselection process that has been the focus ofrecent studies using police records. Second,within the subset appearing in the media, a yetsmaller subset is selected for the constructionof social science data. Thus, in addition to what-ever distortions are introduced by media selec-tion in general, additional selection effects areadded by a focus on one or a few elite media out-lets. This latter selection process is the focus ofthe current study.

In the collection of event data from news-papers, the New York Times is the undisputeddefault source. Just as the Times can dominatepublic news consumption, so it dominatesevent-based social science data. Whether theevents are riots, coups, assassinations, wars,or nonviolent protests, the Times has been thesingular, or else the core, element in manyevent data sets and the hundreds of studiesdrawn from them. As such, the Times has beenalternately attacked and defended as a validand reliable data source. It is clear that theTimes is the best single newspaper source forpolitical event data, because a number of stud-ies have shown that in comparison with someother newspaper, the Times almost alwaysreports more events and provides better detailabout them (Jackman and Boyd 1979; Olzak1982; Olzak 1992; Olzak and Olivier 1994;Paige 1975; Taylor and Jodice 1983).

Unfortunately, the Times is not free from selec-tion effects just because it is the best source.

To assess selection effects in the Times, someanalysts have compared data from it with datafrom a small number of other newspapers (usu-ally one to four) (e.g., Jackman and Boyd 1979;Jenkins and Perrow 1977). The results of thesestudies generally show that the addition of oneor two other sources does not change the over-all picture or the substantive conclusions drawnfrom the Times alone. Moreover, because sub-stantial costs are incurred from the addition ofeven one other source, these studies recom-mend against supplementing the Times. Yet, asreal as these cost-benefit concems are, they donot provide adequate reassurance about selec-tion effects in the Times. The data may not bechanged much by the addition of one supple-mentary newspaper, but this does not mean thatthe Times data are a transparent conduit of infor-mation in the first place. Furthermore, althoughthe gains associated with adding one or twonewspapers may be negligible, consulting alarger number might make a substantialdifference.

In the current study, we took a different strat-egy, comparing the Times record with eventdata created by hundreds of local newspapersources.-* This strategy also has its limits becauseall sources of data are media sources. We can-not detect the selection of events into any mediafrom the universe of events. Thus, we limited ourfocus to the differences between the data carriedby all newspapers and the data carried by theTimes, emphasizing that these selection effectsoccur in addition to those found in recent stud-ies. The first central research question in thisstudy then was this: If an event exists in the localmedia record, what makes it more likely that thisevent will be covered by a national source suchas the New York Times (and thus also appear indata sets derived from it)?

3 Unfortunately, the police data strategy used inrecent studies is not much help because it is work-able only on a local level. For a national or interna-tional study, it would be impossible to locatecomparable police records for the hundreds of loca-tions Involved.

ASSESSING SELECTION

RECORDS OF NEWSPAPER COVERAGE

To construct a nationally comprehensive set ofevents fi-om local media sources, we used thecivil disorder data collected by the LembergCenter for the Study of Violence, which oper-ated from 1965 to 1974 at Brandeis University.The Lemberg Center was established for thepurpose of studying social violence, broadlydefined. But as racial rioting continued tobreak out during the 1960s, the Center realizedthat it was uniquely positioned to record andanalyze race-related civil disorder, and by 1967had focused its activities completely on riot-ing and urban race relations. During its earlyyears, the Lemberg Center was unquestion-ably the nation's premier node for research onracial civil disorder, producing important datathat have been used in a large number of impor-tant social science studies.

One of the most important data-gatheringundertakings of the Lemberg Center was thecompilation of a comprehensive tabulation ofrace-related civil disorders. The Center col-lected these data from 1967 to 1972 usingnewspapers as the primary data source. To castas wide a net as possible, the Lemberg Centerentered into a contract with the Luce PressClippings service, which has a distinguishedhistory of providing clipping services for avariety of research purposes. Throughout tbeperiod in question, trained Luce employeesread every daily newspaper published in theUnited States and extracted clippings accord-ing to guidelines constructed by the LembergCenter. The criteria were broadly defined sothat Luce employees were unlikely to excludepotentially relevant events. Once the clippingswere forwarded to the Lemberg Center, itsstaff reviewed each clipping and determinedwhether the event described met the Center'sestablished criteria for a race-related civil dis-order. A civil disorder was defined as an eventinvolving crowd behavior that resulted in eitherproperty damage or personal injury, defianceof civil authority, or "aggressive disruptionswhich violate[ed] civil law." The latter descrip-tion was primarily intended to capture build-ing seizures. To be considered race-related,events had to include "aggressive or violentbehavior by members of one racial or ethnicgroup against members of another or their

ALL THE RIOTING THATS FIT TO PRINT 523

symbolic equivalents" (see Baskin et al. 1971for further detail). At least four participantswere required for an event to be included in theLemberg Center tabulations.

After the clippings had been collected by theCenter, the Lemberg staff compiled all thearticles related to each event. Using this set ofarticles, the staff wrote a brief summary ofeach event, extracted and recorded key infor-mation (e.g., our intensity indicators), and dis-tributed this information so it could be used inanalyses by other scholars.•* For 2 years, 1968and 1969, tbe Lemberg Center published thesedata—as the Riot Data Review for 1968 and asthe retitled U.S. Race-Related Civil Disordersfor 1969. We were able to uncover additionaldetails about many events in the LembergCenter's archival material. Each event in theLemberg data was reported in at least one localnewspaper, so data for these 2 years are unpar-alleled in terms of completeness and geo-graphic scope. Altogether, the center identified1,114 race-related civil disorder events from1968 to 1969.-̂

To construct the subset of the data to betested in the current study for selection effects,we identified which of the events were report-ed in the New York Times (and secondarily inthe Washington Post). In addition to the clip-ping service, several members of the Lembergstaff monitored the Times and the Post dailyand recorded whether the event was covered inthe two newspapers or not. To ensure accura-cy, we cross-checked the Center's records byexamining the actual news clippings in theCenter's records, conducting a full-text read ofthe New York Times for 1968, and usingLexisNexis to identify candidate articles in

^ The Lemberg newspaper data have been used inwhole or in part in a number of important studies ofracial disorder including those of Baskin et al. (1971),Carter(l983, 1986, 1990), Jiobu (1971), Kelly andIsaac (1984), Kelly and Snyder (1980), Lieske (1978).McAdam {1982), Myers (1997), Myers (2000). andMyers and Buoye (2001), Olzak and Shanahan(1996), Spilerman( 1970, 1971. 1976).

^ Despite the relatively comprehensive nature ofthese data, they are not, of course, bias-free. Becauseour data are derived from local newspapers, all thefactors that have been identified by analysts of localnewspapers and loeal events bear on our data.

S24 AMERICAN SOCIOLOGICAL REVIEW

the Times for 1969, which then were read in fulltext.*" We also conducted a full-text read for 1month of 1969 to ensure tbat the Lexis searchprocedure produced the same results as tbe fiali-text read. Once the sets of data were compiled,we computed selection models using logisticregression to predict coverage from the seriesof hypothesized covariates discussed in the fol-lowing sections.

PREDICTORS OF COVERAGE

EVENT CHARACTERISTICS: INTENSITY. The recentresurgence of attention to newspaper data basidentified a number of factors that affect tbechances tbat a collective event will be coveredby the media. Most of these factors are relatedto the model of media selection presented bySnyder and Kelly (1977; see Mueller 1997b fora more detailed bistory and explication ofthemodel). This model clusters selection variablesinto two general factors: event characteristicsand media sensitivity.

Among event characteristics, Snyder andKelly (1977) identified three aspects of eventintensity that contribute to newswortbiness:size, violence, and duration. The empirical linkbetween intensity and media coverage of protestis well documented (McCarthy, McPhail, andSmith 1996; Oliver and Maney 2000; Oliverand Myers 1999), and because civil disorders area form of protest, a similar relation should existin the current study. In our study, we examinedfive indicators of civil disorder intensity :̂ thenumber killed, tbe number arrested, tbe numberinjured, tbe presence of arson (as a proxy forproperty damage), and the duration of tbe eventin days.*̂ Because these intensity indicators

^ This same procedure could not be used for 1968because the Lexis data base for the New York Timesextended back only to 1969, so an actual full-text readwas conducted.

' Intensity indicators were collected mainly fromthe set of newspaper articles reporting each event.Some intensity information was supplemented bysurveys sent by the Lemberg Center to the policechiefs of cities from which a disorder had been report-ed.

^ For all events, there was complete information forthe number killed, duration, and arson. There weresignificant amounts of missing data for the numberinj ured and the number arrested. As a result, we con-ducted analyses four different ways. First, we used

overlap significantly, and because a single inten-sity indicator would simplify interaction analy-sis, we created a composite index using all fiveintensity indicators. Following Carter (1986)and Myers (2000), we log-transformed the indi-vidual indicators, subjected them to an iteratedprinciple axis factor analysis, and combinedtbem using tbe resulting factor scores.

POLITICAL SIGNIFICANCE: ACTORS AND LOCA-

TIONS. Since Snyder and Kelly (1977), the notionof newswortbiness has expanded to a more com-prehensive notion of news value, which incor-porates parts of what tbey called mediasensitivity. Recent news value definitionsinclude the prominence or importance of tbeactors involved, the level of conflict or contro-versy, physical proximity, buman interest, andthe number of people affected by the event(Oliver and Maney 2000). Ofthe large varietyof news value factors, we examined tbe politi-cal significance of tbe actors, the frequency ofevents, and proximity.

Differences in the significance of actors canbe difficult to ascertain, especially wben par-ticipants are similar across events. Among thecivil disorders we examined, for example, therace, economic status, and class status of par-ticipants varied little fTom event to event. Tbereare, however, two event characteristics tbat mayindicate tbe public significance of the actors.First, botb tbe press and social scientists bavetended to neglect events involving secondaryschool students, viewing their actions as whim-sical and derivative rather than politically pur-posive. Therefore, disorders occurring atsecondary schools or involving primarily sec-ondary school students should be relativelyunderrepresented in the Times.^ In contrast, tbe

only the three indicators for which there were nomissing data. Secondi, we replaced the missing val-ues for arrested and injured with mean values.However, because the data were extracted from news-paper stories, it is plausible that missing values camemainly from events for which there were no arrestsor injuries. Therefore, we also conducted analysesreplacing the missing values with zeros. Becauseresults were not substantively dependent on whichreplacement values were used, our analysis proceededusing the zero-plugged variants.

''A parallel set of all hypotheses will be examinedcombining data from the New York Times and theWashington Post.

A U THE RIOTING THAT'S FIT TO PRINT 535

highest status likely is given to college students,as compared with prison inmates, secondaryschool students, and the residents of urban ghet-tos, tbereby causing a relative overrepresenta-tion in Times-derived data.

Events also can differ in significance becauseof their location. Some cities, for example, aresimply more culturally and politically centra!than others (Myers 2000). Given the intensityofthe nation's focus on race during the periodstudied tbe media may have considered citieswith large black populations to be central botspots for newsworthy events. We bypotbesized,therefore, that events in cities with larger blackpopulations were reported at bigher rates tbanevents in cities with smaller black populations.

The black population is an important variablefor another reason. Almost without exception,prior studies of tbese disturbances show thatblack population size is the single most impor-tant determinant of civil disorder (Carter 1986;Spilerman 1970, 1976; Jiobu 1971; Myers1997). Altbougb recent studies bave added somevariables (Myers 1997; Olzak and Sbanaban1996), black population size and southern loca-tion are well established as tbe core predictorsin the literature. Given its centrality, it is dou-bly important to determine whether black pop-ulation size inflates coverage. If it does, therelation between civil disorder and black pop-ulation size may have been overstated. For thissame reason, we examined a southern regionindicator."* If southern location reduces cover-age, then the depressed rate of disorders in theSouth may have been exaggerated.

NEWS HOLE AND EVENT FREQUENCY. Events of

all types compete witb each otber for mediacoverage. When many newswortby events occurin a short period, tbe chances of any particularevent receiving coverage drops (Hocke 1998;Oliver and Maney 2000; Oliver and Myers1999). Tbis "news hole" effect links increasedlikelibood of coverage to slow news days.Because Mondays are considered slower newsdays and Fridays are considered heavier days,we expect tbat civil disorders occurring on

Mondays will be overreported, as comparedwith tbose occurring on Fridays (Oliver andMyers 1999).

In additional to this general news hole effect,we also expect to find a more specific news holeeffect in whicb an event competes with similarevents for coverage: Wben many otber events ofthe same type have occurred recently (opera-tionalized as the count of disorders in the pre-ceding 2 weeks),'' we expect the likelibood ofcoverage to drop because tbis type of news bolebas already been partially filled.

PROXIMITY. Wbetber rooted in news value ormedia sensitivity, the distance of an event froma newspaper is expected to influence coverage.Snyder and Kelly (1977) reasoned that news-papers are less likely to report distant eventsbecause they are less aware of them and becausethey lack tbe staff to investigate tbem. In addi-tion, distant events may lack coverage becausetbey are less important to tbe newspaper's audi-ence.'^ It is not clear how much distance influ-ences the Times, however, because thetechnology of tbe wire services reduces infra-structure difficulties and the Times' audience isless geographically concentrated than the audi-ences of many newspapers.

Attention to proximity effects has been aweak point of prior studies, if addressed at all.Snyder and Kelly (1977), for example, proxieddistance using two regional dummy variables.Even the most thorough distance tests (e.g.,Mueller 1997a) measure distance at only tbreeor four points, and the results are confoundedby otber cultural difference factors sueh as lan-guage (Barranco and Wisler 1999). Becausethe data we used were drawn from hundreds oflocal newspapers across the United States, wehad a unique opportunity to examine distanceeffects in a more direct way. We bypotbesized,therefore, tbat the chances that any event willappear in tbe Times decrease as tbe distance

'"The definition ofthe South follows Spilerman(1970) and includes Alabama. Arkansas, Florida,Georgia, Kentucky, Louisiana, Mississippi, NorthCarolina. Oklahoma, South Carolina, Tennessee,Texas, Virginia, and West Virginia.

'' We tested a variety of lags ranging from 3 daysto 8 weeks. The results showed that events occurringwithin 2 weeks before an event mattered, whereasthose older than 2 weeks did not.

'̂ Recent investigations into the diffusion of col-lective disorders also suggest that media sourcesemphasize civil disorders that occur in their ownareas (Myers 1997, 2000),

526 AMERICAN SOCIOLOGICAL REVIEW

from the event location to New York Cityincreases.

When geographic sensitivity is considered,simple linear distance may not capture all oftheeffects. The New York Times, af̂ er all, bas a spe-cial obligation (and ability) to provide infor-mation about events occurring in the city. Forthe same reason, it also may provide more cov-erage of events in New York state. We hypoth-esized, therefore, that after controlling fordistance, events in New York City and NewYork state will receive higher coverage. In addi-tion, tbe United States is sometimes character-ized as having a bicoastal nature. That is, thepublic and tbe media care more about the eastand west coasts than about what is in between.This suggests tbat the effect of distance on cov-erage may be curvilinear, decreasing as wemove west, but recovering to higher levels as weapproach the west coast.

INTENSITY-PROXIMITY INTERACTION. Someanalysts have speculated tbat selection based onevent intensity is conditioned by distance:Newspapers are less likely to report low-inten-sity events that occur far away than they are toreport those in their own backyards (Hug andWisler 1998; Mueller 1997a). Evidence is scant,however. Wben comparing American, British,and West German coverage of East Germanevents, Mueller (1997a) did find tbat British andAmerican newspapers were more sensitive tointensity than West German newspapers, but asimilar difference between American and Britishpapers did not exist. Our more detailed distancedata allowed a tborougb test of this interaction.

BASIC PATTERNS OF EVENTS AND COVERAGE

Table 1 displays key descriptive information aboutthe events located by the Lemberg Center.' ̂ Onecharacteristic of these events that distinguishesthem from the nonviolent protests examined inprior studies is the distribution ofthe events acrossdays of the week. Although reported events aresomewhat less likely to occur on Saturdays andSundays, the distribution across days is consid-erably more consistent than the distribution fornonviolent protest events (e.g., Barranco and

Wisler 1999). This is important because someanalysts (e.g., Kreisi et al. 1995) bave advocateda sampling strategy using only Monday editionsof newspapers. Tbis strategy captures events effi-ciently because nonviolent protests are beavilyconcentrated on Saturdays, with the result thatMonday newspaper editions report a large pro-portion of events (Barranco and Wisler 1999). TheMonday strategy would, bowever, bave been lesseffective for tbe current study, because the eventsit investigated were not concentrated on any par-ticular day ofthe week.

Our key question, however, was how welldid the Times and Post cover tbese events?Altogether, 418 (37.5 percent) of tbe eventswere reported in tbe Times. Adding the Postincreased coverage to 498 events (44.7 per-cent).''* It is not surprising, of course, tbat theTimes record is incomplete—the more impor-tant question is whether the satnple provided bythe Times differs substantially from the morecomprehensive data—providing a differentimpression ofthe events to botb scholars and thepublic.

Do these data sets differ with respect to thelocation of reported events? We first calculat-ed the rate of coverage by the Times and Post foreach state (Figure I). We found tbat the bigb-est coverage is clustered on the east coast, andthat coverage generally decreases toward tbewest coast. Of tbe states that had events,California, Arizona, and Oregon all were in thelowest quartile of coverage, with Nevada andWashington in the second lowest quartile. On theeast coast, only Pennsylvania was in the secondlowest quartile. This pattern suggests that dis-tance does indeed have a direct effect on deter-mining what is covered in newspapers.Furthermore., tbe map does not suggest abicoastal coverage pattern. There are few eventsfrom the middle section ofthe country in theTimes and Post, but this seems simply to reflecta lower base rate in tbese areas. For example, noevents in Idabo, Montana, North Dakota, SoutbDakota, Utah, and New Mexico were reported,even in tbe local newspapers of these states.

'̂ See the Appendix for correlations among keyvariables.

''' By itself, the Washington Post reported 22 per-cent of the events. For brevity of presentation andbecause the Post is used more often as a supplementto the New York Times than as a singular data source,separate results for the Post are not presented.

ALL THE RIOTING THATS FIT TO PRINT

Table I. Deseriptive Statistics

Variable

CoverageNew York TimesEither NYT or WP

Distance FaetorsDistance from NYCEvent occurs in NYCEvent occurs in NY StateEvent oceurs in Washington DC

Intensity IndicatorsNumber killedEvent durationArsonNumber injuredNumber arrestedNumber injured"Number arrested^Intensity composite

Loeati on/TypeSecondary sehoolCollege or universityStreet eventPrison event

City CharacteristicsTotal populationProportion blackCity in south

Event DensityNumber of events in prior 2 weeks

Day of the WeekSundayMondayTuesdayWednesdayThursdayFridaySaturday

Mean or Proportion

0.3750.447

7660.060.110.02

0.101.770.329.78

45.06.27

36,5-0.32

0.340.150.500.01

889,1120.230.25

32.6

0.090.150.160.130.180.190.10

SD

0.480.50

7520.230.320.14

0.651.650.46

59.1361

47.53260.83

0,480.350.500.09

1.930,1590.160.43

32.3

0.280.360.3fi0.340.390.390.30

N

11141114

1114111411141114

I I I411141114714902

111411141114

1114111411141114

111411141114

1114

1114111411141114111411141114

Minimum

00

0000

01000

0

0-1.66

0000

646.0001

0

0

0000000

Maximum

11

4964''

11191

11138236111382361.98

1I11

7,894,851.984

1

181

1

" Missing values replaced with zeros (see text for detail).** Without Honolulu. HI the maximum distance is 2575 miles.

Indicators of event severity also provide basicevidence about intensity hypotheses. Forinstance, 85 percent of the events in whichsomeone was killed (N = 47) were covered bythe Times. In bivariate analyses (not shown),each additional death tripled the odds of thatevent appearing in the Times. But even whenmultiple deaths occurred (N = 24), coveragewas not complete. Only 83 percent of thesewere covered. Likewise, events with low num-bers of injuries {fewer than 10) received rela-tively low coverage in the Times (32 percentcovered), whereas more than 80 percent of the

events with 20 or more injuries were covered.Arrests, duration, and arson all showed similarlinks between intensity and higher levels ofcoverage.

MULTIVARIATE SELEcnoN MODELS

These simple relations may be confounded witheach other and with other factors such as citycharacteristics, so we turn to multivariate mod-els. Table 2, presents models examining thechances of coverage in the Times (Models Iand 2) and in either the Times or Post (Models

528 AMERICAN SOCIOLOGICAL REVIEW

Table 2. Logistic Regression Models Predicting Coverage in Newspapers

l.NYT 2.NYT 3. NYT/Post 4. NYT/Post

Distance Factors

Event Intensity

Location/Type

City Characteristics

Event Density

Day ofthe Week(Friday excluded)

Constant

Model x̂ (df)Pseudo R̂

Distance from NYC

Distance from NYC^

Event occurs in NYC

Event occurs in NY State

Event occurs in Washington DC

Composite variable

Intensity x distance

Intensity x distance^

Secondary school

College or university

Street event (excluded)

Prison event

Proportion black (sqrt)

Total population (in N)

City in south

Events in prior 2 weeks (In N+1)

Sunday

Monday

Tuesday

Wednesday

Thursday

Saturday

1

-.00125***(.00033)

2.74x10-'*

( l . lxKT')1.79**(.64)143*. .(.32)

.804***(.10)

-.582***(.18)1.36***

(.22)

-.590(1.0)

1.20**

(-46)

.0131

(.046)

.0798

(.20)

-.227*

(.10).193

(.30)

.600*

(.25).0526

(.26).0988

(.27).326

(.24)

.212

(.28)-.149

(-67)

320.86(18)

.218

-.00113***

(.00033)

2.33 X 10-'*

(1.2x10-')1.75**(.64)1.21***(.32)

.323(.18)

.00111*

(.00045)

2.66 X 1 0 '

(1.9x10^')

-.586***

(.18)

1.39***

(.23)

-.517(1.0)

1.33**

(-46)

.00702

(.046)

.0767(.20)

-.215*

(.10)

.153(-30)

.616*

(.25).0245

(.26)

.110

(.27)

.346

(.25).223

(.29)

-.259

(.68)

333.46 (20)

.226

-.00191***(.00040)

5.27x10"'***(1.5 X 10-')

1.62*

(.64)

1,03***(.32)

1.58*

(.80)

.762***

(.10)

-.829***

(.17)

1.05***

(.22)

-1,13

(1.0).76!

(.46)

.0133

(.044)

.375

(.20)

-.286**

(.10)

.069

(.29)

.452

(.24).0390

(.24)

-.00645

(.26).158

(.24)

-.0528

(.28)

1.01

(.66)

325.31(19)

.212

-.00167***(.00039)

4.46 X 10-'**

(1.5x10-')1.63*(.64).869**

(.32)

1.54

(.79)

.344

(.18)

.000851*

(.00043)

1.72x10-'

(1.8 X to-')

.-.839***

(.17)

1.08***

(.23)

-1.16

(1.0).864

(.47)

.00593

(.045)

.357

(.20)

-.272**

(.10).026

(.29)

.456(.24)

.0123(.24)

.00210

(.26)

.170

(.24)

-.0513

(.28)

.895

(.66)

335.65(21)

.219

Note: SE in parentheses. In = natural log; NYC = New York City; NYT = New York Times; NYT/Post = New York

Times or Washington Post; sqrt = square root.* p < . 0 5 , • * ; ? < . 0 1 , * * * p < . 0 0 1 (two-tailed tests). i. i -

ALL THE RIOTING THAr S FIT TO PRINT

QuartilesNo Events

First

Second

Third

Highest

Figure 1. Coverage in the New York Times and Washington Post by State

3 and 4). The effects of the hypothesized vari-ables are substantial.

First, distance is an important factor in pro-ducing coverage: As the distance from NewYork increases, the likelihood of coverage dropssubstantially. Figure 2 plots the probability ofan event being covered simply on the basis of

distance.'^ Events near New York City havealmost a 60 percent chance of being reported inthe Times, whereas those on the opposite coast

'̂ Dropping the one distance outlier, whichoccurred in Honolulu, did not appreciably change theresults. [

00 5(X) 1000 1500

Distance in Miles

Figure 2. Distance and Declining Probability of Coverage

2000 2500

53O AMERICAN SOCIOLOGICAL REVIEW

have less than a 20 percent chance. The com-bination of the Times and the Post produceshigher coverage overall, but a substantial dis-tance effect is still apparent. The graphs (basedon a polynomial model) also show that there islittle support for the bicoastal hypothesis.Although the negative slope of the curve isattenuated as distance increases, the rate of cov-erage beyond 1,500 miles is essentially con-stant. Adding the Post data increases relativecoverage of the west coast, hut only slightly.

Given that the Times and Post also arelocal-regional papers, it is not surprising thatevents in New York City had a much greaterchance of receiving coverage, as did eventsoccurring in New York state. Even after control-ling for distance and including a New York statedummy, the odds of coverage for New YorkCity events were six times higher than for anyother location. Events in the state had four timesgreater odds of coverage (Model 1, Table 2). Thesame kinds of patterns appeared when we addedthe Post data and a dummy for Washington DC,although the effects were somewhat reduced.'^

Thus, although the Times might be a goodsource for events in New York, it is much lesscomplete for the United States. Adding a sec-ond newspaper, such as the Post, reduces the dis-tance effects slightly, but also contributes specialsensitivity to its local events.

Second, consistent with prior studies, eventintensity was shown to have a strong positiveeffect on coverage (Models I and 3, Table 2).'''As hypothesized eariier, the results also indicatethat the effect of event intensity is conditionedon distance. Models 2 and 4 of Table 2 show asignificant interaction between intensity anddistance (likelihood ratio tests comparing ModelI to Model 2 and Model 3 to Model 4 were sig-nificant at p < .01). Figure 3 displays this rela-tion graphically, showing the effect of distanceat three different levels of intensity: moderate(the mean of the intensity composite), mild (onestandard deviation below the mean), and severe(one standard deviation above the mean). The

'̂ The distances from Washington DC and NewYorkarehighlycoUinear for these data (r= .98), andthus could not be included in the same model.

" We tested additional models using individualintensity indicators and various imputed values formissing data. All the intensity indicators were con-sistently significant except where reduced bycoUinearity with each other, and the other variablesin the models were stable.

0.7

0.6

0.5

£

0.4>oUo•I* 0.33es

0.2

0.1

Severe

\ \ " ^ ^ - - ~ ~ - _ _

Mild / ^ ^

1 1

Moderate

"-"

—1 1

0 500 1000 1500

Distance in Miles

2000 2500

Figure 3. Distance-Severity Interaction

AU THE RIOTING THATS PIT TO PRINT 531

graphs show that distance decreases coverage,that intensity increases coverage, and that theeffect of intensity on coverage increases as dis-tance grows. In other words, the Times alwaysunderreports faraway events and small events,but as the distance grows, it becomes even lesslikely to report small events.

Event types also played a role in coverage.Both newspapers had considerably less interestin secondary school events than street riots, andthe Times emphasized college events. The col-lege effect was reduced to some degree by theaddition ofthe Post data. Besides the low polit-ical significance of secondary school students,these findings also may reflect the politicalleanings ofthe two newspapers or differencesin readership.'**

Bivariate analyses showed that events in larg-er cities had a much greater chance of being cov-ered, but lhe multivariate analysis showed thatthis was primarily because of coUinearitybetween total city size and black populationsize. Once a control is used for the proportionof blacks in the city, the total population size isnot significant in any model, but for the Times,the effect ofthe black population is significant.Given the potential attention cycle bias and the"beat'*-related issues discussed in prior work, itseems likely that newspapers considered blacksto be important newsmakers during the periodstudied—either as political actore or alterna-tively as criminals, hooligans, and threats tosocial order (Jefferies, Tumer, and Morris 1971).If so, the newsmakers paid closer attention toconcentrations of black population and left citieswith smaller numbers relatively less covered.'"*When the Post data were added (Models 3 and

'̂ We also thought that the college effect might bean attention cycle pattern because college protestand collective violence was accelerating during thisperiod. We tested via a dummy variable distinguish-ing 1968 events from 1969 events and an interactionterm for the college indicator and the year. If thisattention cycle process was operating, college eventsshould have received more attention in 1969 than1968. The results did not show such a pattern.

'̂ Some observers have suggested that these kindsof city effects result simply from the distributioti ofreporters. If there is a stritiger or a wire office in thattown, then its events are more likely to be reported.Even if this is true, there still is the question of whyreporters were concentrated in cities with large blackpopulations.

4), neither population indicator was significant,thus demonstrating the one substantial correc-tion achieved by adding the Post data. Finally,southern location did not influence the chancesof coverage.

With respect to news hole effects, the day ofthe week has an effect on coverage, consistentwith prior studies and the slow/heavy news dayexplanation: Monday events have a slight advan-tage over Friday events. Additionally, event den-sity has a significant negative effect both for theTimes and for the combined data: When thereis more competition from similar events, thechances of coverage decrease.^"

SELECTION EFFECTS ONSUBSTANTIVE ANALYSIS OF CIVILDISORDER

These findings about event selection by theTimes and Post reinforce earlier fears about thevalidity of newspaper-based data and doubtsabout findings derived from such data.Nevertheless, no study has yet demonstratedthat these selection effects, in fact, change theoutcomes or interpretation of substantive eventanalyses. Just because the media select rela-tively severe events, for example, does not nec-essarily mean that the event selection distorts therelation between civil disorder and hypothe-sized covariates. ln the second part of our analy-sis, we undertook a direct examination of this

question.I

DATA AND ANALYSIS

To produce some direct evidence about theeffects of omitted events on analyses of civil dis-orders, we used three separate data sets to rean-

^° It is possible that the event density effect cameonly from the week of extreme civil disorder that fol-lowed the assassination of Martin Luther King inApril 1968. Because there was such a large numberof events that week, it may have been impossible fornewspapers to cover them as well, thereby produc-ing the event density effect. Figure 4 (discussed later)shows that this is unlikely because the events in April1968 were covered by both papers at a slightly high-er than average rate. We also inserted a dummy vari-able for the King events into a variety of multJvatiatemodels. When significant, the variable had a slightpositive effect on coverage, and it did not diminishthe event density effect.

532 AMERICAN SOCIOLOGICAL REVIEW

alyze a series of covariates previously found topredict civil disorder. The primary data set con-tained all the events in our data from all news-paper sources. The other two data sets weresubsets of the first: all the events reported in theNew York Times and all the events reported ineither the Times or the Washington Post. Theresults from these latter two data sets representfindings that could have been produced usingthe national newspaper or newspapers strategy.A comparison of results from these three datasets demonstrates what kinds of differences infindings and interpretation can occur if a sin-gle source strategy is used.

For these tests, we selected a series of covari-ates that relate the effects of local conditions tothe occurrence of civil disorder. Admittedly,most attempts to link local conditions with civildisorder have largely been failures (see espe-cially Spilerman, 1970.1971, 1976), but recent-ly, several studies have revived this link (Olzakand Shanahan 1996; Olzak, Shanahan, andMcEneaney 1996; Myers 1997; Myers 2000).In particular, these analyses found that structuralcovariates related to ethnic competition theory,and that diffusion processes were important inproducing disorder. From among the covariatestested in these studies, we selected those thatwere the most robust and produced the moststraightforward findings. The structural indi-cators included the black population of eachcity, a dummy indicating whether the city waslocated in the South, the proportion of the pop-ulation that was foreign bom, the number ofunemployed blacks, and the median blackincome.^' Each indicator was collected from1970 Census tabulations.

We also tested for effects over time and geo-graphic diffusion effects such as those recent-ly found by Myers (1997, 2000). Thefundamental notion driving diffusion analysesis that collective events are clustered systemat-ically in time and space because these events are

^' Many other variables drawn from ideas aboutstructural conditions, inequality, population charac-teristics, and political structures have been testedover the past 30 years, but, as demonstrated inSpilerman (1970, 1976), Myers (1997), and Olzakand Shanahan (1996), these variables perform poor-ly when predicting city-level differences in civil dis-order occurrence.

in some sense "contagious." To test for differ-ences in diffusion effects, we used three covari-ates: the number of events in the precedingweek (temporal clustering), the number ofevents in the preceding week conditioned on theage of each event (temporal clustering withdecaying influence), and the number of eventseach week conditioned on the distance of theevent from each city (spatial clustering). Thefirst diffusion indicator simply means that thegreater the number of recent events, the morelikely there is to be an additional outbreak. Thesecond indicator is similar to the first except thatit recognizes that the imitation-inducing effectof an event wanes daily, in effect, decaying overtime. The third indicator recognizes that eventsfar from a particular city may be considerablyless influential in producing imitation than thoseoccurring nearby.'^

To estimate these models, we used Cox (par-tial-likelihood) regression. Cox regression isone type of survival or event history analysis(Allison 1984; Tuma and Hannan 1984;Yamaguchi 1991). Instead of differentiatingcities on the basis of the presence or the num-ber of disorders, event history analysis focuseson the individual risk or hazard of disorder. Byapproaching the analysis of events in this way,researchers can examine time-invariant covari-ates (e.g., southern location) and time-varyingcovariates (e.g., the number of events in thepast week) in the same models. Although thisadvantage of survival analysis is not particularlyconsequential for examining the city-level struc-tural covariates, it becomes critically importantin the analysis of diffusion processes (Greve,Strang, andTuma 1995; Davis and Greve 1997;Soule 1997; Soule and Zylan 1997; Strang andTuma 1993). The coefficients generated can beinterpreted by taking e**, which produces the

^̂ Because our purposes in this article are to exam-ine the effeets of bias on the estimated relationsamong variables rather than to examine theories ofcivil disorder and difftision, we have given only briefsummaries of these substantive issues and the rele-vant literature. Thorough presentations of linksbetween the economic indicators we used and com-petition theory are available in Olzak and Shanahan(1996) and Myers (1997). Details on the role of dif-fusion in this series of riots and on the calculation ofthe diffusion indicators may be found in Myers (1996,2000).

ALL THE RIOTING THATS FIT TO PRINT 533

amount by which the hazard is multiplied foreach unit increase in the related independentvariable.

Because analysis of collective violencedemands inclusion of repeated observations onthe same unit, and because it is unlikely that esti-mated models account for all sources of intrau-nit dependence, unobserved heterogeneity mayintroduce a downward bias in standard errorestimates. Therefore, a control variable, thenumber of prior events that have occurred in acity, was included in all the models. The inclu-sion of variables that represent the history of theindividual unit is a practical procedure that min-imizes the effects attributable to correlationwithin the same city and does not demand therestrictive assumptions of more general meth-ods used to correct for unobserved heterogene-ity (Allison 1984). I

FINDINGS: THE EFFECTS OFSELECTION ON SUBSTANTIVE MODELS

In the remaining analyses, we report three setsof models, each predicting the city-level hazard

of disorder. Each model is estimated first usingall of our data, then using only those eventsreported in the New York Times, and then againusing events reported in either the Times or theWashington Post. In addition to presenting theCox regression results, each table also com-pares each of the restricted sample models withthe full sample model. Each cell of the tablereports the coefficient, standard error, and pvalue for the respective covariate. The secondand third columns also compare the estimatedcoefficient with the coefficient reported in thefirst column. This comparison is accomplishedby indicating with a dagger those coefficientsjutside the 95 percent confidence interval indi-cated by the results in column one.

Table 3 presents the results of models that usethe city-level covariates to predict the hazard ofa disorder. Model I uses only the two covariatesSpilerman (1970) found to be important: Blackpopulation and the south dummy (plus the dis-order history control variable). Model 2 adds theeconomic/competition variables. The findingsresulting from the more comprehensive dataare consistent with those of prior studies: Strong,

Table 3. Cox Regression of the Disorder Hazard on City CharacterisUes

Model 1Black population (logged)

City in south (dummy)

History of disorder (Control)

Model x^ idf}Psetido R̂

Model 2Black population (logged)

City in south (dummy)

Proportion foreign bom

Black median income (logged)

Proportion blacks unemployed (logged)

History of disorder (Control)

Model x^ (df)Pseudo R̂

All Events

.321(.019)-.605(.073).0245

(.0033)681.03(3)

.0459

.334(.019)-.596(.11)2.15(.58)-.345(-18).309

(.066).0215

(.0034)714.95(6)

.0482

P

.000

.000

.000

.000

.000

.000

.049

.000

.000

In NYT

.418t(.028)-.666(.12).0289

(.0048)436.14(3)

.0802

.414t(.029)-.776(.18)4.49t(.79)-.826t(.30).185

(-13).0214

(.0051)467.55 (6)

.0860

P

.000

.000

.000

.000

.000

.000

.007

.144

.000

In NYT or Post

.40 If(.026)-.604(.11).0277

(.0047)456.92 (3)

.0702

.396t(.027)-.573(.16)3.40t(.78)-.379(.28).184

(.112).0223

(.0049)475.73 (6)

.0731

P

.000

.000

.000

.000

.000

.000

.171

.102

.000

Note: SE in parentheses; p value of .000 indicates p < .0005. NYT = New York Times; Post = Washington Post.t Coefficient is outside the 95% confidence interval indicated by column one.

AMERICAN SOCIOLOGICAL REVIEW

statistically significant relations exist betweenthese predictors and the disorder rate. The blackpopulation size, the proportion of foreign-bomindividuals, black income, and black unem-ployment all have positive relations with disor-der, whereas southern location reducesdisorder.^^

A comparison of these findings with theresults derived from the restricted data sets(Columns 2 and 3) shows some consistenciesand some inconsistencies. On the consistentside, the coefficients all maintain the same signacross the three models. The magnitude andstatistical significance ofthe effects, however,differ substantially In Model 1, for example, theTimes coefficient for black population size ismore than five standard error units larger thanthe estimates given in the all-events data, indi-cating that the importance ofthe black popula-tion variable may have been overstated in pastresearch. The Times data also indicate muchstronger effects of the foreign-born variableand black income (Model 2). Each of thesecoefficients is more than twice those estimatedin the more comprehensive data. Differences ofthis magnitude in coefficients can mean sub-stantial differences in the interpretation oftheresults, and in other circumstances (dependingon the size ofthe original coefficients), couldproduce completely different findings fora vari-able, causing shifts in direction or significance.

Differences in statistical significance are alsoof concern in Table 3. To begin with, most ofthestandard errors are unexpectedly larger in therestricted models. Whereas some upward shiftin the standard errors comes simply from hav-ing fewer events in the data.̂ "* the differences inthis case are substantially higher. In this context,the standard errors in the Times models shouldbe approximately 63 percent higher, and thosein the combined Times and Post models should

^̂ See Myers (1997) and Olzak and Shanahan(1996) fora full interpretation.

''' Although the actual N in this analysis does notchange given that the same number of city-days isexamined in each model, the change in the numberof events recorded for a binary dependent variable(i.e.. the number of zeros versus the number of ones)affects the standard errors such that in this case, therelatively smaller numbers of events increases thestandard errors.

be about 50 percent higher. But across the mod-els shown in Tables 3 and 4, the increase instandard errors usually is higher, and in somecases, higher by factors of two or more.Although this does not mean that the estimat-ed coefficients are biased, it does mean that ourability to detect relations in the data is reduced.̂ ^In fact, we see that although the black unem-ployment effect is significant dX^p value lessthan .001 in column one, it is not significant ineither ofthe restricted-data models. The blackincome effect is significant in the all-eventsmodel and when the Times is used, but it is notsignificant when the Times and Post combina-tion is used. This kind of problem is particularlyimportant in the current empirical contextbecause the riot literature is so strongly domi-nated by null findings (e.g., Spilerman 1970,1976). Given these results, we must consider thepossibility that hypotheses about the effects ofstructural and economic conditions on civil dis-order may have been incorrectly rejected.

One source of the increased standard errorsappears to be the additional collinearity intro-duced by use of the restricted data sets. Forexample, in bivariate analyses (not shown), theblack unemployment variable retains signifi-cance in both the Times and the combined Timesand Post data. Thus, when other variables areadded to these models in Table 3, collinearitybetween them and black unemployment elimi-nates the black unemployment effect, whereasthis does not occur when the ftjll data are used.This kind of effect can occur if the Times reportsmore events in cities where some of the sub-stantive indicators overlap and less events incities where economic indicators are relativelyneglected because economic indicators are moremixed. Perhaps the Times was more vigilant inlooking for events in places where ghetto con-ditions were consistently bad, or more likely toreport disorders that occurred where conditionsseemed consistent with popular understandingsof the riots. In either case, the media processwould hamper attempts by social scientists todiscover the underlying conditions associatedwith racial civil disorder.

^̂ Even those few variables (e.g.. foreign bom) thatproduce changes in standard errors smaller than theexpected increase are problematic because they areoverestimating the significance ofthe results.

Combining the Times and Post data general-ly reduced the differences from the larger dataset, but the change was marginal for most coef-ficients. The most convergence occurred whenthe Times and Post differed in opposite direc-tions from the more comprehensive data, withthe combination canceling out much of the dif-ference. On the one hand, this may testify to theusefulness of combining multiple data sources.On the other hand, to achieve the helpftil effect,data sets that deviate in opposite directions mustbe combined. If the Post differences were inthe same direction as those of the Times, com-bining the two data sets could have exacerbat-ed, rather than reduced, the differences.

Table 4, presents the diffusion analysis. Thecoefficients presented come from models thatalso included the city covariates examined inTable 4. To simplify the presentation, the resultsfrom those covariates are not shown. Onceagain, the news is mixed. The two models inTable 4 show that the effects were all in a con-sistent direction,^^ but restricting the data tothe Times and the combined Times and Postconsistently increased the estimates of all threekinds of difTflision effects: The total number ofrecent events was found to be more influential

*̂ Again, detailed interpretations of the pattern ofcoefficients is available elsewhere (Myers 1996,1997).

AUTHB RIOTING THATS FITTO PRINT 535

(Model 1); these effects were found to decaymore over time (Model 2); and they were foundto be more limited by distance (Model 2). Thedifferences in the time decay and the distancedecay effects were the most extreme because theestimated coefficients were up to 5.5 timesmore than the value found using the completedata set.

These results suggest a somewhat mixed pic-ture for single- or dual-source newspaper data.On one hand, the results show that across the dif-ferent data sets, there is consistency in the direc-tion of significant coefficients. This isencouraging news for those who wish to use sin-gle-source data. On the other hand, if one isconcerned about the magnitude of the effectsrather than direction alone, the results suggestthat substantial distortion can occur across thedifferent event samples. In other circumstances,changes of this magnitude could reverse effectsin different samples, especially if regional vari-ation is strong.

DISCUSSION

In summary, the results in our study suggest thatthe selection problems with single, nationalmedia sources may be more problematic thanpreviously thought. Although we examined onlyone type of American collective violence overa relatively short period, when we add our find-ings to the selection effects other scholars have

Table 4. Cox Regression of the DisordaH HSKdfm Diffusion Indicators

Model 1Total events in prior week

Model x̂ (dfiPseudo R2

Model 2Total events in prior week

Time deeay

Distance decay (I/distance)

Model x^ (dfiPseudo R̂

All Events

.01221 (.0011)

819.46(7).0552

-.0362(.012).0506

(.0066).0277

, (.011), 885.01 (9)' .0596

P

.000

.002

.000

.012

In NYT

.0270t(.0043)

502.28 (7).0924

-.0245t(.063).H7t

(.024).2l4t

(.057)547.38 (9)

.101

P

.000

.000

.000

.000

In NYT or Post

.0235!(.0034)

517.17(7).0795

-.020 It(.050).0917t

{•019).179t

(.046)562.12(9)

.0864

P

.000

.000

.000

.000

Note: SE in parentheses; p value of .000 indicates/* < .0005. Coefficients for city structural covariates not shown.NYT = New York Times; Post = Washington Post.t Coefficient is outside the 95% eonfidence interval indicated by column one.

536 AMERICAN SOCIOLOGICAL REVIEW

documented, it is clear that the use of newspa-per data is not always as straightforward as wewould like it to be, especially when there is areliance on a single source such as the Nevi^York Times. These results also suggest selec-tion problems for other kinds of events besidescivil disorders. The selection mechanisms weexamined (distance, regional focus, event inten-sity, political significance of actors, and newsholes) are relevant across events ranging fromassassinations to riots and protests to earth-quakes. In fact, other events may produce evengreater selection effects than civil disorders.For example, nonviolent protest probably isreported far less, simply because it lacks thedrama of violence and overt conflict (Mueller1997a; Oliver and Maney 2000; Oliver andMyers 1999).

CoRREcnvB METHODOLOGY

What can be done to address these selectionproblems? A number of suggestions for improv-ing the treatment of newspaper data are begin-ning to emerge, focused either on changing howdata is collected or on changing data analysisand interpretation. The most obvious and satis-fying solution is simply to collect more andbetter data by expanding the number of mediasources consulted. Adding just a small numberof sources is not enough, however, becauseadding only one or two sources produces onlymarginal gains and may even make mattersworse.̂ ^ Collecting data from a large number ofsources, however, is neither cheap nor easy, andin some cases (e.g., long-term historical stud-ies), multiple sources that cover the necessarytemporal and geographic scope simply are notavailable. For more recent events, however, thepossibilities for affordable collection of com-prehensive, or at least representative, data growwith the electronic compilation of media sources(e.g., via LexisNexis). Although searching elec-tronic media compilations comes with its ownset of problems that scholars have yet to solve(Maney and Oliver 2001), these sources even-

tually may allow cost-effective comprehensivedata collection.

In the absence of comprehensive data, somehave suggested that a selective reduction of thedata might help. Because, for example, eventintensity is so important in producing coverage,bias may be reduced by analyzing only relativelyintense events (Snyder and Kelly 1977).Although bias might be reduced by this strate-gy, our analysis shows that, at least in the cur-rent empirical circumstances, selection effectswould not be eliminated—the effects in Table2 exist affer control is used for intensity. We fiir-ther tested this strategy by dropping all eventswith less than the median intensity score. Whenwe examined these relatively intense events,we found that both the Times and Post did covermore of the events (45 percent for the Times, and29 percent for the Post). However, when werecalculated the selection analysis, the resultsparalleled Table 3, even for the intensity indi-cators.-̂ ^

In the end, bias in the data sometimes cannotbe avoided, so adjustments to analyses also havebeen considered. Hug and Wisler (1998) havesuggested assessing bias during one brief peri-od, then using that information to produce con-trols in other analyses. Unfortunately, suchstrategies depend on bias being consistent acrosstime. If it is not, the correction may fail or evenovercorrect, thereby obscuring important rela-tions. In the current study, the required consis-tency did not exist. The upper panel of Figure4 plots events over time in all newspapers andin the Times and Post. If such plots are reason-ably parallel, then relatively constant bias isthought to hold. In the current case, the peaksand valleys in the event counts seem reasonablyconsistent.

^' For example, as Table 2 shows, the neglect ofsecondary school events is greater in the analysiscombining the Washington Post and New York Timesdata.

^̂ Another possibility is to limit analyses to largecities. We recalculated our results using only citieswith populations of 25,000 or more, but because wecontrolled for city size in our original analyses, theonly appreciable difference was that the event den-sity coefficient lost significance. As Myers (2000) hasshown, civil disorders in smaller cities tend to followevents in larger cities. Therefore, eliminating smallcities reduces the temporal clustering that drivesevent density and thereby masks social contagioneffects. It is not, therefore, a recommended pnxedure.

ALL THE RIOTING THATS FIT TO PRINT 537

I—I—1—I—I—I—I—1—I—I—I—I I 1 I • • • •

1/68 1/69Note: 198 Events Rcponcd in All Sources in April of lWS

12/69

Times and Post

71/68 1/69 12/69

Figure 4. Number and Proportion of Events Reported in New York Times and Washitigton Post by Month1968-1969

Unfortunately, this representation ofthe datais misleading. If we display the proportion ofevents covered over time instead of raw numbers(the lower panel of Figure 4), we see that cov-erage is extremely inconsistent over time: TheTimes coverage, for example, ranges from morethan 60 percent covered to less than 20 per-cent.̂ ^ This inconsistency makes it impossibleto assess bias in one period and confidently usethat information to correct for unknown bias inother periods.^" Besides these short-term fluc-

^' Both trends deviate significantly from consis-tency: x^ = 47.4; df= 23; p < .01 for the New YorkTimes trend and x^ = 47.8; df= 23; p < .01 for theTimes and IVashington Post trend.

'" Comparison of the Washington Post and theNew York Times shows that coverage patterns are not

tuations in coverage patterns, there also maybe longer-term changes to address. In the cur-rent case, tbe period of 1968 to 1969 wasextraordinarily conflictive, both domesticallyand internationally, and was at the peak of aprotest cycle. Thus, the available news bolemay have been relatively small, as comparedwith that during more tranquil periods, there-by reducing the chances of coverage for anyindividual event relative to a similar event inmore peaceful times. The assumption of con-sistent bias is tberefore quite tenuous, and theusefulness of corrective strategies that dependon it is suspect.

consistent across newspapers either Thus, selectiondetected in one paper is unlikely to provide adequatecorrection for data collected from others.

538 AMERICAN SOCIOLOGICAL REVIEW

Given the uncertainty about the effective-ness of these corrective strategies, event analystsmust approach Times or other single-sourcenewspaper data warily, considering carefullyhow selection effects may influence results andinterpretations. But how serious the problem isfor any individual analysis depends, in part, onwhat kinds of questions are asked. If the analy-sis is aimed primarily at identifying long-termsurges in political contention, then perhaps thereis less reason for concem. Likewise, if Timesevent data is used as an independent variablerepresenting information that has been intro-duced to elite policymakers, it may indeed bean accurate proxy. If, on the other hand, one isinterested in predicting events on the basis ofstructural variables and diffusion influences,the challenge appears to be considerably greater,particularly if one is interested not only in thedirection of effects, but also in their magnitude.

CONCLUSION

Relying on mass media sources for informationinevitably produces a simplified, distorted, andincomplete picture ofthe world around us. Butwhat do the media miss and why are some typesof events neglected more than others? Past mod-els of media selection identify three clusters offactors. First, certain characteristics of eventsmake them important or newsworthy. Theseinclude the intensity and uniqueness of theevent, and the notoriety or political significanceofthe actors. Second, contextual elements canincrease the difficulty of reporting the event.These include proximity, news-gathering rou-tines, and censorship. Finally, attention cyclesreflect changing demands from an audience,which shift from excitement to fatigue as a topicis reported repeatedly.

The current findings clearly support thesepast models. For example, no matter what aspectof riot severity we examine, higher levels ofintensity increase the chances that the eventwill be reported by tbe Times. Thus, the selec-tion process exaggerates the typical amount ofviolence in civil disorders by emphasizingintense outbreaks and ignoring mild ones, lead-ing to a distorted view of protest among poli-cymakers, the public, and scholars. At the sametime, the intensity bias understates the geo-graphically pervasive scope of unrest by ignor-ing events occurring in smaller cities that simply

do not have the critical mass necessary to sus-tain an extremely severe riot. Responses to riot-ing, then, may be targeted more toward endingthe violence (via repression) in a few key loca-tions and less on addressing the pervasivesources of discontent.

The current results also suggest that the rolethe audience plays in producing media selectionhas been underemphasized. A key judgmentthat reporters and editors make is how impor-tant a potential story will be to their audience.The demand for a particular kind of story maybe real, assumed, or even induced by the media,but whatever way the news staff comes to under-stand its audience, this understanding ends updriving its assessment of newsworthiness, andthus its decisions about what will be includedin tbe media record. Past studies, for example,have suggested that proximity is fundamental-ly an infestructure issue: Because of limitedresources, local media will be less sensitive todistant events. But beyond infrastructure diffi-culties, local audiences also are less interestedin distant events, and, as a result, local mediawill not bother to report them, even when theyhave the needed information. Local civil disor-ders, for example, are more salient to the localaudience because they may affect travel in thecity, influence local government action andresource distribution, or even provide opportu-nities for participation in further collectivebehavior. For these and related reasons, atten-tion to international events likely suffers evenmore from selection processes. In addition toeven greater distances and tbe accompanyinginfrastructure challenges, cross-border politi-cal, cultural, and language differences reducerelevance to local audiences, thereby reducingcoverage and comprehensiveness.

One audience effect notion more clearlydeveloped in earlier research is the effect ofattention cycles (McCarthy, McPhail. and Smith1996), which inordinately increases coveragewhen an issue becomes "hot" news and thenneglects related events as the audience losesinterest. In the current study, we identified anevent density effect that may partially offsetthe attention cycles phenomenon. Becausenewspapers have limits on how much protest orrioting they will cover, when there are manyevents available to fill tbis news hole, thechances of any particular event receiving cov-erage decrease. This produces pressure on cov-

ALL THE RIOTING THATS FIT TO PRINT 539

erage exactly opposite that of attention cycles,tending to result in overreporting at the begin-ning and end of a protest cycle and underre-porting in the middle. Thus, the event densityeffect may extend the life of a story despite a rel-atively constricted period of audience interest,and it is a dynamic that may help get protestcycles started at points where they have yet toproduce critical interest and resources.

Whereas attention cycles and event densitymay influence selection decisions in the short-er term, selection wiil change over longer peri-ods because of technological advances. As theability to distribute news over a wider geo-graphic region increases (via satellite trans-mission of broadcasts and newspaper content,live coverage of news events, and Internet newsdelivery), the audience grows and the demandfor broader geographic coverage grows as well.Moreover, as information-gathering technology(wire services, search engines, cellular phones,and networks of personnel) improves, the abil-ity to detect events—even geographically dis-tant events—grows as well. Whereas these twoeffects of technological advance have the poten-tial to reduce the geographic bias in a newspa-per such as the New York Times, they alsoproduce inconsistency in selection pattems overlong periods. Thus, one oft-cited advantage ofnewspaper records, their continuity, may notbe all that it seems: Trends in newspaper-baseddata may be less trends in the empirical eventsand more reflections of changing technology.

Finally, audience interest also helps to deter-mine selection practices by defining who thepolitically and culturally significant actors are.if a political actor's scope of influence overlapswith a newspaper's perceived audience, thenthe newspaper wiil be more likely to report thatactor's words and behavior. In our study, forexample, secondary school students' activitieswere least reported, probably because readers ofthe Times do not typicaily consider the behav-ior of high school students politically signifi-cant. During this same period, however, collegestudents had considerably more notorietybecause of consistent and dramatic campusprotests about many social issues. This notori-ety made their actions more relevant to readersand subsequently increased attention to theircollective behavior.

Notoriety does not necessarily imply positiveattitudes toward the actors, however. But even

when attitudes are negative, notoriety can driveup coverage. Early studies of foreign protests(e.g., Azar et al. 1972) found that Americancoverage focused on regions in which US. for-eign policy interests were at stake, but clearlythis was not because U.S. interests were sup-ported by the protests. More recently, Richardsand McCarthy (2002) found that LouisFarrakhan's Million Man March received muchmore media attention than a very similar event,the Promise Keeper's rally. A major part ofthedifference is attributable to the negative noto-riety of Farrakhan and the Nation of Islam forthe largely white readerships of most US. news-papers. In our study, we found that events incities with larger black populations receivedmore coverage, and we suggest that newspaperswere keying in on blacks as significant politi-cal actors. This does not mean that the news-papers necessarily supported the actions andgoals of protesting blacks, but given the recenthistory ofthe civil rights movement, the destruc-tive violence between blacks and whites in 1967,and the general state of race relations, blackswere undoubtedly significant newsmakers anddrew extra attention to cities in which they werehighly concentrated.

Race also may have driven reporting in otherways. For example, it is possible that reportersmay have selected activities that fit receivedstereotypes. Whites and the white media tend tocouple blacks and violence, consequently over-reporting, for example, black-on-white violentcrime (e.g., Jacobs 2000). Inner-city civil dis-orders in which black participants destroy prop-erty and loot stores are "culturally resonant" forwhites, illustrating, highlighting, and reinforc-ing prior suspicions and stereotypes aboutblacks, and thereby providing a well-fittingnevfs "peg" for white audiences (McCarthy,McPhail, and Smith 1996).

Our analysis shows that media selection is animportant process that must be carefully exam-ined and accounted for in research using news-papers. But in the end, the effects of selectionmay be even more important for their influenceon the social and political wodd. The currentstudy examined a set of events considered news-worthy by some newspapers in the United Statesand tried to understand how it became moreimportant news by virtue of appearing in theTunes or Post. Reported events in tbe two news-papers have large audiences and are more like-

54O AMERICAN SOCIOLOGICAL REVIEW

ly to catch the attention of elites and policy-makers. In fact, this characteristic ofthe Timeshas been offered in defense ofthe data, undertbe logic that the events that really matter areonly those that influence elite policymakers.Although this logic is arguable on a numberof grounds, it reinforces tbe fundamentalnotion that events become more socially

significant by appearing in the Times. Tbeseevents, then, bave greater power to constructand distort social reality tban those that lan-guish in tbeir local newspapers. Moreover,when social scientists subsequently use theselected events as their data, they both acceptand reinforce media standards about what is"fit to print."

APPENDIX:

Table 1. Correlation Coefficients for Variables Used in Coverage Analysis and Disorder Hazard Analysis

Coverage Variables(1) Distance from NYC(2) Event in NYC(3) Event in NY State(4) Event in Washington. DC(5) Intensity composite(6) Proportion black (sqrt)(7) Total population (ln N)(8) City in south(9) Events in prior 2 weeks (In N+I)

(5) Intensity composite(6) Proportion black (sqrt)(7) Total population (In N)(8) City in south(9) Events in prior 2 weeks (ln N+1)

City Characteristics(1) Black population (logged)(2) Proportion foreign bom(3) Black median income (logged)(4) Proportion blacks unemployed (logged)(5) City in south

(1)

1.000-0.254-0.341-0.104

0.065-0.2040.0310.038

-0.002

(6)

—1.0000.1780.245

-0.085

(1)

1.000-0.022-0.117-0.421

0.168

(2)

—1.0000.694

-0.035-0.0810.0180.492

-0.142^.031

(7)

——

1.000-0.284-0.120

(2)

—1.0000.452

-0.020-0.460

(3)

——

1.0000.050

-0.120-0.0660.261

-0.2040.042

(8)

———

1.0000.007

(3)

——

1.0000.067

-0.714

(4)

———

1.0000.0360.3150.114

-0.079-0.106

(9)

———• —

1.000(4)

———

1.000-0.166

(5)

————

1.0000.0510.0380.0240.007

(5)

————

1.000

ln = natural log; NYC = New York City; sqrt = square root.

Daniel J. Myers is Professor and Chair of Sociologyat the University of Notre Dame and Feliow oftheKroc Institute for International Peace Studies. He iscutTenlly studying structural conditions, diffusionpatterns, and media coverage related to U.S. racialrioting and aiso investigating changes in public opin-ion in response to violent events, such as rioting andterrorism attacks.

Beth Schaefer Caniglia is Assistant Professor ofSociology at Oklahotna State University. Her researchfocuses at the intersection of social movements,organization and policy making, especially related tothe environmental movement. Her most recentresearch is examining the impact of science com-

munication on public opinion and environmentalpolicy formation regarding water resources.

REFERENCES

Allison, Paul D. 1984. Event History Analysis.•Regression for Longitudinal Event Data. NewburyPark, CA: Sage Publications.

Azar, Edward E.. Stanley H. Cohen, Thomas O.Jukam, and James M. McCormick. 1972. "TheProblem of Source Coverage in the Use ofInternational Events Data." International StudiesQuarterly 3:3n-SS.

Bandura, Albert and Richard H. Walters. 1963. Social

ALL THE RIOTING THATS FIT TO PRINT S41

Learning and Personality Development. New York:Holt, Rinehart, & Winston.

Banks, Arthur S. 1997. Cross-National Times Serie.sData Archive. Binghatnton, NY: ComputerSolutions Unlimited.

Barranco, Jose and Dominique Wisler. 1999."Validity and Systematicity of Newspaper Data inEvent Analysis." European Sociological Review15:301-22.

Baskin, Jane A.. Joyce K. Hartwig, Ralph G. Lewis,and Lester McCullough. 1971. Race Related CivilDisorders. 1967-69. Waltham, MA: BrandeisUniversity.

Burch, Philip H. 1972. The Managerial RevolutionRea.^sessed. Lexington, MA: Heath.

Carter, Gregg L. 1983. Explaining the Severity of the}960's Black Rioting. Unpublished dissertation.New York: Columbia University.

. 1986. "The 1960s Black Riots Revisited:City Level Explanations of Their Severity,"Sociological Inquiry 56:210-28.

•. 1990. "Colleetive Violence and theProblem of Group Size in Aggregate-LevelStudies." Sociological Focus 23:287-300.

Danzger, M. Herbert. 1975. "Validating ConflictData." American Sociological Review 40:570-84.

Davis, Gerald F. and Henrich R. Greve. 1997."Corporate Elite Networks and GovernanceChanges in the 1980s." American Journal ofSociology 103:1-37.

Franzosi, Roberto. 1987. "The Press as a Source ofSocio-Historical Data: Issues in the Methodologyof Data Collection from Newspapers." HistoricalMethods 20:5~\6.

Galambos, Louis. 1975. Public Image about BigBusiness. 1880-1940: A Quantitative Study inSocial Change. Baltimore. MD: Johns HopkinsUniversity Press.

Gitlin, Todd. 1979. "Prime Time Ideology: TheHegemonic Process in Television Entertainment."Social Problems 26:251-66.

Greve, Henrich R., David Strang, and Nancy Br^donTuma. 1995. Specification and Estimation ofHeterogeneous Diffusion Models. SociologicalMethodology 25:377^20.

Gurr, Ted Robert. 1968. "A Causal Model of CivilStrife: A Comparative Analysis Using NewIndices." American Political Science Review62:1104-24.

Ha2elwood, Leo A. and Gerald T. West. 1974."Bivariate Associations, Factor Structures, andSubstantive Impact: The Source ProblemRevisited." International Studies Quarterly18:317-37.

Hocke, Peter. 1998. "Determining Selection Bias inLocal and National Newspaper Reports on ProtestEvents." Pp. 131-63 in Acts of Dissent: NewDe\'elopments in the Study of Protest, edited byDieter Rucht, Ruud Koopmans, and Friedhelm

Neidhardt. Berlin, Germany: Wissenschafts-zentrum Berlin Fur Sozialforschung.

Hug, Simon and Dominique Wisler. 1998."Correcting for Selection Bias in Social MovementResearch" Mobilization: An JntemationalJoumal3:141-61.

Jackman, Robert W. and William A. Boyd. 1979."Multiple Sources in the Collection of Data onPolitical Conflict." American Journal of PoliticalScience 23:434~5ii.

Jacobs, Ronald N. 2000. Race. Media, and the Crisisof Civil Society: From fVatts to Rodney King.Cambridge, UK: Cambridge University Press.

Jaggers, Keith and Ted R. Gurr 1995. "TrackingDemocracy's Third Wave with the Polity III Data."Journal of Peace Research 32:469-82.

Jefferies, Vincent, Ralph H. T\imer, and Richard T.Morris. 1971. "The Public Perception ofthe WattsRiot as Social Protest." American Sociological^ev;ew 36:443-51.

Jenkins, J. Craig and Charles Perrow, 1977."Insurgency of the Powerless: Farm WorkerMovements (1946-1972)."/^mencon Sociologicalffev/>M'42:249-68.

Jenkins, J. Craig and Craig M. Eckert. 1986."Channeling Black Insurgency: Elite Patronageand Professional Social Movement Organizationsin the Development of the Black Movement."American Sociological Review 5 \ .812-29.

Jiobu, Robert M. 1971. "City Characteristics andRacial Violence." Social Science Quarterly

Kelly, William R. and David Snyder. 1980. "RacialViolence and Socioeconomic Changes amongBlacks in the United States." Social Forces58:739-60.

Kelly, William R. and Larry Isaac. 1984. "The Riseand Fall of Urban Racial Violence in the U.S.:1948-1979" Research in Social MovementsConflict and Change 7:203-33.

Koopmans, Ruud. 1998. "The Use of Protest Data inComparative Research: Cross-NationalComparability, Sampling Methods andRobustness." Pp. 90-110 in Acts of Dissent: NewDe\'elopments in the Study of Prote.st. edited byDieter Rucht, Ruud Koopmans, and FriedhelmNeidhardt. Berlin, Germany: Wissenschafts-zentrum Berlin Fur Sozialforschung.

Kriesi, Hanspeter, Ruud Koopmans, Jan WillemDuyvendak, and Marco G. Guigni. 1995. NewSocial Movements in Westem Europe, Minneapolis,MN: University of Minnesota Press.

Lieberson, Stanley and Arnold R. Silverman. 1965."The Precipitants and Underlying Conditions ofRace Riots." American Sociological Review30:887-98.

Lieske, Joel A. 1978. "The Conditions of RacialViolence in American Cities: A Developmental

542 AMERICAN SOCIOLOGICAL REVIEW

Synthesis." American Political Science Review72:1324-40.

Maney, Gregory M. and Pamela E. Oliver. 2001."Finding Event Records: Timing, Searching,Sources." Sociological Methods and Research29:131-69.

McAdam, Doug. 1982 Political Process and theDevelopment of Black Insurgency, 1930-1970.Chicago: University of Chicago Press.

McCarthy, John D,, Clark McPhail, and Jackie Smith.1996. "Images of Protest: Dimensions of SelectionBias in Media Coverage of WashingtonDemonstrations, 1982 and 1991." AmericanSociological Review 61:478-99.

McCarthy, John D, Larissa Titarenko, Clark McPhail,and Boguslaw Augustyn. 1998. Media Source andTemporal Stability in Selection Bias of NewspaperCoverage of Protest Events. Paper presented atthe Annual Meeting ofthe American SociologicalAssociation, San Francisco, CA.

MinkofF, Debra. 1997. "The Sequencing of SocialMovements." American Sociological Review62:779-99.

Mueller, Carol. 1997a. 'international Press Coverageof East German Protest Events, 1989." AmericanSociological Review 62:820-32.

. 1997b. Media Measurement Models ofProtest Event Data. Mobilization: An IntemationalJournal 2: \65-S4.

Myers, Daniel J. 1996. The Diffusion of CollectiveViolence. Paper presented at the AmericanSociological Association meetings. New York, NY.

— — ^ . 1997. "Racial Rioting in the 1960s: AnEvent History Analysis of Local Conditions."American Sociological Review 62:94-112.

. 2000. "The Diffusion of CollectiveViolence: Infectiousness, Susceptibility, and MassMedia Networks." American Journal of Sociology106:173-208.

Myers, Daniel J. and Alexander J. Buoye. 2001."Campus Racial Disorders and Community Ties,1967-1969." Research in Social Movements,Conflict, and Change 23:297-327,

Oliver, Pamela E. and Daniel J. Myers. 1999. "HowEvents Enter the Public Sphere: Conflict, Location,and Sponsorship in Local Newspaper Coverage ofPublic Events." American Journal of Sociology105:38-87.

Oliver, Pamela E. and Gregory M. Maney. 2000."Political Cycles and Local Newspaper Coverageof Protest Events: From Selection Bias toTdadicInteractions." American Journal of Sociology106:463-505.

Olzak, Susan. 1982. "Ethnic Mobilization inQuebec" Ethnic and Racial Studies 5:253-15.

. 1992. The Dynamics of Ethnic Competitionand Conflict. Stanford, Stanford University Press.

Olzak, Susan and Johan L. Olivier. 1994. The

Dynamics of Ethnic Collective Action in SouthAfrica and the United States: A Comparative Study.Paper presented at the International SociologicalAssociation meetings, Bielefeld, Germany.

Olzak, Susan and Suzanne Shanahan. 1996."Deprivation Race Riots: An Extension ofSpilerman's Analysis." Social Forces 74:931-61.

Olzak, Susan, Suzanne Shanahan, and Elizabeth H.McEneaney. 1996. "Poverty, Segregation, and RaceRiots: 1960 to 1993." American SociologicalReview 6].590-(>\3.

Paige, Jeffery. 1975. Agrarian Revolution: SocialMovements and Export Agriculture in theUnderdeveloped World. New York: Free Press.

Richards, Assata and John McCarthy. 2002. Pie Roleof Controversy and Celebrity in the NewspaperCoverage of the Million Man March and thePromise Keepers Stand in the Gap Rally.Unpublished manuscript. University Park, PA:Pennsylvania State University.

Rummel, Randolph J. 1966, "Dimensions of ConflictBehavior within Nations, 1946-59." Journal ofConflict Resolution 10:65-73.

Russett, Bruce M. 1964. World Handbook of Politicaland Social Indicators. New Haven: Yale UniversityPress.

Shanahan, Suzanne and Susan Olzak. 1999. "TheEffects of Immigrant Diversity and EthnicCompetition on Collective Conflict in UrbanAmerica: An Assessment of Two Movements ofMass Migration, 1869-1924 and 1965-1993."Journal of American Ethnic History 19:40-64

Shorter, Edward and Charles Tilly. 1974. Strikes inFrance. Cambridge, UK: University Press.

Snyder, David and William R. Kelly. 1977. "ConflictIntensity, Media Sensitivity, and the Validity ofNewspaper Data." American Sociological Review42:105-23.

Soule. Sarah A. 1997. "The Student DivestmentMovement in the United States and TacticalDiffusion: The Shantytown Protest." Social Forces75:855-82.

Soule, Sarah A. and Yvonne Zylan. 1997. "RunawayTrain? The Diffusion of State-Level Reform inADC/AFDC Eligibility Requirements,1950-1967." American Journal of Sociology103:733-62.

Spiierman, Seymour. 1970. The Causes of RacialDisturbances: A Comparison of AlternativeExplanations. American Sociological Review35:627-49.

. 197 L The Causes of Racial Disturbances:Test of an Explanation. American SociologicalReview 36:427-42.

. 1976. Structural Characteristics of Citiesand the Severity of Racial Disorders. AmericanSociological Review 41:771-93.

Strang, David and Nancy B. Tuma. 1993. "Spatial and

ALL THE RIOTING THATS FIT TO PRINT 543

Temporal Heterogeneity in Diffusion." ,4mericanJournal of Sociology 99:614-39.

Taylor, Charles L. and David A. Jodice. 1983. WorldHandbook of Political and Social Indicators. 3dedition. New Haven: Yale University Press.

Taylor, Charles L. and Michael C. Hudson. 1972.World Handbook of Political and Social Indicators.2d edition. New Haven: Yale University Press.

Tuma, Nancy B. and Michael T.Hannan. 1984.5bcia/Dynamics: Models and Methods. San Diego, CA:Academic Press.

United States Senate Committee on GovernmentOperations. 1968. Staff Study of Major Riots andCivil Disorders: 1965 through July il. 1968.Washington, DC: U.S. Government PrintingOffice.

Woolley, John T. 2000. "Using Media-Based Data inStudies of Politics." ^mer(canJouma/o//fe//rtc(7/Science 44.\5(>-n.

Yamaguchi, Kazuo. 1991. Event History Analysis.Newbury Park, CA: Sage Publications.