Can News protect Civilians? Media Attention and Civilian Casualties in Syria

1. Introduction and Research Design

Civil wars not only include violence between fighting parties, but also against civilians who are not directly participating in the conflict (“indiscriminate violence”). The suffering of civilian populations during these wars tends to catch the attention of international media. While English-speaking news outlets arguably target an audience that does not directly participate in the conflict, many civil wars involve international actors. The parties who are fighting the war may have an interest to shape their image in a way that increases international support to their struggle.

In that context, one might ask whether conflict parties make an effort to refrain from indiscriminate violence when their war receives a lot of medial attention. Specifically, this short essay asks to what degree the intensity of violence against civilians in civil wars is causally influenced by media attention?

Scholarly literature in this area has previously focused on the effect of medial (human-rights related) attention on both duration of and violence in civil wars.¹ Yet these few studies use yearly country-level data and therefore are unable to track short-term variation in media attention and conflict intensity. Nor can they completely rule out that conflict intensity influences media attention more than the other way around.²

I contribute to that gap here by using daily data on news reports and violent deaths of civilians during the Syrian Civil War in 2012. Media attention is proxied with the number of news articles on the war in English-speaking newspapers.³ Daily data on civilian deaths is provided by the Center for Documentation of Violations in Syria (VDC) and the Syrian Shuhada website.⁴

I examine how the variation in news reports influences the number of civilian deaths, with the hypothesis that:

(H1) Violent deaths of civilians decrease when media attention increases.

The assumption is that decision-makers who conduct military operations are more prone to avoid civilian casualties,⁵ when they feel pressured by international attention to their struggle. The primary reason not to use a simple multilinear regression to conduct this study is that news reports on Syria are likely to be published as a result of violence in the conflict, rather than the other way around (reverse causality):

(H2) Violent deaths of civilians cause media attention to rise.

I thus use a time-series design with lagged measures of media attention to test whether the death count at one time (t) is influenced by news reports before that time (t - “lag”). The primary problem here is to decide how long it takes for (macro-level) media attention to “sink in” into (micro-level) decision-making processes to influence (macro-level) civilian casualties. I assume that this process will take between one and three days, but will test multiple lag structures to determine where the effect is strongest.⁶

I also assume that media attention is only one (out of many, possibly more important) factors taken into account by the conflict party’s decision-makers. I do control for both passed and vetoed UN Security Council resolutions on Syria, since they might spark indiscriminate violence and media attention at the same time. I further use a dummy for the period leading to the US presidential election campaign to control for a potential drop in media attention on Syria, and use seasonal dummies to account for changes in violence related to conflict intensity at temporal stages (see summary statistics in Appendix 1).

2. Analysis

Figure 1 depicts the daily number of civilian deaths and news reports over time: 
 

 
The first thing to note is that the two variables move along similar lines. That media attention on Syria remains relatively low while the death count spikes after July could be a result of the political recess period during the summer in many countries, and the US presidential elections in the fall 2012. Second, since changes in media attention slightly precede changes in death counts, it appears that more attention to the conflict leads to more civilian deaths, and not less as assumed in H1.⁷

I attempt to confirm these visual impressions with a regression analysis in table 1. Models (2) and (3) use the same variables, but a lag of one and four days for media attention, respectively.⁸

Table 1. Influence of Media Attention on Civilian Deaths

DV: Civilian Deaths	(1)	(2)	(3)
Media Attention (log, t-1)	0.27***	0.13*
	(3.60)	(2.44)
Media Attention (log, t-4)			0.17**
			(3.19)
Civilian Deaths (log, t-1)		0.44***	0.43***
		(9.17)	(9.29)
UN Resolution Passed		-0.50*	-0.48*
		(-2.24)	(-2.14)
UN Resolution Vetoed		1.53***	1.52***
		(4.34)	(4.34)
US Presidential Elections		0.13	0.14
		(1.63)	(1.76)
Season April-June		0.078	0.069
		(1.10)	(0.97)
Season July-September		0.59***	0.58***
		(6.77)	(6.66)
Season October-December		0.48***	0.47***
		(5.49)	(5.52)
Constant	3.39***	1.67***	1.57***
	(11.90)	(7.10)	(6.40)
Adj. R-Squared	0.03	0.61	0.60
Observations	365	365	362

Note: Lagged OLS estimates, t-stats in parentheses. ^* p < 0.05, ^** p < 0.01, ^*** p < 0.001.

Source: Syrian Shuhada Database.

The results confirm that media attention has a positive influence on civilian deaths. On average, one additional article on Syria is associated with a 0.13 increase in civilian deaths on the next day, with a chance of 5% that this estimate is a mere coincidence. The effect is slightly stronger and more significant with a lag of four days in model (3).⁹ The problem that previous values of death counts might predict future values is not solved by including the lagged dependent variable, but a more advanced regression model shows that such autocorrelation does not significantly change the results.¹⁰
 
UN Resolutions appeared to result in a slight decrease of violent deaths, and vetoes in a significant increase. The seasonal dummies also confirm the significant time-variant increase in violence starting in July.¹¹ Unsurprisingly, further tests also show clear evidence that media reports increase as a result of violence, confirming the reverse causality hypothesis H2 (see Appendix 2).

3. Conclusion

These results suggest that media outlets anticipate civilian casualties and write more about it before they happen, or that there may be a causal mechanism at work that leads conflict parties to more aggressively target civilians as a result of medial attention. A macro-level study cannot unravel these causal mechanisms, nor is this evidence from a single conflict proof for a causal relationship across civil wars.

This analysis does however show that using year-data to analyze the relationship between media attention and civilian casualties (or conflict intensity in general) is highly problematic. First, changes in conflict intensity can be caused by a number of factors that can’t be controlled for with yearly data. And second, media attention is highly in flux, and is caused by civilian casualties at least as much as it predicts them.

A study that could both attend to these problems and be valid across wars would require daily, micro-level data for many more civil wars, which is unlikely to be available anytime soon.

Appendices

Appendix 1: Summary Statistics

	Mean	Std. Dev.	Min.	Max.	Description
Civilian Deaths Number	102.66	65.79	8.0	419.0	daily number of civilian casualties
News Reports on Syria	51.05	25.23	7.0	162.0	daily number of news articles on the Syrian Civil War
UN Resolution Passed	0.01	0.12	0.0	1.0	dummy variable, 1 if resolution was passed on a day
UN Resolution Vetoed	0.01	0.07	0.0	1.0	dummy variable, 1 if resolution was vetoet on a day
US Presidential Elections	0.16	0.36	0.0	1.0	dummy variable, 1 for every day after US congress recess until election day (Sep-Nov)
Season January-March	0.25	0.43	0.0	1.0	seasonal dummies
Season April-June	0.25	0.43	0.0	1.0
Season July-September	0.25	0.43	0.0	1.0
Season October-December	0.25	0.43	0.0	1.0
Observations	366

Appendix 2: Influence of Civilian Deaths on Media Attention (H2)

DV: Media Attention	(1)	(2)	(3)
Civilian Deaths (log, t-1)	0.20***	0.18***
	(5.76)	(5.63)
Civilian Deaths (log, t-4)			0.060
			(1.69)
Media Attention (log, t-1)		0.66***	0.70***
		(17.67)	(17.53)
UN Resolution Passed		-0.15	-0.18
		(-0.97)	(-1.09)
UN Resolution Vetoed		0.28	0.29
		(1.14)	(1.13)
US Presidential Elections		-0.11*	-0.082
		(-2.10)	(-1.46)
Season April-June		-0.064	-0.046
		(-1.30)	(-0.88)
Season July-September		-0.18**	-0.044
		(-2.95)	(-0.68)
Season October-December		-0.21***	-0.10
		(-3.55)	(-1.61)
Constant	2.94***	0.63***	0.94***
	(19.10)	(3.90)	(5.55)
Adj. R-Squared	0.03	0.61	0.57
Observations	365	365	362

Note: Lagged OLS estimates, t-stats in parentheses. ^* p < 0.05, ^** p < 0.01, ^*** p < 0.001.

Source: Syrian Shuhada Database.

_________________________________________________________________________________

1Burgoon, Brian/Ruggeri, Andrea/Schudel, Willem/Malikkalingam, Ram 2015: From Media Attention to Negotiated Peace. Human Rights Reporting and Civil War Duration, in: International Interactions 00: 1, 1–30.
Ruggeri, Andrea/Burgoon, Brian 2012: Human Rights "Naming & Shaming" and Civil War Violence, in: Peace Economics, Peace Science and Public Policy 18: 3, 1–12.

2Particularly for the effects of media attention on violence, studies struggle with endogeneity problems, and find it difficult to rule out reverse causality. See Ruggeri, Andrea/Burgoon, Brian 2012: Human Rights "Naming & Shaming" and Civil War Violence, in: Peace Economics, Peace Science and Public Policy 18: 3, 1–12.

3As per a query of the LexisNexis database (http://www.lexisnexis.com/), searching for articles related to International Relations and Security (includes tags on human rights and civil wars), and containing the words Syria, Assad, Damacus, Aleppo, Idlib, or Homs. The latter were particularly contested areas of the war. While this multiplies attention if newspapers publish when other’s do, I assume that conflict parties care more about the overall attention (“outcome”), and that a less biased approximation would actually confound the “perceived” media attention of these actors.

4The Syrian Shuhada website (http://syrianshuhada.com/) includes the VDC data. It is considered relatively authentic, considering the inherent reliability problems of casualty reports in (civil) wars. The data is publicly available, includes the names and place of the deceased, and is based on local activist information networks (“Local Coordination Committees”). See: Scharpf, Adam/Schneider, Gerald/Nöh, Anna/Clauset, Aaron 2013: Die Blutspur des Vetos. Eine Prognose zur Gefahr von extremen Massakern in Syrien, in: Zeitschrift für Friedens- und Konfliktforschung 2: 1, 6–31.
The decision to choose the Syrian case in 2012 is primarily informed by the availability of this relatively reliable micro-level data, which can rarely be found for civil wars.

5Whether consciously or unconsciously.

6I also include a lagged measure of death casualties in order to avoid autocorrelation.
Basic formal model:

7Also notable are the spikes (“outliers”) in death counts in February (military offensive on Homs) and August (bombing and massacre in Damascus), which are probably more dependent on other factors than media attention.

8I use a series of vector autoregression tests to determine which time (lag) between media attention and civilian deaths is most appropriate. Most tests suggested that a lag of one day is most appropriate, one test pointed to a four-day lag.

9The cumulative effect of media attention (LTE) equals 0.23 in model 2. The effect is stable for lag structures from one to seven days, and also holds in a Granger causality test. The latter also shows more significant results for the reverse causality hypothesis than for H1.

10Including the lagged dependent variable in models (2) and (3) still resulted in heavy autocorrelation in a Durbin’s alternative test for autocorrelation. However, a Cochrane-Orcutt AR(1) regression shows almost identical, significant results at 1% for a four-day lag structure (model 3), with a transformed Durbon-Watson test statistic of 2.23 (original: 1.15). The one-day lag structure (model 2) reports P>|t| at 0.07, and thus barely looses significance at the 5% level.

11The omitted dummy variable here is the season from January to March, significant increases in violence after July are thus relative to that period.