Does television terrify tourists? Effects of US televisionnews on demand for tourism in Israel

David Fielding & Anja Shortland

Published online: 24 April 2009# Springer Science + Business Media, LLC 2009

Abstract We analyze the impact on US tourist flows to Israel of variations in boththe actual intensity of the Israeli-Palestinian conflict and the intensity implicit in UStelevision news coverage. Conditional on actual events, changes in reported conflictintensity could influence tourists because alternative sources of information arecostly; this explanation is consistent with a rational choice model. However,television news could influence tourist behavior because of its emotional impact, orbecause it causes the conflict to be brought to mind more readily, increasing thesubjective probability of conflict events. We find that tourists respond to variationsin actual Israeli casualties and reported Palestinian casualties; both effects are large.Reports of Israeli casualties and unreported Palestinian casualties have no significantimpact on tourist flows. These asymmetries are consistent with asymmetricinformation costs within a rational choice framework, but are more difficult tosquare with the alternative explanations for media influence.

Keywords Media . Choice under uncertainty . Probability neglect . Cultivationtheory . Israel

JEL D89 . L83 .M3

This paper analyzes the impact of the television media on US tourist decisions aboutwhether to travel to Israel. It reports evidence on the extent to which changes in thereported intensity of the Israeli-Palestinian conflict, as well as changes in the actualintensity, affect tourist behavior. We have two main aims. First of all, we wish toestimate the size of the effect of the conflict, and of reports of the conflict, on the

J Risk Uncertain (2009) 38:245–263DOI 10.1007/s11166-009-9067-z

D. Fielding (*)Department of Economics, University of Otago, Dunedin, New Zealande-mail: dfielding@business.otago.ac.nz

A. ShortlandDepartment of Economics, Brunel University, London, England

Israeli tourism industry and hence the Israeli economy. Secondly, we wish to shedsome light on the reasons why media reporting affects tourist numbers. Some mediainfluence might be consistent with strict neoclassical economic rationality; forexample, alternative sources of information about the risks involved in travelling toIsrael might be very costly. However, there are other explanations for mediainfluence that lie outside the neoclassical model of choice under uncertainty. Forexample, more intense reporting may cause people to bring the conflict to mind morefrequently, and this may increase the subjective probability of being a victim in aconflict event; or television images may have an emotional impact that induces abehavioral response disproportionate to the actual risk involved.

We estimate the size of the impact of the media on US tourists by applyingregression analysis to monthly time-series data on tourist numbers, actual conflictevents and reports of conflict events in the US television news. Our measures ofactual and reported conflict intensity are not very highly correlated, and both areincluded in our regression equations. Thus, our estimates of the impact of televisionnews reports on tourist numbers are conditional on actual events. For the media toinfluence behavior, people must respond not only to the actual monthly changes inconflict intensity, but also to the changes implied by the media coverage.

In measuring conflict intensity, we also make a distinction between Israelicasualties and Palestinian casualties. The threats that tourists typically face (forexample, suicide bomb attacks) resemble the threats to Israelis more closely than thethreats to Palestinians, so the response of tourists to Israeli casualties (actual orreported) may differ from their response to Palestinian casualties. Asymmetries intourist responses to Israeli and Palestinian casualties may also result from politicalsympathies for one side or another, or from asymmetries in the amount ofindependent information tourists have about threats to one side or another—forexample from friends or relatives living in Israel.

We find that variations in conflict intensity have a large impact on touristnumbers, but that there is a marked asymmetry between the impact of Israelicasualties and the impact of Palestinian casualties. Tourist numbers respond tovariations in actual Israeli casualties and to variations in reported Palestiniancasualties. There is no significant response either to reports of Israeli casualties(conditional on actual casualties) or to unreported Palestinian casualties. Thisasymmetry is consistent with a neoclassical explanation of media effects based oncostly information. Tourists may have cheap alternative sources of information onIsraeli casualties (for example, Israeli friends and family) but not on Palestiniancasualties. The asymmetry is harder to reconcile with the alternative explanationsdiscussed above, because it is unlikely that reports of Palestinian casualties willcause the conflict to be brought to mind more often by the average American tourist,or have an emotional impact, while reports of Israeli casualties do not.

1 Conceptual background

The empirical model presented in the next section draws on two strands of economicthought. The first relates to media influence on the perceived risk associated withpotentially dangerous activities. Individual households may have imperfect

246 J Risk Uncertain (2009) 38:245–263

information about the risks associated with, for example, eating certain kinds of foodor visiting certain neighborhoods. When information is costly, they may use theintensity of media coverage of a hazardous activity as an indicator of the level of riskinvolved, interpreting intense media interest as a sign of concern among those whoare better informed. Alternatively, we might abandon strict economic rationality infavour of the psychological model of Tversky and Kahneman (1973), in which thefrequency with which events of a certain type come to mind conditions perceptionsof the incidence of such events in everyday life. More intense media coverage ofdangers increases their perceived frequency, regardless of the true risks involved.Moreover, the emotional impact of reporting extreme violence may result in“probability neglect” (Sunstein 2003), with behavioral responses that are dispropor-tional to the risks involved, especially when the events are easily visualizable.1

The importance of the media in conditioning perceptions is well documented inseveral areas of economics and other social sciences. Criminologists have usedhousehold survey data to examine the extent to which the intensity of coverage ofviolent crime on television news channels affects the perceived risk of being a victim ofcrime. The hypothesis that the television news media conditions people’s perceptions—know as “cultivation theory”—dates back to Gerbner and Gross (1976). Recent studiesprovide substantial evidence for the importance of news coverage of crime inconditioning perceptions. Romer et al. (2003) analyze data from the US General SocialSurvey to show that individuals’ perceptions of the frequency of crime, conditional ona range of socio-economic characteristics, depend not only on actual crime in theirlocality, but also the intensity of crime reporting in the local television media. Thisevidence is reinforced by more detailed surveys of individual cities, showing thatpeople watching more television news perceive there to be a higher level of violentcrime, ceteris paribus (Romer et al. 2003; Gross and Aday 2003; Chiricos et al. 1997).

The focus of attention in criminology is on respondents’ self-reported perceptionsof risk. There is, as far as we are aware, no evidence on the extent to whichtelevision reports of violent crime affect people’s behavior. However, there iseconometric evidence that media reporting of health risks has a direct impact onconsumer behavior. Verbeke and Ward (2001) analyze time-series data on Belgianmeat consumption during the period of the BSE crisis. They show that demand forbeef was highly sensitive to time-series variations in the intensity of media reportsabout the risk of BSE. Other studies reporting a significant media effect on demandfor food during health scares include Mazzocchi (2004) and Kalaitzandonakes et al.(2004). In some papers, such as Piggott and Marsh (2004), the authors note that themedia effect is economically and statistically significant, but short-lived. Oneinterpretation of these results is that media reports produce a Tversky-Kahnemaneffect, but one that lodges only in people’s short-term memory. However, the resultsare also consistent with rational consumers who assume that journalists are betterinformed about short-term variations in the level of risk.

The second strand on which we draw (a large part of which is reviewed by Freyet al. 2007) relates to the economic consequences of violent conflict. Much of the

1 Survey evidence indicates that American attitudes towards the risk of extreme political violence exhibitbiases inconsistent with simple Expected Utility models. See Fischhoff et al. (2003) and Viscusi andZeckhauser (2003).

J Risk Uncertain (2009) 38:245–263 247247

politically motivated violence in the modern world takes the form of low-intensityconflict which might not disrupt economic production in any substantial way.However, it can still have substantial demand-side effects, especially when tourism isa major export industry. Evidence for the effect of political violence on the tourismsector is reported in Enders and Sandler (1991) for Spain, Enders et al. (1992) forAustria, Italy and Greece, Drakos and Kutan (2003) for Greece, Israel and Turkeyand Sloboda (2003) for the US. In related work, Drakos (2004) examines the effectsof 9/11 on various airline companies. One paper that focuses explicitly on Israel isFleischer and Buccola (2002). This paper shows that foreign demand for touristaccommodation is particularly sensitive to violence in the region, whereas localdemand is quite insensitive; these effects are reflected in local financial markets.

Such papers provide rigorous and compelling evidence on the effects of politicalviolence on tourism. However, they leave one key issue unresolved. To the extentthat they employ accurate measures of conflict intensity, rather than the intensitywith which the conflict is reported by the popular media, they do not directly addressthe role of the media in conditioning perceptions of risk.2 In order to analyze the roleof the popular media, we need to distinguish between the impact of actual violenceon tourists’ decisions and the impact of reported violence, as in the criminologyliterature. Our paper addresses this issue using monthly tourism and conflictintensity data for Israel during the most recent Israeli-Palestinian conflict (the AlAqsa Intifada, beginning in October 2000), and corresponding television media datafrom the United States. While not a full-scale civil war, the Intifada represents acontinuous ongoing conflict, rather than a series of isolated conflict events.

2 The structure of the empirical model3

Appendix 1 explains the choice-theoretic framework from which we derive theregression equations used to test hypotheses about the causes of monthly variationsin the volume of tourist flows from the US to Israel. (This derivation relies onassuming the Independence of Irrelevant Alternatives, an assumption that appears tobe appropriate for our data, as discussed in Appendix 2. All appendices are availableon request.) Our regression equations are of the form:4

ln TISt=TEUtð Þ ¼ a1 � ln TISt�1=TEUt�1ð Þ þ a2 � ln TISt�2=TEUt�2ð ÞþP

s qs � hts þ f1 � PSt þ f2 � NYt þ

Pk hk1 � xkt þ hk2 � xkt�1

� � þ "tð1Þ

2 By contrast, Burger and Sturm (2005) construct a model of the German macro-economy conditional onthe number of German television media reports of conflicts around the world. However, they do notcompare this with a model using actual data from the conflicts that are partially reported.3 The regression specification here is similar in spirit to that of Fleischer and Buccola (2002), who analyzetotal foreign demand for Israeli hotel accommodation up to 1999 as a function of a single “terror index,”but differs from their model in some details. There is no significant coefficient on any economic variablein our regression equation.4 Equation (1) includes lags of the dependent variable up to 2; higher lags are theoretically possible, butturn out to be statistically insignificant in all cases we consider.

248 J Risk Uncertain (2009) 38:245–263

where α , θ, � and η represent fixed parameters. TISt represents the number of UStourists visiting Israel in month t; TEUt is a scale variable, measuring US touristflows to a relatively safe location, Europe. Later, we will explore the sensitivity ofour results to using an alternative scale variable. In the underlying theoretical model,ln(TISt / TEUt) is proportional to the difference between the expected utility fromvisiting Israel and the expected utility from visiting Europe. xt

k (where k = 1,…, K)denotes the value in month t of one of K time-varying indicators of conflict intensitythat might impact on the relative attractiveness of Israel as a tourist destination,including indicators of both actual and reported violence. εt is a regression residual.The other elements of Eq. (1) capture seasonal effects: ht

s is dummy variable formonth s of the year, PSt is a dummy variable for the occurrence of Passover inmonth t and NYt a dummy variable for the occurrence of the Jewish New Year.5

Our dependent variable is constructed from two data sources.Monthly Israeli tourismdata are published by the Central Bureau of Statistics and are available online at www.cbs.gov.il. The data used to measure TISt are those for the number of Americanschecking into tourist hotels.6 One alternative data source is the number of Americansrecorded entering Israel by the customs service. However, in 2001 the Israeli customsservice changed its definition of “American” from US resident to US passport holder,so we cannot use customs data to measure US tourist flows for the whole of theIntifada period. For the scale variable TEUt we use monthly figures from the datasetpublished by the ITA Office of Tourism and Travel Industries (http://tinet.ita.doc.gov),which reports the number of American tourists departing for Europe.

The hotels data are available from January 1996, and Fig. 1 depicts the ln(TISt / TEUt)series constructed using these data for the period January 1996–June 2006. It can beseen that there is a pronounced dip in ln(TISt / TEUt) after the onset of the Intifada inOctober 2000, but that Israeli tourism slowly recovers toward the end of the sampleperiod, when there is some diminishment in the intensity of the conflict.

5 NY = 1 if either Rosh Hashanah or Yom Kippur occur in the month; otherwise NY = 0.6 One disadvantage of our data is that they include business visitors in the total. However, no appropriatelydisaggregated data are available at a frequency higher than once every two months.

1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

-11.0

-10.5

-10.0

-9.5

-9.0

Fig. 1 The time series for relative Israeli tourist flows. Relative flows are measured as the log of the ratioof tourists in Israel (TIS) to tourists in Europe (TEU)

J Risk Uncertain (2009) 38:245–263 249249

In order to model ln(TISt / TEUt) we need to make some choices about whichconflict intensity indicators (xt

k) to include in our regression equations. In this paper,we make use of a number of indicators of monthly fatalities in the Israeli-Palestinianconflict. Since the fatality data can be disaggregated in many different ways, thereare many possible alternative forms of the regression equation. There are no a priorigrounds for preferring one form over another, because we do not know how much ofthe detail in the fatality data is used when tourists form their opinions. In the face ofsuch model uncertainty, we take the following approach. First of all, we fit arelatively parsimonious regression equation with only two types of disaggregation:by nationality (Israeli and foreign deaths/Palestinian deaths) and by whether thefatality was reported in the US television news. We then fit three less parsimoniousregression equations in which Israeli deaths are disaggregated. Table 1 lists thealternative models that we consider. The correlation between some of thedisaggregated fatality series is too high and our sample too small for modelselection criteria to give us very powerful tests of which of these models best fits thedata. Therefore, we must remain agnostic about the degree of detail used whentourists form their opinions. We discuss all the results of each model, and theirimplications for the formation of tourist opinion. However, our main focus is onthose results common to all models. The invariance of these results to the specificmodel chosen gives us some confidence in their robustness. We begin with adescription of the most parsimonious model.

2.1 The basic model

Our first model (Model 1) disaggregates fatalities only by nationality and by whetherthe fatality was reported. The data on reported fatalities come from the ReutersFactiva database, which records daily transcripts from the main US television newsshows from January 1997 onwards. The three most popular shows are the NBCNightly News, CBS Evening News and ABC World News Tonight. Our main resultsare based on the World News Tonight data. The comparison of our main results with

Table 1 Alternative specifications for the model of relative Israeli tourist flows

Model 1 Model 2 Model 3 Model 4 Model 5

Reported Palestinian deaths (PLKWNT) × × × × ×

Non-reported Palestinian deaths (PLKNRP) ×

Total Palestinian deaths (PLK) × × × ×

Reported Israeli deaths (IFKWNT) × × × × ×

Non-reported Israeli deaths (IFKNRP) ×

Total Israeli deaths (IFK) ×

Israeli deaths west of the Green Line (IFKISR) ×

Israeli deaths in the West Bank (IFKWBG) ×

Israeli deaths: not suicide bombs (IFKOTH) × ×

Israeli deaths: suicide bombs (IFKSUI) ×

Number of suicide bomb attacks (NSU) ×

250 J Risk Uncertain (2009) 38:245–263

the alternatives—using other news shows instead of World News Tonight, oraggregating over different shows—is discussed later. From the Factiva database weconstruct two monthly time series: PLKWNT and IFKWNT. The first measures thenumber of conflict-related Palestinian fatalities reported on the news show, and thesecond measures the number of Israeli and foreign fatalities. The two times seriesexclude repeat coverage of the same fatalities and retrospective coverage of eventsthat took place months earlier, for which separate time series can be created, asdiscussed later.

Corresponding to PLKWNT and IFKWNT are the series PLKNRP and IFKNRP, thetotals for monthly Palestinian and Israeli fatalities not reported on World NewsTonight. These series are constructed from actual fatality data published by thehuman rights organisation B’Tselem (www.btselem.org) and discussed extensivelyin Fischhoff et al. (2007).7 There is very little discrepancy between the B’Tselemdata and other official sources, such as the Israeli Defense Force (IDF) website,www1.idf.il, or the Palestinian Red Crescent Society website, www.palestinercs.org.Indeed, the actual figures correspond very closely to the numbers that would beproduced from a careful reading of a serious newspaper such as the Jerusalem Post,or even the New York Times. If tourists base their decisions on information in thequality media then there will be no cultivation effect from the television media, andthe effect of PLKWNT (IFKWNT) on tourist flows should be no different from that ofPLKNRP (IFKNRP).

A higher number of casualties among Israelis and foreign visitors represents adirect threat to tourists, and for this reason it may reduce tourist numbers. However,this effect may be mitigated if a higher level of casualties evokes more sympathy forIsraelis and prompts more of the “solidarity” tourism promoted by US groups suchas United Jewish Communities. Palestinian casualties represent less of a direct threatto tourists, but they may increase the perceived level of threat, if there is a perceptionof a “cycle of violence” in which higher Palestinian casualties now lead to a greaterthreat to tourists in the immediate future. Jaeger and Paserman (2006) find noevidence for such a cycle, but there might still be a perception that such a cycleexists. (Moreover, even a statistically literate tourist is likely to be more concernedwith Type II errors than with Type I errors when evaluating the null that there is nosuch cycle.) Alternatively, a rise in Palestinian casualties could prompt touristssympathetic to the Palestinian cause to boycott Israel, as promoted by US groupssuch as the Palestine Solidarity Movement. Moreover, if tourists have to avoid areaswhere Palestinian casualties occur (for example, Bethlehem) then their vacation as awhole may be less enjoyable.

The four time series, PLKWNT, IFKWNT, PLKNRP and IFKNRP, are depicted inFig. 2. It can be seen that the series contain a number of sharp spikes—periods ofintense conflict—along with “quiet” periods during which some types of fatality areequal to zero. Before using these series to model ln(TISt /TEUt), we need to make adecision about whether to employ some functional transformation. It may be that anincrease in, say, reported Israeli fatalities from zero per month to 25 per month hasthe same impact on the expected utility of a trip to Israel as an increase from 50 per

7 The B’Tselem data do not cover violence along the Lebanese border, but this area does not attract such alarge proportion of the foreign tourist market as it does of the domestic one.

J Risk Uncertain (2009) 38:245–263 251251

month to 75 per month. In this case, we should include the raw fatality figures(IFKWNT) in the regression equation. However, it may be that an increase from zeroto 25 has more impact than an increase from 50 to 75. (For example, someAmericans might not travel to Israel unless the level of violence is close to zero.Others who do decide to travel, despite moderate levels of violence, might not

1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

50

100

150

200

WNT

Reported Palestinianfatalities (PLK )

1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

50

100

150

200NRP

Unreported Palestinianfatalities (PLK )

1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

25

50

75 Reported Israelifatalities (IFK )WNT

1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

25

50

75

NRP

Unreported Israelifatalities (IFK )

Fig. 2 Conflict variables

Table 2 Correlations across the explanatory variables

ln(1+PLKWNT) ln(1+IFKWNT) ln(1+PLKNRP) ln(1+IFKNRP)

ln(1+PLKWNT) 1.25 (1.12)

ln(1+IFKWNT) 0.70 (0.62) 1.15 (1.01)

ln(1+PLKNRP) 0.42 (0.07) 0.46 (0.12) 1.60 (0.77)

ln(1+IFKNRP) 0.46 (0.24) 0.40 (0.13) 0.75 (0.42) 1.10 (0.80)

ln(1+IFKISR) ln(1+IFKWBG) ln(1+IFKOTH) ln(1+IFKSUI)

ln(1+IFKISR) 1.19 (0.98)

ln(1+IFKWBG) 0.59 (0.33) 1.06 (0.76)

ln(1+IFKOTH) 0.62 (0.39) 0.94 (0.88) 1.09 (0.80)

ln(1+IFKSUI) 0.91 (0.88) 0.56 (0.33) 0.47 (0.19) 1.19 (1.02)

PLKWNT: Reported Palestinian deaths IFKISR: Israeli deaths west of the Green Line

IFKWNT: Reported Israeli deaths IFKWBG: Israeli deaths in West Bank / Gaza

PLKNRP: Non-reported Palestinian deaths IFKSUI: Israeli deaths in suicide bomb attacks

IFKNRP: Non-reported Israeli deaths IFKOTH: Other Israeli deaths

The sample is 1997(1)–2006(6). Standard deviations are reported on the main diagonal

The figures in parentheses are conditional on an intercept shift in October 2000

252 J Risk Uncertain (2009) 38:245–263

change their plans after a marginal increase in conflict intensity.) In this case, weshould use a convex transformation of the raw data; the simplest such transformationis ln(1 + IFKWNT). The results reported below make use of the second approach,using a logarithmic transformation, which produces a better fit to the data in allversions of the model. The results obtained from the first approach, which arebroadly similar, are discussed in Appendix 3.

Table 2 lists correlation coefficients for the four conflict series used in Model 1: ln(1 + PLKWNT), ln(1 + IFKWNT), ln(1 + PLKNRP) and ln(1 + IFKNRP). There is amoderately high correlation between ln(1 + PLKWNT) and ln(1 + IFKWNT), as there isbetween the corresponding non-reported fatality series. However, correlationsbetween the reported and non-reported series are much smaller: the short-runvariation in reported conflict events is not a particularly accurate reflection of what ishappening on the ground. Moreover, all of the correlation coefficients drop markedlywhen we restrict our attention to the high violence period, 2001–2004.

It can be seen from Fig. 2 that the peak of the violence is in March 2002, when therewere 12 suicide bomb attacks (the next highest number is five). March 2002 is highlyatypical of the rest of the period; at such high levels of violence the response oftourism could exhibit non-linearities not seen at other times, and for this reason Model1 includes a dummy variable for this month. It can also be seen that the onset of theIntifada in October 2000 constitutes a definite structural break. At the time, touristsmay well have wondered whether Israel was going to descend into a full-scale civilwar. For this reason, Model 1 includes a dummy for October–November 2000.

Finally, we must recognise that the Israeli-Palestinian conflict is reported in thecontext of a high level of violence across the region, particularly in Iraq. The Iraqiconflict has been reported virtually every day in the American media since theinvasion of February–March 2003, but we do not include a variable for the numberof casualties reported in Iraq. It is unlikely that tourists perceive there to be a highcorrelation between Iraqi violence (trending upwards over 2005–6) and Israeliviolence (trending downwards over 2005–6). Nevertheless, at the onset of theSecond Gulf War in February–March 2003 tourists may have perceived there to besome risk of a retaliatory attack on Israel. For this reason, we include a dummyvariable for these two months.

We will also report results from fitting a minor modification of Model 1. Becausewe are using logarithmic transformations, a model with reported and non-reportedfatalities is not identical to one with reported and total fatalities. In order to checkwhether our results are sensitive to this difference, Model 2 replaces the two non-reported fatality series with one for total Palestinian fatalities, ln(1 + PLK), and onefor total Israeli fatalities, ln(1 + IFK). In the absence of any media effects, thereported fatality series should be statistically insignificant in this model.

2.2 Extended models

One potential criticism of Models 1–2 is that they do not disaggregate Israeli andforeign fatalities. In Models 3–5 we consider alternative forms of disaggregation.First of all, in Model 3, we distinguish between Israeli and foreign fatalitiesoccurring in the West Bank and Gaza (IFKWBG) and those occurring west of theGreen Line (IFKISR). Fatalities in the West Bank and Gaza may represent less of a

J Risk Uncertain (2009) 38:245–263 253253

threat to tourists than fatalities elsewhere, because Palestinian areas are easy to avoid.(Tourist numbers in West Bank towns such as Bethlehem have fallen almost to zerosince the start of the Intifada.) On the other hand, it is possible that a reduction intourism following a rise in Israeli casualties is partly offset by a solidarity effectamong those with high level of political sympathy for the State of Israel. If casualtiesamong Israeli settlers evoke less sympathy than casualties elsewhere, then the settlercasualty coefficient may be greater in absolute value. Therefore, we do not have anya priori view on the relative sizes of the two coefficients.

In Models 4–5 we distinguish between Israeli and foreign fatalities caused bysuicide bomb attacks and fatalities caused by other types of attack. Suicide bombattacks typically target a particular public location, but many other types of attacktarget specific individuals. Possibly, tourists perceive the former to represent agreater threat than the latter. Model 4 replaces IFK with IFKSUI (the number ofIsraeli and Foreign fatalities in suicide bomb attacks) and IFKOTH (the number ofIsraeli and Foreign fatalities in other attacks). Model 5 uses the number of suicidebomb attacks (NSU) instead IFKSUI. If the number of fatalities in a particular attackis perceived by potential tourists to be largely a consequence of chance, then thenumber of fatalities matters less than the number of attacks.

There are two obvious omissions in Models 1–5. First of all, we do notdisaggregate according to both the location of the attack and the type of attack. Ascan be seen in Table 2 above, there is a very high correlation between ln(1 + IFKSUI)and ln(1 + IFKISR), as there is between ln(1 + IFKOTH) and ln(1 + IFKWBG). Inpractice, the majority of suicide bomb attacks are west of the Green Line, and themajority of other attacks on Israelis are in the West Bank and Gaza. If wedisaggregated fatalities any further, then the t ratios would exhibit a large downwardbias, preventing any sensible inference from the model. Secondly, we do notdisaggregate reported fatalities in the same way as actual fatalities. In order to seewhy, consider the following transcript of a typical World News Tonight report fromthe 4th of March 2002, which follows a description of IDF attacks on Palestinians.

That follows one of Israel’s worst weekends since this conflict began. 21Israelis dead in less than 24 h. Palestinian groups have been hitting on allfronts. Suicide attacks against civilians, shooting attacks against Israeli settlersand soldiers. One Palestinian sharpshooter managed to kill 10 Israelis,including seven soldiers at this checkpoint, and then he got away.

The report is specific about the total number of Israeli casualties, but vague aboutthe proportion occurring in the West Bank and Gaza, and about the proportionoccurring in suicide bomb attacks. We cannot use this report to construct a precisedisaggregation of Israeli fatalities, and neither could the potential tourists whooriginally watched it.

3 Regression results

Before discussing our results, we should note that the validity of any inference from theregressions depends on the stationarity of the time series we are using. Appendix 4discusses stationarity tests, and presents evidence that our time series are indeed

254 J Risk Uncertain (2009) 38:245–263

stationary, at least if we allow for a structural break corresponding to the start of the AlAqsa Intifada.

3.1 The fitted model

Our main results are presented in Table 3, which lists the steady-state regressioncoefficients in each of the five models, that is, [η1

k + η2k]/[1 - α1 – α2] for each

conflict intensity variable k, along with the corresponding t ratios. The full set ofcoefficients on the conflict variables is reported in Appendix 5; in these regressions,standard test statistics reveal no sign of residual autocorrelation, autoregressiveheteroskedasticity or non-normality. For each model, we report two sets ofcoefficients. The first set is based on the whole sample (after taking lags, this is1997(2) –2006(6)), and the second is based on the years of particularly highviolence, 2001(1) –2004(12). The shorter sample is chosen for two reasons. First ofall, we would like to check whether any of our results are sensitive to the structuralbreak in late 2000. Secondly, as shown in Table 2 above, the correlations among theexplanatory variables are generally very much lower over the shorter sample periodthan they are over the whole sample: there is a large rise in all of the series at the endof 2000, but after that the monthly correlation coefficients are not very large. Fittinga model to the shorter sample is a way of investigating whether the insignificance ofa variable in the full sample regression is a consequence of multicollinearity.

Before discussing the results of individual models, we should reiterate that theapplication of standard model selection criteria, such as the Schwartz-BayesianCriterion or the Akaike Criterion, does not reveal any one particular model asobviously preferable to the others. This is not surprising, given the high correlationsamong the alternative regressors. Nevertheless, the results of all the models arebroadly consistent with each other.

Fitting Model 1 to the full sample, we obtain significant negative coefficients onln(1 + PLKWNT), ln(1 + IFKWNT) and ln(1 + IFKNRP). The coefficient on reportedPalestinian fatalities is −0.26, a figure that remains more or less constant across all themodel specifications. This implies that a rise in average monthly reported deaths fromzero to the post-Intifada geometric mean (2.5) will eventually reduce tourist numbersto around 72% of their initial value. The corresponding sample means for reported andunreported Israeli deaths are 2.0 and 6.8 respectively; the estimated coefficients (−0.13and −0.17) imply that a simultaneous rise in both variables from zero to theirrespective sample means would eventually reduce tourist numbers to 24% of theirinitial value. However, the coefficients on reported and unreported Israeli fatalities arenot significantly different from each other, which suggests that Israeli fatalities impacttourist numbers regardless of whether they are reported or not. By contrast, theln(1 + PLKNRP) coefficient is very close to and insignificantly different from zero,suggesting that Palestinian fatalities impact tourist numbers only when they arereported. The results for the 2001–2004 sample are broadly similar: none of theindividual regression coefficients is significantly different from the corresponding fullsample coefficient; this is a feature common to all of the models. In general, ourresults are robust to restricting the sample to the shorter period. However, the shortsample ln(1 + IFKNRP) coefficient is insignificantly different from zero, although it isstill insignificantly different from the ln(1 + IFKWNT) coefficient. (See Appendix 5)

J Risk Uncertain (2009) 38:245–263 255255

Tab

le3

Steady-statecoefficientsfrom

differentmodelsof

relativ

eIsraelitouristflow

s

Model

1Mod

el2

Model

3

fullsample

2001–2

004

fullsample

2001–200

4fullsample

2001–200

4

coef.

tcoef.

tcoef.

tCoef.

tcoef.

tcoef.

t

ln(1

+PLKWNT)

−0.26

−3.44

−0.20

−2.77

−0.27

−3.90

−0.16

−2.52

−0.23

−3.74

−0.17

−2.56

ln(1

+IFKWNT)

−0.13

−2.05

−0.17

−2.88

−0.02

−0.23

−0.03

−0.36

0.04

0.64

−0.04

−0.53

ln(1

+PLKNRP)

0.03

0.60

0.09

0.90

ln(1

+IFKNRP)

−0.17

−2.42

−0.10

−1.02

ln(1

+PLK)

0.04

0.89

0.06

0.63

0.05

1.21

0.04

0.42

ln(1

+IFK)

−0.22

−2.86

−0.33

−2.14

ln(1

+IFKISR)

−0.15

−2.45

−0.14

−1.45

ln(1

+IFKWBG)

−0.23

−3.06

−0.16

−1.58

Oct/Nov

2000

−3.02

−3.56

−2.95

−3.71

−2.51

−3.43

Mar

2002

1.41

2.67

1.32

2.67

1.26

2.73

Feb/M

ar20

03−2

.48

−3.64

−1.91

−3.67

−2.22

−3.50

−1.68

−3.54

−2.02

−3.56

−1.76

−3.56

R2

0.93

0.95

0.93

0.96

0.94

0.96

α0.16

0.15

0.16

0.14

0.16

0.15

PLKWNT:ReportedPalestin

iandeaths

PLK:To

talPalestin

iandeaths

IFKWNT:ReportedIsraelideaths

IFK:To

talIsraelideaths

PLKNRP:Non

-reportedPalestin

iandeaths

IFKISR:Israelideaths

westof

theGreen

Line

IFKNRP:Non-reportedIsraelideaths

IFKWBG:Israelideaths

inWestBank/Gaza

256 J Risk Uncertain (2009) 38:245–263

Tab

le3

(contin

ued)

Model

1Mod

el2

Model

3

fullsample

2001–200

4fullsample

2001–200

4fullsample

2001–200

4

coef.

tcoef.

tcoef.

tCoef.

tcoef.

tcoef.

t

Mod

el4

Model

5

fullsample

2001–200

4fullsample

2001–200

4

coef.

tcoef.

tcoef.

tcoef.

t

ln(1

+PLKWNT)

−0.22

−3.53

−0.14

−2.43

−0.22

−3.65

−0.18

−2.90

ln(1

+IFKWNT)

0.03

0.42

−0.03

−0.41

0.07

0.94

−0.01

−0.08

ln(1

+PLK)

0.04

1.03

0.04

0.52

0.02

0.62

0.08

0.80

ln(1

+IFKOTH)

−0.25

−3.92

−0.20

−2.46

−0.22

−3.62

−0.12

−1.45

ln(1

+IFKSUI)

−0.11

−1.93

−0.15

−2.33

NSU

−0.13

−2.21

−0.12

−1.77

Oct/Nov

2000

−2.39

−3.48

−2.15

−3.33

Mar

2002

1.21

2.76

2.04

3.68

Feb/M

ar20

03−1

.97

−3.49

−1.66

−3.73

−2.19

−4.07

−1.86

−3.95

R2

0.94

0.96

0.94

0.96

α0.16

0.14

0.16

0.15

PLKWNT:ReportedPalestin

iandeaths

IFKSUI:Israelideaths

insuicidebombattacks

IFKWNT:ReportedIsraelideaths

IFKOTH:Other

Israelideaths

PLK:To

talPalestin

iandeaths

NSU

:Num

berof

suicidebo

mbattacks

The

dependentvariable

isthelogof

theratio

oftouristsin

Israel

totouristsin

Europe:

ln(TIS/TEU)

Eachregression

also

includes

month,PassoverandNew

Yeareffects.The

fullsampleis1997(2)−20

06(6)

J Risk Uncertain (2009) 38:245–263 257257

The results of Model 2 are consistent with those of Model 1, in that the twosignificant variables are ln(1 + PLKWNT) and ln(1 + IFK): the figures that matter arereported Palestinian fatalities and total Israeli fatalities, regardless of whether these arereported or not. Although the information criteria for Model 2 are very marginallybetter, there is no strong evidence about which functional form is to be preferred.

In Model 3, which disaggregates IFK by location, there are significant negativecoefficients on both fatalities in the West Bank and Gaza and fatalities west of theGreen Line. They are significantly different from neither each other nor thecoefficient on total fatalities in Model 2. This suggests that the location of deathsdoes not play an important role in conditioning tourist perceptions. The coefficientson ln(1 + PLKWNT), ln(1 + PLKNRP) and ln(1 + IFKWNT) are very similar to those inModel 2, giving us some confidence in the robustness of our previous commentsabout these variables.

In Model 4, which disaggregates IFK by the manner of attack, there aresignificant negative coefficients on both deaths in suicide bomb attacks and deaths inother attacks. The difference between the two coefficients is marginally significant inthe full sample version of the model, and in fact the coefficient on ln(1 + IFKOTH) isthe larger of the two. This seems to contradict the assumption that suicide attackcasualties have at least as much impact on tourist numbers as other types of casualty.However, one explanation for the relatively small coefficient on ln(1 + IFKSUI) isthat it is an imperfect proxy for the number of suicide bomb attacks. In the fullsample version of Model 5, which replaces ln(1 + IFKSUI) with the number ofattacks, NSU, the information criteria are very slightly better than in Model 4,although no statistical significance can be attached to this difference. The NSUcoefficient (−0.13) implies that a single suicide bomb attack every month reducestourist numbers to around 88% of the level they would otherwise reach. Again, thecoefficients on ln(1 + PLKWNT), ln(1 + PLKNRP) and ln(1 + IFKWNT) are very similarto those in Model 2.

The dummy variables are statistically significant in all forms of the model. Thedummy for March 2002 is positive, indicating that the atypically high level ofviolence in that month had a less than proportional effect on tourist numbers. Thedummy for October–November 2000 is negative, indicating that the onset of theIntifada period had a temporary negative impact on tourist numbers in addition tothe actual and reported casualties at this time. The dummy for February–March 2003is also negative, reflecting the impact of the Gulf War on tourist numbers in Israel. Infact, this is the largest single effect of all.8

Given the absence of a significant media effect in the impact of Israeli casualties,there does not seem to be a strong case for arguing that non-rational-choicecultivation effects are relevant to television reporting of the Israeli-Palestinianconflict. Still, reports of Palestinian casualties do have a large and significant impact;altogether, this suggests that viewers are generally rational and well informed, butrely on journalists’ judgement to indicate which events involving Palestiniancasualties are politically significant.

8 The dummy variable for March 2002 is omitted from the model when the 2001-004 sample is used,because it is collinear with the seasonal effects.

258 J Risk Uncertain (2009) 38:245–263

3.2 Dynamics and model simulation

Table 3 does not provide any information about the speed with which changes in thedifferent conflict intensity series impact on tourist numbers. This is implicit in theindividual regression coefficients in Appendix 5, which we will not discuss in detailhere. However, it is worth pointing out that the half-life of most shocks toln(TISt / TEUt) implicit in these coefficients is very short. For example, in Model 5, ashock to ln(TISt / TEUt) arising from a temporary (one-month) rise in ln(1 + PLKWNT)or ln(1 + IFKOTH) has a half-life of three months; the corresponding figure for a rise inNSU is two months. The half-lives implicit in other models are similarly short.Tourists appear to be responding to changes in conflict intensity very quickly. Thisimplies that tourists are prepared to cancel trips to Israel at short notice, if theperceived risk of travelling is too great. Such a response is consistent with someexisting surveys of American tourists. For example, analysis of survey data byFischhoff et al. (2004) indicates a consensus that Israel is the most risky place to visit.Answers to questions about levels of risk are broadly consistent with answers aboutconditions under which a trip would be cancelled, and 85% of respondents wouldcancel a trip to Israel if the level of risk were perceived to be excessive (and if thecancellation were costless).

Another way of exploring the dynamics of the model is to simulate counterfactualtime series for ln(TIS/TEU) under the assumption that (i) there was no reportedviolence over 1997–2006 or (ii) there was no actual or reported violence over thisperiod. Having done so, it is possible to calculate the ratio of the hypothetical levelof TIS to its actual level in both cases. Figure 3 plots the two ratios, based on the fullsample coefficients in Model 5. At the peak of the violence in 2002, tourist numberswould have been about twice as high in the absence of television reports ofPalestinian casualties.9 In the absence of any violence, actual or reported, touristnumbers would have been about five times as high during this period. More recentlythe intensity of the conflict has been relatively low, and the ratio for no actual orreported violence has fallen to around 1.4.

3.3 Alternative model specifications10

The interpretation of the results above depends on the assumption that the violencein Israel has no effect on American tourism in Europe. Europe has experienced someviolence over the sample period, and this might be correlated with the intensity ofthe Arab-Israeli conflict, although the level of European violence is tiny incomparison. One way of testing the robustness of our conclusions is to replace

9 This figure is significant in the light of the Gaza conflict of January 2009, which is outside of our sampleperiod. The IDF prevented foreign journalists from entering Gaza, but Palestinian casualties were widelyreported anyway.10 We also explored the effect of replacing our measure of the number of American tourists visiting Israelwith the number of American tourist person nights (indicating the number of tourists in Israel times thenumber of nights they stayed); this time series is also available on the CBS website. The correlationbetween the two alternative dependent variables is 0.98. The results using the alternative dependentvariable are very similar to the ones reported here; the main difference is in the ln(1 + PLKWNT)coefficients, which are about 3–4 percentage points smaller, but still significant at the 5% level.

J Risk Uncertain (2009) 38:245–263 259259

European tourist data with data from another part of the world. For this reason, welook at the effect on our results of using US tourist departures for Oceania (TOC) asa scale variable in place of TEU. There is a high degree of correlation between ln(TIS/TOC) and ln(TIS/TEU), although of course the seasonal components of the twoseries are rather different. (The correlation coefficient for the two deseasonalizedseries is 0.98.) It is therefore not surprising that the conflict variables in Models 1–5 arealso jointly significant in corresponding models of ln(TIS/TOC). However, the conflictvariables are never remotely significant when we fit a model of ln(TEU/TOC), that is,departures to Europe relative to departures to Oceania. From this we conclude that ourchosen conflict series really do reflect the perceived risk of travelling to Israel.

We also test whether the addition of reported violence series from other newschannels makes a significant difference to our results. There are some differences inthe reporting of the violence across the three major channels. In particular, WorldNews Tonight mentions Palestinian casualties more frequently than the other two.Over the period 1997(1)–2006(6), World News Tonight mentioned a specific numberof Palestinian fatalities in 39.5% of the months; the corresponding figure for theCBS Evening News is only 30.7%. Specific numbers of Palestinian deaths werehardly ever mentioned on the NBC Nightly News. When we add to Model 2the number of fatalities reported on the CBS Evening News (that is, the variablesln(1 + PLKCBS) and ln(1 + IFKCBS)), these additional variables are jointly andindividually insignificant, while ln(1 + PLKWNT) remains significant. The same istrue of Models 3–5. This may reflect the viewing habits of potential tourists: perhapsthey take a particular interest in the channel that provides the most detailedinformation about the conflict (ABC), or in local channels taking news feed fromABC. The theoretical model of Gentzkow and Shapiro (2006) shows that a rationalBayesian consumer will give more credibility to a media source when it confirms herprior expectations. If potential visitors to Israel have a prior belief that it is adangerous place, then they may well pay particular attention to media outlets thatgive the highest profile to the violence there.

1997 1998 1999 2000 2001 2002 2003 2004 2005 20061.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

5.0

No television reports of Palestinian casualties No actual or reported violence

Fig. 3 Ratios of hypothetical tourist numbers to actual tourist numbers. This figure is based on asimulation using the fitted coefficients in Model 5

260 J Risk Uncertain (2009) 38:245–263

Finally, we test whether retrospective reports of the violence have an impact ontourist figures. Sometimes the same fatalities are reported on more than oneoccasion, or a report will mention the total number of fatalities over the last fewmonths or years. Such reports are relatively infrequent: in the case of World NewsTonight, retrospective Palestinian deaths are reported in only 11.4% of the months,and retrospective Israeli and foreign deaths in only 8.8%. Adding these retrospectivefatality series to Models 1–5 does not produce significant coefficients on the extravariables, and the size and significance of the other variables is largely unaffected.Retrospectively reported fatalities constitute a broad, somewhat heterogeneouscategory, and it is possible that greater disaggregation would produce somesignificant coefficients. However, with so few non-zero observations in our sample,it would be difficult to disaggregate further without a serious risk of spurious resultsfrom data-mining.

4 Conclusion

In this paper we have analyzed the impact of monthly variations in the intensity ofthe Israeli-Palestinian conflict on tourist flows from the United States to Israel. Ourfocus has been on the distinction between the actual variation in conflict intensityand the variation implicit in the coverage of the conflict by the US television news.The correlation between actual and reported conflict intensity is low enough for us tobe able to determine whether tourists respond to actual events or to events asreported on TV.

We find that reported conflict intensity does have some effect on tourist flows.This raises questions about what sort of model of choice under uncertainty isappropriate for the tourists whose behavior underlies our data. If we wish to retain aneoclassical rational choice framework, then we can postulate a model withimperfect information. If alternative sources of information are costly, then touristsmay infer the current level of risk in travelling to Israel from the television news.(This approach is more plausible if an assessment of the true level of risk requiresqualitative as well as quantitative information about the conflict. Tourists might inferthat the decision to report a particular conflict event reflects an assessment of itssignificance based on specialist knowledge.) However, there are alternativeexplanations outside this rational choice framework. More intense reporting maycause the conflict to be brought to mind more frequently; this may increase in thesubjective probability of conflict events (Tversky and Kahneman 1973), reducingtourist numbers. Moreover, emotive visual images of the conflict may cause“probability neglect”, and a behavioural response disproportionate to the knownrisks (Sunstein 2003).

We do not test any of these hypotheses directly. However, disaggregation of theconflict data into Israeli and Palestinian casualties generates results that are easier toreconcile with an imperfect information hypothesis than with the alternatives. Wefind that tourist flows respond to actual Israeli casualties and reported Palestiniancasualties. Conditional on these effects, neither reports of Israeli casualties norunreported Palestinian casualties have any significant impact on tourist numbers.Given that the risks to tourists resemble those faced by Israelis more closely than

J Risk Uncertain (2009) 38:245–263 261261

those faced by Palestinians, this asymmetry cannot easily be explained by a Tversky-Kahneman effect or by probability neglect. However, it could be explained by aninformational asymmetry, if tourists have ready access to independent sources ofinformation about Israeli casualties (for example, friends and family in Israel), butless information about Palestinian casualties. In this regard, our results are dissimilarfrom many of those in the criminology literature, which provide some evidence forTversky-Kahneman effects and probability neglect. One possible reason for thisdifference is that the criminology results focus on perceptions about the localneighborhood, not about someone else’s neighborhood. In our case, involvingperceptions about a foreign country, it would not be surprising for informationeffects to dominate other effects.

Our estimates suggest that at the most recent peak in conflict intensity in 2002,tourist flows would have been twice as high in the absence of any reports ofPalestinian casualties, and five times as high in the absence of any actual violence.These results reinforce previous studies of the wider macroeconomic impact of theIntifada, for example Fielding (2003) and Eckstein and Tsiddon (2004). We canexpect even a partial reduction in violent conflict in Israel to increase tourismrevenue, and as returning tourists boost local incomes they may create stakeholdersin the continuation of the peace process, changing the incentives that driveindividual decisions about whether to take up arms (Azam 2005). This could begrounds for optimism regarding a gradual resolution of the conflict. However, theimportance of international tourism as a channel through which violence affectsaggregate demand reinforces the need to consider trans-national externalities whenevaluating public policy (Sandler 2005).

References

Azam, J.-P. (2005). Suicide-bombing as inter-generational investment. Public Choice, 122, 177–198.Burger, H. & Sturm, J-E. (2005). Assessing the Cost of Terrorism. Media Tenor Quarterly 1.Chiricos, T., Eschholz, S., & Gertz, M. (1997). Crime, news and fear of crime. Social Problems, 44, 342–

357.Drakos, K. (2004). Terrorism-induced structural shifts in financial risk. European Journal of Political

Economy, 20, 435–446.Drakos, K., & Kutan, A. (2003). Regional effects of terrorism on tourism in three Mediterranean countries.

Journal of Conflict Resolution, 47, 621–641.Eckstein, Z., & Tsiddon, D. (2004). Macroeconomic consequences of terror: theory and the case of Israel.

Journal of Monetary Economics, 51, 971–1002.Enders, W., & Sandler, T. (1991). Causality between trans-national terrorism and tourism: the case of

Spain. Terrorism, 14, 49–58.Enders, W., Sandler, T., & Parise, G. (1992). An econometric analysis of the impact of terrorism on

tourism. Kyklos, 45, 531–554.Fielding, D. (2003). Counting the costs of the Intifada: consumption, saving and political instability in

Israel. Public Choice, 116, 297–312.Fischhoff, B., Gonzalez, R., Small, D., & Lerner, J. (2003). Judged terror risk and proximity to the World

Trade Center. Journal of Risk and Uncertainty, 26, 137–151.Fischhoff, B., de Bruin, W. B., Perrin, W., & Downs, J. (2004). Travel risks in a time of terror: judgments

and choices. Risk Analysis, 24(5), 1301–1309.Fischhoff, B., Atran, S., & Fischhoff, N. (2007). Counting casualties: a framework for respectful, useful

records. Journal of Risk and Uncertainty, 34, 1–19.

262 J Risk Uncertain (2009) 38:245–263

Fleischer, A., & Buccola, S. (2002). War, terror, and the tourism market in Israel. Applied Economics, 34,1335–1343.

Frey, B., Luechinger, S., & Stutzer, A. (2007). Calculating tragedy: assessing the costs of terrorism.Journal of Economic Surveys, 21(1), 1–24.

Gentzkow, M., & Shapiro, J. (2006). Media bias and reputation. Journal of Political Economy, 114, 280–316.

Gerbner, G., & Gross, L. (1976). Living with television: the violence profile. Journal of Communication,26, 173–199.

Gross, K., & Aday, S. (2003). The scary world in your living room and neighborhood: using localbroadcast news, neighborhood crime rates, and personal experience to test agenda setting andcultivation. Journal of Communication, 53, 411–426.

Jaeger, D. A., & Paserman, M. D. (2006). Israel, the Palestinian factions, and the cycle of violence.American Economic Review, 96(2), 45–49.

Kalaitzandonakes, N., Marks, L., & Vickner, S. (2004). Media coverage of biotech foods and influence onconsumer choice. American Journal of Agricultural Economics, 86, 1238–1246.

Mazzocchi, M. (2004). Modelling Consumer Reaction to Multiple Food Scares. Paper presented to the78th annual meeting of the Agricultural Economics Society, Imperial College, London, England.

Piggott, N., & Marsh, T. (2004). Does food safety information impact U.S. meat demand? AmericanJournal of Agricultural Economics, 86, 154–174.

Romer, D., Jamieson, K., & Aday, S. (2003). Television news and the cultivation of fear of crime. Journalof Communication, 53, 88–104.

Sandler, T. (2005). Collective versus unilateral responses to terrorism. Public Choice, 124, 75–93.Sloboda, B. (2003). Assessing the effects of terrorism on tourism by use of time series methods. Tourism

Economics, 9, 179–190.Sunstein, C. (2003). Terrorism and probability neglect. Journal of Risk and Uncertainty, 26, 121–136.Tversky, A., & Kahneman, D. (1973). Availability: a heuristic for judging frequency and probability.

Cognitive Psychology, 5, 207–232.Verbeke, W., & Ward, R. (2001). A fresh meat almost ideal demand system incorporating negative TV

press and advertising impact. Agricultural Economics, 25, 359–374.Viscusi, W., & Zeckhauser, R. (2003). Sacrificing civil liberties to reduce terrorism risks. Journal of Risk

and Uncertainty, 26, 99–120.

J Risk Uncertain (2009) 38:245–263 263263