Context and data

The human smuggling market in North African port cities is competitive [5]. When smugglers began using inflatable craft that are easily replaced, entry barriers were further reduced. The price for crossing the Mediterranean fell substantially. From the initial surge in 2015 to 2016, prices for the sea leg of the journey were 200 to 250 USD. In June 2017, prices were as low as 60 USD. Following the intervention we study, qualitative assessments suggest prices rebounded [17].

By 2014 when smuggling across the Mediterranean was expanding rapidly, there was little government infrastructure to curtail it. The Libyan Navy and Coast Guard’s vessels were among the targets of the 2011 NATO intervention in Libya. Militias and other armed groups controlled ports and the surrounding water. In the context of a thriving smuggling industry, militias smuggled migrants themselves, charged tolls to other smugglers for passage out of controlled waters, and operated detention centers for migrants who were captured by formal and informal authorities [18].

The journey from departure in Libya or Tunisia to arrival in Italy takes from one to seven days. In a seaworthy boat, the trip takes less than a day [19]. However a journey in a wooden fishing boat or rubber dinghy can take several days. The most detailed estimates of trip duration are from 2013, before there were NGO rescue boats; they suggest smuggled migrants’ trips were 2 to 6 days [20]. This variation is mostly due to time at sea before rescue [20]. During the period we study, NGO boats are rescuing a substantial portion of migrants, and migrants are traveling in their rubber dinghies shorter distances than in the 2013-14 period [21]. Taken together, the qualitative evidence suggests that most journeys take two to three days with some travel taking seven days or more.

To account for this uncertainty in how long cross-sea trips take, we use a measure of sea and political conditions that leverages variation from the seven days prior to arrival. This choice relies on reports from extensive interviewing of migrants upon their arrival in Italy and some smugglers in Libya. Typically, when a migrant in Libya decides to go to Europe, the first step is to contact an intermediary. The intermediary takes the migrant to a holding place, where migrants wait until sea conditions are favorable and there are enough people to fill a boat [22]. Once there are sufficient numbers (80-100), the migrants board a rubber dinghy, and one migrant is charged with piloting the boat for 4-8 hours straight into international waters. Then, the migrants make an SOS call with an included radio [22] and wait to be rescued. Once on a rescue boat, they make their way to a close, designated port. Our arrival data is a composition of estimates from the United Nations High Commissioner for Refugees (UNHCR) and registration from the Italian Ministry of the Interior. These numbers are collected the day of arrival, or within the next 2 days, depending on the port and the timing of UNHCR access to arriving migrants [23]. Econometrically, using variation from a longer lag introduces some noise since the additional lags are only relevant to a (potentially small) subset of migrants. This would cause attenuation bias. If true, this type of bias would lead us to underestimate the true effect (i.e., what we would estimate if we had data on the duration of each individual’s journey).

We study migrant movement into Italy from January 2016 to April 2018. Our data tracks the arrival of 307,056 individuals and comes from UNHCR, through the Operational Portal—Mediterranean. The arrivals were largely processed at the Lampedusa Reception Center, located in the Italian Pelagie Islands. Remaining counts were compiled at other facilities along the Italian coast. These records do not distinguish origins, yielding a daily time series of migration. For more information on origins, we use routes based on qualitative research by the International Organization for Migration (IOM). There are nine primary sea routes from ports in Libya and Tunisia to Lampedusa (see Fig 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1.

Region of Study migration passages and smuggling hubs (a) and area used for calculating sea conditions (b). Map of smuggling routes is recreated based on information from the IOM [24]. Sea condition zone is authors’ projection. Data on sea conditions is from the Copernicus marine environment monitoring service. Boundary files are from GADM [25].

https://doi.org/10.1371/journal.pone.0236646.g001

Migrants’ transit may be influenced by political unrest. To capture variation in political unrest near areas where migrants depart, we construct 25-kilometer buffers around these nine ports. We identify all riots within these buffers collected in the Armed Conflict Location & Event Data Project (ACLED) [26]. We select riots because they occur in urban areas and are more likely to affect labor migrants than organized violence. We expect migrants to respond to riots because they make an area less safe and may be indicative of more systematic violence to come. We anticipate that migrants should be less affected by violence targeting civilians and rebel violence, as migrants may avoid places that are targets of a government crackdown or have ongoing rebel violence. During our sample period, nearly all events recorded within our buffers were georeferenced to the town (spatial precision level 1), giving us more confidence in the assignment of riot events and other instances of violence to ports.

Because sea travel takes up to one week from departure to registration of arrival, we count the number of riot events in the prior week. For migrants with long journeys (6 to 7 days), this implies that there is noise in this measure (since future riots are difficult to anticipate and likely do not effect smuggling). This noise likely causes attenuation in the estimated relationship with riots. We investigate this in the Robustness section below. We follow a similar strategy for investigating violence against civilians and rebel-government violence (S2 Table). We find no consistent evidence that these other forms of violence influence migration as much as riots, though the estimated effects of civilian harm are imprecise. For this reason, our main analysis focuses on riot events.

Migration may also be influenced by sea conditions. To account for this, we gather daily sea wave height data from the EU’s Copernicus marine environment monitoring service [27]. Wave heights were recorded at 0500 local time (approximately sunrise). We calculate wave conditions within a geographic zone derived from the IOM sea routes, enabling us to capture wave heights along and around these passageways (see Fig 1). Similar to our riot data, we construct an average of wave heights from the prior week to account for the lag between departure conditions and arrivals in Italy. For migrants with short journeys (1 to 2 days), this implies the link between migration and wave height will be attenuated (since migration is unlikely to be influenced by historical wave height conditional on current wave conditions). We assess these dynamics below in the Robustness section.

We supplement arrival records with data from the IOM’s Missing Migrants Project [28]. These data are clustered by event and provide estimates of the number of migrants killed or missing. These records include migrant deaths from boats capsizing and more regular occurrences like deaths at sea due to dehydration, illness, or injury. Coordinates are associated with the death if known and otherwise, the place of recovery. The Missing Migrants Project gathers data from official sources like medical examiners and coast guard. They also draw on data from media sources and NGOs that help migrants and their families. The IOM platform is the most comprehensive available and is used by governmental and non-governmental actors for tracking casualties. We collapse these records into a daily time series and, as a robustness check, combine our information on arrivals with this data to calculate a daily total flow (of those who did and did not arrive in Italy).

Time series analysis: Political and environmental risks

We anticipate a tradeoff between political risks (e.g., exposure to riots) and environmental conditions which increase the probability of death at sea (dangerous wave conditions). We exploit the daily time series data above to study this question. Our main specification incorporates political unrest (riots) and sea conditions. We begin by estimating Eq 1: (1) Where arrivals_i is the daily total of arrivals, riots_i indicates the intensity of riots in the prior week, and wave_height_i captures average sea conditions in the prior week. To ease interpretation as elasticities (percentages), we evaluate the log of these three measures. For consistency, we add one to arrivals_i and riots_i before calculating the logarithm. Supplemental models incorporate month_i and day_i. month_i indicates month of year fixed effects, and day_i represents day of week fixed effects. We leverage month_i to capture any seasonal trends in flight, violence, or sea conditions. day_i helps us to account for any systematic variation in arrivals that might be driven by intake operations (e.g., the facility may staff fewer officers on Sundays). In our main specification, we produce heteroskedasticity robust standard errors. To account for the possibility that flows are serially correlated over time, we also produce Driscoll-Kraay temporal autocorrelation robust standard errors.

Migrants may leave Libya in response to riots; however, it is less clear whether the smugglers or the migrants are the decision makers with respect to rough seas. Qualitative evidence suggests that it is the smugglers who decide when the boats leave. Frontex, the European Border Agency, explains that, “Many people who are coming from Africa have never seen the sea in their lives before,” [21, par. 18] which makes it unlikely the migrants are accurately judging the seas. Based on primary data collection in Libya, UNHCR contract researchers concluded that migrants wait “until the smuggler decides that the conditions are right for travel” [22, p. 144]. Economic theory provides a rationale for why smugglers have long term incentives to get their migrants into Europe. Specifically, Mediterranean human smuggling is built on repeated interactions [15]. With repeated interactions smugglers and intermediaries build reputations similar to firms with relational contracts in supply chains [29, 30]. This intuition is similar to Campana 2018 [5], which studies networks of smugglers in Libya. In part, the agency of the smugglers is the impetus for examining the Italian intervention.

Event study: Italy’s 2017 Libya intervention

We leverage an Italian policy intervention backed by EU funding to examine the role of Libyan militias in human smuggling across the Mediterranean. The Italian policy is aimed at disrupting the human smuggling business in Libya, while doing nothing explicitly to address the flow of migrants seeking to travel across the Mediterranean. An important discrete change in July 2017 allows us to examine the role of the militias involved in human smuggling: Italy coordinated with militias engaged in human smuggling and incorporated them into an “anti-smuggling” coast guard force.

At the end of June 2017, Italy facilitated a deal with militias in the key port of Sabratha, west of the capital, Tripoli. The deal turned militias that facilitated smuggling into an “anti-smuggling” coast guard [18, 31]. The deal also gave the militias access to funds and equipment for coast guard activities. Italy brokered a similar agreement in Southern Libya that built a coalition among Southern tribes and the UN-backed government in Libya and gave them access to EU funds [32]. While the Italian Foreign Minister acknowledged Italy “spoke to everyone,” the Italian government has not acknowledged its part in the Sabratha deal [33, 34]. By the beginning of July, the Libyan Coast Guard had refurbished boats, and joint operations between the EU-bolstered Libyan Coast Guard and Italian Navy ensured that 60% of intercepted migrant boats were being returned to Libya [18, 34]. At the same time, Italy announced a draft code of conduct that placed restrictions on NGO rescue ships [18]. Subsequent investigations suggest that the Libyan government used the cash influx from the EU to pay off or co-opt militias who controlled ports, acted as coast guard, and facilitated migrant smuggling and detention [17]. The Italian intervention is ideal for estimation because migrants could not have anticipated the policy. The median migrant spends 6 to 12 months in Libya (based on the IOM Displacement Tracking Matrix) and documentary evidence has Libya turning around most boat within 2 weeks.

Targeting militias in the Central Mediterranean may be particularly important. Libya to Italy sea crossings are distinct among irregular transit because migrants typically use a smuggler [1]. Migrants crossing land borders often make the crossing on their own. Particularly, circular migrants who cross regularly know the way and know how to avoid border enforcement. Those who are less experienced or wish to reduce risk choose to use a smuggler [1, 6]. Because crossing the Central Mediterranean is so difficult, almost all irregular migrants engage a smuggler, which presents an opportunity for Libyan militias to profit.

This intervention differed from other interventions pursued by the EU and Italy. Beyond changing the cost of migrating or the expected utility of moving to Europe, it altered the supply side of the smuggling market. Militias in Sabratha stopped smuggling migrants and began operating as a coast guard force. The intervention, thus, operated through two channels. Like other interventions it decreased the probability a migrant arrived in Europe. The intervention also reduced the number of illicit organizations engaged in smuggling. The next component of our analysis focuses on the impact of the Italian intervention on migrant arrivals.

Italy received orders of magnitude more migrants than any other country in Europe during the period studied. Flows into other European countries therefore cannot serve as reasonable counterfactual trends. Because we lack a viable control group, we implement a synthetic prediction model (using high dimensional time fixed effects) to construct a counterfactual flow of migrants for 2017 based on 2016 migration patterns (the only complete migration season we observe before the intervention). We use calendar-week specific parameters to capture the seasonal trend in migration from 2016. Using this approach to estimate the counterfactual trend enables us to be agnostic about functional form of the seasonal trend. Rather than relying on a linear or higher order polynomial approximation of the trend, we estimate 52 separate slopes. This high dimensional estimation captures the underlying seasonal trend more precisely. We use the 2016 seasonal trend to construct a high resolution counterfactual trend for 2017. We then remove the constructed trend from the realized trend in 2017. This residual measure, our dependent variable in the event study, captures the variation in 2017 that is unexplained by high dimensional seasonal variation in the prior year. The timing of the intervention sets up an ideal shock from which we can estimate a regression model analogous to a difference-in-differences approach using Eq 2: (2) Where is the weekly mean of arrivals (deseasonalized) and Post_Intervention_i is the quantity of interest, the change in migrant flows into Italy following the intervention. The coefficient on Post_Intervention_i captures the deviation after the intervention from the seasonal migration trend (when no comparable intervention took place). This deviation is equivalent to the level difference in a difference-in-differences research design. Following our most conservative time series specification, we produce Driscoll-Kraay temporal autocorrelation robust standard errors (with a bandwidth of four weeks).

Political and environmental risks

Riots in North Africa are positively correlated with migrant arrivals in Italy, and higher wave heights decrease these arrivals. Table 1 reports the results from our main specifications in Eq 1. Summary statistics are presented in S1 Table. Our interpretation of the main effects is consistent with an elasticity between our independent and dependent variables. In Column 1, we present the baseline correlation between arrivals and political unrest in the prior week. We find evidence consistent with our expectation, a strong positive relationship. A 10 percent increase in riot intensity corresponds to a 4.89 percent increase in arrivals. In Column 2, we introduce our measure of travel risk. If our argument is correct, we would expect a negative correlation between wave height and arrival intensity. Here, the evidence is even sharper. A 10 percent increase in wave heights leads to an approximately 27 percent decrease in arrivals. It is possible that sea conditions are linked to riot activity. Jointly estimating these relationships in Column 3 allows us to partial out any residual pair-wise correlation. Our estimates are stable. In Column 4, we introduce month_i fixed effects to the model, which helps to account for seasonal variation in migration, violence, and sea conditions. Our estimate of the impact of sea risks is marginally attenuated, but remains more than five times larger in magnitude than political unrest. It is possible that staffing schedules or intake regulations lead to higher levels of registered arrivals on certain days of the week. To account for this, we add day_i fixed effects to Column 5. Our results are unaffected. It is also possible that the intensity of arrivals, and other conditions, may be strongly correlated over time. If so, our baseline approach to standard errors might overstate the precision of our estimates. Thus, we calculate Driscoll-Kraay temporal autocorrelation robust standard errors in Column 6 using a 14 day window, which we believe is conservative. The precision of our estimates decreases, but the parameters of interest, riot intensity and wave height, remain correlated with migration patterns. In S3 Table we demonstrate that our inferences are consistent for 21, 28, and 56 day bandwidths. We also present results for clustering by month. All results are substantively consistent and precise at conventional levels.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Riots, sea conditions, and migrant flows to Italy.

https://doi.org/10.1371/journal.pone.0236646.t001

Another possibility is that the patterns we observe are driven by survivor bias: during poor sea conditions, fewer migrants survive and are able to make landfall. The negative correlation between risky sea conditions and arrivals, therefore, may be biased in the direction of our argument, overstating the elasticity. If, instead of arrivals, we were able to identify the total flow of migrants, it would address this concern. That would involve combining arrivals with information on the flow of migrants that die or disappear (presumed dead) on their sea passage. To investigate this, we rely on the IOM’s data on missing migrants. We collapse this data into a comparable daily time series and sum arrivals and deaths from this source into a total migrant flow. We replicate our most conservative model specification in Column 7 with this new outcome. The sea conditions coefficient decreases in magnitude, from 2.54 to 2.26. The riots coefficient also declines, which is consistent in relative scale. This suggests that our benchmark results are marginally influenced by survivor bias.

Our data on migrant casualties allow us to verify the mechanism of our argument, that poor sea conditions increase the risk of death. We indeed find higher waves correspond to a higher death rate. To evaluate this mechanism, we calculate a daily death rate. Because this rate is unobserved if migrants neither arrive nor die on a given day, our design is now an interrupted daily time series. We begin by visualizing this relationship in Fig 2. The two trends, death rates and wave heights, covary closely. As waves reach dangerous levels (greater than 1.5 meters for most rubber craft), death rates increase substantially. We next present statistical evidence of the correlation between the death rate and sea conditions in S4 Table, following the main specification in Column 6 of Table 1. We focus first on the correlation between the current death rate and current wave height (Column 1). We explore the effect of the first lag in Column 2. This allows us to account for the possibility that there is a lag between reported death events and the precipitating weather event. The effect of the first lag relative to the day-of measure is large and statistically distinct. In Column 3, we attempt to capture the effect of an extended set of wave height lags; we find compelling evidence that the first lag explains variation in death rates, though the joint significance of the second and third lags is also statistically different from zero (p = 0.0280). This is reflected in Column 4, where we use a three day moving average of wave height and find robust and substantively large effects. For consistency with Table 1, we also demonstrate that the moving average of the week prior to the death event is highly significant and large in magnitude. These estimates confirm that there is a strong positive relationship between sea risk and the percentage of migrants who die or are lost at sea.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Trends in death rates and Sea conditions during sample period.

Daily death rates are calculated using IOM data on migrant deaths and disappearances in the Mediterranean and UNHCR intake data for the Italian coast. Wave heights are drawn from the Copernicus platform.

https://doi.org/10.1371/journal.pone.0236646.g002

Targeted intervention

Fig 3 displays the results of the Italian intervention study. It shows that the intervention reduced migrant flows. Plot (a) traces weekly arrivals in Italy during 2017 (solid line) and the predicted weekly arrivals in Italy using 2016 arrivals to predict 2017 arrivals (dashed line). Prior to the policy intervention, counterfactual predictions based on 2016 weekly arrivals explain 35% of the variation in arrivals in 2017. Migrants flows are highly seasonal, and one year’s arrivals are a good predictor of the next. Italy’s intervention occurs in mid-July, and the drop in arrivals is visible. Following the intervention, the 2016 migrant flows explain only 2% of the variation in arrivals. Plot (b) displays estimated coefficients from the deseasonalized estimation approach in Eq 2. The first point plotted is our preferred specification which marks the intervention at Week 28, the earliest week that we could document the Libyan Coast Guard turning around most boats. We estimate the the policy intervention reduced migrant arrivals by 343 people per day for the remainder of 2017. As robustness checks, we include estimates from three other specifications. The second estimate includes controls for the average number of daily arrivals in the previous week. The third estimate uses an earlier week for the intervention, Week 26 which follows the weekend when Italy is reported to have negotiated with armed groups in the port city of Sabratha. The final estimate uses the earlier treatment week and includes lagged arrivals. Regardless of the specification, the estimates are stable and statistically significant at conventional levels.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Impact of Italian intervention.

(a) Weekly time series of actual (solid) and predicted (dashed) arrivals. Blue vertical bar indicates the week of the intervention. As expected, fit statistics suggest the predicted variation maps well to the pre-treatment period and poorly to the post-treatment period. (b) Regression-based estimates of the reduction in migrant flows following the intervention, relative the predicted flows from the prior year. Reduction is equivalent to 0.6 standard deviations. Outcome mean is -387.1; standard deviation is 595.2. 95% confidence intervals shown. (c) Monthly trends in migrant flows to three receiving countries. Notice large reduction corresponds to minor displacement through Spain and Greece. (d) Monthly trends in arrivals and deaths. Note strong correspondence between two trends (R² = .79).

https://doi.org/10.1371/journal.pone.0236646.g003

It is possible that Italy’s intervention successfully limited flows into Italy while causing other problems; we consider two. First, it may be that when crossing from Libya to Italy became too difficult or costly, migrants moved through alternative routes. The two most natural routes are through the Western Mediterranean into Spain or through the Eastern Mediterranean into Greece. Panel (c) shows that there may have been displacement through these routes into Spain and Greece, although the magnitude of displacement is much lower than the reduction from Italy’s intervention.

Second, even though fewer migrants are crossing the Mediterranean, those migrants who continue their journey during this period are relying on relatively lower quality smugglers and traveling in more dangerous seas. Panel (d) compares the arrivals and deaths before and after Italy’s intervention. The number of deaths tracks the number of migrants consistently (note the different scales). While there is one peak in deaths in November 2017, it is not abnormal for the winter when few migrants cross and conditions are more treacherous.

At the time of Italy’s intervention in Libya’s port cities, no other policies were being introduced to reduce the flow of migrants out of sub-Saharan Africa and into Libya. Indeed, IOM’s estimates of the number of international migrants in Libya in June, July and August 2017 remained stable. The estimated the number of migrants in May (Round 10), before the intervention, to be 393,652. In September (Round 13), the estimate was 416,556. This jump is comparable to jumps earlier in the year [35]. Events that occurred after the Italian intervention, however, may complicate our empirical design. In particular, the battle of Sabratha, which began September 15 and 16 [36], disrupted the major smuggling hub for more than 3 weeks. It is important to demonstrate that the effect we estimate was present before the onset of this battle. We do this in S1 Fig, where we omit the post-battle periods of our sample. The estimated effect is slightly larger in magnitude, suggesting that the reduction in migrants may have been largest right after the intervention. The point estimates from the main analysis and this subsample, however, are not statistically different from one another. Importantly, it is not possible for us to rigorously assess out how events like the battle of Sabratha may have impacted the duration of the effects we estimate. It is plausible that the Italian intervention may have had a lasting effect even in the absence of this port disruption. However, we lack the data needed to more fully explore this possibility.

Overall, Italy’s intervention reduced the number of migrant crossings and deaths in Mediterranean Sea. The event study demonstrates how militias that disengage from smuggling and participate in anti-smuggling efforts can influence the flow of migrants into Europe.