How Much Oil is the Islamic State Group Producing? Evidence from Remote Sensing

Accurately measuring oil production in low-governance contexts is an important task. Many terrorist organizations and insurgencies -- including the Islamic State group, also known as ISIL/ISIS or Daesh -- tap oil as a revenue source. Understanding spatial and temporal variation in production in their territory can help address such threats by providing near real-time monitoring of their revenue streams, helping to assess long-term economic potential, and informing reconstruction strategies. More broadly, remotely measuring extractive industry activity in conflict-affected areas and other regions without reliable administrative data can support a broad range of public policy decisions and academic research. This paper uses satellite multi-spectral imaging and ground-truth pre-war output data to effectively construct a real-time day-to-day census of oil production in areas controlled by the terrorist group. The estimates of production levels were approximately 56,000 barrels per day (bpd) from July-December 2014, drop to an average of 35,000 bpd throughout 2015, before dropping further to approximately 16,000 bpd in 2016.

Accurately measuring oil production in low-governance contexts is an important task. Many terrorist organizations and insurgencies-including the Islamic State group, also known as ISIL/ISIS or Daesh-tap oil as a revenue source. Understanding spatial and temporal variation in production in their territory can help address such threats by providing near real-time monitoring of their revenue streams, helping to assess long-term economic potential, and informing reconstruction strategies. More broadly, remotely measuring extractive industry activity in conflict-affected areas and other regions without reliable administrative data can support a broad range of public policy decisions and academic research. This paper uses satellite multi-spectral imaging and ground-truth pre-war output data to effectively construct a real-time day-to-day census of oil production in areas controlled by the terrorist group. The estimates of production levels were approximately 56,000 barrels per day (bpd) from July-December 2014, drop to an average of 35,000 bpd throughout 2015, before dropping further to approximately 16,000 bpd in 2016.

Introduction
The non-state insurgent organization known as the Islamic State group (also sometimes called and accelerated from early 2014 onwards when the group moved aggressively back into Iraq [28].
For a time, the group was considered the richest jihadist group in the world and was thought to raise money from a variety of sources [15]. In 2014 and 2015 revenue from oil production in areas the group controlled was often cited as its largest potential source of revenue flow, with estimates of weekly oil revenue ranging from "several million" to US$28 million [8,5]. Any reasonable assessment of the organization's long-run survival prospects had to account for these revenues and identify how sustainable they were [21]. Beginning in late of 2015, Daesh steadily lost territory in both Iraq and Syria, but still maintained substantial territory in both countries at of early 2017.
Remotely measuring oil production using satellite imagery provides a way to make a consistent assessment of activity that would have been useful in 2014-15 and to track developments since then with some precision. In our case, combining multi-spectral satellite imagery with available production data enables transparent and reproducible estimation of oil production in Daesh-controlled areas. We find that production levels were approximately 56,000 barrels per day (bpd) from July-December 2014, later dropped to an average of 35,000 bpd throughout 2015, before further sinking to approximately 16,000 bpd in 2016. These estimates are in line with production reported for late-2014 in captured internal Daesh documents, as we discuss in 1 We henceforth use Daesh and Islamic State group/organization interchangeably.

detail below.
Careful study of news outlets' reports, agencies' press releases and Institute for the Study of War (ISW) maps allow us to assign individual oil wells to the Islamic State group's control at the daily level. We use data from the Visible Infrared Imaging Radiometer Suite (VIIRS) sensors deployed on the NOAA/NASA Suomi NPP satellite, to assess the status of each of the sites and then impute the appropriate levels of production.
Our approach relies on the property that the natural gas dissolved in oil underground expands when it reaches the surface. In our study context, the gas is subsequently flared, hence generating heat that is captured by VIIRS sensors. Using pre-war data on oil production obtained from several public and private sources, we estimate the relationship between production and VIIRS detection and infer contemporaneous production of Daesh-controlled oil fields. Alternatively, oil extraction can take place without gas flaring -a process called venting-in which case we assign historic output values as there will not be sensor detection.
Our paper contributes to two streams of literature. The first is the emerging literature that uses remote sensing to assess behavior in extractive industries in low-governance regions. 2 Reliable external measures of resource forestry, mining, and oil production can enable better approaches to a broad range of challenges. In Colombia and Nigeria, for example, insurgent organizations have long controlled territory where oil is produced, and in many regions around the world reliable field-level production numbers are hard to come by. Estimating production remotely can enable governments and international organizations to identify illegal or untaxed production as well as to better understand the role production could play in post-conflict economies and the impact of sanctions, trade restrictions, and other policy interventions. Variants on the approach adopted here could be applied in a much broader set of places.
The second literature is the substantive one that investigates the role of natural resources in shaping conflicts. For instance [36] finds that positive prices shocks to a bulky commodity leads armed groups to create a monopoly of violence to impose taxation and regulate production in Eastern Congo. Along the same lines, [32] use mineral international price changes and data on historical concessions to show that armed groups tend to reduce violence in areas near the mines.
This "protection effect" is consistent with violence reducing economic profitability through higher labor costs. Others find that fighting around diamond mines did not affect civilians in Sierra Leone, but was rather limited to violence among soldiers [3,22]. This result is echoed by [49], which finds that violence against civilians was lower in diamond areas in Angola. Finally, [10] find that price shocks have heterogeneous effects: in labor-intensive sectors, commodity price drops result in higher incentives to join armed groups, while in the capital-intensive sector the rise in the price elicits predatory behavior from armed groups.
The rest of the paper is structured as follows. Section 1 provides some context and describes our methodology, while section 2 presents and discusses the results of the analysis. Early accounts of the group's oil production and the revenues generated indicated that oil was a significant source of financing for the organization. The 2014 Oil Market Report of the International Energy Agency estimates an output of 70,000 barrels per day (bpd) [25]. Other  news outlets give numbers around 50-60,000 bpd yielding an income of US$2.5m per day [33] to more than US$3m per day [1]. Early estimates by the US Departments of State and Treasury put the organization's oil revenues at around US$1m per day [42]. These estimates were then revised down to "a couple million dollars a week" after the U.S. started air-strikes against the organization's assets [4]. Views as to whether Daesh was financing itself through oil, external support, extortion, or taxes then evolved, with higher emphasis put on taxes and extortion as primary sources of revenues over time. Die Zeit for instance reported December 2014 oil revenues to be a mere US$370,000 per day or even lower at US$260,000 [9]. An October 2015 article however gives an estimated output of 34-40,000 bpd, earning the organization an average of US$1.5m per day [16]. In sum, there was no consensus on the production numbers or revenue they created.
On 15 May 2015, U.S. Army special forces killed a senior Daesh leader known as Abu Sayyaf, who, according to the U.S. Department of Defense "helped direct the terrorist organization's illicit oil, gas and financial operations." This raid also yielded significant amounts of intelligence into the Daesh economy, including administrative data providing a retrospective look at Daesh oil production in certain regions. Below, we shall compare our estimates from remote sensing with the administrative data captured during the raid.

Remote sensing estimation of oil production with gas flaring
Our approach relies on the property that the extraction of oil is associated with the liberation of natural gas, primarily methane, which is initially dissolved in crude oil in constant proportions.
The gas is typically collected and flared unless infrastructure exists to either re-inject the gas into the field, utilize the gas on-site, or package and send the gas off to markets [29]. Flaring is the generic method of natural gas disposal in Syria and most of Iraq, as it is in most of the world [47]. 3 Remote sensing from the Visible Infrared Imaging Radiometer Suite (VIIRS) sensor deployed on the NOAA/NASA Suomi NPP satellite [12] allows the use of multi-spectral methods [14,12] to estimate a flare's radiant heat (RH), a measure of the heat released with the combustion of the natural gas. RH is thus an indicator of the volume of gas flared, which in turn is a predictor of the volume of oil extracted. Each point represents the logarithm of yearly oil production (vertical axis) and the logarithm of yearly RH (horizontal axis) for any given oil field in our estimation sample, with black symbols for Syrian fields and grey symbols for Iraqi ones. The black circles with labeled years report the same quantities as for Syria in total. The line shows the slope of the oil-RH relationship estimated by ordinary least squares, while the shaded area is the 95 percent confidence interval. Note: Plot of linear regression of the logarithm of field RH against the logarithm of field oil output on the vertical axis. Each point represents an oil field annual average unless otherwise labeled. Linear regression and 95% confidence intervals also depicted. The fail to reject the null hypothesis that the slope equals one (see Annex A section A.4 for details on estimation sample). Data sources: [45,46,24] and NOAA.
Under the assumption that the oil-RH relationship estimated using data prior to the seizure of oil fields by Daesh still holds thereafter, we can make consistent statistical inference about contemporaneous volume of crude oil extracted from contemporaneous measures of RH.

Remote sensing assessment of oil production with gas venting
At any given moment, a site is in one of three production states: • S1: The site is producing with natural gas flaring Iraq is among the countries with the highest volume of gas flared per unit of oil produced and like Syria, did not have any regulatory limits on flaring before the war [47]. Oil production with flaring is thus the generic production state when a site is producing oil.
• S2: The site is producing without gas flaring (i.e. venting) A site that extracts crude oil could simply vent the accumulated natural gas. Although venting is wasteful and harmful to the environment, methane is lighter than air so that controlled venting (e.g. through a stack) would not constitute a fire or explosion hazard. Venting would happen on a site if (i) the operator voluntarily shuts down the pilot flame in the flare stack, or (ii) long periods of inactivity extinguish the pilot, which is not re-ignited as production resumes.
• S3: The site is inactive A site might become inactive if fighting has damaged productive infrastructure or no qualified personnel are available to operate the site.
Distinguishing between these production states is important since the lack of RH does not necessarily imply an absence of oil production; venting remains a possibility. Because the narrow spectral bands used in the estimation of RH -the near-infrared (NIR) bands M7 and M8, the short-wave-infrared (SWIR) band M10, and the mid-wave-infrared (MWI) bands M12 and M13 -are not sensitive enough to detect low light from low-intensity flaring, we also make use of data from VIIRS' Day/Night band (DNB). DNB is a wide, visible, and NIR imaging spectral band, designed to detect moonlit clouds. The DNB's low detection limits make it possible to detect electric lighting present at the Earth's surface [27,7,48,31,39], which cannot be sensed in the infrared spectral channels; it is even sensitive to moonlight reflection on the Earth. DNB detection, once the effect of the lunar cycle has been accounted for [11], therefore allows discriminating between states S2 and S3 when no infrared signals are detected (see Annex A section A.1 for more details).

Interpreting remote sensing data: case studies
To provide a precise understanding of how trends over time show up in our data, Figure 3  Radiance Radiance

Production Estimates
Analysis of the 42 sites that have ever been under Daesh control reveals that the group's production peaked in July 2014, declined from much of 2015 before peaking again in late-2015 after which it declined precipitously to 16,000 bpd in 2016. Figure 4 shows production estimates with 95 percent confidence intervals (panel A) from January 2014 through November 2016 using the approach and assumptions discussed above. Column (1) shows the results assuming no production when we do not observe a radiant heat signal. That represents our low-end estimate. Column (2) shows the results when we assume that fields without a radiant heat signal are producing via venting on days with DNB detections above the lunar illumination threshold (S2). 4 Several precautions indicate these estimates represent an upper bound on Daesh oil production. First, as gas-to-oil ratio (GOR) gradually increase as oil is extracted, our inferences based on the assumption that the GOR remained constant at its pre-war level will overestimate output when converting RH to oil production. Second, our assumptions are, by design, biased towards over-estimating rather than under-estimating production. For example, the cutoff value we choose to separate electric lights from low-intensity flaring is set at a level that rules out electric lights with probability one, while infrared detection still happens at much lower levels of radiance. Last, we assign cloudy days the average of production at the site over the month.
This method is generous since production is highly sporadic at some sites and even one day of production will be smoothed over the entire month. We performed a validation of this procedure where we randomly assigned clouds to 10% of the observations and obtained higher estimates.
In Annex A section A.4.5, we gauge the robustness of our inferences to alternative assumptions on the way we determine production states and compute RH. Resulting production estimates vary little and stay within a 30 percent range of our baseline specification.
Based on these estimates Daesh oil production peaked at roughly 86,000 bpd in late July  Daesh Production (thousands bpd) J a n 1 4 J u l 1 4 J a n 1 5 J u l 1 5 J a n 1 6 J u l 1 6 J a n 1 7 95% Confidence Interval Point estimate Note: 28-day moving average of production estimates in barrels per day with 95 percent confidence intervals. Our findings, while lower than some accounts provided in the media, are consistent with some estimates for the early period of Daesh control that relied on extrapolation from pre-war field data [8] or captured documents [37]. In particular, administrative data captured from the Abu Sayyaf raid posits daily production ranging from 52,120 bpd to 55,560 bpd in the time frame from roughly June to October 2014. These numbers are well within the bounds of our 95-percent confidence intervals and actually very close to our point estimates ( Figure 5).
Other figures available in the documents seized during the Abu Sayyaf raid, however, suggest potential problems with the Islamic State group's own administrative data. For example, the captured documents indicate significant output from the At Tayyanah field in the Al Khayr Governorate, but for which our remote data detects no visual activity and was understood to be non-active prior to Daesh takeover. Other simple arithmetic inconsistencies also reduce confidence in the accuracy of this administrative data. It is possible that internal political motivations that may bias administrative data upwards. 5 In such circumstances, estimates from remote sensing serve as a complementary estimate of Daesh oil production and further allows real-time monitoring of said production.

Conclusion
Based on our method employing satellite data, we estimate oil production in territory under the Islamic State group's control peaked briefly above 80,000 bpd in July 2014, declined steeply before a short rise in late-2015, after which it dropped steadily until another peak from November  [38], while a subsequent report indicated that prices depend on the field of origin, and that some fields charged $40 to $45 dollars per barrel [17]. At such levels, the annual revenue from oil would be far below many published estimates.
One reason why our estimates differ substantially from many publicly-available ones is likely sampling bias. To our knowledge, prior estimates all relied on what is effectively a survey of Daesh's oil assets. Information was obtained from a few selected sites and at specific dates on the basis of key documents or interviewee self-reports, which were then extrapolated.
In spite of being supplemented with expert opinions, generalizations to the universe of Daeshcontrolled oil facilities are intrinsically imprecise in that the underlying data have observations that are few and might not be representative, therefore leading to imprecise and potentially biased inferences. Updating these estimates over time faces similar methodological challenges.
The approach proposed here instead conducts a real-time census of Daesh oil production facilities with daily temporal resolution. The estimates coming from our analysis are not inferences made from observations on a few selected sites and at a few selected dates but from all sites and in real time. Thus, they have the substantial advantage of enabling less bias than previous estimates of the impact of various kinds of events (e.g. attacks, leadership conflicts, territorial losses, etc.) on oil production.
Our study aims to help bridge the knowledge gap on economic activity in Daesh-controlled territories. The results here can be built on to inform planning for short-term humanitarian assistance and long-term reconstruction. We also provide a methodological contribution in that reliable external measures of oil output can enable better approaches to a broad range of policy challenges. Across the world's poorly-governed states, few report reliable oil production numbers. Yet assessing production is critical for making sound economic policy and can enable governments and international organizations to identify illegal or untaxed production, as well as for assessing the impact of policy changes. We show that combining reliable records for 16 a subset of fields with remote sensing alongside area-specific knowledge can enable reliable production estimates at fine temporal scales in such settings.
In Syria and Iraq, understanding the structure of the economy in Daesh-controlled areas is important as it allows having a better image of the welfare of the local population and anticipating potential humanitarian needs in the region. Furthermore, taking stock of the status of the oil infrastructure will help design reconstruction plans, and inform post-conflict redistributive policy interventions.

A Materials and Methods
A.1 Classification of production states While lower-intensity flares fail to trigger an M10 detection, they always do so on the wideband DNB. Thus, when no infrared signal is detected, we rely on DNB to assess which state a given flaring site is in, since low-intensity flaring could well go undetected by infrared sensors.
However, given the sensitivity of DNB to lunar illuminance, we first proceed by suppressing the radiance effects stemming from moonlight reflection on the Earth or on clouds at night. Such procedure yields the Spike Median Index (SMI), which allows detecting the presence of surface lighting and ranking the brightness of sources over extended periods of time [11]. isolated: three small towns in Syria and a regional airport in Lebanon. The only source of nighttime lighting in these four sites is electric.
As indicated on the histogram, no signal above an SMI value of .6 was detected over the period March 2012 to October 2015, at the exception of a few outliers. We thus consider that SMI levels observed on flaring sites above .6 are associated with low-intensity flaring, so that the site is in state of production with flaring (S1). Conversely, when SMI indicates that no significant night-time light is detected above the lunar cycle for a period of 7 consecutive days, we rule out venting and determine that the site is inactive. Alternative cutoffs for SMI and Finally, when venting cannot be ruled out because either (i) SMI is low enough so that electric lighting could be the sole source of light detected, or (ii) SMI is high enough to rule out site inactivity, we determine that oil is extracted without natural gas flaring (state S2). The DNB signals are thus posited to come from infrastructure or automative electric lighting.
The results of the production state identification exercise are summarized in Table A We look at the robustness of our results by adopting alternative site production classifications (see section A.4.5).
In case of single-spectral infrared detection, Planck curve fitting is thus no longer feasible; we instead set temperature at 1810K, which is the temperature of an object with peak radiant emissions at the M10 wavelength. Alternative temperature assignment rules are considered (see Finally, if we do not have infrared detection and have yet determined that the site is flaring (state S1), we measure DNB radiance and, as above, set temperature to 1810K to apply the Stefan-Boltzmann law and derive the associated RH. Note however that in such instance, the observed radiance aggregates radiance from gas flaring and any other source of lights not filtered out by the SMI such as facility or automotive electric lighting; the resulting measure is then an overestimate of RH from gas flaring.
When a site is deemed inactive (state S3), a value of RH = 0 is then assigned for each day of these 7-day periods.
Finally, when a state of production with venting is identified (state S2), we make the assumption that the site is producing at a level equivalent to the median of its historical production since being taken over by Daesh. If such level is equal to zero or is not defined because there was no detected infrared signal, we instead set venting production at the median pre-war level: we assign the median RH of all measured RH prior to July 2012. were processed with the VIIRS nightfire algorithms [14]. The detections were ingested into a spatial database for analysis. Sites were included when (i) one multispectral band detection 28 occurred over the period March 2012 -October 2015 with a temperature above 1300K, or (ii) two detections occurred regardless of temperature. The resulting set of pixels was then visually inspected and manually edited to remove low temperature sites (under 1300 K) located in agricultural settings, deemed to be the outcomes of biomass burning.

A.3 Identifying oil assets and assigning control
To further ensure that the flaring sites identified by the algorithm described above did correspond to natural gas flaring associated with oil production, we verified the presence of oil images. Included images typically showed production facilities with a pipeline to the flare stack, though even sites with primitive infrastructure, such as simply a stack with a wall, were also included. Excluded sites fell into two categories. Either the site had no infrastructure, suggestive of agricultural burning or bombing, or the site had industrial facilities inconsistent with oil production, such as oil refineries. Figure A.8 shows an example of a site with infrastructure and a site without. 6 The Esri data is available at http://www.arcgis.com/home/item.html?id= 10df2279f9684e4a9f6a7f08febac2a9. The list of providers by region is documented at http: //help.arcgis.com/en/communitymaps/pdf/WorldImageryMap_Contributors.pdf

A.3.2 Delineating Daesh territorial control
Daesh territorial control spanned from the east of Fallujah to the north of the city of Aleppo [26]. Most of the oil assets in Syria under Daesh control are in the eastern part of the country, i.e. in the governorates of Deir Ezzour and Hassaka, and to a lesser extent Raqqa. In Iraq, Daesh controls the Ajil field and oil wells in the Hamrin Mountains, the Qayara and Najma fields, and had access to parts of the Baiji refinery until October 2015. We assign control of fields by determining the date at which a site is taken by or away from Daesh using news reports in both English and Arabic and verifying this assignment with maps published by the U.S. Department of Defense, the New York Times, and the Institute for the Study of War.
Deir Ezzour: In this governorate, Daesh consolidated its control of the western part of the governorate within a few days. CNN reports the seizure of Al Omar field on July 3rd 2014 [13], while Al Arabiya indicates that Tanak field fell under Daesh control on July 4th [4]. The rest of the western fields are not covered individually by news reports. Asharq Al Awsat announced that major fields in the governorate had been seized by July 11th [12]. The hegemony of Daesh over that part of the governorate is confirmed around that time also; we thus assign the fields of Jaffra, Izba, Sijan, Abu Hardan, and others using the date of control of either the closest major field or the district capital. Al Arabiya reports that Al Mayadin city fell under Daesh control by the end of June, while Al Hayat and other sources report that Abu Kamal did so on July 1st [1]. Finally the city of Deir Ezzour was contested for a while, and Al Jazeera announced on July 14th that the group seized control of a number of neighborhoods in this city, chasing out the opposition and other Islamist groups [6]. On July 15th, the Syrian Observatory for Human Rights reported that 90 percent of the governorate had gone under Daesh control [7] and Al Jazeera announced that the group had seized all the fields in the governorate [22]. On the Eastern side, there are only two major fields: local websites and social media feeds report Al Kharrata field falling as soon as June 9th [16], while Al Thayem was reportdly seized in July 2014 and released in January 2015 [5].
Hassaka: The presence of Daesh in this governorate has been mostly restricted to the south and east of Hassaka city. The northern fields are still under the control of Kurdish forces. We date the control of Shadadi, Jebeisse, and Al Hol fields to July 18th, as reported by local news outlets such as Aljomhuria [10].
Raqqa: Individual reports on fields controlled by Daesh in Raqqa governorate are very limited. We instead look at reports on Daesh territorial control within the governorate to assign oil asset control. While Deutsche Welle reports that the governorate was not fully under Daesh control until August 2014 [14], Al Arabiya announced that the group had consolidated its control over the three district capitals by January 13 th [2] and Asharqalarabi reports that all the fields in Raqqa were seized in that month [11]. We thus assign control of Raqqa oil fields to Daesh starting in January 13 2014.
Iraq: We assign control of the fields of Qayarra, Najma, Ajeel and Hamrin to Daesh starting June 24th. Reuters had reported the seizing of two of these fields [18], while International Business Times had announced that the four fields had fallen around the same time in June [17]. Iraqi forces regained control of Ajeel and Hamrin fields on March 4 2015, as reported by Reuters; Daesh forces set fire to the oil wells as they were fleeing the scene [19]. Similar pattern of territorial loss along and sabotage happened in the Qayarrah area, were wells burned for several weeks before engineers were able to restore production [8,21].

A.3.3 Removal of warfare events
We removed from the sample in the main analysis a number of isolated high radiance and low temperature events that are not consistent with oil production. There are two cases that cause these observations: air strikes and sabotage (by fire) by Daesh fighters. Coalition air strikes have damaged infrastructure in Iraq and Syria, and caused fires, while Daesh set fire to some wells to curb the advance of Iraqi forces and deter air strikes. In this section we provide evidence from news, daytime images, and temperature to establish the case for each observation we removed. Al Omar Area: This major field was the target of repeated Coalition and Russian air strikes.
Reports confirm strikes on the 21st of October where B-1 bombers and other allied warplanes hit 26 targets in the field [24,3,25]. Similarly on the 3rd and 12th of November 2015, the Syrian Observatory for Human rights and local news outlets documented air strikes generated fires in the field [23,20]. These reports confirm the existence of large scale fires in the facilities because of such interventions. We also exclude October 31st and November 1st from the sample because the temperatures below 1300 K are not consistent with oil production and the reports of air strikes before that date indicate strikes were successful at destroying the infrastructure needed to operate the field in the short run. No daytime imagery of this site was available in this period.  We interpreted the irregular spike immediately prior to indefinite inactivity as evidence of an air strike or warfare that destroyed infrastructure at the site.
In support of this interpretation, we observed that the radiant heat signals during this period were generally lower than expected for oil production. Most temperatures observed in this period were below 1600K; however, for several days there were no detections in the M10 band and we assumed the temperature was 1810K. Thus the high average temperature generally reflects our generous assumption of activity when the signal was poor. Qualitative evidence of the activity at this location was not available. Landsat images were available November 8, 2015 and November 24, 2015; however, we were unable to observe activity at the site at 30 meter pixel resolution. Thus, we present our results including Al Hussein North for comparison in

A.4 Inferring oil output
To infer oil production we estimate a parametric model predicting pre-war liquid production estimates at the oil-field level with measured radiant heat and other field-level characteristics.
As predicted by theory and documented by [14], there is a linear relationship between R it , the radiant heat measured in period t at site i, and the volume G it of gas flared: there thus exists a constant A such that G it = A · R it . Furthermore, the constant ratio of natural gas dissolved in crude oil can be written G it = Γ it · O it , where Γ i is the gas-to-oil ratio (GOR) that is also field-and time-specific. Substituting implies the following theoretical relationship: O it = A·R it Γ it that can be expressed in logarithmic terms as: Our estimating equation is thus where lower-case notations indicate natural logarithm of corresponding upper-case variables and ε it is the disturbance term that we assume to be independently and identically distributed across sites and time. We examined the estimated ε it to ensure symmetry and normality. The results of the model validation and estimation are reported in A.4.3. A linear relationship between crude oil extraction and radiant heat would imply β 1 = 1.
Under the assumption that the relationship estimated in equation 2 is valid after the onset of the war, estimatesô it of oil production are obtained according tô whereα,β 1 , andβ 2 are estimated by ordinary least squares (OLS). 7 A consistent estimate of the level of oil production needs to adjust for the fact that the expectation of the exponential of a random variable is not equivalent to the exponential of its expectation.

A.4.1 Consolidating oil production and field characteristics data
To construct the dataset used to estimate equation 2, we combine data on field-level characteristics (including oil output and GOR) with flare-level radiant heat measures. In this section, we describe the sources and preparation of calibration data. 7 The statistical software package used to that end is STATA version 14.

38
Information on oil field locations and production were collated from multiple sources. Oil field boundaries for Syria and Iraq were obtained from [27] and [28]. The map lists the field name, operator, operational status, type, and utilization for each field. To supplement this data in Iraq, additional fields were digitized and added from IHS Global Exploration and Production Service's map of Iraq. This map displays the oil and gas field boundaries, field name, field status, as well as oil and gas well points. Yearly field-level production data in barrels per day from 2012 to 2015 in Iraq and Syria were obtained from [27,28,24]. Production data was matched to the field locations using the field name.
A few restrictions were applied to the sample of included fields. Including Syria and Iraq, 45 fields with production data could be assigned to boundaries according to their name. First, to ensure differences in the geology and production technology between oil fields southern Iraq and Daesh-controlled fields did not bias the estimation, 16 oil fields in south Iraq were removed.
South Iraq oil fields were defined as any field below the southern most flare in Syria, 34.64 degrees latitude. Removing south Iraq reduced the sample to 29 oil fields. One field (Miran West) were dropped for having no production data. In total, 28 fields outside of southern Iraq were matched and included data.
Second, two fields were dropped due to inconsistencies in the production data. Ain Zalah (West Butmah) oil field, was removed due to data reliability: oil output was reported invariant during the entire production history, while corresponding RH levels differed significantly. year observations. 8 The estimation employs additional data on the geological characteristics of each flare site to account for heterogeneity in the estimation of the flaring-production relationship. Field-level GOR were obtained from Energy-Redefined, and matched to the field locations according to the field name. 9 Each flare was subsequently assigned the GOR of the nearest field with a known GOR. We do not have field-level GOR data for recent years, so that our statistical inferences assume constant GOR. Field GOR however increases over time as the field ages and crude oil is being extracted. This implies that the assumption of constant GOR will lead to inferences on oil output that are biased upward.

A.4.2 Aggregating RH data
Since the dependent variable, field oil production, is measured yearly and at the field level, we aggregate daily observations on flare-level radiant heat into one yearly field-level measure. The linear relationship between RH and the volume of crude oil extracted allows linear aggregation.
First, each oil field in the production data was linked to gas flares located within its bound- where the site was cloudy were excluded. Last, the yearly averages of all sites linked to a field were summed to obtain the field's yearly radiant heat. If flares from the same field have heterogeneous productivity, the log of the sum does not equal the sum of the logs. To address this concern, we repeated the calibration estimation with the subsample of sites with only one flare so that there was no aggregation.  (2) shows the full specification shown in equation 2.

A.4.3 Estimation results
Column (3) excludes fields with more than one flare site. The error term in the model represents two primary possibilities. First, there exists some measurement errors as both oil output and RH data could not be reported or measured precisely.
Second, the might be model specification errors whereby the relationship between oil output and RH are more complex than what is suggested by equation (2) as other factors (flare stack characteristics, wind conditions, etc.) might affect the RH-oil output relationship.
In Figure A.10 we plot the residuals of the estimation in Table A We observe that while the linear model residuals deviate substantially from the normal distribution, the nonlinear model residuals are a good approximation of the normal distribution. The normality of the nonlinear residuals supports the premise that the model of Table A.8 Column   2 is appropriate for calibrating radiant heat to oil production.

A.4.4 Inferring oil production
To compute the level of oil production for each flare observation, we start with the results of the estimation of equation (2). We impute log production for 200 draws from the estimated joint distribution ofβ 1 andβ 2 . After converting to the level of oil production for each observation for each draw, we compute the daily sum of all draws for all Daesh flares. Last, we compute the expectation (average), 5th, and 95th percentiles of total Daesh production each day over the 200 draws. The implications of this method are notable in the results: confidence intervals are asymmetric with greater mean than median because the exponential of a uniform random variable is positively skewed.  Such procedure allows inferring daily oil production at any site, which we can then aggregate to obtain yearly production as reported in Table A. 9. Column 1 repeats the estimates with the main sample, cited in the text, assuming venting when no radiant heat is observed. We also report our estimates including the November 2015 observations at Al Hussein North in Column 2. The inclusion of Al Hussein North results in an increase of 5,000 bpd on average in 2015.

A.4.5 Alternative site production state classifications
In this section we show the results are not sensitive to the alternative RH assignment rules for sites that are deemed to be venting and discusses the resulting inferences. Figure  sites, which is a rough proxy for total production. Column 3 shows the average RH measured at active sites, which captures in a rough way the intensity of production among working sites.
The top row shows raw values, while the bottom row normalizes by activity levels measured before the war escalated (March-June 2012) as a way to highlight changes over time. As illustrated, Daesh's productive base was small to start with and dropped rapidly long before the group took over (column 1). Moreover, its total productivity has generally fallen since January 2014 (column 2), and the productivity per well after its takeover, as measured by flaring activity, is quite poor compared to that in non-Daesh areas of Iraq and Syria. Interestingly, the share of Daesh sites in each of our three production conditions-S1, the site is producing with natural gas flaring; S2, the site is producing without gas flaring (i.e. venting); or S3, the site is not producing-changes little over time. This is likely because flaring activity had already dropped dramatically at most of those sites before Daesh took over. Fields