Study population and period
The SCI project consisted of seven Primary Health Care (PHC) facilities providing outpatient consultations, first aid, dressing and suturing, the Minimum Initial Service Package (MISP) for reproductive health, referrals, and Basic Emergency Obstetric Care (BEmOC); and one maternity hospital offering Comprehensive Emergency Obstetric Care (CEmOC). The project had a theoretical catchment population of about 250,000, spread across a large portion of Idleb governorate, rural areas of western Aleppo governorate and northern Hama governorate. This analysis covers the period October 2014 to June 2017.
The SCI project developed a bespoke Health Management Information System (HMIS) using a Microsoft SQL server, used to collect, manage and archive individual electronic patient-provider contact records. Four authors (AE, YHA, and HA) were involved in managing this project and developing the HMIS. Unique identifiers were deleted after data aggregation. We excluded records prior to October 2014 as these were collected using a preliminary version of the HMIS, and considered data up to 30 June 2017. The analysis dataset comprised 637,824 patient-provider contacts.
No consolidated database of conflict incidents at the community level was readily available for the study period. We thus constructed such a dataset by systematically consulting two main sources: “Shaam News Network”  and “Ugarit News”: both are local news agencies that feature daily narrative field reports summarising all war incidents in Syria. We verified incidents with casualties by triangulating the information with three other sources: the Syrian Observatory for Human Rights, Al Jazeera and BBC news. More than 1500 daily reports were reviewed for the period between 1 October 2014 and 30 Jun 2017. We extracted meta-data (date, location, type, number of people killed) on all (n = 11,396) war incidents taking place within the project catchment area.
Lists and spellings of settlements (e.g. village names) patients originated from were matched against a master geographic dataset maintained by the United Nations Office for Coordination of Humanitarian Affairs . We excluded from analysis 296/882 ‘settlements’ mentioned on the patient database that could not be matched to this master dataset (these may in fact have been neighbourhoods or streets). We also computed line-of-sight- distances between settlements and their closest health facility based on both locations’ coordinates: 332 settlements > 15 km away from any health facility were also excluded based on average walking distance to access health facilities as per UNHCR standards for primary health care coverage , and based on authors’ experience of travel time limits to realistic accessibility in the Idlib context. In practice, the above exclusions only reduced the total number of patient records from 637,824 to 583,781, as the excluded settlements accounted for very few patients, suggesting most were outside the effective catchment area of the SC project. Lastly, settlements were allocated to the catchment of the health facility to which they sent most patients.
Lastly, it was challenging to get accurate estimates for population figures. The National Population Monitoring (NPM) surveys were used as population denominators . The NPM was a joint initiative by the International Organization for Migration and OCHA in which the populations of most communities in Syria were estimated on a monthly basis starting from December 2015, adjusting for displacement. In the absence of displacement data before this period, the first available (December 2015) estimate was also applied to the period from October 2014 to November 2015.
The primary outcome we looked at was new outpatient consultations. Antenatal care visits, in-facility deliveries and caesarean sections were secondary outcomes. The choice of these outcomes was mainly opportunistic based on data availability.
Exposures: conflict incidents
War incidents were classified as (i) bombardment, defined as any bombing or shelling, either aerial or ground; (ii) explosion, defined as any explosion, detonation or burst caused by explosive devices, cars or remnants; (iii) clashes, defined as any active fighting or clash between any of the warring parties.
All the health facilities, except the maternity hospital, were closed for a couple of weeks in October 2015 due to security threats faced by Save the Children. Also, the maternity hospital was hit by an airstrike on 29 July 2016, and it was closed for less than two weeks following this attack. For these periods, data from the closed health facilities were excluded from the daily analysis. However, it was not excluded from the monthly analysis considering that there were only two incidents of facility closure none lasted for more than two weeks.
For conflict incidents, the database includes variables that indicate scale of incidents, such as number of casualties and injuries. However, considering the unreliability of this type of data we have not used these variables in our analysis.
Patient and public involvement
The study is an observational study of a routinely collected health data, and hence patients were not involved in the study. However, all the data was anonymised and aggregated on facility level before starting the analysis. The findings presented publicly do not include any individual or even facility-level data. Therefore, additional patient consents were not sought; however, further ethical oversight from the organisations involved in running the health facilities was sought.
During the design phase of this study, the organisations that were running the health facilities; Save the Children, Syria Relief and Shafak Syria; were all consulted. They all agreed on the importance of the study questions and their policy and advocacy implications. They also reviewed the proposed methodology, data management plan, and analysis plan to ensure ethical considerations are adequately addressed.
The statistical analysis started with an exploratory and descriptive analysis, data integrity checks, and trend analysis followed by two causal modelling:
Monthly analysis (village—month panel data): A negative binomial regression for longitudinal panel data (monthly outcomes grouped by village).
Daily analysis (village—day panel data): A negative binomial distributed lag model for longitudinal panel data (daily outcomes grouped by village) with a lag period of 30 days. We opted for a 30-day lag period for the daily analysis to be comparable to the monthly analysis.
Some other regression models, such as Poisson and Zero Inflated Negative Binomial, were explored and excluded given the nature of the data and using Likelihood Ration Tests and multicollinearity checks. As Gardner and others suggest, Poisson model for count data can be misleading considering that individual observations are usually over-dispersed than is assumed by the model, and negative binomial regression represent a valid alternative . The model approach took into consideration that the data is count clustered by village, and the data is dispersed and skewed to the right with a high proportion of zero values. Both models used fixed effects, an offset of the estimated population of each village and the following predictors:
The three conflict exposures (bombardment, explosions and clashes);
whether the day/the month falls within Ramadan;
which season the day/the month falls within;
distance from the closest health facility, initially as a linear term;
and, for the daily model: whether the day is Friday.
We used fixed effect models with an assumption that the observations are independent between and within villages. This assumption was based on direct observations during the period of data collection which indicate frequent and massive changes of the study population with relation to population movements and displacement, areas of military control, road accessibility, and availability of health services. All these factors contributed to a greater variation between villages and even within villages leading us to assume independence of the observations.
For confounders, the authors’ own medical work inside Syria suggested that healthcare utilisation decreases during Ramadan. Internal clashes between local armed groups in northwest Syria also tend to decrease during the holy month. The same applies for extreme weather, either in summer or winter when both health seeking behaviour and hostilities decrease. Distance to health facility was another confounder as it affects health seeking behaviour from the one hand, and remote villages far away from any vital infrastructure are less likely to witness war incidents from the other hand.