Insecurity, polio vaccination rates, and polio incidence in northwest Pakistan

Significance

The international spread of poliovirus has been declared a Public Health Emergency of International Concern by the World Health Organization, and Pakistan is one of only three countries that have never stopped poliovirus circulation. We analyzed polio incidence and polio vaccination data from 32 districts of northwest Pakistan and constructed a measure of insecurity based on journalistic reports of deaths and injuries resulting from conflict-related security incidents. Insecurity was associated with reduced vaccinator access, reduced polio vaccination, and increased polio incidence. These findings provide strong scientific evidence that insecurity is an important obstacle to global polio eradication.

Abstract

Pakistan is one of three countries in which endemic transmission of poliovirus has never been stopped. Insecurity is often cited but poorly studied as a barrier to eradicating polio. We analyzed routinely collected health data from 32 districts of northwest Pakistan and constructed an index of insecurity based on journalistic reports of the monthly number of deaths and injuries resulting from conflict-related security incidents. The primary outcomes were the monthly incidence of paralytic polio cases within each district between 2007 and 2014 and the polio vaccination percentage from 666 district-level vaccination campaigns between 2007 and 2009, targeting ∼5.7 million children. Multilevel Poisson regression controlling for time and district fixed effects was used to model the association between insecurity, vaccinator access, vaccination rates, and polio incidence. The number of children inaccessible to vaccinators was 19.7% greater (95% CI: 19.2–20.2%), and vaccination rates were 5.3% lower (95% CI: 5.2–5.3%) in “high-insecurity” campaigns compared with “secure” campaigns. The unadjusted mean vaccination rate was 96.3% (SD = 8.6) in secure campaigns and 88.3% (SD = 19.2) in high-insecurity campaigns. Polio incidence was 73.0% greater (95% CI: 30–131%) during high-insecurity months (unadjusted mean = 0.13 cases per million people, SD = 0.71) compared with secure months (unadjusted mean = 1.23 cases per million people, SD = 4.28). Thus, insecurity was associated with reduced vaccinator access, reduced polio vaccination, and increased polio incidence in northwest Pakistan. These findings demonstrate that insecurity is an important obstacle to global polio eradication.

The Global Polio Eradication Initiative (GPEI) has helped reduce the global incidence of paralytic polio cases by over 99% since its inception in 1988 (1). Only three countries have not stopped endemic poliovirus circulation: Pakistan, Afghanistan, and Nigeria. In certain regions, anecdotal reports have named insecurity as the most important obstacle to polio vaccination (2–5), and polio workers have been targets of violent attacks by local militant groups (6). Despite significant media coverage and commentary, there is little scientific evidence examining the relationships between insecurity, polio vaccinator access, vaccination rates, and polio case incidence.

Polio vaccinations in Pakistan are delivered through several mechanisms, including routine vaccination programs and “supplemental immunization activities” (hereafter referred to as “vaccination campaigns”). Mass vaccination campaigns have been a critical element of the polio eradication strategy in Pakistani regions where routine immunization rates are as low as 25–33% (7).

According to field reports, insecurity can imperil vaccination campaigns by disrupting supply chains, exposing vaccinators to attack, restricting access to vulnerable populations, inducing mass migration, and limiting supervision and oversight (7–9). Furthermore, caregivers in highly insecure areas in Pakistan and Nigeria have demonstrated limited awareness and negative attitudes toward polio vaccination (10). A dramatic example of the effect of insecurity on access to populations was the ban on polio vaccination in North and South Waziristan imposed by the local Taliban in 2012 and 2013 (11).

The purpose of this study was to examine the hypothesis that insecurity is associated with reduced polio vaccination and increased paralytic polio incidence in northwest Pakistan. To inform understanding of the relationship between insecurity and vaccination rates, we also examined the hypothesis that insecurity was associated with reduced vaccinator access to target populations.

Results

There were 645 paralytic polio cases reported in the Khyber Pakhtunkhwa Province (KP) and the Federally Administered Tribal Areas (FATA) from 2007 to 2014. Population, vaccination, insecurity, and polio case data for each district are presented in Table 1.

Table 1.

Population, vaccination, insecurity, and polio case data by district and agency in KP and FATA

District	Population, 1998 census	Target population <5 y, mean*	Polio cases, 2007–2014	Vaccination percentage in campaigns 2007–2009, mean* (IQR)	No. of campaigns 2007–2009	Deaths and injuries due to insecurity, 2006–2014
KP
Abottabad	880,666	185,689	0	98.06 (4.75)	15	0
Bannu	675,667	177,921	16	97.22 (4.72)	26	477
Batagram	307,278	71,191	1	97.06 (5.63)	16	0
Buner	506,048	152,599	7	96.78 (3.15)	17	99
Charsadda	1,022,364	251,140	11	86.87 (20.35)	29	942
Chitral	318,689	70,297	0	100.27 (3.84)	13	0
D.I. Khan	852,995	219,853	2	98.29 (3.37)	27	680
Hangu	314,529	95,574	5	98.00 (7.63)	19	805
Haripur	692,228	156,299	1	97.18 (6.43)	15	66
Karak	430,796	101,054	2	97.25 (2.25)	20	28
Kohat	562,644	162,391	13	92.69 (3.51)	21	477
Kohistan	472,570	69,915	1	96.93 (1.77)	16	0
Lakki Marwat	490,025	131,969	12	98.68 (1.71)	26	319
Lower Dir	717,649	224,632	0	92.25 (6.24)	22	371
Malakand	452,291	109,414	1	96.39 (5.87)	19	90
Mansehra	1,152,839	251,242	0	98.76 (3.34)	17	11
Mardan	1,460,100	369,090	16	94.77 (2.86)	24	258
Nowshera	874,373	250,490	13	96.54 (3.55)	29	237
Peshawar	2,019,118	702,740	76	99.31 (5.13)	29	3,131
Shangla	434,563	126,107	2	93.23 (3.76)	17	97
Swabi	1,026,804	245,784	7	95.10 (2.79)	22	66
Swat	1,257,602	374,955	28	76.73 (40.65)	14	906
Tank	238,216	58,849	11	102.21 (4.64)	27	140
Torghar	500,000	N/A	5	N/A	N/A	0
Upper Dir	575,858	149,463	7	95.59 (5.59)	19	114
FATA
Bajaur	595,227	210,884	37	74.10 (49.46)	26	642
Khyber	546,730	186,274	173	93.02 (4.20)	29	889
Kurram	448,310	129,642	6	84.87 (15.72)	15	786
Mohmand	334,453	111,240	26	79.26 (36.22)	29	366
North Waziristan	361,246	152,258	128	95.82 (4.11)	25	3,498
Orakzai	225,441	79,526	5	96.16 (17.07)	19	342
South Waziristan	429,841	106,900	33	80.40 (53.43)	24	1,436
Total	20,677,160	5,685,382	645	93.35 (12.86)	666	17,273

^*Values are the mean per individual vaccination campaign within each district. D.I. Khan, Dera Ismail Khan.

There were 768 district-level vaccination campaigns planned between 2007 and 2009 in KP and FATA. Excluding 83 campaigns in the Frontier Regions and 19 campaigns that were initially planned but cancelled left 666 campaigns for analysis with a mean target population of 5.7 million children. There was a median of 21 campaigns in each district (minimum: 13; maximum: 29). There was an overall mean vaccination rate per campaign of 93.4% (SD = 13.5). Of 666 campaigns, 316 (47.4%) were categorized as “secure,” and the remainder were allocated evenly into three categories with increasing exposure to insecurity.

Insecurity and Polio Incidence.

Of the 645 polio cases, 542 (84.0%) occurred in a district that had been affected by insecurity in the preceding 12 mo. The unadjusted monthly incidence of polio cases in secure months in districts was 0.13 per one million people (SD = 0.71) compared with 1.23 per one million people (SD = 4.28) in “high-insecurity” months. Time trend adjusted regression estimates demonstrated a 73.0% increase in the incidence of polio cases in “high-insecurity” months in districts compared with secure months (adjusted risk ratio, 1.73; 95% CI: 1.30, 2.31; Table 2).

Table 2.

Results of a two-level Poisson regression model of the association between district-level insecurity and polio case incidence in northwest Pakistan, including a fixed effect for time

Insecurity exposure	Unadjusted monthly incidence of polio cases per million people (SD)	Adjusted risk ratio (95% CI)	P value
Secure	0.13 (0.71)	Reference
Low	0.60 (2.77)	2.24 (1.71–2.95)	<0.001
Medium	0.36 (1.18)	1.87 (1.38–2.52)	<0.001
High	1.23 (4.28)	1.73 (1.30–2.31)	<0.001

Insecurity and Vaccination Campaigns.

Vaccination campaigns exposed to any amount of insecurity had significantly lower rates of vaccination than secure campaigns (Fig. 1). Secure campaigns had an unadjusted mean vaccination rate of 96.3% (SD = 8.6) compared with 88.3% (SD = 19.2) in high-insecurity campaigns. Time trend adjusted regression estimates showed a relative reduction in vaccination rates of 5.3% in high-insecurity compared with secure campaigns (95% CI: −5.2, −5.3; Table 3). This represented an average of 8,720 additional children who were not vaccinated in campaigns affected by high insecurity (95% CI: 8,587, 8,852).

Fig. 1.

Insecurity and polio vaccination rates in district-level vaccination campaigns. Legend: Bars represent 95% CI. n = 666 total campaigns. Mean vaccination percentage per individual campaign within each insecurity category is shown. Vaccination campaigns were categorized according to exposure to insecurity-associated deaths and injuries per 10,000 people per district in the 12 mo preceding the campaign. Category limits were as follows: secure, 0 (n = 316); low, 0.01–0.756 (n = 117); medium, 0.757–1.744 (n = 117); high, >1.744 (n = 116).

Table 3.

Results of multilevel Poisson regression models of the association between insecurity and polio vaccination rates

		Model 1		Model 2		Model 3
Insecurity exposure	Mean unadjusted vaccination rate (SD)	Relative change in vaccination rate	95% CI	Relative change in vaccination rate	95% CI	Relative change in vaccination rate	95% CI
Secure	96.31 (8.55)	Reference		Reference		Reference
Low	91.39 (15.53)	−4.02%	−3.96, −4.08	−3.47%	−3.40, −3.53	−1.80%	−3.4, −0.19
Medium	92.32 (13.85)	−6.20%	−6.13, −6.26	−0.57%	−0.50, −0.65	0.43%	−2.14, 1.28
High	88.28 (19.15)	−12.91%	−12.85, −12.98	−5.26%	−5.18, −5.34	−0.23%	−1.64, 2.10
Inaccessibility						−1.35%	−1.41, −1.30

Model 1 includes insecurity alone; model 2 includes insecurity allowing for time fixed effects; and model 3 includes inaccessibility and insecurity allowing for time fixed effects. The relative change in vaccination rate represents the difference from the reference category (secure). Vaccination campaigns were categorized according to exposure to insecurity-associated deaths and injuries per 10,000 people per district in the 12 mo preceding the campaign. Insecurity category limits were as follows: secure, 0; low, 0.01–0.756; medium, 0.757–1.744; high, >1.744.

Insecurity and Accessibility.

The number of children in regions inaccessible to vaccinators was recorded in campaigns after October 2007 (n = 508). An unadjusted mean of 10.7% of target children were inaccessible to vaccinators in high-insecurity campaigns (SD = 20.6) as opposed to 2.4% in secure campaigns (SD = 8.5). Time trend adjusted regression estimates demonstrated a 19.7% relative increase in inaccessible children in high-insecurity campaigns compared with secure campaigns (95% CI: 19.2, 20.2). When inaccessibility, insecurity, and vaccination rates were modeled together, the negative association between insecurity and vaccination rates was attenuated, whereas the association between inaccessibility and vaccination rates remained statistically significant (Table 3).

Vaccination Rates and Polio Incidence.

The correlation between the mean vaccination rate per campaign in a district and the number of incident polio cases per capita in that district from 2007 to 2009 was −0.73 (P < 0.001).

Sensitivity Analyses.

The results of sensitivity analyses demonstrate that the association between insecurity and vaccination rate persists under various assumptions. Individual district fixed effects and including time as a random effect did not change the overall association between insecurity and vaccination percentage (SI Appendix, Table S2). This association also was not changed by modeling insecurity as a continuous variable rather than categorical. Under this linear assumption, every insecurity-associated death or injury per 10,000 people was associated with a 1.5% reduction in vaccination rates (95% CI: −1.6, −1.5).

There were 19 campaigns that were planned but not conducted, and 14 of these (73.7%) were exposed to medium or high insecurity (SI Appendix, Table S3). When these campaigns were included in the analysis, the relationship between insecurity and vaccination rates was more pronounced. The mean unadjusted vaccination rate in high-insecurity campaigns was 81.0% (SD = 30.2%) and in secure campaigns it was 94.8% (SD = 14.7%). Time trend adjusted regression estimates showed a relative reduction in vaccination rates of 21.0% in high-insecurity compared with secure campaigns (95% CI: −20.9, −21.1; see SI Appendix, Table S4). Similar to the primary analysis, including inaccessibility in the model attenuated the relationship between security and vaccination rates.

In total, 151 out of 666 campaigns (22.7%) had vaccination rates greater than 100%. These were distributed across 26 out of 31 districts (SI Appendix, Table S5). In three districts, more than 50% of campaigns had vaccination rates greater than 100%. Excluding all campaigns with vaccination rates greater than 100%, excluding all campaigns from the above three districts, or setting an upper bound on vaccination rates to 100% did not affect the overall size or statistical significance of the association between insecurity and vaccination rates (SI Appendix, Table S6). Campaigns with vaccination rates greater than 100% and lower than 100% had no statistically significant difference in their distribution across insecurity categories (P = 0.37).

There was no change in the size or statistical significance of the association between insecurity and vaccination rates when the target number of children to be vaccinated was set as the mean, median, minimum, or maximum of all campaigns in each district (SI Appendix, Table S7).

The association between insecurity and vaccination rates was statistically significant when the exposure was assessed at various time periods (SI Appendix, Table S8). Campaigns exposed to security incidents within the preceding 3 mo, between 3 and 6 mo, and between 6 and 12 mo had statistically significantly lower rates of vaccination than those with no insecurity exposure in the preceding 12 mo.

Using 2013 population projections instead of 1998 census data did not substantially change the results of our main analyses (SI Appendix, Tables S9 and S10).

Validity of Insecurity Data.

The validity of the Pakistan Body Count (PBC) insecurity data was tested by comparing it with two other insecurity datasets, the Pak Institute of Peace Studies (PIPS) in 2007 and 2008 and the Bureau of Investigative Journalism (BIJ) from 2006 to 2009. The correlation between PIPS and PBC for the total number of people killed and injured per district per month was 0.895 (P < 0.001). The correlation between PBC and BIJ was 0.942 (P < 0.001).

Discussion

Polio eradication efforts along the Pakistan–Afghanistan border have faced unique security challenges, including targeted attacks on vaccination workers and outright bans on polio vaccination in certain regions (11). Although insecurity is often cited as an important barrier in the final push for global polio eradication, there is very limited scientific evidence demonstrating how insecurity affects polio vaccination or polio incidence. We rigorously studied the effects of insecurity on polio eradication efforts in this historically important context, which offers insights for both polio eradication and for public health science more broadly. This study identifies an association between insecurity and reduced vaccinator access, reduced vaccination rates, and increased polio incidence, thus demonstrating a plausible causal pathway for the relationship between insecurity and poliovirus transmission. Polio vaccination rates were 5.3% lower during high-insecurity campaigns, and polio incidence was 73.0% greater in districts during high-insecurity months compared with secure periods. These analyses may underestimate the effects of insecurity. When we performed sensitivity analyses including campaigns that were cancelled, often due to insecurity, vaccination rates were 21.0% lower in high-insecurity campaigns compared with secure campaigns.

Our study addresses limitations in the existing literature. Previous studies identified preliminary associations between polio and insecurity. Decreased vaccination coverage was linked to reduced population immunity in Pakistan (5), and this was largely attributed to insecurity based on expert opinion and field reports (12). Armed insurgency has been associated with country-level polio incidence after 2011 (13), however country-level comparisons are limited as polio cases are confined to relatively few countries. Studies in Afghanistan, Sierra Leone, and Angola identified an association between insecurity and reduced vaccination or polio incidence (14–17), but they are limited in drawing a causal linkage because insecure areas were designated using expert opinion rather than objective data, vaccination rates and polio incidence were not studied together, vaccinator access was not examined, and/or temporal changes in exposure to insecurity were not captured. Numerous articles have commented on the relationship between insecurity and poliovirus transmission (6, 12, 18–20), but none of these provide empirical evidence to support their anecdotal observations.

We used an objective measure of insecurity and modeled district-level data on polio incidence over an 8-y period and vaccination campaigns over a 3-y period. We used multilevel statistical models with adjustment for both district-level effects and time trends. This allowed us to control for unmeasured confounders that might vary between districts, such as routine healthcare delivery, social development and infrastructure, trust in government agencies, or geography. This also allowed us to control for secular changes over time, for example improvements in vaccination campaign quality. We demonstrate that the effects of insecurity on polio vaccination and incidence are independent of other differences between districts or secular time trends.

Our findings offer several important insights regarding the relationship between insecurity and vaccination campaigns. First, we conducted a mediation analysis, which demonstrated that the relationship between insecurity and reduced vaccination rates was driven by decreased vaccinator access. When the number of inaccessible children was included as a covariate, insecurity had no independent association with vaccination rates. Insecurity may interfere with vaccination efforts in many ways, but limiting vaccinator access has been described as potentially the most important (19). Our findings support this theory and are consistent with results from surveys in which caregivers in insecure zones were more likely to report that vaccinators did not visit their homes (45% versus 7% in more secure zones) (10). It is important to note that our analyses predate the highly publicized vaccination bans in North and South Waziristan. Thus, we observe that vaccinator access is not an “all-or-nothing” concept and it varies over time within districts. The GPEI has a long history of obtaining vaccinator access in insecure times, for example by negotiating ceasefires in war (12). Our findings highlight a unique feature of the polio eradication efforts in Pakistan, in which rapid and unpredictable fluctuations in vaccinator access have necessitated novel approaches, such as conducting continuous vaccination programs with local volunteers and introducing the more efficacious inactivated polio vaccine to boost population immunity during limited access windows (2, 21). Encouragingly, reports suggest that the number of children in inaccessible regions of Pakistan decreased from ∼500,000 in 2013 to 35,000 in 2015 (2).

Second, we demonstrate that insecurity is not a static exposure. Most studies examining insecurity and polio vaccination simply categorize areas as secure or insecure (5, 14–16). However, we found that the relationship between insecurity and vaccination rates was predominantly due to variation within districts over time. In our analyses, 16 of the 31 districts had both vaccination campaigns that were exposed to the highest levels of insecurity and vaccination campaigns that were exposed to the lowest levels of insecurity. Thus, insecurity should be regarded as a longitudinal exposure that varies over time. It may not be accurate to refer to areas as uniformly secure or insecure.

Third, our findings suggest that there is a low threshold in the effects of insecurity on vaccination. Consistent with previous work (14), we found a nonlinear relationship between insecurity and vaccination rates. Exposure to any amount of insecurity was associated with both reduced vaccination rates and increased polio incidence. Achieving 95% population immunity has been identified by the GPEI as a target to interrupt virus circulation in Pakistan (7). In our study, secure campaigns had a mean vaccination rate of 96.1%, whereas campaigns exposed to insecurity had mean vaccination rates less than 95%. It is difficult to estimate how vaccination coverage with oral poliovirus vaccine translates to acquired immunity because the immunogenicity is impacted by the type of vaccine and factors such as the prevalence of diarrheal illness (7). However, our findings suggest that the effects of even a low level of insecurity impair population immunity and result in ongoing poliovirus transmission. This highlights the tenuous nature of polio eradication efforts in Pakistan and emphasizes the importance of maintaining access for vaccinators in insecure times.

Fourth, we find that insecurity may have a lasting effect on vaccination rates. Vaccination rates were reduced in campaigns up to 12 mo after security incidents. This may be due to the occurrence of intervening security incidents that were not captured by our measure of insecurity, such as threats to vaccination workers. Alternatively, it may suggest that insecurity has more persistent consequences, such as erosion of trust (10), which can have longer term effects on vaccination efforts. Thus, our study offers several insights for future research and polio eradication efforts by highlighting the primary role of vaccinator access in mediating the relationship between insecurity and vaccination rates, emphasizing the importance of modeling insecurity longitudinally, identifying threshold effects of insecurity, and suggesting that insecurity may have long-term effects on vaccination efforts.

The most important limitation of this study was poor demographic data. The most recent census was in 1998, however we also conducted a supplementary analysis using 2013 Government of Pakistan population projections and our main findings were unchanged. Sensitivity analyses suggested that campaigns with vaccination rates greater than 100% or changes in the target number of children within a district over time were unlikely to have biased our findings. The district fixed-effects analysis suggested that the relationship between insecurity and vaccination rates was predominantly due to variation within districts over time. Thus, the association between insecurity and vaccination rates was a function of changes in the number of children vaccinated (the numerator) rather than the target number of children to be vaccinated (the denominator). The inability to model spillover effects from one district to another remains an important limitation of this study.

We used vaccinator tally sheet reports to study vaccination rates. An ethnographic study in Pakistan found that vaccination campaign and monitoring records are sometimes falsified, although tally sheet data were felt to be the least prone to falsification (8). If tally sheets were being falsified, campaigns with lower true rates of vaccination would seem more successful because of artificially high reporting. This would minimize observed differences between campaigns and bias toward the null. In the case of erratic reporting, the error would be nondifferential and also bias toward the null. The strong and significant negative correlation between tally sheet vaccination rates and polio incidence suggests that they were a useful indicator of population immunity. Postcampaign monitoring (PCM) by third-party monitors is another source of vaccination coverage data, which was not used because PCM reports may overestimate coverage (5) and areas inaccessible to vaccinators were also inaccessible to PCM, meaning poor coverage in those regions was not captured. In our dataset, PCM was available for 655 campaigns and showed a mean vaccination rate of 96.1% (SD = 2.0), which was 2.7% greater than tally sheet data (95% CI: 1.6, 3.7) and did not correlate with polio case rates (Pearson correlation, −0.10; P = 0.61). Lot Quality Assurance Sampling, which offers more reliable vaccination data, was not performed consistently in the insecure areas of KP and FATA during the study period.

Finally, our analyses were limited to relationships between polio and generalized insecurity rather than targeted violence against polio eradication efforts. We chose this approach because of a lack of systematically collected data pertaining to attacks against polio workers. Preliminary reports involving such data exist (20), but the comprehensiveness or reliability of these data are uncertain. We therefore chose not to include more nuanced polio-specific security data and instead performed a rigorous analysis using systematically collected insecurity data, which are more readily verifiable, thus increasing the validity and replicability of our findings.

Methods

Study Population and Setting.

We focus on two regions in northwest Pakistan, KP and FATA, because they have been troubled by insecurity and are believed to be key to global polio eradication. In 2015, an Independent Monitoring Board of the GPEI described these regions as the “conveyor belt” of polio transmission (2). KP is divided into 25 districts, one of which (Torghar) was newly created in 2011, while FATA consists of seven semiautonomous district-sized tribal agencies. For simplicity, we will refer to all of these as districts. There are also six sparsely populated “Frontier Regions” in FATA (<1.8% of the total sample population). Vaccination campaigns conducted in the Frontier Regions of FATA were excluded because insecurity data were not recorded specifically for these regions. Insecurity in this region has occurred mainly due to incidents between government intelligence and military forces and antigovernment elements and, in some cases, among different militant groups (22).

Data Sources.

Polio case data.

Wild-type paralytic poliovirus case data were obtained for the years 2007–2014 from the national polio eradication surveillance office in Islamabad, Pakistan.

Polio vaccination data.

Polio vaccination campaigns employ volunteers to vaccinate all children under the age of 5 y in a predefined area over a period of several days. For the purposes of our multilevel analysis, we treated the campaign in each district as an individual data point. We use the term “campaign” to refer to individual district-level campaigns, not to be confused with its typical use in denoting national or subnational supplementary immunization activities. Vaccination data were obtained from aggregated tally sheet reports from GPEI vaccination teams for all polio vaccination campaigns conducted in KP and FATA from 2007 to 2009.

Security data.

Data about security incidents were compiled from three organizations: PBC (23), PIPS (24), and BIJ (25). These independent groups use reports from journalistic and government sources to document conflict-related security incidents in Pakistan. The PBC database was used primarily, as this cataloged conflict-related security incidents in Pakistan for the entire duration of this study (2006–2014). Validation of the PBC dataset was performed by determining the Pearson correlation coefficient between PBC data and the other two datasets for the period 2006–2009.

Population data.

The most recent completed census in Pakistan was conducted in 1998. The population of KP and FATA at that time was 20.7 million, which represents 15.6% of the total population of Pakistan. We used the 1998 census data as the primary data for our analyses. The Government of Pakistan Bureau of Statistics has developed projections for district-level population based on the 1998 census data and expected population growth rates. We used the 2013 projected population data in a sensitivity analysis to assess the effect of population changes on our main findings.

Outcome Measures.

The primary outcome variable for the polio case rate analysis was the number of incident paralytic polio cases per capita per month within each district.

The primary outcome variable for the vaccination campaign analysis was the percentage of children vaccinated during a campaign in each district. This was calculated using the number of children vaccinated per campaign according to tally sheet data divided by the target number of children to be vaccinated per campaign. The secondary outcome variable was the percentage of inaccessible children defined as the number of target children in areas that could not be reached by vaccinators. For further details and variable definitions, see SI Appendix, Table S1.

Exposures.

The primary exposure variable was a composite index of insecurity. This was derived by adding the total number of people killed and injured in a given district as a result of conflict-related security incidents in the previous 12 mo. The sum of deaths and injuries was divided by the census population in each district. This insecurity variable was then divided into four categories to allow for nonlinear effects. Campaigns exposed to zero security-related deaths or injuries in the district in the preceding 12 mo were placed in the lowest category (referred to as secure). The remaining campaigns were allocated evenly into three additional categories based on increasing insecurity (“low,” “medium,” and “high”). This index was not intended to represent the only types of security incidents in the region. For example, we were unable to measure direct attacks on polio vaccination teams or threats of violence that did not result in death or injury.

Insecurity can be cyclical, with variations resulting from changes in political climate, weather, or religious occasions (26). To avoid systematic error resulting from these fluctuations and to better capture the enduring climate of insecurity caused by individual incidents, we chose to assess insecurity exposure over 12 mo. Sensitivity analysis was performed to test other time periods, as described below.

Statistical Analysis.

Polio case incidence and vaccination campaigns were repeatedly observed from within the same districts between 2007 and 2014. Given the nested structure of repeated measurements within districts, we used a multilevel longitudinal regression model for analysis (27). A Poisson model was used to account for the count nature of the outcome. The logarithm of the total number of target children was used as the offset for vaccination rates, and the logarithm of the census population in each district was used as the offset for polio case rates. We included time as a fixed effect to account for possible time trends that were unrelated to insecurity. Further details about these models can be found in SI Appendix, SI Methods.

A number of supplementary analyses were performed to test the key findings under modified model parameters and assumptions. (i) Each district was included as a fixed effect in the model predicting vaccination rates to account for observed and unobserved district-level confounders, such as differences in geography, population size, or development indicators. (ii) The relationship between the continuous noncategorized insecurity variable and vaccination rates was tested, assuming linearity. (iii) There were 19 campaigns that were initially planned but cancelled due to either security, weather, or other logistical reasons. These were excluded from the main analysis. In a sensitivity analysis, these campaigns were included and the vaccination rate was set to 0% and inaccessibility rate set to 100%. (iv) The effects of potentially poor-quality population data on vaccination rates were explored. The GPEI set a target number of children to be vaccinated in each campaign, resulting in campaign-to-campaign variation in the population estimate. Because of the difficulty in making accurate population estimates and capturing population movements, in a number of campaigns more people were vaccinated than the projected target, resulting in reported vaccination rates greater than 100%. The following analyses were performed to account for these issues: (i) All campaigns with a vaccination rate greater than 100% were excluded. (ii) Districts were excluded in which more than 50% of campaigns had vaccination rates greater than 100%. (iii) All campaigns with a vaccination rate greater than 100% had the vaccination rate adjusted to 100%. (iv) The distribution of campaigns with a vaccination rate greater than 100% was tested across the four categories of insecurity by χ² analysis to assess whether a skew to the distribution might have systematically biased the association between insecurity and vaccination rates. (v) The relationship between insecurity and vaccination rates was tested with the target number in each district being held constant as the mean, median, minimum, and maximum target values in each district to explore the effect of campaign-to-campaign changes in the target number of children (5). Exposure to insecurity was tested across different time periods. Vaccination campaigns were categorized by exposure to any level of insecurity at four different time periods: within the preceding 3 mo, 3–6 mo, 6–12 mo, and no security incidents within 12 mo (6). The main analyses of the relationship between insecurity and polio case incidence and polio vaccination rates were performed using 2013 population projections to calculate both the conflict index and polio incidence. For all analyses, P values were two-sided, and those of less than 0.05 were considered significant. All statistical analyses were performed using SAS 9.4.

Ethical Review.

The study was reviewed by the University of Oxford Central University Research Ethics Committee and was considered exempt from review because the study was based on an anonymous dataset with no identifiable information and was conducted for the purposes of program evaluation. Because this study was performed retrospectively using anonymous and aggregated routinely collected vaccination and polio case data, informed consent was not obtained.

Conclusion

In northwest Pakistan, higher levels of insecurity were associated with increased polio incidence. Vaccination campaigns exposed to greater insecurity were associated with a lower rate of polio vaccination, which was driven by a higher rate of inaccessible children. There was a strong negative correlation between vaccination rates and paralytic polio case rates. These findings are limited by difficulty with accurate assessments of population data by district but support the conclusion that insecurity is an important obstacle to polio eradication.