Evaluating tuberculosis service accessibility in Pakistan by analyzing population movement patterns from telecom mobility and survey data

Abstract

Background

Geographical distance is an important dimension of access to healthcare and is considered in a wide range of settings to aid strategic planning and evaluation of healthcare interventions. Evidence has shown that in addition to geographic distance, multiple factors influence the travel behavior of people seeking health services. Yet most studies have largely framed the issue from the perspective of how far people should travel, while few have explained how far people actually travel, thereby overlooking potential insights into travel behaviors.

Methods

In this research, we used in-person interviews, telecom mobility data, and spatial analysis to estimate peoples’ accessibility to TB service points in Pakistan. Characteristics of TB service points, including sector (public versus private), urbanicity, socioeconomic class, and administrative province were further analyzed for their association with TB service accessibility. We constructed decay curves to compare the accessibility rates obtained from in-person interviews among people visiting TB service points with those derived from telecom mobility data. We applied generalized linear models to assess whether these measured population movements to TB service points differ by data source, and to explore associations with the TB service point characteristics.

Results

Our results show that significant variations in TB service accessibility are associated with administrative provinces, and the density of TB service points. We found a significant difference between the geographic accessibility measured by the two data sources across distance bands. We also found that relative, and not absolute, TB service accessibility is similar across both data sources. However there is a steeper decay curve in the interview data, reflecting lower accessibility at each successively distal distance band relative to the telecom mobility data.

Conclusions

While telecom mobility data and survey data capture different population movement patterns, both provide insights which may help to align service availability with population needs. These findings are highly relevant for the geoprioritization of TB case-finding services, the efficacy of which hinge on early and efficient screening and diagnosis.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s44263-026-00329-z.

Background

Health systems globally have made enormous progress in expanding access to essential services, yet geographical access is still a challenge in many parts of the world, especially in rural and remote areas [1]. A number of frameworks help to define and describe what access to healthcare services could and should look like [2]. Their dimensions include availability, affordability or financial accessibility, geographical accessibility, accommodation, adequacy, and acceptability. Among these, geographical accessibility (also simply called accessibility) is one of the dimensions which appears in all frameworks [2–4].

Geographic accessibility describes a population’s physical access to resources [5]. It is determined by relating the location of a supply to the location of the consumer [2]. The spatial difference between supply and consumer locations may present accessibility variations in time and space [6]. Geographic accessibility can either be quantified using the travel distance or travel time needed for the consumer to reach a service delivery point [7].

The concept of geographical accessibility has been applied in diverse contexts using several methods. Travel time as a measure of geographic accessibility has been considered through various modes of transportation [8]. Straight-line distances are still used where there is insufficient data on the road network and transport to estimate road distances and times [9]. Recent methods of measuring geographic accessibility include provider-to-population ratios, Euclidean distance, gravity models, kernel density analysis, network analysis and cost-distance analysis [10].

Studies on geographic accessibility have been useful in strategic planning, and monitoring and evaluation of healthcare interventions [9, 11]. However, travel behavior of individuals is mostly unaccounted for or subjectively considered in these methods [12]. As a result, the distance people travel remains largely unexplored, including variation across health services, geographic, and sociodemographic settings. Little empirical data exist to quantify the movement patterns of people seeking healthcare. Accounting for travel behaviors in geographic accessibility studies will help to make the concept of accessibility more empirical and less hypothetical.

In-person interviews are useful to learn how perceptions of healthcare accessibility affect underserved population groups’ decision-making, including how far they usually travel to reach services [13]. Data from interviews capture service utilization, and when standardized to the surrounding population, may serve as a proxy for realized accessibility. Other sources of data can be useful in measuring distance travelled, although not to specific locations, but rather general areas, including telecom mobility data to track population movements [14, 15]. With growing access to satellite imagery and telecom mobility data, actual travel distances can be benchmarked against national and international standards to better align service availability with population needs. Accordingly, analyzing population movement patterns offers a promising approach to assess potential geographic accessibility to healthcare services.

In Pakistan, as in most other countries, knowledge of actual travel behavior to access many primary health services remains largely unknown; however, obstacles which have been shown to limit health-seeking behavior in various settings (due to ability or willingness) are present in Pakistan [16–20]. These include large geographical distances and difficulties in terrain which must be covered to access primary healthcare, and social and economic exclusionary factors [17, 19, 20]. Additionally, while urban settings usually contain a wide array of services offered in both the public and private sector, people living in remote and rural areas may not have the same options and therefore will have to travel longer distances to received preferred care [18].

In the context of tuberculosis (TB) care, this is an important issue for communities which have low geographic access to healthcare providers, as low geographic access has been shown to worsen TB treatment outcomes [21–25]. In 2024, Pakistan had an estimated incidence rate of 266 per 100,000 population and was ranked fifth globally for overall TB burden; therefore, geographic accessibility to TB care is crucial [26]. This study aims to assess the geographical accessibility of TB health services in Pakistan, by analyzing individual and population movement patterns. We consider two categories of TB service points: public-sector TB facilities and private-sector TB drug vendors. The objectives of this research are: (1) to describe distance travelled to TB service points in Pakistan using exit interviews; (2) to assess the factors which may be associated with geographic accessibility to TB service points using exit interview data; and (3) to describe whether measurement of accessibility to TB service points differs by method of data collection, i.e. through telecom mobility data versus exit interview data.

Methods

Study area and sampling of public and private sector TB service points for exit interviews

Two types of TB service points were included in this study: public sector TB facilities, where diagnosis and treatment are free and usually both provided at the same location, and private sector vendors which are distributing rifampicin, which is an anti-TB drug used in almost all first-line treatment regimens for drug-sensitive TB. The definition of the private sector in Pakistan is wide: it includes private general practitioners (GPs), private hospitals and clinics, pharmacies, and shops. Some private vendors may dispense rifampicin without evidence of a confirmed TB diagnosis, and often provide diagnostic and treatment services for a wide range of (sometimes costly) fees [27–29]. Figure 1 shows all of the included public and private TB service points in the study area.

Fig. 1.

Study area and location of all tuberculosis service points (public & private) from which the sample was drawn The grey area of the map representing other provinces such as Gilgit Baltistan and Kashmir that were not included in the study area. Karachi was considered as a distinct area apart from Sindh province given its large population and high TB service point density, but it is included in the Sindh sample in Supplementary Material 1 Tables S1 and S2.

The National Tuberculosis Program (NTP)’s master TB facility list was used as the sampling frame for all public sector TB facilities [30]. We classified each TB facility by its province or subnational area, urbanicity (urban and rural), and socioeconomic classification (SEC): classifications A & B are upper income, C is middle income and D & E are lower income. Definitions of urbanicity and SEC classification are described in Supplementary Material 1. For our sample we aimed for equal representation across all classes of province, urbanicity and SEC by randomly sampling five facilities per province-urbanicity-SEC classification, but where permissions or access were not possible, or existing facilities per class fell below five, fewer were visited. To balance resource availability with sufficient representation across classifications, we aimed to visit approximately 100 facilities in the public sector and 100 private sector vendors. The primary data collected were exit interviews conducted in October and November 2023. The overview of the final sample of public and private TB facilities per class, and individuals interviewed is provided in Supplementary Material 1 (Table S1).

IQVIA Solutions Pakistan (PVT) Limited, otherwise referred to as IQVIA, acts as a data auditor and clinical research organization (CRO) in Pakistan. IQVIA’s list of 27,830 rifampicin distributing vendors for the year 2022 served as the sampling frame for private sector TB vendors, which included pharmacies, private hospitals, clinics, dispensing general practitioners, and shops; however, these were aggregated to the level of a three- to ten-vendor geographic cluster for confidentiality. In total, there were 1,949 clusters in the study area. These clusters were also classified by province, urbanicity and SEC, with the additional classifications of private vendor types present (hospitals only, retail outlets only, hospital and retail outlets) and volume of rifampicin sales (high sales, low sales divided by the overall median). Among these, only 82 classifications actually contained vendors (i.e. actually existed). Our aim was to randomly sample at least one vendor per class (or two based on convenience to reach a sample of 100 vendors). In total, 110 private vendors were randomly selected for exit interviews (Supplementary Material 1 Table S2).

Telecom mobility data

Telecom mobility data were used to measure general population movement in areas surrounding the sampled service points. The telecom mobility data reflected the 6-month aggregate movements (January- June 2023) of Telenor mobile phone subscribers transiting from a maximum 60 km (km) radius into the location of the cluster’s geographic center [31]. It is acknowledged that movements captured by telecom data, which reflect potential accessibility, do not reflect the TB service utilization captured by exit interviews, which we interpreted as a proxy for realized accessibility. This is a limitation of the available data sources. Of the 108 public TB facilities from which primary data were collected, 90 facilities were measured by both exit interview data and telecom mobility data and therefore were included in the analysis of geographic accessibility. Telecom mobility data could be accessed to obtain population movement patterns for 77 of 82 clusters of private facilities, as five clusters had missing telecom mobility data.

Data curation

Outcome measure

We defined the outcome measure as the accessibility of each TB service point (the unit of observation) to the 60 km surrounding population, and we calculated this using two methods: (1) service point exit interview data and (2) telecom mobility data. For both data sources, we calculated the outcome measure by dividing the frequency of visits by the residing population per concentric distance band (0–5 km, 5–10 km, 10–20 km, 20–30 km, 30–60 km) (Eq. 1). These were presented as log rates to support comparability across a wide range of values for regression analyses.

Exit interviews contained questions about the geographic point of origin of visitors at public TB facilities and private TB vendors as a proxy for residence to calculate travel distance and door-to-door travel time, on the given day, regardless of their purpose for visiting the location (Supplementary Material 1). Interviews were conducted at the above-mentioned service points across the study area, and the home addresses of visitors were recorded following their informed consent, primarily by asking visitors to select a point location on a map of a handheld tablet, while a minority of manually entered addresses were later geocoded. Individuals with missing or incomplete addresses were removed from the final study database. The straight-line, as opposed to road network, distance between the service point and respondent’s geocoded residence was estimated using the GPS coordinates of both locations.

The residing population size per distance band for exit interview data was based on gridded population data generated by aiSight.ai’ through a proprietary algorithm, described in Supplementary Material 1, and for the telecom mobility data was based on the active Telenor subscriber database [31, 32]. To ensure the interview and mobility data conformed to normality, these rates were log-transformed. The accessibility per TB service point, i.e. rates of visits per residing population per distance bands, (also referred to as decay curves) were calculated based on the exit interview data and telecom mobility data separately. Rates of accessibility were compared between public and private sector service points, across sociodemographic categories, and between the two data sources.

Equation 1. TB service point geographic accessibility rate.

Covariates

The following covariates were considered for research objective two: urbanicity, SEC, province, service point density, sector, distance band, and data source. The administrative boundaries of provinces in Pakistan were obtained from Global Administrative Areas (GADM), and the public TB master facility list was obtained from Pakistan’s NTP to capture GPS coordinates [33, 34] (Table 1).

Table 1.

List of data sources

Data description	Source
Exit interviews of travel behavior of people visiting TB service points (public TB facilities and private TB vendors)	Primary data collected by IQVIA Pakistan
Telecom mobility data derived from aggregate patterns of movement of Telenor mobile phone subscribers collected over 6 months	Derived from anonymized and aggregated Call Detail Record (CDR) data from Telenor Mobile Operator (31)
Population count at 100 square meter resolution used for exit interview based accessibility indicator denominator	aiSight.ai population grid (Supplementary Material 1)
Urbanicity	Pakistan subnational administrative boundaries, aiSight.ai building footprint analysis and distance to major roads and cities from OSM (Supplementary Material 1)
Socioeconomic classes A - E	Pakistan Social and Living Standards Measurements (PSLM), Multiple Index Cluster Survey (MICS), United Nations Development Programme (UNDP) and night time light data (Supplementary Material 1)
Administrative boundaries	Global Administrative Areas (GADM) (34)
Tuberculosis master facility list	National Tuberculosis Program of Pakistan (30)

Datasets describing the SEC of the TB service points were obtained using night time light data, road density, size of individual houses, distance from main roads, property and rental values, degree of gridization of the built-up area, and socioeconomic data from Pakistan Social and Living Standards Measurements (PSLM) and UNICEF’s Multiple Index Cluster Survey (MICS) with the final SEC estimation generated by INLA modelling techniques [35–38], summarized in Supplementary Material 1. Additionally to inform the INLA model of SEC, aiSight.ai conducted a census survey of close to 800,000 retail establishments between 2021 and 2022.

Urbanicity of TB service points were derived by a Bayesian classification model using INLA on government’s administrative data, density of built-up areas and populations, size of a nucleated community, and distance (Open Source Routing Machine [OSRM] navigation time) to major roads and cities from Open Street Maps (OSM) (Supplementary Material 1).

To determine the effect of the availability of nearby facilities on the geographic accessibility rate, the density of TB service points in a 5 km radius of each service point was calculated. This was calculated as the number of public facilities and private vendors located within a 5 km straight line distance of each individual TB service point divided by the surface area of a 5 km radius circle. The facility density was then stratified into three quantiles: high, medium and low.

Lastly, distance band and the data source of the geographic accessibility outcome were included as covariates. Distance band was coded as an ordinal variable with the five aforementioned distance band categories, and data source was coded as a binary variable (interview or telecom mobility). All covariates were joined to the respective service point observation in the dataset.

Data analysis

Descriptive analysis

Using the exit interview data, the mean and median distance travelled by the interviewees to visit each service point were calculated and mapped using a free and open-source geographic information system (GIS) software known as QGIS [version 3.40.12] [39]. The median distances to each TB service point were summarized by province, SEC, urbanicity, sector, and 5 km service point density. We also produced histograms of the number of visitors and geographic accessibility rates of TB service points across the decay curves based on exit interview data, and compared the geographic accessibility rates of TB service points by source of data using R software [version 4.6.0 (2026-04-24 ucrt)] and R Studio [version 2025.09.2 + 418 “Cucumberleaf Sunflower” Release, Posit Software, PBC] [40].

Regression analyses

Regression analysis was conducted in R Studio [version 2025.09.02] and used to identify associations with TB service point geographic accessibility as the outcome variable. The factors tested for associations with TB service point accessibility (as defined in Eq. 1) included SEC, urbanicity, province, five kilometer service point density, and distance bands. We applied two Gaussian generalized linear models (GLM), one for public TB facilities and another for private TB vendors, to understand the specific sub-covariates that might explain variations in the rate of geographic accessibility. A third GLM (binomial distribution with a logit link) was conducted to determine whether telecom mobility based accessibility can be used to predict interview-based accessibility of public TB facilities after controlling for the aforementioned factors. This third GLM could only be performed for public TB facilities due to a lack of sufficient spatial overlap of private sector vendor data between collected telecom mobility data and interview data. As there were no sampled public TB facilities in FATA or FCT, these provinces were not included in the GLM regression analyses on public sector data. A check for multicollinearity was performed on all regressors. Variance inflation factors (VIF) ranged from 1.5 to 2.7, indicating low to moderate collinearity.

Results

Analysis of the travel distance to TB service points based on in-person exit interviews

First we present a description of travel distances, because this was the basis for our outcome variable: TB service point geographic accessibility. Due to the right-skewed outcome of travel distance (Supplementary Material 1 Figures S5, S6, Table S3) and large differences between mean and median distances travelled across most stratifications in the dataset, we chose to compare median distances in the stratified analyses. Across all 200 TB service points visited by exit interview respondents, the mean of the median distance travelled to a TB service point was 6.77 km (km) (95% CI 4.81, 8.74) while the median of the median travel distance was 2.66 km (95% CI 2.20, 3.48) (Table 2).

Table 2.

Summary statistics of distance traveled to tuberculosis service points according to exit interviews

Covariates	Number of sites (N = 200)	Number of visits (%)	Variation in median distance travelled by interviewees to TB service points (kilometers)
Province	N (%)	N (%)	Min	Max	Range	Mean	Median
Balochistan	37 (18.50)	1102 (20.91)	0.36	43.23	42.88	5.23	3.67
Federally Administered Tribal Areas (FATA)	1 (0.50)	24 (0.46)	20.96	20.96	0.00	20.96	20.96
Federal Capital Territory (FCT)	2 (1.00)	24 (0.46)	0.28	7.11	6.83	3.69	3.69
Karachi	11 (5.50)	153 (2.90)	0.50	8.51	8.00	2.72	1.85
Khyber Pakhtunkhwa	67 (33.50)	1606 (30.47)	0.28	145.91	145.63	9.28	2.69
Punjab	47 (23.50)	1295 (24.57)	0.54	54.34	53.80	6.31	3.26
Sindh	35 (17.50)	1067 (20.24)	0.13	66.68	66.55	5.28	1.48
Socioeconomic class (SEC)
A/B (higher income)	48 (31.58)	1280 (24.28)	0.28	29.22	28.94	4.09	2.76
C (middle income)	99 (65.13)	2629 (49.88)	0.13	54.34	54.21	6.86	2.95
D/E (lower income)	53 (34.87)	1362 (25.84)	0.21	145.91	145.69	9.03	2.26
Urbanicity
Urban	122 (61.00)	3259 (61.83)	0.21	71.94	71.72	5.47	2.66
Rural	78 (39.00)	2012 (38.17)	0.13	145.91	145.78	8.82	2.65
Sector
Public TB facilities	90 (45)	2496 (47.35)	0.13	71.94	71.81	6.65 s	2.88
Private drug vendors	110 (55)	2775 (52.65)	0.21	145.91	145.69	6.87	2.48
5-km TB service point density
High	67(33.5)	1973 (37.43)	0.13	145.91	145.78	9.51	2.47
Medium	67(33.5)	1619 (30.72)	0.21	43.23	43.02	6.15	3.67
Low	66(33)	1679 (31.85)	0.43	32.42	31.99	4.63	2.30
Total	200 (100)	5271(100)	0.13	145.91	145.78	6.77	2.66

Figure 2 presents the median distances traveled by people to the nearest TB service point including both public facilities and private vendors. Figure 2 shows that median travel distances appear high in the northern province Khyber Pakhtunkhwa, while Table 2 shows larger median distances traveled in Balochistan (3.67 km) and Punjab (3.26 km) and that the largest median travel distance observed was at the single TB service point sampled in Federally Administered Tribal Areas (FATA) (20.96 km). Median travel distance was lowest in non-Karachi Sindh at 1.48 km.

Fig. 2.

Median distance travelled by interviewees to sampled tuberculosis service points

Stratified by SEC and urbanicity, median travel distances were quite similar, although highest at 2.95 km in middle income (class C) areas, and with largest variance in lower income (class D/E) and rural areas. Distance traveled to public TB facilities was higher compared to private vendors (2.88 km and 2.48 km respectively), which is expected given the broader distribution and availability of vendors. By 5 km service point density, the shortest distance travelled was observed in low density areas (2.30 km), and was largest in medium density areas (3.67 km).

Assessment of factors associated with TB service point geographic accessibility using exit interview data

Figure 3a shows that the majority of visitors to TB facilities and vendors traveled from a distance of 0–5 km (approximately 18 and 16 visitors respectively), and that increasingly fewer visitors traveled from larger distances: only three to four visitors traveled from a distance of 5–10 km, two visitors from 10 to 20 km, and one visitor from farther distances. Figure 3a and b show an exponentially decaying relationship between residence distance band and accessibility of TB services. For both public TB facilities and TB drug vendors, the geographic accessibility rate decreased with increasing distance bands. Across the two most proximal distance bands, the observed accessibility rate was higher for private vendors than for public facilities.

Fig. 3.

Comparison of tuberculosis service point geographic accessibility between public and private service points. a Mean number of people who visited TB service points per distance band. b Mean TB service point geographic accessibility rate within distance bands

More extreme negative log rates (for example − 4.4 in non-Karachi Sindh compared to -1.42 in Khyber Pakhtunkhwa) reflect relatively lower accessibility, also compared to the reference category; whereas in the case of positive log rates, higher values reflect higher accessibility. The highest accessibility was observed in Balochistan (the reference category), although only significantly higher than that of non-Karachi Sindh province (p < 0.01). No significant differences in accessibility in either the univariate or multivariate analyses were observed by SEC, urbanicity, or 5-km TB service point density. Accessibility was highest within the 0–5 km distance band of public TB facilities, and significantly higher than each of the farther distance bands in both univariate and multivariate analyses (Table 3).

Table 3.

Effect of geographic and socioeconomic factors on accessibility of public TB facilities

Covariate		Univariate			Multivariate
	Covariate category	Log rate of accessibility (95% confidence interval)	t value #	p-value	Log rate of accessibility (95% CI)	t value #	p-value
Province (Ref: Balochistan)	Karachi	-0.47 (-11.53,10.6)	-0.08	0.93	-0.2 (-9.17,8.76)	-0.04	< 0.001
	Khyber Pakhtunkhwa	-1.08 (-4.46,2.29)	-0.63	0.53	-1.42 (-4.25,1.41)	-0.98	0.96
	Punjab	-0.65 (-3.9,2.61)	-0.39	0.7	-0.45 (-3.16,2.25)	-0.33	0.33
	Sindh (non-Karachi)	-3.94 (-7.44,-0.44)	-2.21	0.03	-4.40 (-7.29,-1.52)	-2.99	< 0.001
Socioeconomic classification (Ref: A/B: higher income)	C (middle income)	0.84 (-2.17,3.86)	0.55	0.58	0.56 (-2.23,3.35)	0.39	0.69
	D/E (lower income)	0.29 (-2.92,3.51)	0.18	0.86	0.86 (-2.25,3.97)	0.54	0.59
Urbanicity (Ref: Rural)	Urban	-1.49 (-3.82,0.84)	-1.25	0.21	-1.81 (-3.9,0.28)	-1.70	0.09
5-km TB service point density (Ref: High)	Medium	-0.60 (-3.44,2.24)	-0.41	0.68	-1.21 (-3.56,1.14)	-1.01	0.31
	Low	-0.64 (-3.48,2.21)	-0.44	0.66	-0.01 (-2.70,2.68)	-0.01	0.99
Distance bands (Ref: 0–5 km)	5–10 km	-5.90 (-8.82,-2.98)	-3.96	< 0.001	-5.90 (-8.8,-3.00)	-3.99	< 0.001
	10–20 km	-11.16 (-14.08,-8.24)	-7.50	< 0.001	-11.16 (-14.06,-8.26)	-7.54	< 0.001
	20–30 km	-17.67 (-20.58,-14.75)	-11.87	< 0.001	-17.67 (-20.57,-14.77)	-11.94	< 0.001
	30–60 km	-20.50 (-23.42,-17.59)	-13.78	< 0.001	-20.50 (-23.4,-17.6)	-13.86	< 0.001

# t-values are included as they represent the number of standard errors by which the coefficient diverges from zero, and therefore its significance.

Table 4 shows that among private sector vendors, the highest accessibility was observed in Balochistan, although only significantly higher than that of Karachi (p = 0.02). No significant differences in accessibility are observed by SEC or urbanicity. However, accessibility was significantly higher in medium compared to high service point density areas. Accessibility was highest within the 0–5 km distance band of private drug vendors, and significantly higher than each of the farther distance bands.

Table 4.

Effect of geographic and socioeconomic factors on accessibility of private TB vendors

Covariate		Univariate			Multivariate
	Covariate category	Log rate of accessibility (95% confidence interval)	t value #	p-value	Log rate of accessibility (95% CI)	t value #	p-value
Province (Ref: Balochistan)	FATA†	1.1 (-9.92,12.11)	0.20	0.85	-1.09 (-10.31,8.12)	-0.23	0.82
	Federal Capital Territory (FCT)†	-4.86 (-12.85,3.13)	-1.19	0.23	-4.46 (-11.01,2.09)	-1.34	0.18
	Karachi	-4.69 (-8.91,-0.46)	-2.17	0.03	-4.71 (-8.59,-0.83)	-2.38	0.02
	Khyber Pakhtunkhwa	1.97 (-1.1,5.04)	1.26	0.21	-0.2 (-3.06,2.67)	-0.14	0.89
	Punjab	0.01 (-3.47,3.49)	0.01	0.99	-1.79 (-4.91,1.33)	-1.12	0.26
	Sindh (non-Karachi)	4.57 (0.75,8.39)	2.34	0.02	2.89 (-0.42,6.21)	1.71	0.09
SEC (Ref: A/B; higher income)	C (middle income)	1.64 (-0.86,4.15)	1.29	0.2	0.74 (-1.69,3.17)	0.6	0.55
	D/E (lower income)	2.75 (-0.4,5.91)	1.71	0.09	0.62 (-2.29,3.52)	0.42	0.68
Urbanicity (Ref: Rural)	Urban	-2.09 (-4.4,0.22)	-1.77	0.08	-2.27 (-4.81,0.27)	-1.75	0.08
5-km TB service point density (Ref: High)	Medium	2.03 (-0.55,4.62)	1.54	0.12	2.91 (0.24,5.59)	2.13	0.03
	Low	-3.18 (-5.76,-0.59)	-2.41	0.02	0.12 (-3.07,3.32)	0.08	0.94
Distance bands (Ref: 0–5 km)	5–10 km	-7.81 (-10.6,-5)	-5.44	< 0.001	-7.81 (-10.54,-5.08)	-5.61	< 0.001
	10–20 km	-13.64 (-16.4,-10.82)	-9.51	< 0.001	-13.64 (-16.36,-10.91)	-9.79	< 0.001
	20–30 km	-17.81(-20.6,-15)	-12.42	< 0.001	-17.81 (-20.54,-15.08)	-12.79	< 0.001
	30–60 km	-18.86(-21.7,-16.05)	-13.16	< 0.001	-18.86 (-21.59,-16.14)	-13.55	< 0.001

# t-values are included as they represent the number of standard errors by which the coefficient diverges from zero, and therefore its significance.† Federally Administered Tribal Areas (FATA) was represented by a single private TB facility and Federal Capital Territory (FCT) by two; findings are not generalizable.

Comparison of TB service point accessibility as measured by different sources

Figure 4 presents TB service point accessibility measurements per data source: exit interviews conducted at public TB facilities, and telecom mobility data. Both sources of data show that as distance bands around public TB facilities get larger, accessibility to the respective service point reduces. The accessibility rates per 100,000 population for interviews were 4.78 visits (0–5 km); 1.19 visits (5–10 km); 0.17 visits (10–20 km); 0.10 visits (20–30 km); and 0.03 visits (30–60 km). The accessibility rates per 100,000 population based on telecom mobility data were 4282.54 visits (0–5 km); 1975.58 visits (5–10 km); 849.18 visits (10–20 km); 298.48 visits (20–30 km); and 211.55 visits (30–60 km).

Fig. 4.

Tuberculosis service point accessibility of public tuberculosis facilities by data source at given distance bands

The binomial GLM of interview-based accessibility is presented in Table 5. A one-unit increase in the log rate of telecom mobility-based accessibility was significantly associated with an increase of 0.10096 in the log rate of interview-based accessibility rates after adjusting for geographic and demographic factors, also reflecting the decay curve patterns shown in Fig. 4.

Table 5.

Association Between Telecom Mobility and Interview-Based Measures of Public TB Facility Accessibility: Binomial GLM Findings

Covariate	Covariate category	Log rate of accessibility (95% confidence interval)	z value	p-value
	Intercept	-9.46 (-9.72, 9.20)	-71.15	< 0.001
Telecom mobility accessibility log rate	Telecom mobility accessibility log rate	0.10 (0.07, 0.14)	5.81	< 0.001
Province (Ref: Balochistan)	Karachi	-2.07 (-2.69, -1.45)	-8.24	< 0.001
	Khyber Pakhtunkhwa	0.21 (0.02, 0.39)	2.19	0.029
	Punjab	0.35 (0.17, 0.53)	3.71	< 0.001
	Sindh (non-Karachi)	0.67 (0.48, 0.87)	6.66	< 0.001
Socioeconomic classification (Ref: A/B; higher income)	C (middle income)	1.05 (0.89,1.21)	12.87	< 0.001
	D/E (lower income)	1.25 (1.07,1.43)	13.69	< 0.001
Urbanicity (Ref: Rural)	Urban	-0.17 (-0.28,0.06)	-2.92	0.003
5-km TB service point density (Ref: High)	Medium	-0.01 (-0.13,0.10)	-0.23	0.822
	Low	-1.16 (-1.31,-1.00)	-14.28	< 0.001
Distance bands (Ref: 0–5 km)	5–10 km	-2.19 (-2.34,2.04)	-29.19	< 0.001
	10–20 km	-3.59 (-3.78,-3.40)	-36.90	< 0.001
	20–30 km	-4.78 (-5.07,4.49)	-32.17	< 0.001
	30–60 km	-6.09 (-6.40,-5.77)	-38.11	< 0.001

Urban areas showed lower accessibility than rural areas: this effect was not seen in the combined data source public sector gaussian GLM in Table 3. SEC and distance band patterns were also retained between the gaussian and binomial GLMs. The binomial GLM showed lowest accessibility in Sindh province, as opposed to Table 3’s Gaussian GLM without the telecom mobility accessibility. Service density quantiles also showed a different pattern, with low density quantiles showing the lowest accessibility.

Discussion

We have compared the distance travelled and geographic accessibility of visitors accessing TB service points across urbanicity, socioeconomic classes, geographic regions, TB service point sector, 5 km service point density, distance bands, and source of travel behavior data. Our results show that travel behavior to the service point is negatively associated with distance to the service point, as shown elsewhere [2, 3, 15, 41, 42]. Geographic accessibility varied by geographic region and urbanicity after adjusting for data source. Telecom mobility data show higher accessibility of TB service points compared to in-person interviews across all distance bands. Notably, geographic accessibility of urban public TB facilities was significantly lower than rural public TB facilities.

Telecom mobility data measured a significantly higher rate of accessibility compared to interview data, and this effect was observed across all distance bands. Other studies showed mixed results. Differences in the health system and study population limit direct comparability, yet are noteworthy as studies comparing self-reported and modelled travel time are limited. Similar to our findings, one study comparing self-reported with modelled travel time showed longer self-reported travel time (i.e. lower accessibility) among people seeking family planning services in Kenya [43]. A study of elderly travelers in Sweden showed that self-reported accessibility could both under- and over-estimate more objective measures of accessibility [44], while another study showed higher self-reported accessibility to reproductive health services compared to modelled accessibility in Pakistan. However this modelling did not utilize telecom mobility data [18]. The difference in accessibility measurements by data source in our study can be explained by the fact that telecom mobility data utilized travel data of populations owning cell phones traveling into the cell tower coverage area nearest to the respective health facilities, resulting in a much higher numerator (number of visitors) per population denominator. The interviews, however, defined visitors as the number of individuals visiting each of the 200 service points in a single day. Interviews and telecom mobility data showed a similar pattern in geographic accessibility across the decay curve, however the difference between the two data sources increased with the larger distance bands, as reflected in the much steeper decay curve for interview data compared to the telecom mobility data. This suggests that the interview and mobility data may be capturing more comparable population movement patterns among proximally located populations than among distally located populations. It may suggest that interviewed people under-report longer travel distances, and it clearly shows that mobility data capture a very different process, which is movement of large numbers of people into a commercial area as opposed to toward a specific building.

These results do not suggest that telecom mobility data are a representative measure of service point accessibility, but because they do quantify behavior of people traveling from distal to proximal areas surrounding a service point, these data can be used to describe the number of people who could, in theory be targeted by a public health campaign or screening event in the vicinity of the service point, as suggested elsewhere [45, 46]. The data could also be valuable in determining how accessible a facility or prospective facility is to the daytime commuting population, given the assumption that a visit can be made to the service point en route to or from the commercial area. Telecom mobility data may also be useful to understand the broad pattern of accessibility across distance bands, with the caveat that the actual site-specific or health-specific travel of interest may have a steeper decay curve. While more precise location information for individual mobility can be pulled from smart phone usage, this is generally more expensive, and comes with additional legal and ethical concerns [14].

Characteristics of health-seekers such as income level, SEC, health concern and health awareness play important roles in the decision-making or behavior of people seeking healthcare, including TB care [3, 13, 47–49]. However, the social environment of TB service points, as measured by their urbanity and SEC, did not appear to be associated with the travel behavior of visitors before adjusting for the data source in our study. However, after adjustment for other factors, geographic accessibility was significantly lower among urban TB service points compared to rural service points, which contrasts with other studies showing that health-care seekers in metropolitan areas had improved access [48, 50]. This observation is likely due to the large size of the total catchment area (60 km) used to calculate accessibility. This is to say that more health care seekers per population may be traveling from longer distances to access rural service points, compared to the per capita visitation rate of farther distance bands to urban service points. If a rural service point is the single or among few options servicing the rural health seekers within a 60 km radius, this will be visited by a larger proportion of the catchment population, resulting in higher accessibility relative to the urban service point, as per our definition of accessibility. Therefore these counterintuitive findings should not be interpreted to inform urban TB control planning; rather, they reflect the challenges in defining appropriate catchment population areas to study geographic accessibility.

Meaningful interpretation of the province stratified analysis of accessibility is challenging due to the large variance in urbanicity, SEC and service point density within a province, most of which span an enormous geographic area. Nevertheless, accessibility rates showed significant differences in all models, although the geographic pattern shifted in the binomial GLM including the telecom mobility term. Accessibility was significantly lower in non-Karachi Sindh compared to Balochistan among public TB facilities in the gaussian GLMs, also after adjusting for data source and data source-distance band interactions. This is notable given that in Balochistan the median travel distance was relatively longer compared to most other provinces as shown in previous literature [16]. This might be explained by the fact that Balochistan is predominantly rural, where we have theorized accessibility may be higher due to higher per capita visitation to rural service points. However in the binomial GLM, Karachi showed the lowest accessibility, as the model accounts better for the population denominator.

The association between TB service point density and accessibility appeared to vary between public and private service points. Amongst public TB facilities, accessibility was virtually the same for facilities located in high and low service point density areas, and lower in medium density areas, although insignificant. Amongst private drug vendors the situation was reversed, with significantly higher accessibility in medium density areas compared to the high-density areas. These differences were not observed after adjustment for data source in the public sector (Supplementary Material 1 Table S1). Evidence appears to be limited on how geographic accessibility to health services varies across service point density, although in an Indonesian setting there appeared to be no relationship between health care utilization rates and service point density [42]. Among public TB facilities in high density areas, there may have been a sufficient variety of and/or quality of services, thereby explaining higher accessibility or utilization, while in low density areas, few options mean higher utilization which translates to our definition of accessibility. Results of the binomial GLM challenge this hypothesis however, showing a more intuitive pattern of higher accessibility in higher service density areas.

Low to moderate multicollinearity was observed in the regression analyses, and it is possible that true urbanicity differences were somewhat absorbed by SEC. As this was an exploratory analysis of how accessibility varies across space, social contexts, and data sources, multicollinearity did not affect our conclusions; however, it is important to note that regression coefficients should not be interpreted as independent effects. It is also important to note that because the regression analyses were not specified at the time of data collection, the ability to detect statistically significant associations was not factored into sample size or power calculations, potentially leading to non-significant findings.

It may be argued that our outcome variable is more appropriately defined as the rate of utilization, which it could also be argued can be better explained by availability and relative quality of nearby service points [51, 52] rather than true geographic accessibility. Furthermore, while we used the geographic accessibility rate per 100,000 population as the outcome variable for the regression analysis, median distance travelled to each service point would have also been a valid choice. A sensitivity analysis which could be considered is to compare accessibility rates when the outcome variable is changed to median distance traveled to the facility. However geographic accessibility is a function of utilization as much as the median distance travelled. And with our method we were able to account for the residing population around the service points as a proxy of demand. We also used a 60 km cut-off, which excludes some visitors who travelled further, to exclude errors during data collection for both sources. It should also be noted that while the geographic accessibility population denominator data captured residential data, Telenor and interview-based data may have better reflected daytime movement, which does not perfectly correspond with residential denominators. Also, while Telenor is one of the largest telecom providers in Pakistan, access to additional provider data would have improved representation of travel behavior.

While road network distances more closely reflect actual travel distance than straight-line distances, literature suggests that straight-line distance is sufficient to establish relative differences in geographic access [7]. Future research could consider characteristics of visitors by including more descriptive questions in exit interviews to capture information such as age, gender, income level, and mode of transportation as potential variables in regression analyses. The latter point is especially salient given the influence of mode of transport on travel behavior of individuals [41, 53, 54].

Conclusions

This research provides insights for TB case-finding intervention development and policy planning. The finding that accessibility rates were higher in rural areas is counterintuitive and less likely a reflection of geographic accessibility than of service utilization, i.e., that many people will travel very long distances to reach the care they need. As TB health outcomes and geographic accessibility (as shown in this study) are negatively associated with increasing distance from service points, closing TB treatment coverage gaps requires measures to bridge long travel distances. Remote and less well-connected communities should be plugged into the TB care environment with appropriately matched community health interventions, sputum transportation networks, and additional diagnostic services. As presented here, estimating the number of potential beneficiaries for new services may be possible by using aggregate population movement data into commercial areas, from telecom providers. Finally, while new evidence on global estimated travel times is also available to support such planning [55], our findings show that it is important to contextualize these estimates based on the method of measurement, as modelled travel time may show different decay curves from self-reported travel behavior.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

CM and JI contributed equally to writing the main manuscript text. All authors reviewed the manuscript, and YTN, FC and ER provided substantial feedback to the manuscript. JI prepared all figures. Data were provided and curated by PS, RS, YTN, US and BB. Quantitative analyses were conducted by JI and reviewed and interpreted by CM, JI and ER. All authors read and approved the final manuscript.

Funding

YTN: Global Fund to Fight AIDS, Tuberculosis and Malaria.

Data availability

Aggregate data supporting the findings of this study are available at KIT Institute. Please contact c.mergenthaler@kit.nl to request these data or for generic data inquiries. R code generated for data analysis are available https://doi.org/10.5281/zenodo.21871922. For questions about access to proprietary private sector provider data, please contact Yee Theng at yeetheng.ng@iqvia.com. For questions about access to proprietary population grid, socioeconomic data and telecom mobility data, please contact Bilal Butt at ops@surveyauto.com.

Declarations

Consent for publication

Not applicable.

Ethical approval and consent to participate

This study was approved in October 2023 by the National Institutes of Health, Health Research Institute, National Bioethics Committee for Research (NBC-R) with Reference number 4–87/NBC- in ‘NBC-996 Approval Letter 12-10-2023’. Because of the aggregate nature of the telecom data, personal data protection was ensured, and did not require informed consent as is the case with smartphone app-based individual level GPS data. Telenor shared no information on the number of movements from any location, number of home locations or any individual phone movement data. This research conforms to the principles of the Helsinki Declaration.

Competing interests

IQVIA: YTN and RS are employed by IQVIA SOLUTIONS ASIA PTE LTD and IQVIA SOLUTIONS PAKISTAN (PVT) LTD respectively, which provide proprietary pharmaceutical sales data on a commercial basis. AiSight.ai: US and BB work at aiSight.ai, which provides proprietary data analytical services on a commercial basis. The remaining authors declare that they have no competing interests.