The sins of the church: The long-term impacts of Christian missionary praxis on HIV and sexual behaviour in Zambia

ABSTRACT

This study examines the long-term effect of Christian missionary exposure on HIV infection and related sexual behaviour in Zambia. I use distance to a historical missionary church and health facility as proxies for missionary exposure. I constructed a geocoded data set combining information on the historical locations of churches and missionary health centres with contemporary individual-level data. I find that individuals who live close to a historical missionary church have a higher likelihood of being infected with HIV. I find no significant effect of proximity to a missionary health centre on HIV. Considering that heterosexual transmission is the main channel of HIV transmission in Zambia, I analyse the effect of missionary exposure on sexual behaviour. The following patterns emerge: individuals who live close to a Protestant church are less likely to engage in premarital sexual abstinence; they also have their first sexual encounter at an earlier age, with the effect being stronger for men than women. Living near a Catholic church is associated with having a higher number of sexual partners.

KEYWORDS:

• Zambia
• Christian missionary exposure
• HIV
• sexual behaviour

JEL CODES:

• N01
• I1
• Z12
• N97

1. Introduction

In recent years, there has been a burgeoning interest in understanding the impact of exposure to Christian missionaries on various outcomes in Africa today (Becker and Woessmann 2008; Cagé and Rueda 2020; Calvi and Mantovanelli 2018; Cantoni 2015; De Haas and Frankema 2018; Meier zu Selhausen, Van Leeuwen, and Weisdorf 2018; Montgomery 2017; Nunn et al. 2014; Woodberry 2012). While the enduring effect of historical missionary education on educational outcomes has received considerable attention, less is known about the influence of missionaries on health outcomes and behaviours (Baten et al. 2021; Bolt and Bezemer 2009; Chiseni and Bolt 2020; Fourie and Swanepoel 2015; Frankema 2012; Gallego and Woodberry 2010; Kelly 1999; Montgomery 2017; Nunn 2010; Nunn et al. 2014).

Christian missionaries played a pivotal role not only in providing education but also in pioneering medical services during the colonial era in Africa (Doyle, Meier zu Selhausen, and Weisdorf 2019; Gallego and Woodberry 2010; Hardiman 2006; Kalusa 2014). Recent micro-level studies have explored the effects of historical exposure to medical missions on current health outcomes and related behaviours, revealing a significant impact on health, Human Immunodeficiency Virus (hereafter HIV) prevalence, and sexual behaviours in Sub-Saharan Africa and India (Cagé and Rueda 2020; Calvi and Mantovanelli 2018; Mantovanelli 2013; Menon and McQueeney 2020). However, due to cultural variations and differences in religious contexts across Africa, there remains a need to investigate the region-specific effects of historical exposure to Christian missionaries on health outcomes. This study aims to contribute to the religion and HIV discourse by examining the historical and contemporary influence of Christianity on sexual behaviour and HIV prevalence in Zambia (previously known as Northern Rhodesia). Zambia serves as an ideal case study due to its severe HIV epidemic and the extensive presence of missionary establishments during the late nineteenth and mid-twentieth centuries, making it one of the most missionized countries in Africa (ARHAP 2006; Gann 1968; Nakazwe et al. 2019). The fundamental question this paper seeks to address is: Did historical Christian missionary exposure affect contemporary HIV prevalence and related sexual behaviour in Zambia?

In the fight against HIV, experts in HIV policy have emphasized the importance of changing sexual behaviour as a primary preventive measure (Kalunde 1997). Studies from Uganda and Ethiopia have demonstrated the correlation between a shift in sexual behaviour and the decline in HIV prevalence (Alsan and Cutler 2013; Green et al. 2006; Mekonnen et al. 2003). Faith-based organizations (FBOs) in Zambia have played a crucial role in shaping societal values surrounding sexual behaviour through the propagation of fundamental Christian ethos (Kalunde 1997; Nakazwe et al. 2019). However, conflicting evidence exists regarding the effectiveness of condom usage, with some studies suggesting that religious groups opposing condom use may increase the risk of HIV infection (Feldman et al. 1997; Mash and Mash, 2013). Additionally, scholars have examined the relationship between colonial legacies and current HIV prevalence in sub-Saharan Africa, with some arguing that exposure to historical Catholic missionaries is associated with a decline in today's HIV infection rate (Cagé and Rueda 2020; Mantovanelli 2013), while others suggest that the transatlantic slave trade contributed to the contemporary diffusion of polygyny and related sexual behaviours that increase the risk of HIV infection (Bertocchi and Dimico 2019).

To investigate the historical and contemporary influence of Christian missionary exposure on HIV prevalence and sexual behaviour in Zambia, this study employs unique data sources. In contrast to previous studies that utilized world missionary atlases from Beach (1903) and Streit (1929), I utilize information obtained from British colonial ecclesiastical census returns for mission station locations and British colonial medical reports for missionary health centre locations. This approach improves the accuracy and comprehensiveness of missionary station data, addressing potential underreporting and underestimation of missionary influence found in previous studies (Jedwab, Meier zu Selhausen, and Moradi 2022). Using the geocoded locations of missionary stations and health centres in Zambia for the years 1948 and 1953, respectively, I measure the Euclidean distances from cluster locations provided in the Demographic Health Survey (DHS) for 2007, 2013–2014, and 2018–2019 to the nearest mission station and health centre. The analysis focuses on HIV infection rates, lifetime sex partners, age at first sex, premarital abstinence, and condom use at first intercourse as outcome variables.

The findings indicate that HIV infection rates are higher in regions near historical missionary churches, while proximity to missionary health centres does not exhibit a significant impact on HIV infection. The inculcation of Christian ethos at weekly gatherings in missionary churches may explain the stronger effect observed. Moreover, proximity to historical missionary stations is associated with a higher number of lifetime sex partners and earlier age at first sex, driven by proximity to Catholic and Protestant churches, respectively. Also, individuals living closer to a historical missionary church are less likely to practice premarital abstinence, particularly influenced by proximity to a Protestant church. Results on condom use at first intercourse are not statistically significant. It is important to note that self-reported information may have influenced these results. To mitigate potential endogeneity concerns inherent in the empirical model due to the non-random assignment of missions, I control for confounding factors such as economic and geographic variables that may have influenced missionary settlement. Additionally, I employ a method suggested by Oster (2019) to assess the extent of omitted variable bias, which indicates a smaller impact of unobservables relative to observables in the empirical estimation.

The findings of this study are consistent with those obtained by Cagé and Rueda (2020) in terms of the positive association between proximity to a missionary station and higher HIV infection likelihood. However, unlike their study, I do not find a significant impact of proximity to a mission health centre on HIV prevalence. This discrepancy may be attributed to the lower correlation between historical missionary health facilities and current health facilities in Zambia, indicating a lack of persistence in historical health services. Furthermore, this study demonstrates that the missionary atlases used by Cagé and Rueda (2020) underreport missionary stations by about 91%, resulting in an underestimation of the impact of Christian missionary exposure on HIV. More specifically, I find that when using data from the missionary atlases, the coefficient on the missionary exposure variable decreases by 93 percentage points. Essentially, by utilizing the collected and geocoded missionary station data, this study significantly enhances the measurement of the effect of Christian missionary exposure on HIV. These findings reinvigorate the concerns raised by Jedwab, Meier zu Selhausen, and Moradi (2022) that world missionary atlases used by studies in the Christian missionary discourse underestimate the impact of Christian missionary exposure. One of the primary contributions of this study to the discourse on Christian missionaries is the significant improvement in estimating Christian missionary exposure through the use of country-specific mission stations. As researchers increasingly strive to measure the impact of Christian missionary exposure on various developmental outcomes in Africa, it becomes imperative to geocode mission stations specific to each country. By doing so, we can enhance the accuracy and reliability of our estimates, allowing for more robust and nuanced analyses of the effects of Christian missionary presence.

The remainder of this paper is organized as follows: Section 2 provides a brief historical background on missionary settlement and influence in Zambia, along with a discussion on the history of FBOs and HIV/AIDS. Section 3 presents the novel data used in this analysis, describes the descriptive statistics, and outlines the empirical strategy. Section 4 discusses the results from the empirical estimations, and Section 5 concludes the discussion by suggesting plausible interventions.

2. Historical background

2.1. Missionary influence in Zambia

From the late nineteenth century into the mid-twentieth century, there was a rapid and wide expansion of missionary establishments across Northern Rhodesia (Gann 1968). By 1900, there were only about 21 mission churches, being run by about six different missionary societies. By the mid-twentieth century, close to 200 mission churches had been established by about 25 mission societies across Northern Rhodesia (Gann 1968; Rotberg 1965). The missionaries were not only establishing churches but were to a great extent responsible for providing education and health care (Gelfand 1961; Snelson 1974).

For the missionary enterprise in Northern Rhodesia, the enduring commitment to provide education and health care was not purely altruistic; education and health care were used to proselytize the locals (Snelson 1974). For example, to keep with their evangelistic mission, missionary schools emphasized that for scholars to comprehend the Bible’s message fully, they should be able to read, perform some simple arithmetic and write (also known as the Three Rs) (Snelson 1974). The missionary approach to education can be summed up in the words of Reverend William Chapman of the Primitive Methodists at Nkala mission school:

For a long time, we had been anxious to train the young in the principles of Christianity. By instructing them in reading, writing and arithmetic, we cultivate their intelligence and prepare their minds for the religious truths we wish them to retain. Our reading books are scripture stories, a passage of scripture serves as a copy, and a paragraph for a lesson in dictation. Indeed, all our teaching is conducted with the view to the development of Christian character and conduct. (Ragsdale 1986, 31)

The ability to cure various ailments using Western medicine endowed missionaries with significant hegemony over Africans (Vaughan 1991). The reputation of missionary medicine drew many natives to their health centres, and missionaries utilized such opportunities to preach the Gospel. Essentially, every patient who visited any missionary health centre was considered a potential convert (Vaughan 1991). This strategy became popularly known as the ‘Gospel of the Syringe’ (Good 1991). Hardiman (2006) noted that while patients were waiting to seek medical attention, missionary doctors and their assistants would take this time to inculcate the essential Christian doctrines in the patients.

The churches, schools, and health centres were all used as platforms to disseminate the Gospel. To a great extent, the missionaries’ message was focused on pointing out the sin in some of the traditional customs practised by the locals and imploring them to live the right way as prescribed by the Bible (Rotberg 1965). Although the missionaries abhorred numerous African traditions, a great emphasis was placed on forsaking polygyny as essential for salvation (ibid).¹ Polygyny was a common practice and the ideal type of marriage in many African societies, and as pointed out by early missionaries, it became a great impediment to African conversion (Amanze 1998; Hastings 1967; Webster 2018). Upon conversion, missionaries implored the Christian men to forsake all other wives and commit to only one (Falen 2008). For the case of Uganda, Meier zu Selhausen, Van Leeuwen, and Weisdorf (2018) point out that, although men swore to commit to one wife before the Christian marriage rites were performed, many Christian men still went and traditionally took on one or more wives, essentially going against the teachings of the missionaries. In Northern Rhodesia, as in Benin, Falen (2008) and Rotberg (1965) assert that as the African converts read the Bible, they began to question the teachings of the missionaries on polygyny since the Bible did not explicitly discourage congregants from engaging in polygamous relationships but rather encouraged only church leaders to be husbands of one wife. Dr Walter Fisher, a medical missionary at Kalene, implored other Protestant missionaries not to deny the sacraments. The conflicting views on marriage between the locals and missionaries ultimately led to various exceptions among the various missionary societies, more so for the Protestants than the Catholics (Falen 2008). Barrett (2018) even postulated that the exceptions on marriage made the Protestant missionaries more attractive to the Africans than the Catholic missionaries.

2.2. FBOs and HIV/AIDS in Zambia

In the early 1980s, there was an increase in the incidence of Kaposi’s sarcoma, an increase associated with poor treatment responses among patients with the condition (Bayley et al. 1985).² The manifestation of poor treatment responses among patients diagnosed with Kaposi’s sarcoma was not unique to Zambia; surrounding regions in Central Africa were also making similar observations among patients with the condition (Setel, Lewis, and Lyons 1999). Surgeons at the University Teaching Hospital in Zambia later associated the atypical manifestations of poor responses to sarcoma with the presence of HIV-1 antibodies (ibid). By 1984, Zambia had recorded its first cases of AIDS (Bayley et al. 1985). Towards the end of the 1980s, Zambia had not yet conducted any population-based sampling of HIV incidence; the Ministry of Health had based its known numbers of HIV incidence on estimates extrapolated from varying cohort studies by medical scholars (Fylkesnes, Brunborg, and Msiska 1994). Researchers carried out the first population-based samples of HIV between 1995 and 1996. With the help of these samples, the Ministry of Health determined the HIV incidence around urban areas, specifically Lusaka and Kapirimposhi (Setel, Lewis, and Lyons 1999).

To complement the localized serological surveys, the Zambian government carried out sentinel site surveys between 1994 and 1996; the findings from these site surveys indicated that HIV prevalence in Zambia during the survey ranged from 1.6% to 31.9% (Webb 1996). The surveys also showed that HIV prevalence was higher in rural areas than in peri-urban areas and higher among females than men (ibid). By 1997, it was estimated that from a population of about 800,000, 17% were HIV positive, with a total of about 150,000 symptomatic cases (Setel, Lewis, and Lyons 1999). Figure 1 illustrates the development of HIV prevalence between 1990 and 2019 in Southern Africa. As can be seen, Zambia has had one of the highest HIV prevalences in the region; in 1990, the HIV prevalence among those between the ages of 15 and 49 years stood at 8.9%, ranking as the second highest in the region, following Zimbabwe at a 15.3% prevalence rate. In the last three decades, although Zambia’s HIV prevalence has become one of the lowest in the region, in relation to the rest of the world, the HIV prevalence rate remains very high at 11.5% as of 2019, with the highest in the region being Eswatini at 27% (World Bank 2020).

Figure 1. HIV prevalence in Southern Africa from 1990 to 2019. Data source: World Bank Data (2020).

In the initial stages of HIV in Zambia, the government was not very vocal about the dangers of HIV. Nevertheless, as cases of HIV/AIDS continued to soar, the government saw fit to increase its intervention programmes and efforts in sensitizing people about the nature of the virus (Setel, Lewis, and Lyons 1999). But with limited resources, government interventions to fight the spread of HIV were limited; ultimately, there was a great reliance on international donor communities to aid in the fight against the virus (ibid). As part of its intervention in the fight against HIV, the Ministry of Health emphasized control of sexually transmitted diseases (STDs) through prenatal blood screening, an increase in public campaigns against HIV targeting the youth, large-scale distribution of condoms, and incorporation of traditional healers into their AIDS care and prevention programmes. Furthermore, because Christianity was and still is a prominent facet of most people living in Zambia, the government and the international donor communities began to call on churches to assist them in the fight against aids. The approach of the FBOs in Zambia has somewhat differed from that used by the government and other non-governmental organizations. While the latter have advocated for abstinence, faithfulness and condomization (ABC), FBOs have mostly focused on encouraging their congregants only to abstain and be faithful. As earlier pointed out, FBOs fear that if they put forward the message of condomization, this strategy will encourage sexual promiscuity among congregants (Mash and Mash 2013). Although the government and non-governmental organizations clash with the FBOs on condomization because of their pronounced influence, FBOs are still considered essential in the fight against HIV/AIDS in Zambia (Wiginton, King, and Fuller 2019).

3. Data and empirical strategy

3.1. Data

3.1.1. Contemporary data

To study the long-term impact of Christian missionaries on sexual behaviour and HIV prevalence in Zambia, this study uses three waves of the Zambian Demographic Health Survey (hereafter, ZDHS), specifically the 2007, 2013 and 2018–2019 surveys. The ZDHS is a nationally representative sample of about 13,625 households. It includes women aged 15–49 and men aged 15–59 from the selected households and captures various individual and household characteristics. The household characteristics include the type of dwelling unit, water sources accessed by the household, type of toilet facility used, and possession of durable goods. The individual characteristics of household respondents include background characteristics such as age, marital status, religion, sex, occupation, and education level. From the year 2000 onwards, DHS began to collect biomarkers on STDs; these are biological measures of an individual’s health condition. Essentially, the data used in this study on HIV status are not self-reported data but rather based on actual HIV test results.³ The DHS collects dried blood samples (DBSs), which are then sent to the laboratory for HIV testing.

As of 2007, the ZDHS provided geo-referenced information for the household groupings that participated in the survey, referred to as clusters. Figure 2 shows the distribution of surveyed clusters; there are 318 from the 2007 DHS data, 720 from the 2013 DHS data, and 544 from the 2018 DHS data. In total, there are 1582 surveyed clusters across all 73 districts in Zambia. To ensure respondent confidentiality, global positioning system coordinates for urban clusters have 2 km of positional error, and rural clusters have 5 km of error. The cluster coordinates are imperative for this study’s empirical strategy; these coordinates allow for calculating Euclidean distances from a respondent’s place of residence to the nearest missionary establishment. Here, only respondents who indicated that the place where they were surveyed was their usual place of residence are included in the analysis; ultimately, the final data set after pooling all three ZDHS waves contains 63,224 individuals coming from 1572 distinct locations.

Figure 2. Demographic Health Survey (DHS) cluster locations.

3.1.2. Historical data

I obtained historical information on Protestant and Catholic mission station locations from the ecclesiastical reports recorded in the Northern Rhodesia Colonial Blue Books for 1948. By 1948, there were 212 Protestant and Catholic churches established in Northern Rhodesia, with about 61 churches belonging to the Catholic missionaries and 137 to Protestant missions. Using Geographic Information System (hereafter, GIS), I geocoded the exact locations of these missionary churches; Figure 3 shows the distribution of Protestant and Catholic churches across Northern Rhodesia. Although we see a slight degree of clustering among Protestant and Catholic missionaries, there is a significant dispersion between Catholic and Protestant missionaries to a greater extent. Because of different doctrinal positions, there was little to no cooperation between the Catholic and Protestant missionaries, and cooperation mostly existed among Protestant missionaries. I extracted information on the location of Protestant and Catholic mission health centres⁴ from the 1953 colonial medical report contained in the Northern Rhodesian Colonial Blue Books. Using GIS, I geocoded the locations of the Protestant and Catholic mission health centres in Northern Rhodesia for 1953, and Figure 4 shows the distribution of these health centres. There were approximately 70 missionary health centres spread across Northern Rhodesia, 20 of which belonged to Catholic missionaries, and the rest were operated by Protestant missionaries. Ultimately, I combine the DHS cluster locations from the three waves with missionary church and health centre locations in GIS and calculate the Euclidean distance from each DHS cluster location to a historical missionary church and health centre.

Figure 3. Protestant and Catholic missionary churches in Northern Rhodesia, 1948.

Figure 4. Protestant and Catholic mission health centres in Northern Rhodesia, 1953.

3.1.3. Geographic data

Using GIS, I created 30 km buffers around each DHS cluster location. Utilizing data on land elevation obtained from DIVA-GIS, I then calculated the mean land elevation in each 30 km buffer. I digitized the historical railway line in Zambia using the information on contemporary rail data from DIVA-GIS. The railway line running from Livingstone to the Copperbelt was constructed in three phases: the first phase running from Livingstone to Monze was completed in 1905, the second phase from Monze to Broken Hill was completed in 1906, and the last phase from Broken to Kitwe was completed in 1909. Using the new historical rail data, I calculated the shortest Euclidean distances from each DHS cluster location to the railway. I obtained geographic data on rivers in Zambia from Open Knowledge International (OKI). Using this data, I measured the Euclidean distance from each DHS cluster location to the nearest river. Additionally, using explorer route data from Nunn (2008), I measured the Euclidean distance from each DHS cluster location to the closest historical explorer route.⁵ I then constructed binary variables indicating whether the DHS clusters are within 10 km of a river, 50 km of the railways or 50 km from an explorer route. Several studies have established that these geographic features influenced missionary settlement; if not controlled for in analysing the long-term impact on our outcome variables, they could potentially bias the estimated impact of missionaries (Jedwab, Meier zu Selhausen, and Moradi 2022; Johnson 1967; Mantovanelli 2013).

3.2. Descriptive statistics

Table A1 in the Appendix provides the summary statistics for the outcome and explanatory variables used in the empirical estimations. Noticeably in Table A1, 13% of respondents in the sample tested HIV positive, and HIV infection is higher among females relative to males. Regarding condom use at first intercourse, 30% of the individuals reported using a condom during their first intercourse; there is no large difference between males and females. Using the individual’s age at first sex and age at first marriage, I created a new variable, premarital abstinence, that indicates whether an individual waited until marriage for their first sexual encounter. As reported in Table A1, 22% of the sampled individuals reported having practised premarital abstinence, with more females than males indicating having engaged in it. On average, males had more lifetime sexual partners than females. Almost 81% of the surveyed population belong to a Protestant church, and 82% of historical missionary churches are Protestant.

Table 1. Christian missions and Human Immunodeficiency Virus (HIV) infection.

Dependent variable: HIV status
	1	2	3	4
Distance to church	−7.81e-05*	−0.00159**	−0.00151**	−0.00155**
	(4.20e-05)	(0.000637)	(0.000700)	(0.000650)
Distance to health centre	−0.0255**	−0.0175	−0.0169	−0.0199
	(0.0130)	(0.0125)	(0.0132)	(0.0130)
Protestant	−0.0216	−0.00238	−0.0195	−0.00586
	(0.0243)	(0.0242)	(0.0267)	(0.0265)
Age	0.113***	0.113***	0.126***	0.113***
	(0.00679)	(0.00680)	(0.00795)	(0.00762)
Age^2	−0.00127***	−0.00128***	−0.00140***	−0.00128***
	(9.38e-05)	(9.40e-05)	(0.000108)	(0.000104)
Married	0.0606*	0.0481	−0.00754	0.00716
	(0.0359)	(0.0357)	(0.0378)	(0.0361)
Female	0.173***	0.176***	0.249***	0.266***
	(0.0184)	(0.0183)	(0.0204)	(0.0254)
Urban	0.425***	0.374***	0.380***	0.329***
	(0.0266)	(0.0299)	(0.0332)	(0.0328)
Lifetime sex partners				0.0177***
				(0.00204)
Age first sex			−0.0258***
			(0.00364)
Asset index
Lowest index			0.132**	0.103**
			(0.0542)	(0.0515)
Low index			0.181***	0.163***
			(0.0459)	(0.0427)
Medium index			0.133***	0.124***
			(0.0425)	(0.0400)
High index			0.157***	0.155***
			(0.0436)	(0.0410)
Education level
No education			−0.115*	−0.0480
			(0.0622)	(0.0636)
Primary education			0.0219	0.0571
			(0.0454)	(0.0491)
Secondary education			0.105**	0.129***
			(0.0425)	(0.0457)
Constant	−3.448***	−3.897***	−3.516***	−3.803***
	(0.124)	(0.178)	(0.233)	(0.220)
Provincial FE	Yes	No	Yes	Yes
Occupational FE	No	No	No	Yes
Geographic controls	No	Yes	Yes	Yes
Observations	56,350	55,946	42,235	46,374

Note: Probit model estimates. Clustered robust standard errors at the DHS cluster level are given in parentheses. I use multiple correspondence analysis to calculate an asset index based on whether an individual owns a television set, bicycle, radio, refrigerator, motorcycle, truck, or electricity. The index ranks from lowest to highest depending on what the individual owns. The highest ranking is given to an individual who owns the most assets and the lowest is assigned to those who own the fewest; the comparison group in this analysis is individuals with the highest number of assets. The comparison group for education level is individuals with the highest level of education. Distance to the health centre is log-transformed. Geographic controls include whether a cluster is within 50 km of an explorer route, 50 km of the railway, or 10 km of a river, and mean land elevation in logarithmic form. For all the models, I confine the regressions only to those who answered that where they were living was their usual place of residence. Then I also include occupational and provincial fixed effects. ***p < .01, **p < .05, *p < .1.

3.3. Empirical strategy

To study the long-term impact of early Christian missionaries on HIV infection, I estimate the following equation: (1) $Pr\left(Y_{\mathrm{ij}}=1\right)=\sigma +{\text{β}}_1\mathrm{Church}\mathrm{Distanc}{e}_{\mathrm{ij}}+{\text{β}}_2\mathrm{Hospital}\mathrm{Distanc}{e}_{\mathrm{ij}}+{\text{β}}_c{X}_{\mathrm{ij}}+\theta V_j+{\mathrm{D}}_{\mathrm{c}}\mathrm{Province}+{\mathrm{D}}_{\mathrm{r}}\mathrm{Occupation}+{\in }_{\mathrm{ij}}$(1)

The subscript i represents the surveyed individual and j their cluster location; in Equation (1) $Y_{\mathrm{ij}}$ is a dichotomous dependent variable representing the HIV status of individual i belonging to cluster j, which takes on the value 1 if an individual tested HIV positive and 0 otherwise. $\mathrm{Church}\mathrm{Distanc}{e}_{\mathrm{ij}}$ measures the minimum Euclidean distance in kilometres from a historical missionary church to the individual i’s location. The variable $\mathrm{Hospital}\mathrm{Distanc}{e}_{\mathrm{ij}}$ is the Euclidean log distance from a historical missionary health centre to individual i’s cluster j location, which allows approaching the change in $Y_{\mathrm{ij}}$ given a percentage change in the minimum distance from a historical health centre. The vector $X_{\mathrm{ij}}$ captures various individual characteristics such as age, age at first sex, number of lifetime partners, asset index, marital status, and education level attained by the individuals. I control for vector $V_j$, which includes a host of geographical variables; these include a dichotomous variable that captures individuals living within 10 km of a river, 50 km of a historical railway line, or 50 km of a historical explorer route. Additionally, in the geographic vector, I control for the average land elevation for each 30 km cluster buffer. To account for any omitted variable bias that may arise due to provincial unobserved heterogeneity, I include province, which captures the provincial fixed effects in the model. Further included in the model are occupational fixed effects represented by the variable occupation.

Furthermore, I analyse the impact of Christian missionaries on sexual behaviour by estimating variations of the following equations: (2) $Pr({\ddot{Y}}_{\mathrm{ij}}=1)=\mathrm{\partial }+{\text{β}}_3\mathrm{Church}\mathrm{Distanc}{e}_{\mathrm{ij}}+{\text{β}}_4\mathrm{Hospital}\mathrm{Distanc}{e}_{\mathrm{ij}}+{\text{β}}_r{X}_{\mathrm{ij}}+\theta V_z+{\mathrm{D}}_{\mathrm{k}}\mathrm{Province}+{\mathrm{D}}_{\mathrm{r}}\mathrm{Occupation}+{\vartheta }_{\mathrm{ij}}$(2) (3) ${\gamma }_{\mathrm{ij}}=\mathrm{\varnothing }+{\text{β}}_5\mathrm{Church}\mathrm{Distanc}{e}_{\mathrm{ij}}+{\text{β}}_6\mathrm{Hospital}\mathrm{Distanc}{e}_{\mathrm{ij}}+{\text{β}}_n{X}_{\mathrm{ij}}+\theta V_m+{\mathrm{D}}_{\mathrm{L}}\mathrm{Province}+{\mathrm{D}}_{\mathrm{r}}\mathrm{Occupation}+{ϵ}_{\mathrm{ij}}$(3) In Equation (2), ${\ddot{Y}}_{\mathrm{ij}}$ is a dichotomous variable capturing condom use at first intercourse, which takes on the value 1 if an individual i belonging to cluster j used a condom during their first sexual experience and 0 otherwise. In the second estimation of Equation (2), ${\ddot{Y}}_{\mathrm{ij}}$ is a dichotomous variable representing premarital abstinence, which takes on the value 1 if an individual i belonging to cluster j waited until marriage for their first sexual encounter and 0 otherwise. Similarly, I estimate two variations of Equation (3), with ${\gamma }_{\mathrm{ij}}$ being a continuous outcome variable capturing either age at first sex, which is the age at which individual i from cluster j had their first sexual encounter; or number of lifetime sex partners – that is, the number of lifetime sexual partners that individual i of cluster j has had in their lifetime. I include in the variants of Equation (2) and (3) some of the explanatory variables used in Equation (1).

4. Results

4.1. Historical Christian missionaries and contemporary HIV

To estimate the impact of distance to a missionary church and health centre on HIV infection, I estimate Equation 1 with HIV status as the dependent variable. The results are presented in Table 1, columns (1) to (4). In column (1), I control for provincial fixed effects but do not include occupational fixed effects and geographic controls; the results show that distance to a historical church is negatively related to HIV infection at the 10% significance level; however, the magnitude of the impact of a historical church on HIV infection is minute. The distance to a historical missionary health centre is also negatively correlated with HIV infection but is not statistically significant. In columns (2) to (4), I gradually include geographic controls and occupational fixed effects; the results show that proximity to a historic church is still negatively correlated with HIV infection, and the absolute value of the coefficient on distance to a church increases in magnitude as I control for geographic controls and occupational fixed effects; the coefficient is statistically significant at the 5% significance level. With additional controls in columns (2) to (4), the coefficient on the logarithmic distance to a historical mission health centre remains negative but statistically non-significant. The results regarding the impact of distance to a historical church on HIV infection suggest that a 1 km increase in distance from a historical church reduces an individual’s probability of getting infected with HIV by about 0.0002; the results are consistent across columns (2) to (4). Essentially, an increase by 10 km in distance from a mission station would result in a 0.002 reduction in the likelihood of being infected with HIV. The results with regards to the impact of proximity to a mission station are similar to those obtained by Cagé and Rueda (2020) on proximity to a missionary settlement. By contrast, Cagé and Rueda (2020) find that proximity to a missionary health facility reduces the likelihood of HIV infection, while in this study I find that proximity to a missionary health facility has no impact on HIV. Additionally, in Table 2 I decompose the main results by rural and urban areas; the results show that the impact of missionary exposure on HIV is only negative and significant in rural areas.

Table 2. Christian missions and HIV infection in urban and rural areas.

Dependent variable: HIV status
Variables	Urban	Rural
Distance to church	0.000385	−0.00243***
	(0.000983)	(0.000841)
Distance to health centre	0.00461	−0.0256
	(0.0169)	(0.0193)
Protestant	−0.0280	−0.000585
	(0.0379)	(0.0372)
Age	0.138***	0.0918***
	(0.0121)	(0.00982)
Age^2	−0.00160***	−0.00101***
	(0.000169)	(0.000132)
Married	−0.0206	0.0191
	(0.0493)	(0.0499)
Female	0.361***	0.170***
	(0.0397)	(0.0323)
HIV knowledge	0.0131	0.0163**
	(0.00958)	(0.00678)
Lifetime sex partners	0.0179***	0.0162***
	(0.00357)	(0.00232)
Lowest asset index	0.144*	0.286***
	(0.0870)	(0.0847)
Low asset index	0.207**	0.145***
	(0.0829)	(0.0534)
Medium asset index	0.237***	0.0810*
	(0.0842)	(0.0452)
High asset index	0.274***	0.106**
	(0.0863)	(0.0469)
No education	−0.0551	−0.0455
	(0.103)	(0.113)
Primary education	0.145**	−0.0409
	(0.0608)	(0.103)
Secondary education	0.137***	0.0769
	(0.0516)	(0.102)
Constant	−4.003***	−3.446***
	(0.336)	(0.291)
Provincial FE	Yes	Yes
Occupational FE	Yes	Yes
Geographic controls	Yes	Yes
Observations	18,302	27,476

Note: Probit model estimates. Clustered robust standard errors at the DHS cluster level are given in parentheses. I use multiple correspondence analysis to calculate an asset index based on whether an individual owns a television set, bicycle, radio, refrigerator, motorcycle, truck, or electricity. The index ranks from lowest to highest depending on what the individual owns. The highest ranking is assigned to individuals who own the most assets and the lowest is assigned to those who own the fewest; the comparison group in this analysis is individuals with the highest number of assets. The comparison group for education level is individuals with the highest level of education. Distance to health centre is log-transformed. Geographic controls include whether a cluster is within 50 km of an explorer route, 50 km of the railway, or 10 km of a river, and mean land elevation in logarithmic form. For all the models, I confine the regressions to those who answered that where they were living was their usual place of residence. Then I also include occupational and provincial fixed effects. ***p < .01, **p < .05, *p < .1.

I re-estimate Equation (1), differentiating between distance to a Protestant versus a Catholic church; the results in Table A2 show that the distance to a Protestant or a Catholic church has a negative and significant impact on HIV infection at the 5% and 10% significance level, respectively. The negative coefficient on distance to a Protestant or Catholic church suggests that a 1 km increase in this distance decreases one’s likelihood of being infected with HIV by 0.0003. Essentially, an increase of 10 km in distance from a mission station would result in a 0.003 reduction in the likelihood of being infected with HIV. The coefficient on the logarithmic distance to a mission health centre is negative and statistically significant when controlling for distance to a Protestant church. The results in Table A2 show that a 1% increase in logarithmic distance from a Protestant missionary health centre controlling for distance to a Protestant church decreases the probability of being infected with HIV by 0.004.

I explore the heterogeneous impact of proximity to a missionary church and health centre on HIV infection; I estimate Equation (1) for females and for males. The results in Table A3 suggest that for both males and females, distance to a missionary church is negatively related to HIV infection and is statistically significant at the 5% and 10% significance level, respectively. These results imply that for females, a 1 km increase in distance from a historical church reduces the probability of HIV infection by 0.0003. The results for males suggest that a 1 km increase in distance from a historical church decreases the probability of HIV infection also by 0.0003. Distinguishing between males and females, I find no significant effect of distance to a missionary health centre on HIV infection.

In Table 1, I find no impact of proximity to health care on HIV. I postulate that this could be because of a lack of persistence in historical health facilities in Zambia. I use a grid cell approach to understand how historical missions correlate with current health services. In GIS, I partition Zambia into 7134 grid cells. In each cell, I determine whether there is a historical mission station or health centre. In the same fashion, I also ascertain whether a cell contains a contemporary health centre or not. Ultimately, with the new data set, I calculate simple correlations to determine whether the presence of a historical mission station or health centre in a cell correlates with the presence of a current hospital in that specific cell. Figure 5 presents simple correlations between current and missionary health centres and missionary stations. As shown in the figure, the correlation between current health centres and missionary health centres and stations is very low.

Figure 5. Correlation between current health centres and missionary health centres and stations in Zambia.

Consequently, I postulate there is no strong persistence in missionary health centres; therefore, this may explain why proximity to a missionary health centre has little to no significant effect on HIV and related sexual behaviour. Although I have not been able to obtain data on current churches in Zambia, I assert that there is a stronger persistence of missionary churches, hence the significant effect of proximity to a historical missionary church on HIV and related sexual behaviour. Cheyeka, Hinfelaar, and Udelhoven (2014) conducted a case study on Christianity in Bauleni; Bauleni is a township in Lusaka, Zambia. They estimated that by 2010 there were 82 churches in Bauleni township alone, serving a population of about 26,000 individuals, which is almost one church per 318 people. Additionally, the persistence of Christianity in Zambia is also reflected in the proportion of the population that adheres to Christianity, an estimated 85% (Phiri 2019) and in the 1991 constitutional adoption of Christianity as the national religion (Cheyeka, Hinfelaar, and Udelhoven 2014).

4.2. Christian missionaries, HIV, and related sexual behaviours

The main channel of HIV transmission that has been identified in Zambia, as in other countries within the region, is heterosexual contact; essentially, recommendations by HIV experts imply that a change in sexual behaviour would significantly help in the fight against HIV (Kalunde 1997). For this reason, many FBOs have attempted to help in the fight against HIV by influencing and shaping societal values surrounding sexual behaviour, through the promulgation of Christian praxis (Duflo, Dupas, and Kremer 2015; Mash and Mash 2013; Nakazwe et al. 2019). Primarily, contemporary Christian FBOs, like early Christian missionaries, still advocate for abstinence before marriage, commitment and faithfulness to one partner, and early marriage rather than condom use (Mash and Mash 2013). To understand the mechanisms through which distance to a missionary church and health centre may affect sexual behaviour, I use the following proxies for societal values propagated by FBOs to estimate equations (2) and (3): the number of lifetime sex partners, age at first sex, age at first marriage, and condom use at first intercourse. It should be emphasized that unlike the measure for HIV status that captures the individual’s actual HIV status, the variables relating to sexual behaviour are self-reported; therefore, these are imperfect proxies and should be interpreted with caution.

4.2.1. Christian missionaries and lifetime sex partners

In the fundamental Christian praxis, a strong emphasis has been placed on faithfulness to one life partner. In light of this, we expect to find a lower propensity for multiple sexual partners for individuals who live near a mission station or health centre than for those who live farther away. Table 3 presents the results of the impact of distance to a historical church on the number of lifetime sexual partners an individual has had. Contrary to the expectations, in column (1), the coefficient on the distance to a historical church is negative and statistically significant at a 5% significance level, suggesting that individuals who live closer to a missionary station tend to have more sexual partners than those who live farther away. Cagé and Rueda (2020) find an opposite effect of proximity to a mission station on lifetime sex partners; they find that those who live close to a mission station tend to have fewer sexual partners. The difference in results may be due to the fact that the sexual behaviours are not observed but are rather self-reported and thus may succumb to self-reported bias. Additionally, attitudes towards sex and the stigma related to it may differ from country to country; Cagé and Rueda (2020) estimate the average sexual behaviours in Africa, unlike the present study that looks at country-specific sexual behaviours, and the differences in units of analysis may produce varying results.

Table 3. Impact of proximity to a mission church on sexual behaviour.

	(1)	(2)	(3)
	All missions	Protestant	Catholic
Lifetime sex partners	−0.00351**	−0.00266	−0.00713***
	(0.00162)	(0.00184)	(0.00265)
Age at first sex	0.00223*	0.00216*	0.00165
	(0.00114)	(0.00121)	(0.00255)
Premarital abstinence	0.00110**	0.00142**	0.000906
	(0.000527)	(0.000585)	(0.00113)
Condom use	0.000876	0.000426	0.00302
	(0.00101)	(0.00110)	(0.00261)
Provincial FE	Yes	Yes	Yes
Occupational FE	Yes	Yes	Yes
Geographic controls	Yes	Yes	Yes
Individual controls	Yes	Yes	Yes

Note: Probit model estimates. Clustered robust standard errors at the DHS cluster level are given in parentheses. The comparison group for the asset index is individuals with the highest index. The comparison group for education level is individuals with the highest level of education. Distance to health centre is log-transformed, and distance to church is given in kilometres. Geographic controls include whether a cluster is within 50 km of an explorer route, 50 km of the railway, or 10 km of a river, and mean land elevation in logarithmic form. For all the models, I confine the regressions to those who answered that where they were living was their usual place of residence. Individual controls: age, age squared, married, female, condom knowledge, condom access, urban, P. ***p < .01, **p < .05, *p < .1.

Noticeable from Table 4, column (1), distance to a missionary health centre does not significantly impact the number of lifetime sexual partners an individual engages with. I further explore the heterogeneous effects of distance to Protestant and Catholic missions on the number of lifetime sexual partners; Table 3, columns (2) and (3) present these results. The findings demonstrate that distance to a Protestant mission station is not significantly related to the number of lifetime sexual partners. Surprisingly, results show that those who live near a Catholic historical church tend to have more sexual partners than those who live farther away. These results imply that the effect of distance to a historical church on the number of lifetime sex partners in column (1) is mainly driven by distance to a historical Catholic church.

Table 4. Impact of proximity to a mission health centre on sexual behaviour.

	(1)	(2)	(3)
	All missions	Protestant	Catholic
Lifetime sex partners	−0.0708	−0.0414	−0.0687
	(0.0448)	(0.0482)	(0.0627)
Age at first sex	−0.00927	−0.0175	−0.0349
	(0.0255)	(0.0276)	(0.0696)
Premarital abstinence	0.00274	−0.00191	−0.00305
	(0.0106)	(0.0124)	(0.0237)
Condom use	0.00952	0.00819	−0.0263
	(0.0222)	(0.0230)	(0.0749)
Provincial FE	Yes	Yes	Yes
Occupational FE	Yes	Yes	Yes
Geographic controls	Yes	Yes	Yes
Individual controls	Yes	Yes	Yes

Note: Probit model estimates. Clustered robust standard errors at the DHS cluster level are given in parentheses. The comparison group for the asset index is individuals with the highest index. The comparison group for education level is individuals with the highest level of education. Distance to health centre is log-transformed, and distance to church is given in kilometres. Geographic controls include whether a cluster is within 50 km of an explorer route, 50 km of the railway, or 10 km of a river, and mean land elevation in logarithmic form. For all the models, I confine the regressions to those who answered that where they were living was their usual place of residence. Individual controls: age, age squared, married, female, condom knowledge, condom access, urban, protestant, lifetime partners, asset index, education level. ***p < .01, **p < .05, *p < .1.

I conducted a further inquiry into the varying effects of Christian missionaries on lifetime sexual partners by gender; the results are presented in Table A4. As seen in Table A4, columns (1) and (2), the effect of distance to a Protestant mission on lifetime sexual partners remains non-significant for both males and females. In columns (3) and (4), it is noted that the coefficient on the distance to a Catholic mission station continues to be negative and significant, entailing that females and males living closer to a Catholic mission station tend to have more lifetime sexual partners than those who live farther away. For all results presented in this section, I find that distance to a missionary health centre continues to have no significant effect on the number of lifetime sexual partners. It is intriguing to find that individuals who live closer to a Catholic church tend to have more lifetime sexual partners because, as pointed out earlier, both Catholics and Protestants tend to emphasize the importance of faithfulness to one partner; consequently, I expected individuals who live near a Catholic or Protestant church to have fewer lifetime sexual partners. In light of the findings, I note that Tiruneh (2009) puts forward an interesting argument that even though Christian organizations encourage commitment to one life partner and premarital sexual abstinence, these views are not strictly enforced, so there is a high recalcitration likelihood among congregants.

4.2.2. Christian missionaries and premarital abstinence

Another fundamental Christian doctrine emphasized by churches is that sex before marriage and outside marriage is a sin;⁶ indeed, church leaders encourage their congregants to remain sexually abstinent until marriage. Considering this, I expected to find a negative relationship between proximity to a missionary church and premarital abstinence; principally, an individual’s likelihood to abstain from premarital sex should increase with proximity to a missionary church. Contrary to perceived expectations, Table 3 column (1) reveals that there is a positive and significant relationship between distance to any historical missionary church and premarital abstinence, entailing that individuals who live farther away from a historical missionary church are more likely to abstain from premarital sex than those living nearby.⁷ In columns (2) and (3) in Table 3, I consider the variegated effects of distance to a Protestant or Catholic church on premarital abstinence; the differences are noticeable, and results suggest a positive and significant relationship between distance to a Protestant church and premarital abstinence and a positive but non-significant relationship between proximity to a Catholic church and premarital abstinence. These findings imply that individuals who live closer to a Protestant church are less likely to abstain than those who live farther away. As shown in Table 4, although negative, the coefficient on the distance to a missionary health centre remains non-significant, suggesting no effect.

I further decompose the effect of historical missionary churches by Protestant and Catholic at the gender level. In Table A5, columns (1) and (2), we see that the relationship between distance to a Protestant church and premarital abstinence across gender remains positive, meaning that women and men who live closer to a Protestant church are less likely to abstain before marriage relative to those who live farther away. The effect is stronger for men compared to women. Results presented in column (3) show a negative relationship between distance to a Catholic church and premarital abstinence for women, implying that women who live closer to a Catholic church are more likely to engage in premarital abstinence; however, the results are non-significant. For men in column (4), the coefficient on the distance to a Catholic church is positive and significant at a 10% level of significance, suggesting that males who live closer to a Catholic church are less likely to abstain before marriage. Comparing the effect between distance to a Protestant church or Catholic church and premarital abstinence, I find that the effect is stronger for men who live close to a Protestant church than a Catholic church. Women who live near a Protestant church are less likely to abstain before marriage than women who live near a Catholic church. Decomposing our results by distance to a Protestant or Catholic church by gender, the coefficient on the distance to a missionary health centre remains non-significant, suggesting that distance to a health centre does not affect the likelihood of abstaining before marriage.

4.2.3. Christian missionaries and age at first sex

Considering that Christian maxims accentuate premarital abstinence as an important aspect of a Christian lifestyle, I expected to see higher age at first sex among individuals who live closer to missionary churches. Table 3 presents the results for the impact of Christian missionaries on age at first sex; in column (1), the results suggest a positive relationship between distance to any historical church and age at first sex, contrary to expectations.⁸ The results indicate that individuals who live farther away from a missionary church tend to postpone their first sexual encounter to a later age than those living near a historical church. In columns (2) and (3) in Table 3, I distinguish between distance to a Protestant church and a Catholic church, respectively. The results in column (2) suggest that there is a positive and significant relation between age at first sex and distance to a Protestant church at the 10% significance level, suggesting that individuals who live close to a Protestant church tend to have their first sexual encounter at an earlier age compared to those who live farther away. The results for the impact of distance to a Catholic church and age at first sex are presented in column (3); they suggest no significant effect.

Additionally, I explore the mixed impacts of distance to a historical church on age at first sex for males and females; the findings are reported in Table A6. The findings reported in column (1) show that there is no significant relationship between distance to a Protestant church and age at first sex for women; noticeable in column (2), the coefficient on distance to a Protestant mission for males is positive and significant at the 1% significance level, entailing that men who live closer to a Protestant historical church tend to have their first sexual encounter at an early age relative to those who live farther away. In columns (3) and (4), I find that there is a negative relationship between distance to a Catholic church and age at first sex, meaning that women and men who live closer to a Catholic church tend to postpone their first sexual experience to a later stage; however, the results are not statistically significant. For all estimations, I consider the impact of the distance to any missionary health centre on age at first sex. As in previous results, I find that proximity to a health centre does not significantly affect the age at first sex.

4.2.4. Christian missionaries and condom use at first intercourse

As earlier noted, FBOs do not advocate for or preach condom use. In their attempt to help in the fight against HIV, they have mostly focused on abstinence and faithfulness. Essentially, we expect to see lower condom usage and knowledge among Christian men and women. To try and capture the relationship between condom use and proximity to a missionary church, I rely on the variable condom use at first intercourse, which indicates whether an individual used a condom during their first sexual encounter or not. The findings reported in Table 3 show that there is no significant relationship between distance to any mission church and condom use at first sex (hereafter ‘condom use’); even after differentiating between distance to a Protestant and Catholic church, I find that there is still no significant effect of distance to either a Protestant or a Catholic church on condom use. Cagé and Rueda (2020) also do not find any significant relationship between distance to a mission and condom use.

I further analyse the varying effects of Christian missionaries on condom use at first sex across gender. The results are presented in Table A7; columns (1) and (2) suggest that there is no significant effect between proximity to a Protestant church and condom use for both males and females. In columns (3) and (4), the results suggest that males and females who live closer to a Catholic mission are less likely to use a condom during their first sexual encounter; however, these results are not statistically significant. For all specifications, I consider the impact of proximity to any historical missionary health centre on condom use at first intercourse; results are presented in Tables 4 and A7. I find that distance to a historical missionary health centre has no significant effect on condom use at first intercourse.

Regarding sexual behaviour, I find differences between those living near Protestant versus Catholic missionaries. I contend that these differences could be due to differences in doctrine and attitude towards sexual praxis. As Barrett (2018) and Falen (2008) indicated, the Protestant missionaries had more liberal views on marriage and sexuality than the Catholics. This also made the Protestant church a more attractive option, especially for Africans who wanted to continue certain practices such as polygyny. The differences in the doctrines and attitudes surrounding sexuality between Catholics and Protestants may account for observed differences. Tiruneh (2009) argues that although Christian institutions encourage their congregants to observe a certain code of conduct regarding their sexuality, because their views are not strictly enforced, there is a recalcitration likelihood among congregants.

With regards to HIV knowledge, DHS asks individuals a set of six questions; for each question, I observe whether an individual answered correctly, gave a wrong answer or does not know. If an individual answers correctly, I assign a score of 1, wrongly −1, and if they do not know the answer, 0. The score ranges from best knowledge (6) to worst knowledge (−6). Using this score, I found that the Protestants are more knowledgeable than Catholics about HIV. But, as Tiruneh pointed out, having knowledge does not entail the knowledge being put into practice.

4.3. Robustness checks

4.3.1. Omitted variable bias and endogenous measurement error

In econometric estimation, endogeneity refers to a situation where there is a correlation between the error term and the independent variables (omitted variable bias) or when the explanatory variables are measured with error (Depken et al. 2019). In effect, endogeneity thwarts the econometric model’s capacity to capture the true effect of the explanatory variables on the outcome variable. According to Jedwab, Meier zu Selhausen, and Moradi (2022), when estimating the long-run impact of missionaries on contemporary development outcomes, omitted variable bias may emanate from not understanding the factors that influenced missionary expansion. Measurement error may be introduced into the model because of the nature of the world missionary atlases used to map missionary stations’ historical locations. They postulate that the widely used world missionary atlases underreport mission stations significantly.

When using historical missions and health facilities as a proxy to measure the long-run impact of Christian missionary exposure on contemporary development, there may be concerns about omitted variable bias. This is because the establishment and expansion of Christian missionaries was not stochastic but was influenced by various factors. The older literature has qualitatively shown this (Johnson 1967), and more recently, Jedwab, Meier zu Selhausen, and Moradi (2022) for Africa and Ghana and Chiseni (2022) for Zambia have empirically shown that Christian missionaries favoured certain geographical locations over others. For Zambia, Chiseni (2022) has shown that, because of the difference in the timing of establishment, some factors that were important for mission station establishment did not significantly influence the expansion of missionary health facilities. This study shows that the factors that equally influenced the expansion of both mission stations and health facilities include, rivers, the railway, altitude and explorer routes. Jedwab, Meier zu Selhausen, and Moradi (2022) show that, as an initial disease prevention measure, Christian missionaries favoured locations that were at a higher altitude relative to those at a lower altitude.

Additionally, Johnson (1967) indicates that rivers played a crucial role in inland missionary expansion; missionaries essentially used rivers for transportation and irrigational purposes. As Ragsdale (1986) and Buxton (1840) indicate, the missionary impetus was to proselytize the natives and equip them with necessary agriculture skills. Hence, missionaries needed to establish themselves in areas near water sources for easy access to irrigation water; also, to transport agricultural produce, and for trade purposes, missionaries preferred locations that were near the railway (Chiseni 2022; Jedwab, Meier zu Selhausen, and Moradi 2022; Johnson 1967). The literature has also shown that to penetrate the African interior, Christian missionaries greatly relied on routes followed by earlier European explorers; this ultimately meant that regions that were close to explorer routes were preferred to those that were farther away (ibid). Consequently, if some of these factors are not controlled for when estimating long-run impacts, they could potentially bias the estimated impact of Christian missionaries on outcome variables. To ameliorate the incidence of omitted variable bias, in all empirical estimations, I control for factors that were important for the expansion of both mission stations and health facilities – that is, explorer routes, rives, historical railways, and altitude.

Moreover, studies have also shown that missionary-preferred locations received more missionary-led development than those that were less preferred (Alpino and Hammersmark 2021; Jedwab, Meier zu Selhausen, and Moradi 2022). This literature has also shown that individuals living in areas close to missionary locations are more likely to be more educated (Baten et al. 2021; Chiseni and Bolt 2020) and, therefore, more likely to have better occupations. The contemporary literature has shown that there is no conclusive causal link between education level and HIV prevalence; however, the relationship between education level and sexual behaviour is apparent (Aggarwal and Rous 2006; Filmer 1998; Lagarde et al. 2001; Waithaka and Bessinger 2001). For example, the literature documents a positive relationship between the level of education and condom use and HIV knowledge (ibid). For this reason, to further ameliorate the incidence of omitted variable bias when estimating the impact of Christian missionary exposure on current HIV and related sexual behaviour, I control for various demographic characteristics, such as level of education. Furthermore, I include in all models occupational fixed effects, the household asset index and a dichotomous variable that captures whether the individual lives in an urban or rural area. Additionally, I control for provincial fixed effects to control for any unobserved regional-specific factors.

While it may be desirable to include ‘many’ confounders to further reduce the omitted variable bias, it is also important to not overfit the regression by including more regressors than needed in the model, as this may lead to a reduction in statistical power and may lead to multicollinearity. According to the principle of parsimony, it is imperative to include only regressors in the model that are necessary for modelling, and nothing more. For example, if a regression model contains two regressors that adequately explain the variation in the outcome variable, then not more than these two explanatory variables should be utilized (Babyak 2004; Hawkins 2004).

To test the extent of the omitted variable bias, given the explanatory variables that have been included in my estimation, I use the test suggested by Oster (2019).⁹ Oster (2019) developed a Stata module, psacalc; the module assesses how large the magnitude of the bias due to unobserved confounders should be compared to that due to observables to explain away the main effect of missionary exposure on HIV. Oster suggests that the selection on unobserved explanatory variables needs to be at least four times as large as the selection on observed explanatory variables to explain away the long-run impact on HIV (in the baseline model). The calculated ratio is referred to as $\delta$.

To obtain the value of $\delta$, Oster (2019) suggests there is a need to hypothesize the value of $R_{\max }$, which is the R² value that controls for observables and unobservables. Although Oster acknowledges the difficulty in obtaining an empirically grounded value of $R_{\max }$, Oster (2013) initially suggested scaling the R² from the model with all controls by a factor of 2.2. to obtain $R_{\max }$. In her more recent work, Oster (2019) revised her findings and proposed that researchers may calculate a bias-adjusted treatment effect bound using a value of $R_{\max }$ that is 1.3 times more than the R² from the model with full controls. In this paper, to obtain $\delta$, I use both the initially suggested factor 2.2 and the more recently suggested factor 1.3 for our field and an additional assumed factor of 3.3 to observe the sensitivity of the results with different levels of R_max.

If $\delta$ = 1, this entails that observed factors are of equal importance to the unobserved confounders, a value of $\delta$>1 implies that there is a larger effect of omitted variables relative to the control variables, and a value of $\delta$<1 indicates a smaller impact of unobservables compared to control variables. The results for the Oster test are presented in Table A8. As can be seen from the results, the Oster coefficient of proportionality $\delta$ is less than 1 for our Equation (1) estimations, suggesting that there is a smaller impact of unobservables relative to the observables. However, the results should still be interpreted with caution, as stability in the model does not entail causality.

4.3.1.1. Endogenous measurement error

The key source of endogenous measurement error is the missionary atlases that underreport missionary station locations. For example, Jedwab, Meier zu Selhausen, and Moradi (2022) cite that for Ghana, 91% of the missionary stations reported in the ecclesiastical 1900 reports are missing in the world atlases, and 98% for Zambia. The missionary atlases used by Cagé and Rueda (2020), Beach (1903) and Streit (1929) captured 18 mission stations; using the ecclesiastical reports, I determined that there were 212 mission stations across Zambia by 1948. Thus, approximately 91% of the mission stations are missing from the Beach (1903) and Streit (1929) atlases. Figure 6 compares the spatial distribution of the newly geocoded mission stations and the missionary stations utilized by Cagé and Rueda (2020).

Figure 6. Cagé and Rueda missions (2020) versus new geocoded (1948) missions for Zambia.

Additionally, the missionary atlases do not capture the dynamic increase in mission stations over the twentieth century. In essence, if a study is measuring the impact of missionaries on contemporary development in Zambia using the Beach (1903) and Streit (1929) atlases as a source of mission stations, they would omit the mission stations that were established after 1929; in effect, 91% of the mission stations would be omitted from the analysis. Because some mission stations are omitted in the atlases, a measurement error will be introduced in the model. When using a continuous variable to proxy for missionary exposure, i.e. distance to a mission station with calculations being based on missionary station information obtained from the missionary atlases, due to missing information in the atlases, the proxy for missionary exposure will underestimate the effect of Christian missionaries on the outcome variable (Hyslop and Imbens, 2001). In this study, I geocode the exact location of mission stations and health centres using information obtained from the British Colonial census reports; the use of detailed missionary locations ameliorates the incidence of endogenous measurement errors. To estimate the effect of using world missionary atlases to estimate the effect of missionary exposure on HIV in Zambia, I use distances to the mission stations in the Beach (1903) and Streit (1929) atlases as a proxy for missionary exposure to estimate the impact of missionary exposure on HIV. The results are presented in Table A9; as can be seen, when these missionary atlases are used to measure the impact of missionary exposure on HIV, specifically the atlases utilized by Cagé and Rueda (2020), the effect decreases by about 93%. Thus, essentially, using the world missionary atlases underestimates the effect of missionary exposure.

5. Conclusion

This study aimed to investigate the long-term effect of exposure to missionaries on current HIV infection and related sexual behaviours in Zambia. To this aim, I built a new geocoded data set that combines information on historical missionary church and health centre locations with current individual-level data. Using distance to a historical church as a proxy for missionary exposure, I show that living close to a historical missionary church increases an individual’s likelihood of HIV infection.

The results show that proximity to a historical missionary health centre has no significant effect on HIV infection today. I also verify that it is highly unlikely that unobservables are driving the results on the effect of proximity to a historical church and missionary health centre. Analysing sexual behaviour in regions close to a historical church and missionary health centre reveals an interesting pattern of sexual behaviours for individuals close to a Protestant or Catholic church. Individuals who live close to a Protestant church are less likely to engage in premarital abstinence. They are more likely to have their first sexual encounter at an earlier age, especially for men relative to women. In contrast, individuals who live close to a Catholic church tend to have more lifetime sexual partners, particularly for men as opposed to women. I find no significant impact of proximity to a Protestant or Catholic church on the individual’s likelihood of having used a condom during their first intercourse.

In conclusion, this study has shown that historical Christian missionary exposure has had enduring effects on HIV and related sexual behaviour in Zambia. However, the effects vary depending on whether an individual is exposed to Protestant or Catholic missionaries.

Acknowledgements

I thank Jutta Bolt, Ellen Hillbom, Jeanne Cilliers, Alexandra Lopez Cermeno, Felix Meier zu Selhausen, Shane Doyle, Sarah Walters, Martin Andersson, Christian Thibon, the participants of the Laboratory for The Economics of African Past (LEAP) webinar 2020, and two anonymous reviewers for their valuable comments on the previous drafts of this paper. I also acknowledge the financial support from the Knut and Alice Wallenberg Foundation (KAW 2016.0184) to complete this project.

Disclosure statement

No potential conflict of interest was reported by the author.

Additional information

Funding

This work was funded by the Knut and Alice Wallenberg Foundation (grant number KAW 2016.0184).

Notes

1 Polygyny is a type of polygamy that allows men to have multiple wives (Falen 2008).

2 Kaposi’s sarcoma is a rare type of cancer caused by a virus. Kaposi’s sarcoma is mostly seen in people with an advanced HIV infection; see https://www.nhs.uk/conditions/kaposis-sarcoma/.

3 More information on biomarkers can be found at https://dhsprogram.com/.

4 The missionary health centres include both dispensaries and hospitals.

5 All Euclidean distances are measured in metres and then converted to kilometres in Stata.

6 In 1 Corinthians 7:2, it can be seen that sex is only acceptable within the confines of marriage, and any sex outside the confines of marriage is considered a sin: ‘But since sexual immorality is occurring, each man should have sexual relations with his own wife, and each woman with her own husband’, see also 1 Thessalonians 4:3–4, 1 Corinthians 5:9–11.

7 See the argument put forward by Tiruneh (2009) in section 4.2.1.

8 See footnote 7.

9 For a technical explanation of how the Oster method works, see the Appendix.

Appendix

Technical explanation of the Oster method

According to Oster (2019), we can determine the extent to which the results are driven by omitted variable bias if the following assumptions are met. If the relationship between the variable of interest and observed control variables is proportional to the relationship between the variable of interest and the omitted variables (proportional selection), this entails that the size of the changes in the coefficient of interest and the R² value after including control variables are informative about the size of the omitted variable bias. To formalize the assumptions around the covariance, I use the following equation. (A1) $Y=\text{β}Z+M_1+M_2+ϵ$(A1) In Equation (A1), $\text{β}$ is the coefficient; Z is the variable of interest (in this study, the proximity to a mission); the vector $M_1$ is a set of control variables multiplied by their true coefficients; $M_2$ is a vector of unobserved control variables multiplied by their coefficients; and $ϵ$ is the error term. The $R^2$ obtained from this regression is considered as the $R_{\max }^2$. It is imperative to be cognizant that $R_{\max }^2$ < 1 if Y is measured with an error or when there are some components of the variation in Y that are orthogonal to Z, $M_1$, and $M_2$. An additional assumption is made that $\mathrm{Cov}(M_1,M_2)=0$, $\mathrm{Cov}(M_1,ϵ)=0$, and $\mathrm{Cov}(M_2,ϵ)=0$. When these assumptions hold, the proportional selection assumption is therefore denoted by the equation: (A2) ${\displaystyle \frac{\mathrm{Cov}(M_1,Z)}{\mathrm{Var}\left(M_1\right)}={\displaystyle \frac{\mathrm{Cov}(M_2,Z)}{\mathrm{Var}\left(M_2\right)}}}$(A2)

If the proportional selection assumption holds for $\delta$> 0, Oster (2019) postulates that it is then probable to estimate $\text{β}$ by using: (I) the coefficients on Z with or without controls for observed variables, (II) $R^2$ values from controlled and uncontrolled regressions; (III) an assumption about the $R^2$ of a hypothetical regression that controls for Z and both unobserved and observed variables ($R_{\max }^2$), and (IV) a value of the degree of proportionality $\delta$. In practice, the value for $\delta$ is unknown; essentially, Oster (2019) suggests that we can calculate the bounding values for $\text{β}$ given assumptions about $\delta$ and $R_{\max }^2$. Additionally, to obtain bounds on the coefficient $\text{β}$, Oster proposes using values $\delta \in \left(0,1\right)$ and $R_{\max }^2$(2.2 ${\hat{R}}^2$), where ${\hat{R}}^2$ is the $R^2$ from the regression excluding all controls. Essentially, Oster suggests an upper bound of $\delta$= 1; we then obtain the bound on $\text{β}$ as $\text{β}$*(1); if the bound for $\text{β}$ that is $\widetilde{\text{β}}$ is obtained by $\text{β}$*($\delta$), the values of $\delta$<1 would scale down the coefficient to obtain the lower bound. Essentially, if the values of $\widetilde{\text{β}}$ exclude zero, this implies robustness. To have an idea of the magnitude of $\delta$, If $\delta$= 1 this entails that observed factors are of equal importance to the omitted variables in determining Z, a value of $\delta$>1 implies that there is a larger effect of omitted variables relative to the controls and a value of $\delta$<1 indicates a smaller impact of unobservables compared to the controls.

Table A1. Descriptive statistics for geocoded DHS data.

	Observations	Mean	Std. Dev.	Min.	Max.
Overall outcome variables
HIV positive	63,882	0.13	0.33	0	1
Condom use at first intercourse	17,312	0.30	0.46	0	1
Premarital abstinence	63,224	0.22	0.41	0	1
Lifetime sex partner	54,166	3.86	5.48	1	90
Age at First Sex	48,416	16.9	3.11	5	46
Female outcome variables
HIV positive	33,904	0.15	0.35	0	1
Age at first sex	23,702	16.5	2.6	7	39
Condom use first intercourse	9,777	0.30	0.46	0	1
Premarital abstinence	33,446	0.25	0.43	0	1
Lifetime sex partner	29,003	2.09	1.85	1	70
Male outcome variables
HIV positive	29,978	0.10	0.31	0	1
Age at first sex	24,714	17.3	3.52	5	46
Condom use first intercourse	7,535	0.29	0.45	0	1
Premarital abstinence	29,778	0.18	0.39	0	1
Lifetime sex partners	25,163	5.90	7.3	1	90
Explanatory variables
Distance to church	64,033	26.69	25.37	0.11	142.51
Log distance to health centre	64,033	3.47	1.00	−2.19	5.11
Protestant	62,939	0.81	0.39	0	1
Protestant church	64,033	0.82	0.38	0	1
Age	64,033	29.25	10.63	15	59
Age squared	64,033	968.59	698.07	225	3481
Married	58,139	0.62	0.49	0	1
Female	64,033	0.53	0.50	0	1
Urban	64,033	0.43	0.50	0	1
River 10 km	64,033	0.44	0.50	0	1
Explorer route 50 km	64,033	0.53	0.50	0	1
Rail 50 km	64,033	0.35	0.48	0	1
Mean log land elevation	64,033	1.31	0.16	0.73	1.87
Province	64,033	5.26	2.75	1	10
Asset index	63,224	2.86	1.30	1	5
Highest education level	64,015	1.49	0.71	0	3
Occupation	63,007	3.09	2.83	0	9

Table A2. Protestant and Catholic missionaries and HIV infection.

Dependent variable: HIV status
	Protestant	Catholic
Distance to church	−0.00182**	−0.00227*
	(0.000733)	(0.00126)
Distance to health centre	−0.0268*	0.0203
	(0.0145)	(0.0245)
Provincial FE	Yes	Yes
Occupational FE	Yes	Yes
Geographic controls	Yes	Yes
Individual controls	Yes	Yes
Observations	45,882	10,106

Note: Probit model estimates. Clustered robust standard errors at the DHS cluster level are given in parentheses. The comparison group for the asset index is individuals with the highest index. The comparison group for education level is individuals with the highest level of education. Distance to health centre is log-transformed, and distance to church is given in kilometres. Geographic controls include whether the cluster is within 50 km of an explorer route, 50 km of the railway, or 10 km of a river, and mean land elevation in logarithmic form. For all the models, I confine the regressions to those who answered that where they were living was their usual place of residence. Individual control: age, age squared, married, female, urban, asset index, education level. ***p < .01, **p < .05, *p < .1. FE = fixed effects.

Table A3. Christian missions and the heterogeneous impact on HIV infection.

	Female	Male
Distance to mission	−0.00145*	−0.00164**
	(0.000784)	(0.000819)
Distance to health centre	−0.0208	−0.0197
	(0.0155)	(0.0184)
Provincial FE	Yes	Yes
Geographic controls	Yes	Yes
Occupational FE	Yes	Yes
Individual controls	Yes	Yes
Observations	23,868	22,506

Note: Probit model estimates. Clustered robust standard errors at the DHS cluster level are given in parentheses. The comparison group for the asset index is individuals with the highest index. The comparison group for education level is individuals with the highest level of education. Distance to health centre is log-transformed, and distance to church is given in kilometres. Geographic controls include whether a cluster is within 50 km of an explorer route, 50 km of the railway, or 10 km of a river, and mean land elevation in logarithmic form. For all the models, I confine the regressions to those who answered that where they were living was their usual place of residence. Individual controls: age, age squared, married, female, urban, Protestant, lifetime partners, asset index, education level. ***p < .01, **p < .05, *p < .1.

Table A4. Heterogeneous effect of Christian missions and lifetime sex partners.

	(1)	(2)	(3)	(4)
	Women	Men	Women	Men
Distance to Protestant mission	−0.000748	−0.00484
	(0.000733)	(0.00370)
Distance to health centre	−0.0135	−0.0776	0.000952	−0.162
	(0.0191)	(0.102)	(0.0247)	(0.122)
Distance to Catholic mission			−0.00214*	−0.0126**
			(0.00119)	(0.00500)
Provincial FE	Yes	Yes	Yes	Yes
Occupational FE	Yes	Yes	Yes	Yes
Geographic controls	Yes	Yes	Yes	Yes
Individual controls	Yes	Yes	Yes	Yes
Observations	19,937	18,736	4,354	4,292
R^2	0.038	0.058	0.059	0.064

Note: OLS model estimates. Clustered robust standard errors at the DHS cluster level are given in parentheses. The comparison group for the asset index is individuals with the highest index. The comparison group for education level is individuals with the highest level of education. Distance to church is given in kilometres. Geographic controls include whether a cluster is within 50 km of an explorer route, 50 km of the railway, or 10 km of a river, and mean land elevation in logarithmic form. For all the models, I confine the regressions to those who answered that where they were living was their usual place of residence. Individual control: age, age squared, married, female, urban, Protestant, asset index, education level. ***p < .01, **p < .05, * p < .1.

Table A5. Heterogeneous effects of Christian missionaries on premarital abstinence.

	(1)	(2)	(3)	(4)
	Women	Men	Women	Men
Distance to Protestant mission	0.00139*	0.00151**
	(0.000772)	(0.000714)
Distance to health centre	−0.00564	−0.000670	0.0157	−0.0195
	(0.0162)	(0.0147)	(0.0335)	(0.0330)
Distance to Catholic mission			−0.000249	0.00246*
			(0.00136)	(0.00135)
Provincial FE	Yes	Yes	Yes	Yes
Occupational FE	Yes	Yes	Yes	Yes
Geographic controls	Yes	Yes	Yes	Yes
Individual controls	Yes	Yes	Yes	Yes
Observations	23,542	22,434	5,110	5,023

Note: Probit model estimates. Clustered robust standard errors at the DHS cluster level are give in parentheses. The comparison group for the asset index is individuals with the highest index. The comparison group for education level is individuals with the highest level of education. Distance to health centre is log-transformed, and distance to church is given in kilometres. Geographic controls include whether a cluster is within 50 km of an explorer route, 50 km of the railway, or 10 km of a river, and mean land elevation in logarithmic form. For all the models, I confine the regressions to those who answered that where they were living was their usual place of residence. Individual control: age, age squared, married, urban, Protestant, asset index, education level. ***p < .01, **p < .05, *p < .1.

Table A6. Heterogeneous effect of Christian missionaries on age at first sex.

	(1)	(2)	(3)	(4)
	Women	Men	Women	Men
Distance to Protestant mission	−8.96e-05	0.00469***
	(0.00122)	(0.00175)
Distance to health centre	−0.00379	−0.0303	−0.0219	−0.0470
	(0.0293)	(0.0441)	(0.0499)	(0.106)
Distance to Catholic mission			5.95e-05	0.00238
			(0.00258)	(0.00384)
Provincial FE	Yes	Yes	Yes	Yes
Occupational FE	Yes	Yes	Yes	Yes
Geographic controls	Yes	Yes	Yes	Yes
Individual controls	Yes	Yes	Yes	Yes
Observations	16,452	18,491	3,364	4,093
R^2	0.229	0.169	0.156	0.157

Note: OLS model estimates. Clustered robust standard errors at the DHS cluster level are given in parentheses. The comparison group for the asset index is individuals with the highest index. The comparison group for education level is individuals with the highest level of education. Distance to church is given in kilometres. Geographic controls include whether a cluster is within 50 km of an explorer route, 50 km of the railway, or 10 km of a river, and mean land elevation in logarithmic form. For all the models, I confine the regressions to those who answered that where were living was their usual place of residence. Individual control: age, age squared, married, urban, Protestant, asset index, education level. ***p < .01, **p < .05, *p < .1.

Table A7. Heterogeneous effect of Christian missionaries on condom use at first intercourse.

	(3)	(4)	(5)	(6)
	Female	Male	Female	Male
Distance to health centre	0.0158	−0.00524	−0.0892	0.0489
	(0.0278)	(0.0343)	(0.101)	(0.0786)
Distance to Protestant mission	0.000261	0.00115
	(0.00129)	(0.00175)
Distance to Catholic mission			0.00524	0.00158
			(0.00365)	(0.00326)
Provincial FE	Yes	Yes	Yes	Yes
Occupational FE	Yes	Yes	Yes	Yes
Geographic controls	Yes	Yes	Yes	Yes
Individual controls	Yes	Yes	Yes	Yes
Observations	3,806	3,342	725	627

Note: Probit model estimates. Clustered robust standard errors at the DHS cluster level are given in parentheses. The comparison group for the asset index is individuals with the highest index. The comparison group for education level is individuals with the highest level of education. Distance to health centre is log-transformed, and distance to church is given in kilometres. Geographic controls include whether a cluster is within 50 km of an explorer route, 50 km of the railway, or 10 km of a river, and mean land elevation in logarithmic form. For all the models, I confine the regressions to those who answered that where they were living was their usual place of residence. Individual controls: age, age squared, married, condom knowledge, condom access, urban, Protestant, lifetime partners, asset index, education level. ***p < .01, **p < .05, *p < .1.

Table A8. Oster test for Christian missionary impact on HIV infection.

Dependent variable: HIV status
R-max: 1.3R^2
Distance to church	−0.0002***
	(0.00008)
Distance to health centre	−0.0037*
	(0.00181)
Oster coefficient	0.57
R-max: 2.2R^2
Distance to church	−0.0002***
	(0.00008)
Distance to health centre	−0.0037*
	(0.00181)
Oster coefficient	0.14
R-max: 3.3R^2
Distance to church	−0.0002***
	(0.00008)
Distance to health centre	−0.0037*
	(0.00181)
Oster coefficient	0.08
Individual controls	Yes
Provincial FE	Yes
Occupational FE	Yes
Observations	46,374

Note: Probit model estimates. Clustered robust standard errors at the DHS cluster level are given in parentheses. The comparison group in this analysis is individuals with the highest number of assets. The comparison group for education level is individuals with the highest level of education. Distance to health centre is log-transformed. Geographic controls include whether a cluster is within 50 km of an explorer route, 50 km of the railway, or 10 km of a river, and mean land elevation in logarithmic form. For all the models, I confine the regressions to those who answered that where they were living was their usual place of residence. Then I also include occupational and provincial fixed effects. ***p < .01, **p < .05, *p < 0.

Table A9. Comparison between author geocoded missions and Cagé and Rueda (2020) missions.

Dependent variable: HIV status
Variables	New geocoded missions	Cagé and Rueda missions
Distance to church	−0.00188***	−0.000128
	(0.000646)	(0.000198)
Constant	−3,944***	−4,031***
	(0.218)	(0.213)
Provincial FE	Yes	Yes
Occupational FE	Yes	Yes
Geographic controls	Yes	Yes
Individual controls	Yes	Yes
All other controls	Yes	Yes
Observations	45,264	45,215

Note: Probit model estimates. Clustered robust standard errors at the DHS cluster level are given in parentheses. The comparison group for the asset index is individuals with the highest index. The comparison group for education level is individuals with the highest level of education. Distance to health centre is log-transformed, and distance to church is given in kilometres. Geographic controls include whether a cluster is within 50 km of an explorer route, 50 km of the railway, of 10 km of a river, and mean land elevation in logarithmic form. For all the models, I confine the regressions to those who answered that where they were living was their usual place of residence. Individual controls: age, age squared, married, condom knowledge, condom access, urban, Protestant, lifetime partners, asset index, education level. ***p < .01, **p < .05, *p < .1.