Free access to a broad contraceptive method mix and women’s contraceptive choice: Evidence from sub-Saharan Africa

Abstract

Financial barriers may restrict women’s ability to use their preferred contraceptive methods, especially long-acting reversible contraceptives (LARC). Providing free access to a broad contraceptive method mix, including both LARC and short-acting reversible contraceptives (SARC), may increase contraceptive use, meet women’s various fertility needs, and increase their agency in contraceptive decisions. Linking facility and individual data from eight countries in sub-Saharan Africa, we use a propensity score approach combined with machine learning techniques to examine how free access to a broad contraceptive method mix affects women’s contraceptive choice. Free access to both LARC and SARC was associated with an increase of 3.2 percentage points (95% confidence interval: 0.006 – 0.058) in the likelihood of contraceptive use, driven by greater use of SARC. Among contraceptive users, free access did not prompt women to switch to LARC and had no effect on contraceptive decision-making. The price effects were larger among older and more educated women, but free access was associated with lower contraceptive use among adolescents. While free access to contraceptives is associated with a modest increase in contraceptive use for some women, removing user fees alone does not address all barriers women face, especially for the most vulnerable groups of women.

Introduction

Access to a wide variety of contraceptive methods is a fundamental element of quality of care and a critical dimension in multiple frameworks of family planning services (Hardee et al. 2014; Bruce 1990). Long-acting reversible contraceptives (LARC), including implants and intrauterine devices (IUDs), are 20 times as effective as shorter-term methods, have much higher continuation rates, and suit women’s growing needs for limiting births (Hubacher et al. 2017; Benova et al. 2017; Curtis and Peipert 2017; Van Lith, Yahner, and Bakamjian 2013). However, a number of demand- and supply-side factors continue to limit women’s uptake of LARC, such as upfront costs that often include informal fees, lack of trained providers that could perform both insertion and removal of LARC, and concerns about side effects (Darroch and Singh 2013; Bertrand et al. 2014; Adedini, Omisakin, and Somefun 2019). Free access to both LARC and short-acting reversible contraceptives (SARC) may increase contraceptive use by meeting women’s various fertility needs in low- and middle-income countries (LMIC).

Free access to a broad contraceptive method mix also gives women greater autonomy in contraceptive decision-making, especially for those from lower socioeconomic backgrounds. Previous research has shown that women from the poorest wealth quintiles in sub-Saharan Africa (SSA) are less likely to use LARC and such wealth-related inequalities have increased in several countries (Creanga et al. 2011; Adedini, Omisakin, and Somefun 2019). Although the costs per couple-year served are similar or even lower for LARC compared to SARC, higher upfront out-of-pocket costs, including both the commodity prices and additional consultation charges or insertion/removal fees, may be a barrier to LARC use (Duvall et al. 2014; Tibaijuka et al. 2017; Tumlinson et al. 2013). Since men usually control household spending and often have different contraceptive or fertility preferences compared to women, providing both LARC and SARC free of charge removes the concern of costs from the decision-making process and may increase a woman’s ability to use any contraceptive method without her partner’s knowledge or approval (Wulifan et al. 2016; Ashraf, Field, and Lee 2014).

Previous literature is mixed on whether free access to contraceptives is effective in increasing contraceptive use or optimal for the sustainability of health systems. User fees for family planning were introduced in the 1980s in many LMIC to ensure the quality of health services and sustainability of health systems, but charging a price for basic health services may prevent women from choosing the method that best suits their needs and preferences, especially for those on the lower end of the wealth spectrum (Korachais, Macouillard, and Meessen 2016; Fedor, Kohler, and McMahon 2016; Ross and Stover 2013; Ross 2015). For public health products, field experiments have shown that any increase in price can considerably dampen demand in SSA (Comfort and Krezanoski 2017; Fischer et al. 2019; Ashraf, Berry, and Shapiro 2010; Cohen and Dupas 2010; Dupas 2014; Chang et al. 2019). Specifically for LARC, recent studies showed that removing the cost barriers substantially increased LARC use among low-income women and adolescents in the United States (Kelly, Lindo, and Packham 2019; McNicholas et al. 2014; Lindo and Packham 2017).

Contrary to these two strands of literature, previous reviews suggest that the effects of price on contraceptive use in LMIC have been largely inconclusive due to lack of recent studies, weak study designs, and challenges in measuring supply and demand simultaneously (Lewis 1986; Korachais, Macouillard, and Meessen 2016; Janowitz and Bratt 1996; Belaid et al. 2016; Bellows et al. 2016; Lissner and Ali 2016). For example, providing free vouchers for contraceptives did not increase contraceptive use or shift method choice among postpartum women in Kenya in a randomized controlled trial (McConnell et al. 2018) although quasi-experimental studies have shown that vouchers were effective in increasing contraceptive use in Cambodia, Pakistan, and Madagascar (Burke et al. 2017; Bajracharya et al. 2016; Ali et al. 2019). In addition, program evaluations of service expansion have demonstrated uptake and acceptability of LARC, such as the multi-component LARC expansion initiative (including free or subsidized family planning services) implemented by Marie Stopes International across 14 countries in SSA (Ngo et al. 2017).

Further evidence is needed to understand whether free access to a broad contraceptive method mix is an effective and empowering strategy to meet women’s family planning needs in SSA. To inform pricing policies for governments and international donors, this study estimates the effects of providing both LARC and SARC free of charge on women’s contraceptive use and agency.

Methods

Linking data from health facility and individual surveys, we estimated whether free access to both LARC and SARC was associated with an increase in women’s contraceptive use in eight countries in SSA. Among current contraceptive users, we estimated whether free access was associated with increased use of LARC and greater autonomy in decisions about contraceptive methods. We combined propensity score weighting with machine learning techniques to account for differences in selection into the treatment and estimated differential treatment effects among key subgroups.

Data and sample.

The data source is Performance Monitoring and Accountability 2020 (PMA2020) harmonized by the IPUMS-PMA project (Boyle, Kristiansen, and Sobek 2018). PMA2020 collected data annually from 2014 to 2017 from households, women, and service delivery points (i.e., health facilities) in 11 countries that have pledged to the Family Planning 2020 effort. These data are collected by trained enumerators using mobile devices and include key indicators of family planning use by individuals as well as provision of reproductive health services by healthcare facilities. To identify public facilities, the survey team consulted district or local authorities to locate public sector facilities designated to cover the residents of each enumeration area, which may be located outside of the community. PMA2020 surveyed all secondary and tertiary public facilities that serve an enumeration area along with the primary facility located within the area (Zimmerman 2017). For private facilities, PMA2020 randomly sampled three facilities located within each enumerator area (Zimmerman 2017).

The sample came from the most recent PMA2020 surveys that were conducted in 2016/2017 and are representative at the national level in eight countries in SSA, including Burkina Faso, Cote d’Ivoire, Ethiopia, Ghana, Kenya, Niger, Nigeria, and Uganda (we excluded countries that do not have nationally-representative data or are not in SSA). Individual data excluded pregnant women or women who expressed the desire to have a child soon. The analysis dataset linked individual data with service delivery points data by enumeration area. For the main outcome, use of modern contraceptives, the sample included all women of reproductive age. For the other outcomes, use of LARC and autonomy in contraceptive decision-making, the sample consisted of women of reproductive age who reported using a modern contraceptive method at the time of the survey.

Measures.

The main policy variable of interest, free access to a broad contraceptive method mix, is a binary variable coded to 1 if both LARC (implants or intrauterine devices) and SARC (injectables that include Depo-Provera and Sayana Press, pills, male/female condoms, diaphragms, spermicide, and n tablets) are routinely provided and offered free of charge in at least one facility that serves an enumeration area. Diaphragms, spermicide, and n tablets collectively make up less than 1% of the total method mix within our sample. We did not include permanent methods because we expect the decision-making process to be different from those for reversible methods. Similarly, we did not include emergency contraception because we expect a different user profile for emergency contraception compared to other SARC.

Free access was constructed based on two categories of questions in the Service Delivery Points Questionnaire to measure routine provision and charge of family planning services. First, routine provision is defined by whether a facility provides a contraceptive method, has trained personnel and supplies, and has the devices in stock. Service provision is measured by the question “which of the following methods are provided to clients at this facility?”. A facility is considered to routinely provide LARC only if it has trained personnel that are able to insert implants/IUDs as measured by the question “on days when you offer family planning services, does this facility have trained personnel able to insert implants/IUDs?”. A facility also needs to have the supplies necessary to perform the procedures (clean gloves, antiseptic, sterile gauze pad, anesthetic, implant pack for implants and speculums, forceps, and tenaculum for IUDs), as measured by the question “Does this facility have the following supplies needed to insert and/or remove implants/IUD?”. Moreover, a facility needs to have the LARC or SARC method in stock on the day of the survey, regardless of whether this is based on interviewers’ observation or providers’ response. Second, a LARC or SARC method was provided free of charge if a facility did not charge any consultation fee or method-specific fee. This is assessed by two questions – “Do family planning clients need to pay any fees in order to be seen by a provider even if they do not obtain a method of contraception” and “are clients charged for obtaining any of the following methods at this facility?”.

The three outcome variables are based on questions included in the Female Questionnaire for women of reproductive age. Use of modern contraceptives is measured by the pre-coded binary “modern contraceptive user” variable that indicates current use of a modern family planning method. Use of LARC is a binary variable coded to 1 if the response to current use of either implant or IUD is yes. Contraceptive method choice is measured based on the question “during that visit, who made the final decision about what method you got?” and is coded to 1 if the answer is respondent alone, respondent and provider, or respondent and partner.

The Female Questionnaire has a wide range of individual-level covariates. These include women’s socio-demographic characteristics (e.g., age, education, marital status, rurality, and wealth), sexual and reproductive history (e.g., age of first sex, age of first birth), fertility preference (e.g., when to have another child), knowledge about different family planning methods (e.g., ever heard of a specific method), and sources for family planning information (e.g., read about family planning in newspaper). Using a principle component analysis approach, we constructed two indices based on all knowledge or information source variables, with higher scores indicating women’s greater knowledge of different contraceptive methods or exposure to various information sources about family planning.

Statistical analyses.

The study’s main hypothesis is that free access to a broad contraceptive method mix is associated with women’s greater overall contraceptive use, LARC use, and autonomy in decision-making. We used the propensity score approach because differences between individuals who live in communities with good access to family planning services and those that do not might affect contraceptive use and women’s agency. For example, women who prioritize health might live closer to well-funded facilities and are more likely to use effective modern contraceptives. Conventional propensity score methods reduce confounding by accounting for observed characteristics that predict treatment and often apply propensity weights in treatment effect models (Rosenbaum and Rubin 1983; Hirano, Imbens, and Ridder 2003). However, such propensity score models assume comparability of unobserved pre-treatment characteristics between groups, are likely to be misspecified, and might omit covariates that are important to treatment selection (Karim, Pang, and Platt 2018; Lee, Lessler, and Stuart 2010; McCaffrey, Ridgeway, and Morral 2004).

To strengthen the propensity score approach, we used generalized boosted modeling (GBM) to estimate propensity scores. GBM is a non-parametric machine learning technique that adds together a collection of simple regression tree models to fit a nonlinear surface and is effective in producing probability estimates with a large number of covariates (McCaffrey, Ridgeway, and Morral 2004; Westreich, Lessler, and Funk 2010). Similar machine learning algorithms have increasingly been applied in health and medical studies (Bisaso et al. 2017; Cruz and Wishart 2006). GBM is particularly tuned to produce well-calibrated probability estimates and has been shown to outperform standard logistic regression and covariate balancing propensity methods in complex models for non-linear relationships (Setodji et al. 2017; McCaffrey, Ridgeway, and Morral 2004).

The propensity score approach consists of two steps. First, we used GBM to estimate probability of having free access using all available covariates at the individual level (Table A1). The algorithm was stopped at the number of iterations that minimized the average standardized absolute mean difference in the covariates. A dummy missing variable approach was used to create a missing category for all factor variables so that the algorithms would use the missingness in the predictions.

Then, weights generated from the GBM propensity scores were applied using inverse weighting to the logistic regression model specified below:

$$\mathit{\text{Logit}}(Y_{ijc})=\alpha +\beta {\mathit{\text{Access}}}_{jct}+z_c+X_{ijc}{}^{\prime }\gamma +{\epsilon }_{ijc}$$

where Y_ijc is the outcome of interest for individual i of community j in country c, which include contraceptive use, LARC use, and contraceptive decision-making, Access_i is a dummy equal to one if the community provides access to both LARC and SARC free of charge, z_c is a full set of country fixed effects to control for time-invariant country characteristics, and X_ijc is a vector of individual characteristics including age (continuous variable), education (categorical variable for highest level of school attended), marital status (categorical variable), rurality (binary variable), household wealth index quintile (categorical variable), family planning knowledge score (continuous variable), family planning information exposure score (continuous variable), fertility preference (binary variable), and provider type (categorical variable for LARC use and decision-model only). The coefficient of interest is β, which represents the estimated effects of free access to LARC and SARC. The results are presented as average marginal effects.

In addition to the primary models, we conducted additional sensitivity and subgroup analyses. First, we used standard logistic regression models to generate propensity scores and compared the results with the GBM-based models. Second, to examine whether the price effects vary by contraceptive type, we assessed the effects of free access to both LARC and SARC as well as free access to only SARC on method-specific use. Third, we used an alternative definition for the treatment variable and examined the effects of access to contraceptives without removing user fees. Specifically, we used multinomial logit models to generate propensity scores for a categorical treatment variable with three levels: no access, access with some fee, and free access. Lastly, we draw on the recursive partitioning approach to examine heterogeneity in the effects of free access among subgroups of women (Athey and Imbens 2016; Atkinson and Therneau 2000). We first split the data into two randomly chosen sub-samples, one sample to identify the sources of heterogeneity (training subsample) and the other to estimate the treatment effects and confidence intervals (estimation subsample). The machine learning algorithm created a decision tree that aimed to correctly classify individuals’ outcome status by splitting the training subsample into high-dimensional and mutually exclusive groups. Based on the structure of the decision tree and the variable importance measure, we identified the subgroups and estimated heterogeneous treatment effects using the estimation subsample.

Statistical tests were 2-sided and the statistical significance was set at p<.05. Analyses were performed using R v3.5.1 (the R Foundation, packages “rpart” and “twang”) (Therneau, Atkinson, and Ripley 2019; Ridgeway, McCaffrey, and Morral 2020) and Stata, version 15.1 (StataCorp LLC).

Results

The full sample contained 29,833 individuals, of which 15,998 (53.6%) had free access to a broad contraceptive method mix and the remaining 13,835 (46.4%) did not (Table 1). Overall, 28.8% of individuals were using a modern contraceptive method. Among these contraceptive users, 29.4% were using LARC and 91.2% reported having participated in the decisions about which method to use. Unweighted, 32.5% of women with free access to both LARC and SARC were using a modern contraceptive method compared to 24.4% among those without access. The proportion of LARC users did not differ by access status while more women with free access participated in contraceptive decision-making (93.3% vs. 87.8%). Most other covariates differed by treatment status.

TABLE 1.

Descriptive statistics

Variable	All	Treat	Comparison	p-value
N	29833	15998	13835
Modern contraceptive user	8579 (28.8%)	5207 (32.5%)	3372 (24.4%)	<0.001
Use of LARC	2520 (29.4%)	1522 (29.2%)	998 (29.6%)	0.72
Decided contraceptive method	7820 (91.2%)	4860 (93.3%)	2960 (87.8%)	<0.001
Age, mean (SD)	28.0 (9.3)	27.9 (9.3)	28.0 (9.3)	0.42
Urban	13744 (46.1%)	6920 (43.3%)	6824 (49.3%)	<0.001
Highest level of education				<0.001
never attended	7224 (24.2%)	3043 (19.0%)	4181 (30.2%)
primary/middle school	10232 (34.3%)	6122 (38.3%)	4110 (29.7%)
secondary/post-primary	9486 (31.8%)	5097 (31.9%)	4389 (31.7%)
tertiary/post-secondary	2891 (9.7%)	1736 (10.9%)	1155 (8.3%)
Marital status				<0.001
never married	9971 (33.4%)	5352 (33.5%)	4619 (33.4%)
currently married	15200 (51.0%)	8096 (50.6%)	7104 (51.3%)
currently living with partner	2178 (7.3%)	1000 (6.3%)	1178 (8.5%)
divorced or separated	1666 (5.6%)	1085 (6.8%)	581 (4.2%)
widow or widower	818 (2.7%)	465 (2.9%)	353 (2.6%)
Married once or more than once				0.22
never	9971 (33.5%)	5352 (33.5%)	4619 (33.5%)
once	17484 (58.7%)	9348 (58.5%)	8136 (59.0%)
more than once	2311 (7.8%)	1280 (8.0%)	1031 (7.5%)
Partner has other wives	4655 (27.1%)	2088 (23.3%)	2567 (31.3%)	<0.001
Wealth score quintile				<0.001
lowest quintile	6048 (20.3%)	3062 (19.1%)	2986 (21.6%)
lower quintile	5851 (19.6%)	3020 (18.9%)	2831 (20.5%)
middle quintile	5530 (18.5%)	2902 (18.1%)	2628 (19.0%)
higher quintile	5503 (18.4%)	2929 (18.3%)	2574 (18.6%)
highest quintile	6901 (23.1%)	4085 (25.5%)	2816 (20.4%)
Ever given birth	20059 (67.3%)	10678 (66.8%)	9381 (67.8%)	0.050
Age at first sex, median (IQR)	17.0 (15.0, 19.0)	17.0 (15.0, 19.0)	17.0 (15.0, 19.0)	0.054
Age at first birth, mean (SD)	19.9 (4.2)	19.9 (4.1)	20.0 (4.3)	0.28
Prefer no more children	7768 (26.0%)	4484 (28.0%)	3284 (23.7%)	<0.001
Months to wait before another child, mean (SD)	48.3 (38.9)	50.8 (39.7)	45.6 (37.9)	<0.001
Family planning knowledge score, mean (SD)	0.0 (2.1)	0.2 (2.1)	−0.2 (2.2)	<0.001
Family planning information exposure, mean (SD)	0.0 (1.4)	0.0 (1.4)	−0.0 (1.3)	<0.001
Country				<0.001
Burkina Faso	2650 (8.9%)	67 (0.4%)	2583 (18.7%)
Ethiopia	5617 (18.8%)	4711 (29.4%)	906 (6.5%)
Ghana	2427 (8.1%)	35 (0.2%)	2392 (17.3%)
Kenya	4801 (16.1%)	4281 (26.8%)	520 (3.8%)
Niger	1830 (6.1%)	1339 (8.4%)	491 (3.5%)
Nigeria	7425 (24.9%)	2794 (17.5%)	4631 (33.5%)
Uganda	3234 (10.8%)	2674 (16.7%)	560 (4.0%)
Cote d’Ivoire	1849 (6.2%)	97 (0.6%)	1752 (12.7%)

^*The analysis dataset excludes pregnant women and women who would like to have another child soon. LARC: long-acting reversible contraceptives, including implants and intrauterine devices; SD: standard deviation; IQR: interquartile range.

We used the machine learning algorithm GBM and all variables listed in Table A1 to generate propensity weights, including missing values as dummy variables. Weighted standardized differences in covariates between treatment and comparison groups were reduced to below .03 on average and below .06 in all variables except for one country. For comparison purposes, we also used the standard logistic regression approach to generate propensity weights, including a limited set of covariates. A comparison on covariate balance between the GBM approach and the logistic regression approach were presented in Figure A1. The distribution of propensity scores generated from GBM was presented in Figure A2.

The main results are presented in Table 2. We applied the propensity weights built from GBM, our preferred approach, and the weights from logistic regression models to the main specification and presented average marginal effects. Free access to a broad contraceptive mix was associated with an increase of 3.2 percentage points (95% confidence interval [CI]: 0.006 – 0.058) in the likelihood of contraceptive use among all women, representing an increase of 13.1% from the 24.4% contraceptive use in the control group. Among current contraceptive users, free access had no effect on LARC use or women’s role in contraceptive method decision-making. The specifications using propensity weights from logistic regression models produced similar effects except for the LARC use outcome: free access was associated with a decrease of 3.8 percentage points (95% CI: −0.075 – −0.000) in the likelihood of LARC use.

TABLE 2.

Effects of free access to a broad contraceptive method mix

	(1)	(2)	(3)	(4)	(5)	(6)
	All Women		Current Contraceptive Users
	Modern Contraceptive Use		LARC Use		Decision-Making
	GBM	Logit	GBM	Logit	GBM	Logit
Treat (free access)	0.032* [0.006,0.058]	0.029* [0.002,0.055]	−0.029 [−0.066,0.007]	−0.038* [−0.075,−0.000]	0.013 [−0.021,0.047]	0.012 [−0.021,0.045]
Age	0.003 [−0.022,0.028]	0.010 [−0.016,0.036]	−0.008 [−0.048,0.033]	−0.005 [−0.047,0.037]	0.018 [−0.013,0.049]	0.022 [−0.010,0.053]
Urban	−0.002*** [−0.003,−0.001]	−0.002*** [−0.003,−0.001]	0.000 [−0.003,0.003]	0.000 [−0.003,0.004]	0.002** [0.001,0.004]	0.002** [0.001,0.004]
Education (base = no schooling)
Primary school	0.047*** [0.026,0.069]	0.032* [0.003,0.061]	0.011 [−0.025,0.047]	0.014 [−0.025,0.053]	−0.010 [−0.041,0.021]	−0.012 [−0.043,0.018]
Secondary school	0.048** [0.018,0.078]	0.030 [−0.006,0.067]	−0.033 [−0.076,0.011]	−0.022 [−0.063,0.020]	0.012 [−0.016,0.040]	0.012 [−0.016,0.039]
Tertiary school	0.045** [0.011,0.079]	0.025 [−0.017,0.066]	0.019 [−0.039,0.077]	0.025 [−0.039,0.089]	0.000 [−0.036,0.036]	0.001 [−0.035,0.036]
Marital status (base = never married)
Married	0.270*** [0.240,0.299]	0.257*** [0.224,0.290]	0.084** [0.023,0.146]	0.087** [0.029,0.144]	0.029 [−0.015,0.072]	0.029 [−0.015,0.072]
Cohabiting	0.243*** [0.205,0.280]	0.239*** [0.200,0.279]	0.060 [−0.010,0.130]	0.074 [−0.003,0.151]	0.060* [0.004,0.117]	0.063* [0.004,0.122]
Divorced or separated	0.108*** [0.073,0.143]	0.095*** [0.055,0.134]	0.159*** [0.076,0.241]	0.158*** [0.077,0.240]	0.094*** [0.050,0.138]	0.094*** [0.049,0.138]
Widow	−0.006 [−0.042,0.030]	−0.019 [−0.057,0.018]	0.053 [−0.065,0.171]	0.089 [−0.049,0.226]	0.055 [−0.012,0.121]	0.056 [−0.010,0.121]
Wealth index quintile (base = lowest)
Lower quintile	0.059*** [0.033,0.084]	0.064*** [0.037,0.091]	−0.046* [−0.089,−0.003]	−0.033 [−0.075,0.010]	−0.033 [−0.076,0.011]	−0.035 [−0.074,0.004]
Middle quintile	0.052*** [0.021,0.083]	0.050** [0.018,0.083]	−0.052* [−0.102,−0.002]	−0.042 [−0.094,0.010]	−0.023 [−0.055,0.009]	−0.030* [−0.060,−0.000]
Higher quintile	0.093*** [0.057,0.129]	0.097*** [0.059,0.134]	−0.036 [−0.087,0.014]	−0.033 [−0.086,0.020]	−0.053* [−0.096,−0.010]	−0.060** [−0.104,−0.017]
Highest quintile	0.078*** [0.045,0.112]	0.081*** [0.046,0.116]	0.020 [−0.039,0.079]	0.025 [−0.036,0.087]	−0.019 [−0.054,0.016]	−0.028 [−0.062,0.006]
Family planning knowledge score	0.051*** [0.045,0.058]	0.054*** [0.047,0.061]	0.019** [0.007,0.030]	0.016* [0.004,0.028]	0.006 [−0.002,0.015]	0.006 [−0.002,0.015]
Family planning information	0.006 [−0.000,0.013]	0.006 [−0.000,0.013]	−0.001 [−0.012,0.010]	0.000 [−0.011,0.011]	−0.001 [−0.010,0.007]	−0.001 [−0.010,0.007]
Fertility preference (base = prefer more children)
Prefer no more child	0.017 [−0.004,0.038]	0.014 [−0.007,0.034]	0.032* [0.003,0.061]	0.034* [0.005,0.064]	−0.015 [−0.036,0.006]	−0.011 [−0.033,0.010]
Family planning provider (base = public sector)
private sector	--	--	−0.290*** [−0.317,−0.263]	−0.281*** [−0.309,−0.252]	−0.033* [−0.062,−0.004]	−0.031* [−0.059,−0.002]
NGO	--	--	0.083 [−0.094,0.260]	0.085 [−0.089,0.260]	−0.057 [−0.177,0.064]	−0.059 [−0.178,0.061]
other	--	--	−0.347*** [−0.385,−0.308]	−0.348*** [−0.385,−0.311]	−0.072* [−0.137,−0.008]	−0.087* [−0.162,−0.012]
Number of observations	29833	29833	8464	8464	8464	8464
Control mean	0.244	0.244	0.303	0.303	0.890	0.890

95% confidence intervals in brackets

LARC: long-acting reversible contraceptives, including implants and intrauterine devices. Decision-making: whether respondent participated in the decision on contraceptive method. Logistic regression models report average marginal effects and include country fixed effects. Propensity weights generated from generalized boosted models (GBM) or logistic models were applied to adjust for observed differences by treatment status.

^*p < 0.05,

^**p < 0.01,

^***p < 0.001

To further investigate the price effects on contraceptive use, we estimated the effects of free access to different types of contraceptives on women’s contraceptive use by method-type (Table 3). Table A2 shows that 77.7% of all women and 84.0% of current contraceptive users who had free access to SARC also had free access to LARC.

TABLE 3.

Effects of free access on contraceptive use by method type

	(1)	(2)	(3)
	Modern Contraceptive Use	LARC Use	SARC Use
Panel A: among all women
Free access to both LARC & SARC	0.032* [0.006,0.058]	0.003 [−0.010,0.016]	0.025* [0.003,0.047]
Control mean	0.244	0.072	0.170
Free access to SARC only	0.041* [0.007,0.074]	0.016 [−0.000,0.031]	0.023 [−0.003,0.049]
Control mean	0.258	0.081	0.175
Number of observations	29833	29833	29833
Panel B: among current contraceptive users
Free access to both LARC & SARC	--	−0.026 [−0.066,0.013]	0.015 [−0.027,0.056]
Control mean	--	0.296	0.697
Free access to SARC only	--	0.014 [−0.030,0.058]	−0.020 [−0.065,0.025]
Control mean	--	0.314	0.679
Number of observations	--	8579	8579

95% confidence intervals in brackets

Logistic regression models report average marginal effects and include country fixed effects. Propensity weights generated from generalized boosted models were applied to adjust for observed differences by treatment status. LARC: long-acting reversible contraceptives, including implants and intrauterine devices. SARC: short-acting reversible contraceptives.

^*p < 0.05,

^**p < 0.01,

^***p < 0.001

Panel A of Table 3 shows that free access to both LARC and SARC was associated with an increase in overall contraceptive use as well as an increase of 2.5 percentage points in SARC use (95% CI: 0.003 – 0.047) but had no effect on LARC use. Free access to SARC only was associated with an increase of 4.1 percentage points (95% CI: 0.007 – 0.074) in overall contraceptive use and had no effect on LARC use. Although free access to SARC was not associated with any statistically significant effect on SARC use, the effect size is similar to that of free access to both LARC and SARC with a p-value of 0.079. These results suggest that LARC use was not responsive to the removal of user fees, or at least not as sensitive as SARC use.

Turning to women who were contraceptive users, Panel B of Table 3 shows that neither free access to both LARC and SARC nor free access to SARC alone was associated with any statistically significant effect on method-specific use. This indicates that among women who were already using a modern contraceptive method, removing user fees did not shift their preference for contraceptives defined by the two general contraceptive types.

Using an alternative definition of the treatment variable, Table 4 assesses the effects of access to contraceptives with and without user fees. Among all women, access alone without removing user fees did not affect contraceptive use. By comparison, free access was associated with an increase of 4.7 percentage points in contraceptive use (95% CI: 0.019 – 0.074), which underscores the importance of removing financial barriers in addition to making family planning services available. Among contraceptive users, neither access alone nor free access affected LARC use or decision-making.

TABLE 4.

Effects of access and free access to a broad contraceptive method mix

	(1)	(2)	(3)
	All women	Current contraceptive users
	Modern Contraceptive Use	LARC Use	Decision-Making
Access to both LARC and SARC	0.021 [−0.007,0.049]	0.055 [−0.007,0.116]	0.014 [−0.020,0.049]
Free access to both LARC and SARC	0.047** [0.019,0.074]	0.004 [−0.055,0.062]	0.014 [−0.020,0.047]
p-value	0.004	0.027	0.663
Control mean	0.252	0.282	0.906
Number of observations	29802	8462	8462

95% confidence intervals in brackets

LARC: long-acting reversible contraceptives, including implants and intrauterine devices. SARC: short-acting reversible contraceptives. Decision-making: whether respondent participated in the decision on contraceptive method. Propensity weights generated from multinomial logistic models were applied to adjust for observed differences by treatment status. The p-value row reports the two-sided p-value from an F-test of equality of the treatment effects of access and free access.

^*p < 0.05,

^**p < 0.01,

^***p < 0.001

Heterogeneity.

We draw on the machine learning recursive partitioning approach to identify subgroups that best predict modern contraceptive use (Athey and Imbens 2016; Atkinson and Therneau 2000). The machine learning algorithm tried to predict which distinct groups of women would use modern contraceptives based on a list of observed characteristics. The training sub-sample used in the exercise was a randomly generated half sample of the full dataset. The other half sub-sample was used in estimating treatment effects. This “honest” approach helps to avoid identifying spurious relationships by overfitting the model (Athey and Imbens 2016; Atkinson and Therneau 2000). The decision tree generated by the training sub-sample correctly classified contraceptive use for 76.5% of the observations in the estimation sub-sample.

In the final tree structure (see Figure A3), the candidates for the primary splits for the first node include marital status, family planning knowledge, and age. These variables are also the three highest ranked variables according to the variable importance measure (see Table A3), which indicates how much a model uses a given variable to make accurate predictions. We also included education and wealth as potential source of heterogeneity based on the literature and the fact that they had higher or similar variable importance values compared to the treatment variable. We used the splitting rules as cutoff values to create subgroups in the estimation sub-sample.

Free access to a broad contraceptive method mix was associated with greater contraceptive use among older women who were at least 20 years of age (see Table 5 Panel B). By comparison, adolescents used contraceptives less when LARC and SARC were offered for free. Free access also had greater effects on overall contraceptive use among women who were relatively knowledgeable about family planning, had at least some schooling, and came from an above-the-lowest wealth quintile, although the result on wealth was not statistically significant (p = 0.061). For current contraceptive users, free access was associated with a 9.9 percentage point decrease (95% CI: −0.194 – −0.003) in LARC use among women who came from the lowest wealth quintile, although we could not reject the null hypothesis that this is different from the effect among women from a wealthier background. The effects on LARC use and decision-making did not differ in any of the other subgroups we examined.

TABLE 5.

Effects of free access to a broad contraceptive method mix among key sub-groups

	(1)	(2)	(3)
	All women	Current contraceptive users
	Modern Contraceptive Use	LARC Use	Decision-Making
Panel A: Heterogeneity by marital status
Treat × Marstat1 (married, cohabiting, & divorced)	0.028 [−0.002,0.059]	−0.016 [−0.059,0.028]	0.034 [−0.011,0.080]
Treat × Marstat0 (never married & widow)	0.055* [0.001,0.108]	−0.112 [−0.235,0.011]	−0.016 [−0.091,0.059]
p-value	0.428	0.120	0.247
Panel B: Heterogeneity by age
Treat × Older (age > 19.5)	0.066*** [0.038,0.094]	−0.025 [−0.068,0.019]	0.025 [−0.017,0.067]
Treat × Younger (age < 19.5)	−0.183*** [−0.238,−0.128]	−0.079 [−0.169,0.011]	−0.004 [−0.069,0.062]
p-value	0.000	0.195	0.275
Panel C: Heterogeneity by knowledge
Treat × High (knowledge score > −1.14)	0.042*** [0.020,0.065]	−0.025 [−0.070,0.020]	0.019 [−0.028,0.066]
Treat × Low (knowledge score < −1.14)	−0.043 [−0.118,0.033]	−0.022 [−0.133,0.088]	0.019 [−0.045,0.082]
p-value	0.029	0.967	0.997
Panel D: Heterogeneity by education
Treat × Edu1 (at least some schooling)	0.043*** [0.021,0.065]	−0.026 [−0.073,0.021]	0.022 [−0.027,0.072]
Treat × Edu0 (no schooling)	−0.047 [−0.109,0.015]	−0.048 [−0.140,0.044]	0.009 [−0.057,0.075]
p-value	0.000	0.371	0.670
Panel E: Heterogeneity by wealth
Treat × Wealth1 (above lowest quintile)	0.029* [0.005,0.054]	−0.022 [−0.067,0.023]	0.027 [−0.021,0.075]
Treat × Wealth0 (lowest quintile)	0.003 [−0.064,0.071]	−0.099* [−0.194,−0.003]	−0.038 [−0.099,0.023]
p-value	0.061	0.115	0.171
Number of observations	14917	4229	4229

95% confidence intervals in brackets

LARC: long-acting reversible contraceptives, including implants and intrauterine devices. Decision-making: whether respondent participated in the decision on contraceptive method. This table reports the average marginal effects of the treatment (free access to a broad contraceptive method mix) on subgroups of women indicated in each panel. The subgroups were identified by a recursive partitioning machine learning analysis. A random half of the full data were used to identify the source of heterogeneity and the other half were used to estimate treatment effects reported in this table. Logistic regression models included individual covariates and country fixed effects. Propensity weights generated from generalized boosted models were used to adjust for observed differences by treatment status. The p-value rows report the two-sided p-value from an F-test of equality of the treatment effects for the two sub-groups indicated in each panel.

^*p < 0.05,

^**p < 0.01,

^***p < 0.001

Discussion

Free access to a broad contraceptive method mix may enhance women’s agency in family planning, but existing literature is lacking on whether removing user fees is an effective approach to increase women’s contraceptive use and autonomy in low-resource settings. Linking facility and individual-level data from eight countries in SSA, this study suggests that free access was associated with a modest increase in overall contraceptive use, did not prompt current users to switch contraceptive types, and did not seem to reach the most vulnerable women that need access to services the most.

The results showed that free access to both LARC and SARC was associated with a 13% increase in modern contraceptive use, indicating that demand for contraceptives is sensitive to price in contexts with limited resources, in line with recent field studies (Thompson et al. 2016; Jarvis et al. 2018; Burke et al. 2017). In addition, the increase in contraceptive use was only observed when contraceptives were provided free of charge, indicating that only making family planning services available to women without removing cost barriers would not increase women’s contraceptive use. However, this is a modest effect as it only represents an increase of contraceptive use from 24.4% to 27.3% among women who had no immediate plan to have a child. The lack of stronger effect might be explained by frequent stockout of contraceptive commodities in this setting (Mukasa et al. 2017; Douglas-Durham, Blanchard, and Higgins 2015), which can prompt providers to prioritize women who are older, married, and with children (Solo and Festin 2019). The finding that the price effect was driven by an increase in SARC use might reflect women’s preference for short-term methods because of their convenience, privacy, and fewer side effects compared to LARC as well as the ability to discontinue SARC without assistance from a willing and skilled provider (Tibaijuka et al. 2017; Britton et al. 2020; Senderowicz 2019). Moreover, by definition, free access to SARC indicated that a facility does not charge any consultation fee for family planning visits. Consultation fees can be an important access barrier in addition to method-specific charges and can add up as women need more frequent visits for short-term methods (Tumlinson et al. 2013). These findings suggest that removing user fees alone might not have a transformative impact on contraceptive uptake. In particular, women from underserved communities might not be aware of the availability of free family planning services even when they are offered for free (Tumlinson, Gichane, and Curtis 2020). For this reason, family planning promotion programs, such as the Special Family Planning Days programs that provide free contraceptive services in East Africa, often select high volume locations in the community, such as market places, and conduct community mobilization through community health workers to inform women about upcoming service schedules (High Impact Practices in Family Planning 2019). On the other hand, when resources are limited, partial subsidies that target the methods women prefer could increase overall contraceptive use.

The finding that free access was not associated with an increase in LARC use suggests that other factors might have played a larger role in LARC use than financial barriers. A recent review suggests that different promotion strategies, including vouchers, could raise uptake of IUDs, but such effects did not translate into any impact at the national level due to providers’ preference for methods that take less time and women’s discomfort with having an IUD insertion in facilities that lack the space for privacy or by a male doctor (Cleland et al. 2017). In addition, some countries have policies that restrict the use of IUDs by unmarried women or adolescents, further limiting access to LARC for certain subgroups (Ali, Folz, and Farron 2019). Even when LARC are provided for free, women might not want to use them due to the concern for removal, reliance on health providers, and less familiarity through social network (Tibaijuka et al. 2017; Mumah et al. 2018). Addressing these supply- and demand-side factors may have a larger effect on LARC use than relying on pricing strategies alone.

Additionally, among women from the poorest households, free access was associated with decreased LARC use, indicating that women who were using LARC switched to non-LARC methods when LARC were offered for free. While we need more detailed information on women’s preferences and access to understand this switching pattern, we offer one potential explanation: the treatment variable free access is defined as having free access to at least one method under two general categories of contraceptive method-types (i.e., LARC and SARC), which does not measure accurately whether a woman has free access to her preferred methods. For example, injectables, a short-term method, are the most popular contraceptive in SSA (Department of Economic and Social Affairs at United Nations 2019). In Kenya, all public sector users were supposed to receive contraceptives for free, but users of injectables were more likely to pay compared to users of implants, IUDs, pills, or condoms, indicating women’s strong preference for injectables (Radovich et al. 2019). If free access to LARC and SARC reflected more resources to cover LARC and some previously not-covered SARC methods, such as injectables, LARC users might switch to injectables when LARC were provided free of charge as defined by the treatment variable. Given the wide differences in women’s preferences for contraceptive methods (Department of Economic and Social Affairs at United Nations 2019), more detailed analysis is needed to examine whether women’s preferences and access align and how price affects demand for different contraceptive methods in specific contexts.

Older women, women who were more educated or more informed about contraceptive methods, and women from wealthier economic backgrounds were more likely to use contraceptives when LARC were offered for free. By comparison, women from more disadvantaged backgrounds might face more non-financial barriers that limit their contraceptive use, such as low health literacy, provider bias, and time costs (Tibaijuka et al. 2017). In particular, free access was associated with lower contraceptive use among adolescents. One explanation is that adolescents might be less likely to receive services from public facilities where contraceptives are provided free of charge but providers are reluctant to offer them to young unmarried women due to personal biases (Tibaijuka et al. 2017; Solo and Festin 2019). In the analysis dataset, non-public facilities rarely offer both LARC and SARC for free (see Table A4) and there is a moderate negative correlation between age (under 20) and visiting public-sector provider for the most recent family planning method (r = −0.213, p < 0.001), suggesting that adolescents were less likely to obtain contraceptives from public facilities where youth-friendly services are often lacking. Thus, the association between free access and lower contraceptive use among teenagers might indicate that other factors, such as provider bias, are more prominent barriers for young women’s access to contraceptive services. This is alarming as young women are less able to afford user fees and more susceptible to the negative consequences of unintended pregnancies, such as worse maternal and infant health outcomes, compromised educational prospects, and fewer economic opportunities (Ganchimeg et al. 2014; UNFPA 2013). To provide better access to contraceptives for all women, especially those who are most vulnerable and in need of family planning services, more research is needed to examine the barriers these women face and test the effectiveness of other interventions combined with financial subsidies, which include but are not limited to targeted information campaigns, community-based distribution strategies, and trainings that address provider biases.

Removing cost barriers did not enhance women’s role in contraceptive decision-making based on women’s response to a single standardized survey question. This might not be a surprise as 91% of women who were current contraceptive users in the dataset participated in contraceptive method decisions, leaving relatively little space to improve. In addition, this specific survey question might not capture the complex decision-making process or reflect women’s agency in whether or not to use any contraceptive. Nevertheless, this finding suggests that removing costs alone is unlikely to have any large effect on women’s autonomy in contraceptive method decisions for those already using family planning services.

This study has several limitations. First, the main treatment variable free access is defined based on the charges reported in the Service Delivery Points Questionnaire, which might not be an accurate measure due to informal fees (Tumlinson et al. 2013). We used data from the Female Questionnaire to validate facility-reported contraceptive charges. Specifically, women were asked whether they paid any fees for family planning services in the past 12 months. There was a negative correlation between user-reported fees for contraceptives and free access defined by facility surveys, r = −.19 (p < .001), providing some assurance for the validity of the treatment variable.

Second, women who live in a community where contraceptives are available free of charge might still have limited access due to travel and time costs, especially since public facilities can be located outside of the communities they serve. Women might also be denied free services due to provider biases based on age, parity, or marital status (Solo and Festin 2019). On the other hand, health facility characteristics, such as distance to the nearest health facility, was not consistently associated with contraceptive use (Zimmerman et al. 2019). While we defined access based on availability of family planning services in the community a woman lives in, some women might prefer to visit a facility that is farther away for more privacy, shorter wait time, or better quality of care, and such preferences could vary by age or rurality (Shiferaw et al. 2017; Asiimwe et al. 2014; Abate and Tareke 2019). Future studies could use detailed service utilization data to link individuals with facilities and examine the importance of costs relative to other barriers for women’s contraceptive uptake.

Third, contraceptive stockout is usually high in the study settings but was not common based on the data from the Service Delivery Points surveys (Mukasa et al. 2017; Douglas-Durham, Blanchard, and Higgins 2015). For example, among facilities that routinely provided implants and IUDs, about 90% had them in stock and this was verified by enumerators on the day of the survey. Although the Family Planning 2020 Initiative might have brought more resources and strengthened supply chain systems in these countries, it is still likely that data from these facility surveys might not capture whether women had routine access to family planning services. Thus, women who did not have free access might have been misclassified in our analyses, which would underestimate the treatment effects.

Fourth, this study pooled data from eight countries in SSA that have pledged to the Family Planning 2020 Initiative to estimate the association between access and contraceptive use in this region. However, these eight countries vary greatly among themselves, as indicated by the contraceptive prevalence rates in our data, and our analyses did not provide country-specific estimates because the sample size would be too small. Meanwhile, these eight countries might not be representative of all countries in SSA, limiting the generalizability of the study findings. For example, these countries might have higher government commitment to financing family planning services, lower contraceptive prevalence before the Initiative, or different social norms about the use of LARC versus short-term methods. Fifth, the analysis sample does not capture all facilities that serve a community (Zimmerman 2017). If facilities included in the study were different from those excluded from the sample, such sampling bias might limit the generalizability of study findings.

Lastly, unobserved factors could have confounded the estimated associations between free access and contraceptive use. Future studies that randomly assign treatment status would provide stronger evidence on the effects of price on contraceptive use. While researchers need to overcome substantial financial, operational, and regulatory barriers to conduct experimental studies in LMIC (Alemayehu, Mitchell, and Nikles 2018), price experiments have been done on other public health topics in this setting, such as malaria control and HIV prevention, and have provided important evidence for policy decisions (Comfort and Krezanoski 2017; Fischer et al. 2019; Ashraf, Berry, and Shapiro 2010; Cohen and Dupas 2010; Dupas 2014; Chang et al. 2019).

Conclusions

Free access to a broad contraceptive method mix was associated with a modest increase in contraceptive use, driven by higher use of short-term methods, based on data from eight countries in SSA. Among current contraceptive users, free access to LARC did not prompt more women to switch to LARC and had no effect on women’s autonomy in contraceptive decision-making. Governments and international donors should continue using financial subsidies to accelerate contraceptive uptake, but removing user fees alone does not necessarily increase access or choice, especially for women from more disadvantaged segments of the population who need family planning services the most.