Hormonal contraception and hemoglobin levels in urban Malawi: a longitudinal analysis.

We use three waves of longitudinal data that were collected as part of a randomized controlled trial of a family planning intervention in urban Malawi that was conducted between 2016 and 2019; a more detailed description of the study design and data collection approach is presented elsewhere [ 31 , 32 ]. As part of the trial, 2,143 women who were either pregnant or immediately postpartum (within 6 months of their last live birth), between the ages of 18 and 35, and living in Lilongwe, the capital, were recruited in 2016. Following a baseline survey, which was conducted from 2016 to 2017, women were individually randomized to either an intervention arm or control arm. Women assigned to the intervention arm received a multicomponent package of family planning services over a two-year period. Two annual follow-up surveys were conducted with all women from 2017 to 2018 and from 2018 to 2019, respectively. Our data are measured using a range of validated metrics and instruments. In particular, our baseline and follow-up survey instruments are comprised of modules from the household and women’s questionnaires from the Malawi Demographic and Health Survey (DHS) and includes information on fertility, family planning, reproductive health, and maternal health [ 33 ]. In each survey wave, women are asked to report their contraceptive use, history of contraceptive use or non-use since the last interview, and the type of contraceptive method(s) that they are currently using. We classify a woman to be using hormonal contraception if she reports using either the contraceptive implant (Jadelle), injectable contraception (Depo-Provera), or the oral contraceptive pill at the time of interview. In our study setting, only a nonhormonal IUD brand (specifically, the copper IUD) was available to women at the time when the study was conducted; we therefore assign a woman who reports using an IUD at the time of interview to be using a nonhormonal method. For women who were enrolled in the study, anthropometric data that included height, weight, and hemoglobin measures, as an indicator for anemia, were collected in all three survey waves. The study team measured women’s hemoglobin levels (in g/cL) using a rapid, on-site blood diagnostic test, the HemoCue 201 + point-of-care system, and followed the standard DHS protocols for the collection and management of blood biomarker data. For our study, we analyze the role of women’s contraceptive use, and particularly hormonal contraception, on measured hemoglobin levels at the time of the survey (in g/cL). Throughout the study, hemoglobin measurement rates among eligible women are low. Two factors drive these low rates. First, this study was conducted in a densely urban environment where women were often not available at home and therefore had to often be interviewed by phone; anthropometric and hemoglobin measurements were therefore not collected for these women. Second, we observe a consistently low rate of consent among women across our surveyed waves, which may, in part, have been driven by religious beliefs and mistrust around the collection and handling of blood. 1 As a result, we were forced to halt all on-site blood testing for anemia, out of concern for the safety of our field research team. Our main empirical models exploit the longitudinal study design. Specifically, we run linear panel models with two-way (woman-level and survey wave-level) fixed effects as follows: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{it}=\beta\:Metho{d}_{it}+{\boldsymbol{X}}_{it}\boldsymbol{\gamma\:}+{\delta\:}_{i}+{\eta\:}_{t}+{\epsilon\:}_{it}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{it}$$\end{document} is the measured hemoglobin level for woman \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document} at survey wave \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t$$\end{document} . \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Metho{d}_{it}$$\end{document} is a categorical variable that identifies the type of contraceptive method used (with non-use coded as 0); we also run specifications that classifies a woman’s method use as a categorical variable, taking on a value of 2 if the woman reported using a hormonal contraceptive method at survey wave \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t$$\end{document} , 1 if the woman reported using a non-hormonal contraceptive method at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t$$\end{document} , and 0 otherwise. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{X}}_{it}$$\end{document} represents a vector of time-varying covariates, including women’s work status (working / not working), whether the women own agricultural land (yes / no), body mass index (BMI) at the time of the survey, educational attainment (none, primary, secondary, or higher), and parity, measured by the total number of children to whom the woman has given birth. To account for unobserved woman-level and temporal variation, we include individual woman-level fixed or random effects ( \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\delta\:}_{i}$$\end{document} ) as well as survey wave fixed effects ( \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\eta\:}_{t}$$\end{document} ); finally, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\epsilon\:}_{it}$$\end{document} represents the error term. To compare our panel estimates with previous cross-sectional approaches, we run a linear ordinary least-squares (OLS) regression specification using cross-sectional data from the third (2018–2019) survey wave as follows: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{i}=\alpha\:+\beta\:Metho{d}_{i}+{\boldsymbol{Z}}_{i}\boldsymbol{\gamma\:}+{\nu\:}_{i}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{Z}}_{i}$$\end{document} represents a vector of individual-level covariates that were collected for each woman \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document} in that wave, including women’s work status (working / not working), whether the women own agricultural land (yes / no), body mass index (BMI) at the time of the survey, educational attainment (none, primary, secondary, or higher), ever use of contraception (yes / no), age (in years), pregnancy status at the time of the survey (yes / no), and parity, measured by the total number of children to whom the woman has given birth. We note that in the cross-sectional model, the error term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\nu\:}_{i}$$\end{document} is equivalent to the sum of individual random effect \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\delta\:}_{i}$$\end{document} and the random error term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\epsilon\:}_{it}$$\end{document} in the panel models. The fixed effects model is more general than the random effects approach. With fixed effects, we allow for the unobserved woman-level characteristics \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\delta\:}_{i}$$\end{document} to be correlated with a woman’s contraceptive method use \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Metho{d}_{it}$$\end{document} ; specifically, women who use different methods may have different unobserved characteristics. In contrast, the random effects model is more restrictive in that it assumes that a woman’s contraceptive use is uncorrelated with any unobserved characteristics for the woman. While the random effects model is less general, it is a more efficiently estimated model if the assumption that the unobserved characteristics are random is indeed true. This is due to the fact that the fixed effect model requires the estimation of each woman-level fixed effect, which reduces the model’s degrees of freedom and leads to wider confidence intervals for the estimated parameters of interest. We test if the assumptions of the random effect model hold relative to the fixed effects model by calculating the Wu-Hausman F-statistic for parameter equality between the parameter estimates of the fixed effects and random effects models [ 35 , 36 ]. Rejection of equality using this test is evidence that the random effects parameter estimates are biased and the fixed effect estimates are consistent. In contrast, failure to reject parameter equality indicates that the random effects assumption may be correct and the fixed effects estimates, though also consistent, may be less efficient. In addition to testing for parameter equality overall, we can test each parameter individually for equality across the two models to observe which parameters seem to be biased if the results differ.

Table  1 presents descriptive statistics by survey wave, and Appendix Fig. 1 presents a flowchart as to how the analytic sample was obtained after accounting for observations with missing data in our key outcomes and other variables of interest. We observe an increase in measured hemoglobin levels in our sample between the first survey and subsequent waves, which is indicative of postnatal recovery in our sample of pregnant and immediate postpartum women from when they were initially recruited at baseline. Figure 1 presents the rightward shift in the distribution of measured hemoglobin in our sample from baseline (2016) to endline (2018), which provides additional evidence of this recovery over time. Similarly, anemia prevalence in our sample decreases over time from 57% in 2016 (when women were either pregnant or postpartum at recruitment) to 29.1% in 2018. Table 1 Descriptive Statistics, by survey wave 2016 2017 2018 Mean SD Mean SD Mean SD Outcomes     Hemoglobin (g/cL) 110.196 17.278 125.767 15.272 126.054 14.559     Hemoglobin, Preg. Adjusted (g/cL) 115.628 16.351 125.958 15.191 126.431 14.442 Anemia Status     None (adjusted greater than 120 g/cL) 0.430 0.695 0.709     Low (adjusted between 110 and 120 g/cL) 0.243 0.186 0.172     Moderate (adjusted between 80 and 110 g/cL) 0.307 0.112 0.118     Severe (adjusted less than 80 g/cL) 0.020 0.008 0.002     Anemic (1 = Yes) 0.570 0.305 0.291 Covariates     Currently Pregnant (1 = Yes) 0.543 0.137 0.038     Currently Using FP (1 = Yes) 0.217 0.370 0.858 Current Method Use     None 0.783 0.163 0.142     Female Sterilization 0.000 0.026 0.028     IUD 0.001 0.008 0.008     Injectable 0.173 0.572 0.499     Implant 0.034 0.166 0.239     Pills 0.003 0.036 0.054     Condoms 0.006 0.018 0.017     Standard Days Method 0.000 0.005 0.005     Withdrawal 0.001 0.007 0.008     Hormonal Method (1 = Yes) 0.210 0.774 0.792 Educational Attainment     None 0.568 0.525 0.523     Primary 0.407 0.448 0.457     Secondary 0.025 0.026 0.019     Higher 0.000 0.001 0.002     Currently working (1 = Yes) 0.086 0.190 0.227     Owns agricultural land (1 = Yes) 0.347 0.320 0.301     Age (years) 24.264 4.474 25.103 4.728 26.033 4.654     Total number of children 1.641 1.305 2.176 1.262 2.243 1.273     BMI 24.685 3.970 24.288 4.293 25.135 4.437     N 1447 1049 635 The unit of observation is a woman. Standard deviations are presented for non-binary variables Descriptive Statistics, by survey wave The unit of observation is a woman. Standard deviations are presented for non-binary variables Fig. 1 Histogram, Hemoglobin Levels (g/cL) in 2016 and 2018 Histogram, Hemoglobin Levels (g/cL) in 2016 and 2018 Contraceptive use in our sample increases considerably over the study period, from 21.7% in 2016 to 85.8% in 2018. Finally, we note that hormonal method use makes up a significant majority (accounting for more than 90%) of all contraceptive use by women in our sample in any year. Table  2 presents estimates from our cross-sectional OLS, fixed effects, and random effects panel regressions, respectively. Across all three specifications, we note a positive association between hormonal contraceptive method use, particularly the injectable and the implant, and hemoglobin levels in women. Our cross-sectional results (column 1) indicate that compared to non-users of contraception (the reference group), injectable and implant use among women in 2018 is associated with a 4.7 g/cL and a 7.3 g/cL increase in hemoglobin levels, respectively. In contrast, our random effects results (column 3) indicate that a change in women’s method use from non-use to injectable and implant use (i.e. adoption of the method) between 2016 and 2018 is associated with a 3.7 g/cL and 5.7 g/cL respective increase in their hemoglobin levels. Estimates from our fixed effects model (column 2) show a similar, but smaller positive association between these hormonal methods and hemoglobin in women over our study period. In comparing the coefficients from the random effects and fixed effects models (Table  3 ), we find significantly larger estimates in the random effects model relative to the fixed effects model for the injectable and the pill. In addition, a formal Wu-Hausman test of parameter equality shows that we fail to reject parameter equality between the two models, indicating that the random effects model, as the more efficient model, would be preferred. Table 2 Contraceptive method used and hemoglobin levels (1) Cross Section OLS (2) Panel Fixed Effects (3) Panel Random Effects Outcome Hemoglobin (g/cL) Hemoglobin (g/cL) Hemoglobin (g/cL) Pregnancy Status (Ref: No)     Pregnant (1 = Yes) -6.745 -9.028*** -9.923*** (-15.10–1.614) (-13.37 - -4.690) (-11.70 - -8.148) Method Use (Ref: No Use)     Female Sterilization 4.103 -1.214 -2.022 (-2.671–10.88) (-8.369–5.941) (-6.924–2.879)     IUD 5.198 3.020 7.038* (-1.511–11.91) (-7.898–13.94) (-1.116–15.19)     Injectable 4.742** 1.570 3.684*** (0.0125–9.471) (-0.681–3.822) (2.060–5.309)     Implant 7.285*** 3.828** 5.705*** (2.259–12.31) (0.644–7.012) (3.532–7.878)     Pill 1.783 -1.197 2.169 (-5.299–8.866) (-5.977–3.583) (-1.519–5.856)     Condom -0.638 -0.504 -0.219 (-9.589–8.314) (-7.774–6.766) (-5.300–4.861)     Standard Days 8.142 -1.578 3.283 (-4.621–20.91) (-17.31–14.15) (-7.218–13.78)     Withdrawal -0.154 1.216 2.625 (-12.50–12.19) (-9.152–11.58) (-5.519–10.77)     Constant 118.5*** 107.7*** 108.5*** (106.8–130.2) (95.30–120.1) (104.7–112.3) Test of Fixed Effect versus Random Effect models     Wu-Hausman F-statistic (p-value) 15.26 (0.4325)     Observations 635 3,131 3,131     Number of Women 635 1,732 1,732     R-squared 0.155 *** \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.01$$\end{document} , ** \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.05$$\end{document} , * For all columns, the unit of observation is a woman. Column 1 presents OLS results from an adjusted linear model using the wave 3 cross-sectional data, with 95% confidence intervals in parentheses. Columns 2 and 3 respectively present results from adjusted panel woman fixed effects (FE) and woman random effects (RE) regression models with 95% confidence intervals in parentheses. Time-varying covariates in all models include the total number of children ever born, whether the woman works, the woman’s BMI, and whether the woman’s household owns any agricultural land. Time-invariant covariates in the OLS model include the woman’s educational attainment (none, primary, secondary, or higher), age of the woman (in completed years), and her ever use of contraception. For the OLS model, area and survey wave fixed effects are included, and heteroscedasticity-robust standard errors are presented in all regressions. For the panel FE and RE models, survey wave fixed effects are included in all regressions Contraceptive method used and hemoglobin levels *** \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.01$$\end{document} , ** \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.05$$\end{document} , * For all columns, the unit of observation is a woman. Column 1 presents OLS results from an adjusted linear model using the wave 3 cross-sectional data, with 95% confidence intervals in parentheses. Columns 2 and 3 respectively present results from adjusted panel woman fixed effects (FE) and woman random effects (RE) regression models with 95% confidence intervals in parentheses. Time-varying covariates in all models include the total number of children ever born, whether the woman works, the woman’s BMI, and whether the woman’s household owns any agricultural land. Time-invariant covariates in the OLS model include the woman’s educational attainment (none, primary, secondary, or higher), age of the woman (in completed years), and her ever use of contraception. For the OLS model, area and survey wave fixed effects are included, and heteroscedasticity-robust standard errors are presented in all regressions. For the panel FE and RE models, survey wave fixed effects are included in all regressions Table 3 Test of parameter equality on methods for fixed effect and random effect models FE RE Diff. (FE – RE) SE Test Statistic Comparison of coefficients Table  3 , Columns 2 and 3     Pregnant (1 = Yes) -9.028 -9.923 0.895 2.01 0.445 Current Method Use     Female Sterilization -1.214 -2.022 0.808 2.64 0.306     IUD 3.02 7.038 -4.018 3.671 -1.095     Injectable 1.57 3.684 -2.114 0.789 -2.679**     Implant 3.828 5.705 -1.877 1.178 -1.593     Pills -1.197 2.169 -3.366 1.536 -2.191**     Condoms -0.504 -0.219 -0.285 2.632 -0.108     Standard Days Method -1.578 3.283 -4.861 5.934 -0.819     Withdrawal 1.216 2.625 -1.409 3.24 -0.435     Currently working (1 = Yes) 0.579 0.694 -0.115 0.73 -0.158     Owns agricultural land (1 = Yes) -0.474 -0.444 -0.03 0.691 -0.043     BMI 0.149 0.275 -0.125 0.155 -0.806     Total number of children 2.189 -0.261 2.451 2.095 1.17 Survey Wave     2017 8.158 8.089 0.069 0.59 0.117     2018 8.556 8.287 0.269 0.711 0.378 Wu-Hausman F-Statistic (p-value) 15.26 (0.4325) *** \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.01$$\end{document} , ** \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.05$$\end{document} , * \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.1$$\end{document} Test of parameter equality on methods for fixed effect and random effect models *** \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.01$$\end{document} , ** \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.05$$\end{document} , * \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.1$$\end{document} In Table  4 , we present estimates from our three models that use a coarser categorical indicator of contraceptive use (non-use, non-hormonal method use, and hormonal method use). Our findings are qualitatively consistent with the estimates that we obtain from the models that examine methods separately. With this said, we note that the estimated association between hormonal method use and hemoglobin levels in the fixed effects model are smaller and insignificant relative to the random effects model, which highlight a 3.9 g/cL associated increase in hemoglobin levels with hormonal method use. Findings from the Wu-Hausman test (Table  5 ) indicate significant differences in the fixed effects model estimates relative to the random effects model, to the extent that we would reject the null hypothesis of parameter equality in favor of the fixed effects model. Table 4 Hormonal contraceptive use and hemoglobin levels (1) Cross Section OLS (2) Panel Fixed Effects (3) Panel Random Effects Outcome Hemoglobin (g/cL) Hemoglobin (g/cL) Hemoglobin (g/cL) Pregnancy Status (Ref: No)     Pregnant (1 = Yes) -6.978* -9.052*** -9.988*** (-15.26–1.299) (-13.37 - -4.730) (-11.76 - -8.217) Method Use (Ref: No Use)     Non-Hormonal Method (1 = Yes) 2.135 -0.239 0.251 (-3.394–7.664) (-4.761–4.283) (-2.985–3.487)     Hormonal Method (1 = Yes) 4.931** 1.729 3.917*** (0.466–9.397) (-0.452–3.911) (2.350–5.483)     Constant 118.3*** 108.0*** 108.7*** (106.9–129.8) (95.62–120.4) (104.9–112.5) Test of Fixed Effect versus Random Effect models     Wu-Hausman F-statistic (p-value) 20.97 (0.0.0128)     Observations 635 3,131 3,131     Number of Women 1,732 1,732     R-squared 0.144 *** \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.01$$\end{document} , ** \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.05$$\end{document} , * For all columns, the unit of observation is a woman. Column 1 presents OLS results from an adjusted linear model using the wave 3 cross-sectional data, with 95% confidence intervals in parentheses. Columns 2 and 3 respectively present results from adjusted panel woman fixed effects (FE) and woman random effects (RE) regression models with 95% confidence intervals in parentheses. Time-varying covariates in all models include the total number of children ever born, whether the woman works, the woman’s BMI, and whether the woman’s household owns any agricultural land. Time-invariant covariates in the OLS model include the woman’s educational attainment (none, primary, secondary, or higher), age of the woman (in completed years), and her ever use of contraception. For the OLS model, area and survey wave fixed effects are included, and heteroscedasticity-robust standard errors are presented in all regressions. For the panel FE and RE models, survey wave fixed effects are included in all regressions Hormonal contraceptive use and hemoglobin levels *** \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.01$$\end{document} , ** \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.05$$\end{document} , * For all columns, the unit of observation is a woman. Column 1 presents OLS results from an adjusted linear model using the wave 3 cross-sectional data, with 95% confidence intervals in parentheses. Columns 2 and 3 respectively present results from adjusted panel woman fixed effects (FE) and woman random effects (RE) regression models with 95% confidence intervals in parentheses. Time-varying covariates in all models include the total number of children ever born, whether the woman works, the woman’s BMI, and whether the woman’s household owns any agricultural land. Time-invariant covariates in the OLS model include the woman’s educational attainment (none, primary, secondary, or higher), age of the woman (in completed years), and her ever use of contraception. For the OLS model, area and survey wave fixed effects are included, and heteroscedasticity-robust standard errors are presented in all regressions. For the panel FE and RE models, survey wave fixed effects are included in all regressions Table 5 Test of parameter equality on hormonal method use for fixed effect and random effect models FE RE Diff. (FE – RE) SE Test Statistic Comparison of Table  5 , Columns 2 and 3     Pregnant (1 = Yes) -0.129 0.172 -0.301 0.376 -0.801 Current Method Use (Ref: No Use)     Use Non-Hormonal Method -0.119 0.103 -0.222 0.411 -0.54     Use Hormonal Method -0.144 -0.492 0.348 0.162 2.148**     Currently working (1 = Yes) -0.361 -0.183 -0.178 0.166 -1.072     Owns agricultural land (1 = Yes) 0.188 0.026 0.162 0.158 1.025     BMI -0.03 -0.046 0.016 0.033 0.485     Total number of children -0.568 0.052 -0.621 0.381 -1.63 Survey Wave     2017 -0.92 -1.015 0.095 0.121 0.785     2018 -1.082 -1.086 0.005 0.144 0.035     Wu-Hausman F-Statistic (p-value) 20.97 (0.0128) *** \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.01$$\end{document} , ** \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.05$$\end{document} , * \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.1$$\end{document} Test of parameter equality on hormonal method use for fixed effect and random effect models *** \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.01$$\end{document} , ** \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.05$$\end{document} , * \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p<0.1$$\end{document}

Most methods of contraception were developed with the goal of preventing unintended and mistimed pregnancies in mind. However, a growing body of research has also identified potential non-contraceptive benefits from contraceptive method use, particularly from the use of hormonal, reversible contraceptive methods such as the intrauterine device (IUD), the contraceptive implant, injectable contraception, and the oral contraceptive pill [ 1 , 2 ]. Some studies have examined the extent to which the use of oral contraceptives and IUDs may reduce the risk of gynecological cancers, while others have documented the effectiveness of these methods as treatments for a wide range of conditions, including irregular and heavy menstrual bleeding, dysmenorrhea, premenstrual syndrome, and endometriosis, among others [ 3 – 6 ]. In addition, a number of studies have explored the role of hormonal contraception in reducing the risk of non- gynecological conditions, such as severe asthma [ 7 ] or colorectal cancer [ 8 – 10 ]. At the same time, studies have also identified potential risks and side effects associated with hormonal contraceptive use, including thromboembolic events [ 11 ], breast cancer [ 12 ], and some sexually transmitted infections [ 13 ]. Given the physiological pathways through which hormonal contraception may interact with menstrual flow, anovulatory bleeding, and other related conditions, a growing body of work has documented the extent to which hormonal method use might be associated with reductions in the risk of anemia in women. Anemia, which is characterized by low hemoglobin levels in the bloodstream, affects more than 2.2 billion people globally and an estimated 4 in 10 people in low- and middle-income countries [ 14 , 15 ]. Anemia is more commonly observed in children and in women of reproductive age; however, while the primary cause of anemia in children is insufficient iron in the diet, anemia in women can also be due to blood loss during menstruation and pregnancy [ 16 , 17 ]. Low concentrations of blood-level hemoglobin have been linked to adverse physical and cognitive health outcomes (fatigue, low energy, poor productivity), and anemia during pregnancy has also been associated with increased risks of maternal mortality and morbidity [ 18 , 19 ]. A number of studies, including several from low- and middle-income countries, have assessed the correlation between hormonal contraceptive use and increased hemoglobin levels, leading to a reduction in women’s risk of anemia [ 20 – 25 ]. While these findings are promising, the scope for causal inference from these studies is restricted by their cross-sectional design, which limits the extent to which bias from individual-level confounding can effectively be mitigated. In cross-sectional studies, it is difficult to distinguish if a correlation between contraceptive use and hemoglobin levels is due to an effect of the contraceptive, or whether women of a particular type, who tend to use a particular contraceptive method, are more likely to have elevated or depressed hemoglobin levels independently of their use. This can be partially addressed by adding control variables, but it is difficult to ensure that all relevant factors have been accounted for so as to eliminate confounder bias. In addition, we note that some studies have examined the broader categorical effect of hormonal versus non-hormonal contraceptive methods on anemia, while other studies have documented the effect of each specific method separately. The separation of effect size by method may be important because different hormonal methods (oral contraceptives, implants, injectables, and IUDs) use different types of hormones at different concentrations and have different time patterns of hormone release [ 26 , 27 ]. In this study, we assess the relationship between women’s use of hormonal contraception and changes to their risk of anemia using panel data on women from urban Malawi. Our study setting presents a valuable opportunity to study the relationship between women’s hormonal contraceptive use and anemia status during and following pregnancy. Anemia in pregnancy is a persistent public health challenge in Malawi [ 28 , 29 ]; findings from the Malawi Demographic and Health Survey (2015–2016) indicate that 33% of women of reproductive age are anemic, including 45% of pregnant women and 30% of breastfeeding women. Moreover, anemia prevalence in women is slightly higher in urban areas (36%) compared to rural areas (30%) [ 30 ]. At the same time, hormonal contraceptive use is both high and stable in Malawi. Contraceptive prevalence among married women increased from 43% in 2010 to almost 60% in 2015—one of the highest rates in Sub-Saharan Africa—while the method mix has remained relatively unchanged, with injectable contraceptives (about 30%) and intrauterine devices (about 11.5%) being the most widely used methods. We document the extent to which women’s adoption of hormonal contraceptive methods following pregnancy relates to changes to their hemoglobin level. The panel data allows us to control for individual woman fixed effects, capturing all women specific characteristics that are time invariant. More directly, the cross-sectional approach asks if using a contraceptive method is associated with a higher hemoglobin level, while the fixed effects approach asks if a woman’s hemoglobin level changes when she changes to a new contraceptive method. Our longitudinal data structure thus allows us to provide more rigorously identified estimates of these relationships. An important issue that we examine is to determine whether using the panel data approach with fixed effects changes the results of the analysis relative to the cross-sectional estimates; this comparison would indicate the extent to which the existing body of evidence that is based on cross-sectional data may be reliable. We make this comparison by testing whether models with individual women fixed effects yield different coefficient estimates than models that assume that any individual woman-level effects are random and are therefore uncorrelated with contraceptive use and do not confound the relationship. If we find that the panel data approach with individual fixed effects does not significantly affect coefficient estimates, it will indicate the reliability of the results from studies that use cross-sectional approaches.

We assess the relationship between Malawian women’s hormonal contraceptive use and their risk of anemia, measured by hemoglobin levels, using a range of panel and cross-sectional approaches. We find that hormonal method use, particularly implant use and (to a lesser degree) injectable use, is associated with increased levels of hemoglobin in women. These estimates are consistent across model specifications. Our main method-specific analyses yield two principal findings. First, we note that our results from adding individual woman-level fixed effects in a panel data approach are not significantly different from those that we obtain when using a random effects model. Second, we do not find sufficient evidence to conclude that the cross-sectional association between contraceptive method use and hemoglobin levels is due to unobserved women characteristics; specifically, we note the overlap in the confidence intervals between our cross-sectional OLS estimates and panel models. In contrast, we do find that compared to the cross-section results, the fixed effects and random effects approaches give significantly different results when we aggregate contraceptive methods into hormonal and non-hormonal method groups. In this case, the estimated coefficient on hormonal methods using the fixed effects approach is smaller and less significant than that the coefficient that we estimate using random effects and cross-section approaches. However, we do not think that this observed difference in estimates is necessarily due to differences in how individual heterogeneity is accounted for in each model, but rather due to variable misspecification in the models in which all hormonal methods are grouped together. This is because the cross-sectional models estimate the associations between the mix of hormonal methods being used and hemoglobin levels, whereas the panel models with fixed (or random) effects estimate the within-woman associations between changes to hormonal method use through either adoption or discontinuation and changes to hemoglobin levels. The patterns in current hormonal method use across women and the patterns in adoption and discontinuation among these methods within each woman may be inherently different, which would likely yield different effect size estimates if different hormonal methods have different effects on hemoglobin. In this study, we focus on the effect of contraceptive method use on hemoglobin levels rather than on anemia. Anemia is a binary outcome that is based on cut points in the hemoglobin level to delineate levels of severity [ 37 ]. The analysis of panel data using fixed effects for binary or categorical outcomes can still be carried out using the conditional logit model but is problematic because the fixed effect is not identified for any women who do not change their anemia status; as a result, these women are dropped from the estimation [ 38 ]. This leaves us with a very small sample size and limited power for examining anemia using the fixed effects approach, and it becomes difficult to draw conclusions from this approach.

Our study reinforces the evidence base that identifies an effect of hormonal contraceptive use on hemoglobin levels using cross-sectional data, and our findings suggest that the cross-sectional results may be reliable. However, we caution that grouping all hormonal methods together, thereby treating them as having a common effect size, may be misleading and may give rise to differences in estimated effect sizes as the composition of individual methods within the method group change. Our findings have significant implications for health policies and programs in Malawi, where the Government of Malawi has prioritized reducing maternal morbidity and mortality, and addressing anemia in pregnancy is central to these efforts [ 39 ]. At the same time, strengthening family planning programs and improving access to services remains a cornerstone of Malawi’s reproductive health strategy [ 40 , 41 ]. Taken together, our findings highlight potential maternal health benefits from hormonal contraceptive use that extend beyond their primary function to prevent pregnancy. These benefits, in turn, have significant implications as to how women are counseled on contraception both within family planning and nutrition programs. At the very least, awareness of these potential advantages during contraceptive counseling would help women to make more informed reproductive health choices.

Supplementary Material 1. Supplementary Material 1.