Prevalence and determinants of unintended pregnancy in Mchinji district, Malawi; using a conceptual hierarchy to inform analysis

Background In 2012 there were around 85 million unintended pregnancies globally. Unintended pregnancies unnecessarily expose women to the risks associated with pregnancy, unsafe abortion and childbirth, thereby contributing to maternal mortality and morbidity. Studies have identified a range of potential determinants of unplanned pregnancy but have used varying methodologies, measures of pregnancy intention and analysis techniques. Consequently there are many contradictions in their findings. Identifying women at risk of unplanned pregnancy is important as this information can be used to help with designing and targeting interventions and developing preventative policies. Methods 4,244 pregnant women from Mchinji District, Malawi were interviewed at home between March and December 2013. They were asked about their pregnancy intention using the validated Chichewa version of the London Measure of Unplanned Pregnancy, as well as their socio-demographics and obstetric and psychiatric history. A conceptual hierarchical model of the determinants of pregnancy intention was developed and used to inform the analysis. Multiple random effects linear regression was used to explore the ways in which factors determine pregnancy intention leading to the identification of women at risk of unplanned pregnancies. Results 44.4% of pregnancies were planned. On univariate analyses pregnancy intention was associated with mother and father’s age and education, marital status, number of live children, birth interval, socio-economic status, intimate partner violence and previous depression all at p<0.001. Multiple linear regression analysis found that increasing socio- economic status is associated with increasing pregnancy intention but its effect is mediated through other factors in the model. Socio-demographic factors of importance were marital status, which was the factor in the model that had the largest effect on pregnancy intention, partner's age and mother's education level. The effect of mother's education level was mediated by maternal reproductive characteristics. Previous depression, abuse in the last year or sexual abuse, younger age, increasing number of children and short birth intervals were all associated with lower pregnancy intention having controlled for all other factors in the model. This suggests that women in Mchinji District who are either young, unmarried women having their first pregnancy, or older, married women who have completed their desired family size or recently given birth, or women who have experienced depression, abuse in the last year or sexual abuse are at higher risk of unintended pregnancies. Conclusion A simple measure of pregnancy intention with well-established psychometric properties was used to show the distribution of pregnancy planning among women from a poor rural population and to identify those women at higher risk of unintended pregnancy. An analysis informed by a conceptual hierarchical model shed light on the pathways that lead from socio-demographic determinants to pregnancy intention. This information can be used to target family planning services to those most at risk of unplanned pregnancies, particularly women with a history of depression or who are experiencing intimate partner violence. Mchinji District is a rural district in the central region of Malawi. It borders Zambia and Mozambique and has a population of around 500 000 (local data, unpublished). Malawi has a prevalence of unintended pregnancy of 45% [30] but there are no data on variations either by district or by women’s characteristics. Previous research divided the district into 49 geographical areas of approximately equal population [31]; from this sampling frame a random sample of 25 areas were selected to take part in research into pregnancy intention and maternal and neonatal outcomes. Using the pre-existing district-wide surveillance system, all pregnant women that were identified and were aged 15 and over in these 25 areas between March and December 2013 were invited to participate. 4,244 interviews were conducted with pregnant women in Mchinji District. These data were collected as part of a study into the relationships between pregnancy intention and pregnancy outcome. For this study a sample size calculation was conducted assuming 41% of pregnancies would be unplanned [32], a prevalence of 15% for adverse pregnancy outcome, [30, 33–35] a relative risk of 1.25, for 80% power at the 0.05 significance level; we thereby estimated that 3,737 pregnancy outcomes were needed. One of 25 trained data collectors visited the pregnant women at home and conducted a 20-minute interview using a questionnaire programmed using CommCare ODK software on a smartphone. Women were asked demographic and obstetric history questions considered to be potential determinants of pregnancy intention on the basis of previous literature. including questions about the father of the child and previous episodes of depression or intimate partner violence (IPV). The variables considered as potential determinants of pregnancy intention are shown in Table 2. A principal components analysis (PCA) was conducted to generate an asset-based measure of socio-economic status (SES). In addition to ownership of assets such as a bicycle and radio, variables included in the PCA were household characteristics, such as floor and roof materials and household density, and access to water and sanitation facilities. The distribution of SES score from this PCA was grouped into quintiles to create an ordered categorical variable for SES. Pregnancy intention was measured using the validated Chichewa version of the LMUP [36]. In the absence of a validated tool for assessing previous depression, we worked with experts in the field to devise four questions about previous periods of low mood or anhedonia and whether these lasted for more than two weeks. These were used to categorise women as to the extent of possible previous depression. Women who experienced both low mood and anhedonia for more than two weeks were considered most likely to have experienced previous depression. Those who experienced only one of these, or who experienced them for a period of less than two weeks, were considered less likely to have experienced previous depression and women who reported neither of these were the least likely to have experienced previous depression. IPV was assessed using the Abuse Assessment Screen [37]. This asks about experience of abuse ever, in the last year or while pregnant as well as experience of sexual abuse. GPS readings of the location of the interview were taken and were used to calculate the distance to the nearest health facility. To assess the determinants of pregnancy intention, each potential determinant was considered in a univariate linear regression analysis before developing a multiple regression model. Collinearity was examined prior to the selection of variables for inclusion in the hierarchical model selection process. Where variables were collinear (e.g. marital status and living arrangements) only one variable was included in the model selection process. Categorical variables were entered into the models as sets of dummy variables in the standard way. For example, socio-economic quintile had the poorest quintile as baseline and the effects of the other quintiles were assessed relative to the baseline. Several regression models were considered and a linear regression model with robust standard errors was selected as a good fit for the data. A random effects model was used to account for the clustering of our participants within geographical areas. The conceptual hierarchical model was used to inform the creation of the final multiple linear regression model [29]. The aim of the model was to understand the effects of various potential determinants of unintended pregnancies. All analyses were conducted in Stata version 13. The multiple linear regression models were created using the ‘reg’ command with the ‘robust’ suffix, with ‘xtreg’ used for the random effects models. The conceptual hierarchical model (Fig 1) was developed based on the literature and temporal considerations with variables grouped into the five hierarchical levels as shown in Table 3. It starts with the most distal determinant, SES, as measured by the asset-based measure that had taken multiple variables into account in Level One at the top, and works down through increasingly more proximate determinants. Variables higher in the hierarchy influence those below them either indirectly, through their effect on the variables in other levels in the model, or directly. For example, SES may affect pregnancy intention indirectly through its effect on education (a Level Two variable, pathway a in Fig 1), previous depression (Level 3, pathway b), number of children (Level 4, pathway c) or gestation which is the most proximate determinant of pregnancy intention in our conceptual hierarchical model (Level 5, pathway d). SES may also have a direct effect on pregnancy intention (pathway e). Socio-demographic variables were included in the second level of the hierarchy. Geographical area was included in Level Two, when it was introduced as a random effect to acknowledge the clustering of study participants within geographical areas. It was added at Level Two because areas do vary in their distribution of SES and we were only interested in the effect of area that was not due to differences in SES. Previous depression and intimate partner violence were introduced in Level Three, before maternal reproductive factors in Level Four, because we are looking at experience of these factors in the year prior to becoming pregnant and this may have played a role in becoming pregnant and the intendedness of this pregnancy. Gestation was included in Level Five as a marker of the time since conception and as the most proximate variable. Although the development of a conceptual hierarchical model is based on the literature it is, nonetheless, subjective. It could be argued that educational achievement is a past exposure that determines current SES and should therefore be considered higher than SES, contrary to our hierarchy. Equally a high number of children may be a reason for previous episodes of depression rather than depression preceding a high number of children, as in our hierarchy. In recognition of this we conducted a number of sensitivity analysis to test the robustness of the findings to decisions we had made in constructing our conceptual hierarchical model. Only variables that were associated with pregnancy intention at p<0.10 in the univariate analysis were considered for inclusion in the conceptual hierarchical model. Each variable or group of variables in a level was introduced simultaneously into the multiple regression model and the coefficients inspected. After each level was introduced, any of the new variables with p-values of > 0.10 were excluded in a manual backwards stepwise manner starting with the variable with the largest p-value. After each variable was excluded the significance of the remaining variables in the same level was examined. Once the removal of the variables was completed, since all remaining new variables in the level were p<0.10, the next level of variables was added to the regression model. Once a variable had been accepted into the model it was not subsequently removed, even if the inclusion of variables from lower levels in the conceptual hierarchical model resulted in it being no longer statistically significant. The first regression model considers only the relationship between SES and pregnancy intention. This analysis shows the overall effect of the SES alone and not (improperly) controlled for the proximate factors that are partly determined by SES. Level Two of the conceptual hierarchical model contains socio-demographic variables, such as education level and marital status, which are added to Model One. Level Two variables with a p-value of >0.10 were excluded using manual backwards stepwise regression to create Model Two. The coefficients for the remaining socio-demographic variables tell us their effect having (properly) controlled for SES. The new coefficients for the Level One variable of SES in Model Two give the estimate of its effect that is not mediated through the Level Two socio-demographic variables (pathway e). The variables of Level Three in the conceptual hierarchical model were then added to the Model Two. Level Three contains previous depression and experience of IPV which are influenced by the factors in the Levels above and which can affect pregnancy intention either through Levels Four and Five (pathways j and k) or directly (pathway l). Level Three variables with p-values of >0.10 were excluded using manual backwards stepwise regression to create Model Three. The coefficients for the previous depression and remaining IPV variables in Model Three tell us their effect on pregnancy intention adjusted for the confounding roles of the socio-economic and socio-demographic variables in Levels One and Two. The new coefficient for the Level One variable, SES, gives an estimate of its effect that is not mediated through socio-demographic factors, previous experience of depression or IPV (pathway e) and the new coefficients for the socio-demographic factors are estimates of their effects that are not mediated through previous experience of depression or IPV (pathway i). Next the maternal reproductive characteristics of Level Four of the conceptual hierarchical model were simultaneously added to Model Three. Level Four contains factors such as the number of live children a woman has and the time since the last birth that are affected by the determinants distal to it in Levels One to Three. Maternal reproductive characteristics may influence pregnancy intention through the final level of the model, Level Five, (pathway m) or directly (pathway n). Level Four variables with p-values of >0.10 were excluded using manual backwards stepwise regression to create Model Four. The coefficients for the remaining maternal reproductive characteristics tell us the effect of each factor on pregnancy intention adjusted for the confounding effects of the variables in Levels One to Three. The new coefficients for the variables in Levels One to Three are estimates of their effect on pregnancy intention that are not mediated through the variables at the lower levels of the hierarchy (pathways e, i and l). Finally, the Level Five variable of the conceptual hierarchical model, gestation, was added to the Model Four. Gestation is a marker of the time since conception, that is the time that we are primarily interested in, and was included to account for any possible differences in reported level of pregnancy intention that are due to the timing of the assessment. Gestation influences pregnancy intention through pathway o. The complete model now tells us: the residual effect of socio-economic status on pregnancy intention that is not mediated through socio-demographic factors, previous depression, IPV, maternal reproductive characteristics or gestation (pathway e); the residual effect of socio-demographic variables on pregnancy intention that is not mediated through previous depression, IPV, maternal reproductive characteristics or gestation (pathway i); the residual effect of previous depression and IPV on pregnancy intention that is not mediated through maternal reproductive characteristics or gestation (pathway l); the residual effect of the maternal reproductive characteristics that is not mediated through gestation (pathway n) and the unconfounded effect of gestation on pregnancy intention (pathway o). The residual effects may be either direct effects or effects that are mediated through other determinants that are not included in the model. The University College London Research Ethics Committee and the College of Medicine Research Ethics Committee at the University of Malawi granted ethical approval for this research (approval numbers 3974/001 and P.03/12/1273 respectively). Ethical approval was given to include pregnant women aged 15 and over. Field workers were trained to assess competency for consent in those aged below 18; these women gave their own written consent, no proxy was used. All women gave written informed consent to participate, by thumbprint if necessary, after they had read the information sheet and/or had the study explained to them. The participants retained the information sheet and one copy of the signed consent form; a second copy of the signed consent form was stored in a lockable cabinet in the main study office. Both ethics committees approved this consent procedure. Local approval to conduct the research in Mchinji District was given by the District Health Officer and the District Executive Committee.