Data

Data were taken from the 2003 and 2008 Turkish Demographic and Health Survey (TDHS 2003 and 2008). The TDHS is a national representative cross-sectional survey, based on standardized DHS survey questionnaires that have been applied worldwide, that has been repeated every five years since 1968. The survey has two questionnaires: (i) a household questionnaire and (ii) an ever-married women questionnaire. The household questionnaire lists all usual members of selected households (Rutstein & Rojas, Reference Rutstein and Rojas2006), and gathers information on level of wealth, age, gender, education, employment status, marital status, region and neighbourhood for each person. The prime objective of the household questionnaire is to detect the women eligible for individual interview in the ever-married women questionnaire (Rutstein & Rojas, Reference Rutstein and Rojas2006).

The ever-married women questionnaire seeks to obtain information about the respondent’s background characteristics (and those of her husband), including migration history, current marriage and marriage history, pregnancy, birth history and fertility preferences, knowledge and use of contraceptive methods, antenatal and postnatal care, breast-feeding, nutrition, women’s current employment status and employment history and anthropometric measurement of women and children under five (TDHS, 2008). The focus of the survey is women of reproductive age (15–49) (Rutstein & Rojas, Reference Rutstein and Rojas2006), so the survey fieldwork is structured to identify eligible women respondents. Thus the present study was based on a ‘base file’ created by merging the datasets from the two questionnaires. As a result the characteristics of a respondent’s household were assigned to every women respondent in the ever-married women questionnaire.

The 2003 survey included 10,836 households (response rate: 92.9%) and 8075 individual interviews with ever-married women within these households (TDHS, 2003). The 2008 survey included 10,525 households (response rate: 88.4%) and 7405 individual interviews with ever-married women within selected households (TDHS, 2008).

Variables

The summary statistics of the study sample are shown in Table 1. Although dummy variables were created to illustrate age groups, continuous age variables fitted better in the models. Accordingly, respondent’s age and a squared function of this were employed, referring to the non-linear relationship between age and the outcome of interest. In addition, respondent’s age at first marriage was included as previous research has shown this to affect fertility (Kirdar et al., Reference Kirdar, Dayioglu-Tayfur and Koc2011; Güneş, Reference Güneş2013).

Table 1 Summary statistics (mean levels of each variable) of socioeconomic variables, ever-married women aged 15–49 (N=15,480), 2003 and 2008 TDHS

Four different dummy variables were generated for education level, varying from no education to higher education. The durations of educational stages (primary, secondary and higher education) were modified in 1997 and 2012. These were coded according to the regulations prior to the 1997 modification since most of the respondents were educated according to these regulations. Therefore, the ‘no education’ variable indicates women with no education at all; ‘primary education’ indicates up to the first five years of education; secondary education refers to the next six years of education after primary education (reference category); and higher education denotes twelve years or more in education.

Dummy variables for ethnic background were used: Turkish (reference category), Kurdish, Arabic and other ethnic minorities. All observations were for ever-married women, so three dummy variables (married, divorced (reference category) and widowed) were generated to indicate current marital status.

There are seven official regions in Turkey, and observations were classified within five different regional groups according to the TDHS (2003, 2008) classification: the five region dummies were north, south, west, east (reference category) and central.

It was intended that income would be used in the analyses but there were insufficient observations in the 2003 survey, so a variable related to family wealth was used instead, with a wealth score being derived from ownership of assets (such as a car or computer) and housing characteristics (such as building material, source of water). Observations were classified into five dummy wealth groups ranging from the poorest to the richest (the middle-wealth group was the reference category), with a higher wealth score indicating a more affluent household. It is important to note that the wealth scores were calculated for households, and therefore each member of a family has the same score. This may raise allocation issues within a household; however, scarcity of data regarding allocation did not allow this to be tested. The determination of the wealth score and the classification were made by TDHS (2003, 2008).

At the time of the surveys, five different public social security schemes were operating in Turkey. The Social Insurance Organization (SIO) covered blue collar workers and private sector employees. The Social Insurance Agency of Merchants, Artisans and the Self-employed (BAG-KUR) covered self-employed people and their dependants, unemployed people (those who could pay their premiums), housewives and agricultural workers. The Government Employees’ Retirement Fund (GERF) covered retired civil servants and their dependants, retired military personnel and retired parliamentarians. The Active Civil Servants Scheme covered civil servants, military personnel and parliamentarians who were currently working. In the data, the Active Civil Servants Scheme was combined with the Government Employees’ Retirement Fund by the nature of the survey. Finally, the Green Card covered all those who could not afford health care services, with the condition being that the applicant must: (i) be a citizen of the Republic of Turkey, (ii) be uncovered by any of the other insurance schemes; (iii) earn less than one-third of the minimum wage (after tax) per month. It is important to keep in mind that the GERF is the most privileged social security group, while Green Card is the least beneficial (Liu et al., Reference Liu, Celik and Sahin2005; Buğra & Keyder, Reference Buğra and Keyder2006; Tatar et al., Reference Tatar, Ozgen, Sahin, Belli and Berman2007; Yardim et al., Reference Yardim and Cilingiroglu2010; Erus & Aktakke, Reference Erus and Aktakke2012). Additionally, BAG-KUR and SIO are more beneficial than Green Card (when the coverage and quality of care are considered) even though their benefits are limited (compared with those provided by GERF) (Buğra & Keyder, Reference Buğra and Keyder2006; Erus & Aktakke, Reference Erus and Aktakke2012). Individuals covered by the Government Employees’ Retirement Fund is the reference category, and dummy variables were generated for each of the insurance schemes. In addition, two dummy variables were created to identify individuals with private insurance and those with no insurance at all.

Abortion was measured by two variables: (i) an unbounded variable, i.e. the number of induced abortions, and (ii) a bounded variable, i.e. whether or not have had an induced abortion. The question related to the bounded abortion variable was ‘Have you ever had a pregnancy that ended in an induced abortion?’, while the continuous abortion variable related to ‘In all, how many induced abortion have you had?’.

Concentration indices

Concentration indices were calculated to measure income-related inequalities in abortion. The concentration index is an indicator of health inequality in relation to the socioeconomic position of an individual (Erreygers, Reference Erreygers2009a, Reference Erreygersb). It was introduced by Kakwani (Reference Kakwani and Bank1980) and Wagstaff et al. (Reference Wagstaff, Paci and van Doorslaer1991). The value of the concentration index is twice the area between concentration curve and the diagonal (Wagstaff et al., Reference Wagstaff, Paci and van Doorslaer1991); the concentration curve is obtained by plotting the cumulative proportions of the population, ranked by socioeconomic status beginning with the most advantaged (well-off), against the cumulative proportions of the health variable (Wagstaff et al., Reference Wagstaff, Paci and van Doorslaer1991). The concentration index takes values of between ${{1{\minus}n} \over n}$ and ${{n{\minus}1} \over n}$ (where n is the number of observations) and ${{1{\minus}n} \over n}$ and ${{n{\minus}1} \over n}$ approach −1 and 1, respectively, when the number of observations increases. The concentration index takes positive values when health inequalities favour the well-off (e.g. the variable of interest is more concentrated among the well-off) and vice versa (Wagstaff & van Doorslaer, Reference van Doorslaer and Wagstaff1997). If the value of the concentration index is 0, this means the health variable is equally distributed (Wagstaff et al., Reference Wagstaff, Paci and van Doorslaer1991; Wagstaff & van Doorslaer, Reference van Doorslaer and Wagstaff1997). The concentration index can be calculated using the formula:

(1) $$C\left( h \right)\,\equiv\,1{\minus}{{\sum_{i=1}^n \left( {2\lambda _{i} {\minus}1} \right)h_{i} } \over {n^{2} \mu _{h} }}$$

where C(h) is the concentration index of the health variable, λ is the rank of the individual, h _i is the health of individual i, n is the number of observations, and $\mu _{{\rm h}} $ is the mean of the health variable. As mentioned before, the value of the concentration index is bounded to −1 and 1 (Wagstaff et al., Reference Wagstaff, Paci and van Doorslaer1991). However, when the variable of interest is binary, the mean of the binary variable further bounds the value of the concentration index (Wagstaff, Reference Wagstaff2005). In this case the concentration index can take values between $\mu {\minus}1{\plus}\left( {1/n} \right)$ and $1{\minus}\mu {\plus}\left( {1/n} \right)$ instead of −1 and 1 (ibid.). As the mean of the variable of interest increases, the range of the values that the concentration index can take gets smaller. To overcome this, Wagstaff (Reference Wagstaff2005) suggested normalizing the concentration index by dividing it by its upper bound ( $1{\minus}\mu {\plus}\left( {1/n} \right)$ ):

(2) $$W\left( h \right)={{2\left( {b_{h} {\minus}a_{h} } \right)} \over {n^{2} \left( {b_{h} {\minus}\mu _{h} } \right)\left( {\mu _{h} {\minus}a_{h} } \right)}}\mathop{\sum}\nolimits_{i=1}^n {z_{i} h_{i} =C(h)/\left( {1{\minus}\mu } \right)} $$

where b _h and a _h are the upper and lower bounds of the variable of interest, $\mu _{h} $ is the mean of the variable of interest, n is the number of observations, $\lambda _{i} $ is the rank of individual i, z _i equals ${{n{\plus}1} \over 2}{\minus}\lambda _{i} $ for individual i and the expression (1/n) approaches zero when the number of observations increases.

This normalization (Wagstaff, Reference Wagstaff2005) was specific to the case of the binary variable of interest (Wagstaff, Reference Wagstaff2009). However, Erreygers (Reference Erreygers2009a) generalized the normalization and introduced a corrected concentration index that facilitated the remedy of the bounds issue (Erreygers, Reference Erreygers2009a, Reference Erreygersb). Accordingly, the corrected concentration index can be written as:

(3) $$E\left( h \right)={8 \over {n^{2} (b_{h} {\minus}a_{h} )}}\mathop{\sum}\nolimits_{i=1}^n {z_{i} h_{i} } $$

where:

(4) $$z_{i} ={{n{\plus}1} \over 2}{\minus}\lambda _{i} $$

where b _h and a _h are the upper and lower bounds of the variable of interest, $\mu _{h} $ is the mean of the variable of interest, n is the number of observations and $\lambda _{i} $ is the rank of individual i.

Therefore the concentration index (Wagstaff et al., Reference Wagstaff, Paci and van Doorslaer1991) was employed for the unbounded outcome variable. Due to the existing bounds issue of the concentration index (Wagstaff, Reference Wagstaff2005), Wagstaff’s normalization (Reference Wagstaff2005) and Erreygers’s correction (Reference Erreygers2009a) were preferred for that bounded variable of interest.

Concentration indices (Wagstaff et al., Reference Wagstaff, Paci and van Doorslaer1991; Wagstaff, Reference Wagstaff2005; Erreygers, Reference Erreygers2009a) can be used to measure levels of inequality; however they are unable to highlight the individual factors leading to the inequalities (Wagstaff et al., Reference Wagstaff, van Doorslaer and Watanabe2003; O’Donnell et al., Reference O’Donell and van Doorslaer2006). Instead, decomposition methods provide the opportunity of unpacking the individual contributions of socioeconomic factors to the outcome of interest (Wagstaff et al., Reference Wagstaff, van Doorslaer, Paci and Joshi2001, Reference Wagstaff, van Doorslaer and Watanabe2003; Wildman, Reference Wildman2003; O’Donnell & van Doorslaer, Reference O’Donell and van Doorslaer2006). Therefore, in this study, the decomposition of the income-related inequalities for 2003 and 2008 was performed using the methods of Wagstaff et al. (Reference Wagstaff, van Doorslaer and Watanabe2003) in order to understand the factors leading inequalities in abortion. Accordingly, for a linear regression model of selected outcome such as:

(5) $$y=a{\plus}\mathop{\sum}\limits_k {\beta _{k} x_{k} {\plus}{\varepsilon}} $$

the concentration index for y can be written as:

(6) $$C=\mathop{\sum}\limits_k {\left( {{{\beta _{k} \bar{x}_{k} } \over \mu }} \right){\rm }C_{k} {\plus}{{GC_{{\varepsilon}} } \over \mu }} $$

where $\mu $ is the mean of the variable of interest (y), x _k variables are the explanatory variables (k regressors), $\bar{x}_{k} $ is the mean of x _k , C _k is the concentration index for x _k , ${\varepsilon}$ is the error term and $GC_{{\varepsilon}} $ is the generalized concentration index for the error term. Accordingly, observed inequalities in the outcome of interest (y) can be decomposed into explained and unexplained parts (Wagstaff et al., Reference Wagstaff, van Doorslaer, Paci and Joshi2001). The explained part is equal to the weighted sum of the concentration indices of explanatory variables (captured by the first term in eqn (6)), where the weight for each variable (x _k ) is the elasticity of outcome (y) with respect to relevant variable (x _k ) (Wagstaff et al., Reference Wagstaff, van Doorslaer and Watanabe2003; van Doorslaer & Koolman, Reference van Doorslaer and Koolman2004; O’Donnell et al., Reference O’Donell and van Doorslaer2006). The unexplained part (captured by the last term in eqn (6)), which should approach zero for a well-specified model, is the inequality in the outcome of interest that is not explained by the variations in its determinants (ibid.). Therefore, each contribution is produced by (i) the sensitivity of outcome with respect to the related factor ( $\beta _{k} (\bar{x}_{k} /\mu )$ ) and (ii) the degree of income-related inequality in the related factor (C _k ) (Wagstaff et al., Reference Wagstaff, van Doorslaer and Watanabe2003; O’Donnell et al., Reference O’Donell and van Doorslaer2006). Hence the larger elasticity ( $\beta _{k} (\bar{x}_{k} /\mu )$ ) and more unequally distributed x _k (across the income range; i.e. larger C _k ) the greater its importance in explaining the inequalities in the outcome of interest (Wagstaff et al., Reference Wagstaff, van Doorslaer, Paci and Joshi2001).

The changes in income-related inequalities over time were then decomposed using an approach proposed by Wagstaff et al. (Reference Wagstaff, van Doorslaer and Watanabe2003). It was derived by applying Oaxaca-type decomposition (Oaxaca, Reference Oaxaca1973) to the aforementioned technique, which decomposes the concentration index (Wagstaff et al., Reference Wagstaff, van Doorslaer and Watanabe2003). The changes in the inequalities over time can be decomposed as:

(7) $$\Delta C={\rm }\mathop{\sum}\limits_k {\eta _{k} \left( {C_{{kt}} {\minus}C_{{kt{\minus}1}} } \right)} {\plus}\mathop{\sum}\limits_k {C_{{kt{\minus}1}} \left( {\eta _{{kt}} {\minus}\eta _{{kt{\minus}1}} } \right){\plus}\left({{GC_{{{\varepsilon}t}} } \over {\mu _{t} }}\right)} $$

where t is time (t=2008, t−1=2003 for this study) and $\Delta $ indicates the first difference, $\eta _{{kt}} $ is the elasticity of y with respect to k regressor at time t, C _kt is the concentration index of the k regressor at time t and $GC_{{\varepsilon}} $ is the generalized concentration index for the error term. Accordingly, the changes in the inequalities over time were separated out: (i) the changes in inequality in the determinants of the outcome interest, and (ii) the changes in health effects of the determinants over time (Wagstaff et al., Reference Wagstaff, van Doorslaer and Watanabe2003; O’Donnell et al., Reference O’Donell and van Doorslaer2006). Therefore, the first part of the formula indicates the changes in concentration indices due to the differences in inequalities in the determinants of abortion. The second bit indicates the changes in the effects of abortion determinants and the last part captures the unexplained part of the changes. Hence, this allows one to understand the extent to which changes in income-related inequalities in abortion are due to changes in inequality in its determinants, rather than to changes in its elasticity and vice versa (Wagstaff et al., Reference Wagstaff, van Doorslaer and Watanabe2003).