Background

Some traces of fertility behavior differences in the population can be found early in the literature (Bongaarts and Potter Reference Bongaarts and Potter1983; Knodel Reference Knodel1987), and population and mortality heterogeneity can be found even earlier (Keyfitz and Littman Reference Keyfitz and Littman1979; Land and Rogers Reference Land, Rogers, Land and Rogers1982), however less attention has been paid to assessing the variation and heterogeneity of fertility courses. Indeed, there are several notable examples of the shortcomings of the demographic transition framework wherein the non-homogeneous nature of fertility decline is rarely discussed (Coale Reference Coale, Behrman, Corsa and Freedman1969; Cleland and Wilson Reference Cleland and Wilson1987; Kohler et al. Reference Kohler, Billari and Ortega2002; Kreager and Bochow Reference Kreager, Bochow, Kreager and Bochow2017). With second demographic transition (SDT) literature, attention has been drawn for the first time to the diversification of fertility outcomes. Lesthaeghe (Reference Lesthaeghe2010) points out that the second demographic transition (SDT) results in non-stationary populations with a multitude of living arrangements, sometimes characterized by a “convergence to diversity”, and further claims that fertility cannot be studied without a framework that reflects changing lifestyle preferences. Pesando (Reference Pesando2019) refers to this heterogeneity as “persistent diversity with development”. Indeed, it is possible to see a later upsurge in the literature on behavioral diversities among individuals. The rise of the life-course approach (Huinink and Kohli Reference Huinink and Kohli2014), the literature on social interaction effects on fertility (Rossier and Bernardi Reference Rossier and Bernardi2009) and an emphasis on the decision-making process of individuals (Hakim Reference Hakim2003) show how fertility can be examined from a more integrated perspective. Pesando (Reference Pesando2019) examines the persistent diversity of global family change, noting that divergent demographic trajectories of fertility have begun to characterize high-income societies. All these recent works focus on the fact that fertility has no monolithic and homogeneous structure, especially in populations where changes in fertility size and structure continue. In light of these studies, it is necessary to understand constituent and more homogeneous fertility trajectories and their transition, in order to correctly understand structural change in fertility.

Contrary to what was experienced at the beginning of the first demographic transition, variations in fertility structure are based on fragmented and more fluid behavioral changes. As suggested by the SDT, cultural shifts and changes trigger individualization in demographic behavior (Lesthaeghe Reference Lesthaeghe2010). Higher education is linked to a higher likelihood of accepting new family values and predicts new ideas and behaviors that originate among young, highly educated, and less traditional people in urban settings (Vitali et al. Reference Vitali, Aassve and Lappegård2015). Accordingly, interpersonal connections and networks come to the fore in examining these changes.

In Turkey, changes in fertility have taken place alongside rapid societal transformations. The total fertility rate in Turkey declined steadily in the second half of the 20th century, from a total fertility rate of 5 children in the 1970s to around replacement level in the 2010s (HUIPS 2019). In parallel with the rest of the world, neoliberal economic policies have been adopted and Gross Domestic Product (GDP) per capita has more than doubled in 50 years (World Bank 2021a). Following the First Five-Year Development Plan in 1963 (State Planning Organization (SPO) 1963), anti-natalist policies such as the authorization of family-planning methods and the easing of laws banning abortion were adapted under the new population law in 1965 (Population Planning Law 1965). In 1983, with the legalization of abortion, a new population law was accepted and the family planning-oriented approach was continued (Population Planning Law 1983). Despite these regulations, the population of Turkey, which was 40 million in 1975, doubled to 80 million in 2017. Accordingly, a similar increase was also seen in the analyzed female cohorts. While there were approximately 2.5 million women aged 40-49 in 1990 (TurkStat 2010), the number of women aged 40-49 is over 5.5 million according to the 2018 Address Based Population Registration System results (TurkStat 2019).

The most significant societal transformations in Turkey during the fertility decline period are urbanization and the increase in the education level of women. While the population living in the urban areas was around 40% in the 1970s, changing economic structures required more workers for the growing urban settlements. Beginning in the 1980s, the influx of migrants from rural to urban settlements has continued unabated and as a result, the urban population in Turkey went from 65% in 2000 to over 75% today (World Bank 2021c). In Turkey, where urban fertility levels are always lower than rural areas (HUIPS 2019), the findings of Kavas and Thornton (2019) confirmed that most of the urban population acknowledges the relationship between development and low fertility. Parallel to urbanization, the education level of women increased significantly in the second half of the 20^th century. The percentage of ever-married women aged 15-49 with high school or higher education increased from 14 in 1998 to 32 in 2018 (HUIPS 1999; HUIPS 2019), and the literacy rate of women 15 years of age and over has increased from 45% in 1975 to 93% in 2017 (World Bank 2021b). In addition to the temporal contiguity of social changes, there have also been changes more directly related to fertility. Although the median age at first marriage increased from 19.5 in 1998 to 21.4 in 2018 for women in the 25-49 age group, the relatively early marriages and universality of marriage behavior among women means the pattern of having the first birth shortly after marriage in Turkey is retained, and the median age at first birth is calculated as 23.3 years (HUIPS 2019). In accordance with the age at first birth pattern, the currently childbearing group peaks with the 25-29-year-olds in Turkey, although that peak used to belong to the 20-24 age group. The change in the fertility of women aged 40-49, who have mostly passed their reproductive ages, has been in parallel to the overall decline of fertility. The mean number of children ever born to women aged 40-49 decreased from 4.3 children for the 1949-58 cohort to 2.7 children for 1969-78 cohort. (HUIPS 1999; HUIPS 2019). On the other hand, contraceptive prevalence has increased from 63% to 70% in 1993 to 2018 period (HUIPS 2019).

Fertility in Turkey has been largely studied as an extension of global fertility research trends and is mainly focused on language groups and regional differences that are indicators of ethnicity in that country. The study by Koç, Hancıoğlu and Çavlin (Reference Koç, Hancioglu and Çavlin2008) shows the demographic differentials and integrational aspects of Turkish and Kurdish populations in Turkey. Their results indicate that strong demographic differentials exist between Turkish and Kurdish populations and the convergence of the two groups is not yet apparent. Yavuz (Reference Yavuz2006) also investigated the fertility decline in Turkey according to main language groups. His findings suggest that parity progression intensities of Turkish speaking mothers are lower than Kurdish speaking mothers, which implies that the fertility decline started much later for the latter group. Gore and Carlson (Reference Gore and Carlson2010) stated in their study that besides ethnicity, education also influences marriage patterns and therefore fertility patterns. The results of their study showed that although low-educated Kurdish women married earlier than Turkish women, the difference was reversed among educated women. In addition, the study of Greulich et al. (Reference Greulich, Dasre and Inan2016) concludes that differences in female education are the driving force behind the regional heterogeneity of fertility in Turkey. As well, regional differentiation is apparent in other research in Turkey (Yüceşahin and Özgür Reference Yüceşahin and Özgür2008; Caarls and de Valk Reference Caarls and de Valk2018). Although studies on Turkey’s fertility refer to various rates and phases of the fertility transition in spatially distant population groups (Duben and Behar Reference Duben and Behar2002), differences in fertility are usually related to predefined observed variables. These approaches only evaluated fertility by linking it to diversity in the demographic structure and did not go beyond this; however, the increase in the mean age at first birth in postponed marriages, together with the slowdown in fertility decline, gives clues that the change in Turkey may not be uniform.

Data and Methods

The data source of this study is the Turkey Demographic and Health Surveys (TDHS), which is part of the global DHS series. Turkey Demographic and Health Surveys are household-based nationally representative sample surveys designed to provide information on fertility, infant and child mortality, family planning, and maternal and child health. The surveys are carried out by Hacettepe University Institute of Population Studies (HUIPS). This study is based on three quinquennial TDHS datasets; 1998, 2008 and 2018, which contain the complete birth histories of women aged 15-49. For the purposes of the research, we limited the focus of the study to women aged 40-49, as near-complete fertility histories were needed to properly analyze women’s fertility trajectories. Furthermore, since the 2008 study were conducted on ever-married women, all analyses were carried on ever-married women (Table 1). The vast majority of births in Turkey take place within marriage, so the exclusion of never-married women is negligible when analyzing fertility trajectories. However, since women who have never been married can be assumed as childless, overall childless women may have been underestimated. The proportion of women aged 40-49 who have never been married is 2, 1 and 4 percent for the 1949-58, 1959-68 and 1969-78 cohorts respectively. All ever-married women (currently married, divorced and widowed) were analyzed for their fertility structure with the sequence analysis. However, we excluded remarried women (5, 4 and 5 percent for the 1949-58, 1959-68 and 1969-78 ever-married cohorts respectively) from distance analysis and multinomial regression since there was more than one group of variables related to the husbands and marriage characteristics.

Table 1. Number of unweighted observations of ever-married women aged 40-49

The analysis in this study was carried out in three steps. For the first part, we used the sequence analysis approach originally proposed by Abbott (1995) for ever-married women 40-49 separately in each dataset. The sequence analysis method can be used to describe the quantum and tempo of interrelated events and their sequencing (their order of happening) (Di Giulio et al. Reference Di Giulio, Impicciatore and Sironi2019). This strategy emphasizes the holistic nature of trajectories, and rather than handling them as a point in time, treats every observation as a life-course trajectory. By focusing on the analysis of entire trajectories rather than single events, sequence analysis considers the interrelation between multiple events (Barban and Sironi Reference Barban, Sironi and Schoen2019). After separating the childless 40-49 women from the data, childbearing trajectories of women are constructed as sequences of 29 states from ages 12 to 40, where each age between 12 and 40 represents a parity-related state. Since the children ever born to women aged 40-49 in the cohort with the highest fertility was 4.8 (HUIPS 2019), the alphabet constructed for sequence analysis contained six mutually exclusive states according to children ever born, namely; “no birth”, “one birth”, “two births”, “three births”, “four births” and “five or more births”. Then, women in each age can be represented with a state according to their number of children ever born in the corresponding age. After that, we employed the TraMineR package available in R to construct sequences in each dataset and used the optimal matching (OM) method (with insertion/deletion cost as 1, and the transition rates between states observed in the sequence data as substitution costs) to calculate distances between sequences and form the dissimilarity matrix for each dataset. We used the results of hierarchical clustering (calculated with Ward algorithm) as initial medoids (fertility trajectories in each cluster whose sum of dissimilarities to all the trajectories in the cluster is minimal) in a PAM (partitioning around medoid) algorithm. Since PAM algorithm assign each fertility trajectory to the closest medoid, in each cluster, women are closer to their cluster medoid than the medoids of the remaining clusters. Based on the clustering analysis and the weighted average silhouette width (ASWw) of the clusters, the optimal number of clusters were selected for each dataset. The ASWw value is a measure of coherence of assignments. High coherence indicates high between-group distances and strong within-group homogeneity. For this reason, the fact that a woman is in a particular cluster according to her fertility trajectory indicates that her fertility is more similar to the fertility behavior of this cluster than the others. Two women falling into the same cluster regardless of their last parity results from having similar fertility behaviors for ages 12-40 in terms of tempo and quantum. In other words, clustering by fertility trajectories allows us to consider the amount of the reproductive period spent at each parity. Although women in a certain final parity come to the fore in the resulting clusters, women in different final parities can coexist in the same clusters. Therefore, we interpreted clusters considering similar fertility behaviors rather than same final parities. While it seems to contradict orthodox categorization practices, this illustrates the importance of considering the timing of fertility. The clustering analysis resulted in five clusters in each data set and after combining the previously separated childless women, there were six clusters in total.

In the second part, we performed distance analysis in order to interpret the fertility behaviors in the light of the differences regarding the basic characteristics before the fertility period. We measured the similarity of women in each cluster to calculate the heterogeneity of background characteristics on three dimensions; background characteristics of women, their husbands and their marriages. In order to measure dissimilarities, we calculated the heterogeneity scores $\phi \left( P\right)$ using Hamming distances between observations, for six clusters of three surveys. On the axis of urbanization and increase in educational levels, the background characteristics of women and men consist of the place of residence where they spent their childhood, their educational status and their mother tongue (Table 2). In addition to these, since it contains cultural codes, the marriage characteristics dimension was created from the variables of age at first marriage of women, kinship with her husband, arrangement of marriage and marriage ceremony. The background characteristics used in this study to understand heterogeneity are the features that women and men acquire mainly in the pre-fertility period. Therefore, background characteristics are not only related to the heterogeneity of fertility trajectories, but also constitute the foundation of this heterogeneity. Even though other aspects such as religiosity contain invaluable insights to the fertility, the available data are insufficient to provide these variables to make retrospective comparable analyses. We used the selected variables with binary categories to give equal weights to each variable in a dimension and so as to preserve the difference in the categories of the variable for all 3 cohorts in distance analysis.

Table 2. Variables for distance analysis and multinomial logistic regression

The Hamming distance ${d_H}\left( {{x_1},{x_2}} \right)$ is defined as the number of variables at which the two observations ${x_1}$ and ${x_2}\;$ are different. The variables are recoded with binary categories and the Hamming distance between two observation according to these variables is hypothesized as the theoretical distance between women in that dimension. Since the Hamming distance between two observations can be measured, it is also possible to calculate the pairwise distance of a group of observations. The pairwise Hamming distance H between n observation would then be,

$$H = \mathop \sum \limits_{k = 1}^{k = n - 1} \mathop \sum \limits_{k' = k + 1}^{k' = n} {d_H}\left( {{x_k},{x_{k'}}} \right)$$

The sum of all possible pairwise distances gives the pairwise distance of observations. When the pairwise distance is divided to number of distances, the average pairwise distance in a group of observation is calculated. The average pairwise distance is,

$${H_{avg}} = {{2*H} \over {n*\left( {n - 1} \right)}}$$

In order to calculate average Hamming distances, the algorithm introduced by Morrison (Reference Morrison2004) were used. His algorithm first calculates the centroid (moment of inertia) of the observations. The $i$ -th coordinate of the centroid of equally weighted points is,

$${c_i} = {{\mathop \sum \nolimits_{j = 1}^{j = n} {x_{ij}}} \over n}$$

where ${x_{ij}}$ are the values of observations (in this case ${x_{25}}$ indicates the 5^th observation in second variable). Then, total pairwise Hamming distance becomes the sum of the moments of inertia about their centroid (Morrison, Reference Morrison2004),

$$H = n*\mathop \sum \limits_{i = 1}^{i = m} \mathop \sum \limits_{j = 1}^{j = n} {\left( {{x_{ij}} - {c_i}} \right)^2}$$

where $m$ is the number of variables and $n$ is the number of observations. However, this calculation of the distance does not take case weights into consideration. When the case weights are introduced to the above equations, the weighted total pairwise Hamming distance becomes,

$${H^w} = \mathop \sum \limits_{j = 1}^n {w_j}*\mathop \sum \limits_{i = 1}^{i = m} \mathop \sum \limits_{j = 1}^{j = n} {w_j}*{\left( {{x_{ij}} - {c_i}} \right)^2}$$

where ${w_j}$ is the case weight of observation $j$ . Since the weights are introduced, the weighted total pairwise Hamming distance can be divided by the sum of pairwise products of weights to calculate the average of the total distance. The average weighted pairwise Hamming distance becomes,

$$H_{avg}^w = {{{H^w}} \over {\mathop \sum \nolimits_{i = 1}^{i = n - 1} \mathop \sum \nolimits_{j = i + 1}^{j = n} \left( {{w_i}*{w_j}} \right)}}$$

$H_{avg}^w$ takes values between $0$ (minimum heterogeneity) and ${m \over 2}$ (maximum heterogeneity) where m is the number of variables and n is sufficiently large. For example, for the TDHS, if all observations have the same value at every background characteristic of women, which means the population is extremely homogenous, the indicator will take a value of 0. On the other hand, for a sufficiently large n, if the observations are distributed evenly to all possible categories, which means the population is at maximum heterogeneity, the indicator will take a value of 3/2 (since m=3 for the background of women). In order to normalize the indicator and generate heterogeneity scores $\phi $ ,

$$\phi \left( P\right) = \left( {{{H_{avg}^w\left( P\right)} \over {{H_{max}}\left( P\right)}}} \right)*100$$

is used where ${H_{max}}$ is the maximum heterogeneous distribution of the population $P$ . For these scores, higher values indicate a more heterogeneous distribution of women. The relative heterogeneity scores were calculated as ${\rm{\Delta }}{\phi _c} = \;{\phi _c}\left( P\right) - {\rm{\;}}{\phi _t}\left( P\right)$ , the difference between the heterogeneity score of a cluster and the score of the whole cohort. Relative heterogeneity scores are used to understand the similarities of background characteristics of the clusters with the overall cohort. Positive values represent more heterogeneous nature of the cluster related to the cohort overall and negative values show less heterogeneity.

In the last part, we further analyzed the fertility trajectories of cohorts using multinomial logistic regression using the same variables as the distance analysis but using more detailed categories (Table 2). The predicted cluster membership probabilities of “ideal types”, i.e. an average educated/uneducated woman where other independent variables were kept in their group means, are calculated with the help of two variables: women’s education and childhood place of residence. The probability distribution of educated and uneducated average women in clusters and the change of these probabilities through cohorts not only help to understand past experiences, but also contain clues about future changes. Distance measures are calculated with R and regression analyses and marginal effects were calculated with SPost13 package in Stata (Long and Freese Reference Long and Freese2014).