Estimation of the adolescent pregnancy rate in Thailand 2008–2013: an application of capture-recapture method

Adolescent pregnancy is an important health and social issue which affects both individual and social well-being. Pregnancy related conditions are the leading causes of death among young women [1, 2], and also increase risks of preterm delivery, low birth weight and a number of maternal and neonatal complications [3‐5]. A large proportion of pregnancies in young women are unintended and pose a risk of unsafe abortions [6]. Furthermore, adolescent pregnancy also increases socio-economic problems in society such as having poor achievement in education, being a single mother, unemployed and living in poverty [7].

Adolescent pregnancy is a global concern, so it was included in the global health agenda for the Millennium Development Goals (MDGs) in the years 2000 to 2015 and into the Sustainable Development Goals (SDGs) in the years 2016 to 2030. Adolescent birth rate, the number of births per 1000 women aged 15–19 years, was the MDGs indicator under Goal 5B, which was aimed to improve maternal health within 2015 [8]. Reducing adolescent birth rate is currently SDGs indicator number 3.7.2, which is used to improve sexual and reproductive health and the social and economic well-being in adolescents [9]. Unfortunately, adolescent birth rate does not represent the total number of adolescent pregnancies if registry data of abortions and stillbirths are incomplete. Estimation of the total number of adolescent pregnancies is reliable only in the countries with complete data on abortion [10‐12]. In countries with restrictive abortion law, underreporting is mainly due to the missing data from induced abortions. Some approaches and indirect estimations have been developed [13‐15] to estimate the abortion rates in countries with incomplete abortion statistics, although the most appropriate estimation method is still inconclusive.

Estimation of prevalence or incidence of event or disease condition such as adolescent pregnancy, using complete enumeration of all relevant cases is costly and thus rarely possible, particularly where data registry is not well developed. The indirect estimation method by combining multiple sources of information and deleting duplicated cases always has some degree of undercounting, and thus some adjustment is needed. The capture-recapture (CRC) method has been widely used to estimate population size especially in “hard to reach” populations with incomplete registered data [16‐19]. This method can take into account the undercounting of disease/condition using the recapture information, i.e., intersection or overlapping sources, to estimate the number of missing cases under proper assumptions. Although Thailand has well established birth registration, this database includes only live births whereas data for abortion, stillbirth, and miscarriage are not included. A more accurate estimate of adolescent pregnancy rate should lead to better situation analysis and strategic planning for policy makers. We therefore applied the CRC technique to indirectly estimate adolescent pregnancy rate using multiple incomplete data sources.

An application of CRC method was conducted using three cross-sectional data sources, which were the national birth registrations, the Ministry of Public Health (MOPH) standard health databases, and hospital-based survey data during the years 2008 to 2013. The study was approved after full review by the Committee on Human Rights Related to Research Involving Human Subjects of the Faculty of Medicine Ramathibodi Hospital (ID 12–55-01) and the Department of Health, the Ministry of Public Health (ID 027). All data owners officially granted access to databases. Pregnant women were included in our study if they were aged 15 to 19 years at delivery. The outcomes of interest were live births and non-live births. The live birth was defined as a complete expulsion or extraction of a product of conception from mother after 22 weeks of gestation with sign of evidence of life or breath. The non-live births included miscarriage, induced abortion, stillbirth, and other abnormal pregnancies which were defined as follows: Abortion, which included induced abortion and miscarriage, which was defined as any delivery which occurred before 22 completed weeks of gestation. Stillbirth was defined as fetal death after 22 completed weeks of gestation. Abnormal pregnancy included ectopic pregnancy, molar pregnancy and others.

Total numbers of 741,084, 290,922 and 25,478 records from Source1, Source2, and Source3 were respectively eligible yielding 772,036 pregnancy records for further data analysis, see Fig. 1 and Additional file 1 (Fig. A1-A4).

Among them, 122,292 (15.8%) episodes were excluded due to non-OPS hospitals leaving a total of 649,744 episodes of OPS hospitals for CRC consisting of 627,453 and 22,291 pregnant episodes of live-birth (group 1) and non-live birth (group 2), respectively. Numbers of still birth, miscarriage, induced abortions, and abnormal pregnancies are described in Additional file 1-Table A5. Distributions of data were described by sources and time for live birth (group1, Table 1) and non-live birth (group 2, Table 2).

For group 1, the best model contained all possible two-way interactions with the AIC and BIC of 596.7 and 643.6, respectively, see Table 3. The missing numbers of pregnancies ranged from 25,819 to 30,218 given the observed numbers of live births of 98,791 to 112,003. The live birth rates were further estimated, which ranged from 52.7 to 59.2 per 1000 adolescent women, see Table 4.

For non-live births, a total of 22,291 observations from only Source2 and Source3 were used for CRC analysis, see Table 2. The estimated total number of non-live births ranged from 5445 to 26,897 with pregnancy rates of 2.3 to 11.2, see Table 4. Finally, the total count of numbers of non-live births were then combined with numbers of live births, yielding a total number of pregnancies of 133,551 to 169,119, which gained about 5445 to 26,898 more pregnancies compared with estimated numbers by live birth only. The adolescent pregnancy rate trended to increase significantly from 56.3 to 70.3 during the years from 2008 to 2012 (Chi-square for trend = 3.54, p = 0.009), but it decreased to 63.0 in 2013, see Table 4. Adolescent pregnancy rates were estimated by CRC and actual observed data were compared indicating higher estimated rates by CRC than Source1 only (adolescent birth rate), Source1 plus Source2, and Source1 plus Source2 plus Source3 with the corresponding case detection rates of 75.9–89.0%, 81.5–90.8% and 81.9–91.4%, respectively, see Fig. 2.

This study was conducted applying CRC analysis to estimate adolescent pregnancy rate in Thailand using a log-linear model approach which indicated a significant trend of increasing rate during the years 2008 to 2012, but declining in the year 2013. In addition, the estimated rates were higher in CRC method than the actual observed data by the Public Health Statistics. Estimation of adolescent pregnancy rate is still challenging in many countries, particularly where abortion is still restricted and thus only minimal cases estimations were mostly reported [22, 23].

Although birth registration in Thailand covers as high as 99% of all births [24], data for abortion, stillbirth, and miscarriage are incompletely registered with some degree of underreports. Applying CRC with a log-linear model for estimation of these numbers yielded many advantages as follows: first, all models were constructed under unified statistical framework, and model selection criteria were available for comparing models; second, dependence of data sources could be incorporated by adding interactions between each pair of data sources; and third, the covariates could be taken into account by adding in the model; and all inferences are within the statistical framework [17, 19].

However, the following limitations which might violate assumptions for performing CRC, were difficult to avoid [17, 19, 25]. First limitation was about the assumption that study population should be in closed system during the study period. Although we studied only subjects with Thai nationality, immigration still occurred and could not be avoided. The second limitation was from the assumption which stated individual subjects should be matched from capture to recapture. This refers to the correctness of the identification of subjects and matching them between different data sources, so each individual subject has positive probability to be ascertained by any data source, i.e., missing from any data source should not be a ‘structural zero’ or missing due to impossibility [19]. We were strongly concerned about this issue and performed two steps of CRC, i.e., predicted numbers of live births using three sources of data and non-live births using only Source2 and Source3. Only data from OPS hospitals were selected to keep the probability of pregnant women being identified from each data source to not be zero. The third limitation was from assumption concerning independence of data sources. Source independence can be accounted for by adding the interaction between the pair of data sources into the models. However, the highest order interaction must be assumed to be zero to allow identifiability, which could not be avoided for 2-source CRC in non-live birth group.

The fourth limitation was from assumption about capture homogeneity, which states each individual has the same chance to be ascertained by each data source. Heterogeneity among individuals may induce sources of dependence which can be partially reduced by stratified analysis. The fifth limitation was early pregnancy loss which would not require hospitalization and so could not be included in the samples and thus the estimation procedure.

Our CRC estimates yielded higher adolescent pregnancy rates than estimates based on actually observed data sources, particularly in non-live birth which was the consequence of adding Source2 to Source1. Therefore, we encourage applying CRC to provide more accurate estimation of adolescence pregnancy rate particularly in countries with restrictive abortion law. This will lead health care providers and policy makers to allocate resources properly. However, Source1 and Source2 are needed to improve the quality of the data, especially the identification using CID. The hospital-based survey should be performed regularly depending on feasibility and available funding and applying CRC method to provide more accurate estimation. For the non-livebirth group, the two-source CRC analysis has many theoretical limitations, so the third or fourth data sources should be sought to improve performances of CRC analysis and thus provide more valid results.

CRC method indicated that estimated adolescent pregnancy rates were much higher than the adolescent birth rates reported in Public Health Statistics. These two indicators should be used altogether for country situation analysis and strategic planning. This method can be applied not only in Thailand, but also other countries with similar contexts.