Women’s use of long-acting reversible contraception for birth timing and birth stopping

Abstract

The use of long-acting reversible contraceptive (LARC) methods—intrauterine devices (IUDs) and implants—has recently expanded rapidly in the U.S., and these methods together approach the contraceptive pill in current prevalence. Research so far on LARCs has analyzed their use to reduce unintended pregnancies, but not their use to enable intended pregnancies. Knowledge of both is necessary to understand LARC’s potential impacts on the reproductive life courses of U.S. women. We combine data from two nationally-representative surveys to estimate women’s likelihood and timing of subsequent reproductive events, including births resulting from an intended pregnancy up to nine years after discontinuing LARC use. We estimate that 62% of women will give birth, and 45% will give birth from an intended pregnancy. Additionally, 18% will have a new LARC inserted, and 13% will transition to sterilization. Most of these reproductive events occur within 2 years after discontinuing LARC use. Births from an intended pregnancy are especially common when no intervening switch to another contraceptive method occurs. We infer that women’s motives for using LARC are varied, but include to postpone a birth, to postpone a decision about whether to have a(nother) birth, and to transition definitively to the completion of childbearing.

INTRODUCTION

The twenty-first century has seen rapid growth in the use of long-acting reversible contraceptive (LARC) methods—intrauterine devices (IUDs) and implants—in the United States. LARC use among U.S. women has increased from close to 1% of contraceptive use in 1995 to 6% in 2008 to 16% in 2017–2019 (Daniels and Abma 2020; Hubacher et al. 2011; Kavanaugh and Jerman 2018; Kavanaugh et al. 2011). In 2014, IUDs (12%) were the fourth most used contraceptive method after the pill (25%), female sterilization (22%), and the male condom (15%), whereas implant use accounted for 3% of contraceptive use, up sharply from 0.5% in 2008 (Kavanaugh and Jerman 2018). IUDs and implants have much lower contraceptive failure rates than all other reversible methods (Sundaram et al. 2017; Winner et al. 2012). In addition, the methods have been shown to have higher continuation rates (Peipert et al. 2011; Simmons et al. 2019) and higher user satisfaction (Peipert et al. 2011; Sittig et al. 2020). Foster et al.’s (2015) survey of over 100 contraceptive researchers found median and modal projections of U.S. women’s combined IUD and implant use in the absence of common access barriers in the range of 25–29% of contraceptive use.

Quantitative research into LARC’s impacts until now has largely been from a public health perspective that has emphasized LARCs’ potential for reducing women’s risk of unintended pregnancy (e.g., Parks and Peipert 2016). The use of highly-effective and reversible contraception, however, should enable women to better achieve a broader set of reproductive goals than that of reducing the risk of unintended pregnancy (Dehlendorf et al. 2018). We argue that a major criterion against which a new contraceptive method should be evaluated is whether it increases women’s ability to have births of their desired number and timing over the reproductive life course, that is, to have intended births. Hence the present study’s principal focus is on women’s success in achieving intended pregnancies resulting in births following LARC use and subsequent discontinuation of use. Additionally, we are interested in women’s use of LARC as an intermediate method that may be a precursor to her transitioning to an equally effective, but permanent method (female or male sterilization). Our investigation of reproductive events following LARC use thus allows for a shift of focus from contraception as pregnancy prevention to contraception as a means towards achieving a woman or couple’s family-planning goals.

We first use a multinomial regression model to study the associations of a woman’s socio-demographic characteristics with experiencing any of three reproductive events in a given year of exposure up to nine years following discontinuation of LARC use. The three events are: giving birth; new LARC insertion; and initiating contraceptive sterilization. Second, we use a synthetic-cohort, life table methodology (similar to that used to describe, for example, the length of time until divorce, Martin 2006, or until remarriage, McNamee and Raley 2011) to trace out the cumulative implications for her reproductive life course of a woman’s annual rates of experiencing each of the three reproductive events. We estimate not only the proportion of LARC discontinuations that end in a birth, but also the proportion that end in a birth from an intended pregnancy. Our simulated reproductive life-course measures thereby include a cumulative fertility rate, consisting of the proportion of women who discontinue LARC use and who then subsequently have a next birth, and a cumulative intended fertility rate, consisting of the proportion of women who discontinue LARC use and have a next birth that results from an intended pregnancy. Also as a proportion of all women who discontinue LARC use, we simulate a cumulative rate of new LARC insertion and a cumulative rate of initiating contraceptive sterilization. Our simulation estimates these outcomes after up to nine years of exposure to reproductive events among women aged up to 34 years old at the time of LARC discontinuation.

Our data are from women who have used and then stopped using a LARC device. Only recently, however, have LARC devices been used to any substantial extent in the U.S. among the current generation of women of reproductive age. Therefore relatively few women have had any years of post-LARC-use exposure to observe, and even fewer have had even moderately long periods of fertility exposure following LARC use. We address this data challenge by pooling observations from two nationally representative data sources, the 2006 to 2017 years of the National Survey of Family Growth (NSFG) and the 2002 to 2017 waves of the National Longitudinal Survey of Youth 1997 (NLSY97). We assess the appropriateness of pooling observations across the two surveys using formal statistical methods.

To anticipate our findings, we estimate that within nine years of discontinuing LARC, 62% of women will have a birth, and 45% of women will have a birth resulting from an intended pregnancy. Additionally, 31% of women will either have a new LARC inserted or begin sterilization as their method of contraception within nine years of discontinuing LARC. Hence, we infer substantial successful use of LARC to realize women’s plans for their birth timing, and also substantial use of LARC as part of women’s or couples’ decision-making about whether and when to definitively end their exposure to childbearing.

LITERATURE REVIEW

Compared with other low-fertility countries, until very recently (Hubacher and Kavanaugh 2018) the U.S. has had one of the lowest rates of LARC use (Buhling et al. 2014; Eeckhaut et al. 2014). The historical background of LARCs in the U.S.—first-generation IUDs such as the Lippes Loop—goes back to around 1960 (Thiery 1997). However, their take-up was shortly followed by steep declines in usage when serious safety concerns regarding the Dalkon Shield IUD in the early 1970s resulted in a prolonged period of litigation (Sonfield 2007). IUD use was revived in the United States after the levonorgestrel-releasing IUD was approved for medical use in 2000 (Kavanaugh et al. 2011). Although IUDs remain the dominant LARC in the U.S., as they do in most other developed countries (Eeckhaut et al. 2014; U.N. 2013), implant use has also increased rapidly in recent years, doubling from 1.3 percent of current contraceptive users in 2012 to 2.6 percent in 2014 (Kavanaugh and Jerman 2018). The first implant—Norplant—was approved in 1990, but there were early gaps in availability of the method. Norplant was withdrawn from market in 2002, and the second implant method, Implanon, was not approved until 2006, and then replaced by its newer version, Nexplanon, in 2011 (Strasser et al. 2016). The high up-front cost of LARCs has been slow to be covered by health insurance. LARC devices and their insertion cost were, until the introduction of the Affordable Care Act in 2010, omitted from many health insurance plans (Bearak et al. 2016; Sonfield 2007). Reforms to Medicaid policy have provided coverage for immediate postpartum (IPP) LARC insertion across an increasing number of states (Moniz et al. 2015; Rodriguez et al. 2014). In addition, several localities and states have seen efforts to increase LARC use more broadly, including the Contraceptive CHOICE Project in St.-Louis, Missouri, the Colorado Family Planning Initiative (CFPI) in Colorado, the HER Salt Lake Contraceptive Initiative in Utah, and the Delaware Contraceptive Access Now (DEL-CAN) Initiative in Delaware. At the same time, however, a growing body of reproductive justice-based work has issued warnings about the potentially negative impact on women’s reproductive autonomy of efforts aimed at increasing LARC use, rather than increasing access to the full range of contraceptive methods (Dehlendorf et al. 2018; Gomez et al. 2014; Higgins 2014).

Arguments for, and research on, LARC access and use in the U.S. has centered on its potential to reduce unintended pregnancy (Karpilow and Thomas 2017; Lindo and Packham 2017; Parks and Peipert 2016; Peipert et al. 2012; Ricketts et al. 2014; Secura et al. 2014; Trussell et al. 2013; Wu and Mark 2018). Reducing the high level of unintended pregnancy among U.S. women has been a public health priority for decades (Cutright and Jaffe 1976). Despite shortcomings in terms of describing or predicting the behavior of any particular individual (Higgins et al. 2012; Potter et al. 2019), and failure to capture affective dimensions (Geist et al. 2019; Higgins 2017), pregnancy intendedness has been shown to be a valuable concept at the population level (Kost and Zolna 2019; Santelli et al. 2003). Pregnancy intentions have been found to show extraordinary consistency at the aggregate level, and constitute a valuable predictor of population level fertility as well as a key indicator of the need for contraceptive and abortion services. Hence their importance in terms of assessing progress towards improving women’s reproductive health and well-being.

The U.S. has a substantially higher rate of unintended pregnancy than that in other high-income countries. Sedgh et al. (2014) showed in a cross-national comparison of unintended pregnancy around 2012 that North America’s (the U.S. and Canada’s) 51% unintended fraction was as much as 50% higher than in other high-income countries: for example, 38% and 34% of pregnancies respectively in Northern and Western Europe, and 37% of pregnancies in East Asia and Oceania. Unintended pregnancies are associated with adverse birth and maternal health outcomes (Kost and Lindberg 2015; Lindberg et al. 2015). Therefore, increasing access to and use of LARC in the U.S. to reduce unintended pregnancies has been promoted as a public health goal (ACOG 2009, 2015; Wu et al. 2018). The use of LARC, however, should enable women to achieve their reproductive life course goals in ways that go beyond those linked to reducing the risk of unintended pregnancy (Dehlendorf et al. 2018). Crucial among its potential purposes is the extent to which LARC use increases women’s ability to have births of their desired number and timing over the reproductive life course, that is, to have intended births. A related reason for attending to the impact of LARC use on intended fertility that should be of major salience to demographers is that, unless reductions in unintended fertility achieved through use of LARC are counter-balanced by positive impacts on intended fertility, overall U.S. fertility will decrease markedly with increased LARC use. The U.S. total fertility rate (TFR), at 1.73 in 2018 (Martin et al. 2019), stands well below replacement (though see also Wu and Mark 2019). With unintended fertility accounting for approximately 35% of all U.S. births (Hayford and Guzzo 2016), a simple subtraction of all unintended births would leave the TFR at 1.12, squarely within the range of ‘lowest-low’ fertility (Kohler et al. 2002).

We argue that efforts to increase use of LARC methods in the U.S. need to be grounded in a fuller understanding of their impact on women’s reproductive life courses. We note here that LARC devices are marketed as being a fully-effective contraceptive method for three (implant and Skyla hormonal IUDs), five (Mirena, Liletta, and Kyleena hormonal IUDs), or ten years (Paragard copper IUD) (ACOG 2015; U.S. FDA 2018). Following a reproductive life-course perspective, use of LARC may increase intended fertility in two ways: (1) a birth that would have occurred earlier than desired (and thus resulted from an unintended pregnancy) may instead occur later, as a birth resulting from an intended pregnancy; and (2) a birth that would not have occurred at all because the woman opts to rely on sterilization (which she would have come to regret), may instead occur and result from an intended pregnancy. That is, some women who would have eliminated their risk of unintended pregnancy by using a permanent method (contraceptive sterilization) might, following a period of LARC use in place of sterilization, experience a change in fertility goals and go on to have (further) intended pregnancies. As many as one in four women in the U.S. who are sterilized subsequently report that they would like to have their sterilization reversed (Eeckhaut et al. 2018).

On the other hand, use of LARC could also work to reduce intended fertility if LARCs’ comparatively low risk of contraceptive failure works to prevent intended, in addition to unintended, pregnancies. First, research has shown that ‘only’ 68 percent of pregnancies resulting from contraceptive failure are reported as unintended (Trussell et al. 1999), and therefore reducing contraceptive failure will potentially prevent some number of pregnancies that would subsequently be reported as intended. Second, barriers to LARC removal experienced by women who may either have been ambivalent, or who had changed their intentions, about wanting to become pregnant (Amico et al. 2020; Strasser et al. 2017) could have the effect of delaying or even preventing pregnancies that would be reported as intended.

We know of no previous study that has investigated the fraction of women that go on to have any birth or an intended birth following LARC discontinuation. Kramer et al. (2018), however, found that among current LARC users, 57% of Black women, 45% of Hispanic women, and 38% of non-Hispanic White women report either desiring future children or not being sure about their intentions. This suggests considerable potential for intended fertility following LARC discontinuation. We know of no direct evidence on a role for LARC use in increasing births resulting from an intended pregnancy through substituting LARC for contraceptive sterilization. Suggestive evidence, however, is found in the substantial increases in LARC use between the late-2000s and mid-2010s that coincided with substantial decreases in sterilization (Kavanaugh and Jerman 2018). Moreover, Eeckhaut et al. (2014) found in a nine-country study of couples’ contraceptive use in the mid-late 2000s that those countries that had the lowest prevalence of LARC use (Australia and the U.S.) had by far the highest prevalence of contraceptive sterilization. Greater use of LARC in place of sterilization might therefore be associated with more intended pregnancies before a permanent end to a woman’s exposure to pregnancy, where that permanent end to exposure may occur either through the woman’s aging out of her reproductive years or through the woman’s initiating contraceptive sterilization at a later age, still inside her reproductive years.

METHODS

Data for this study were drawn from the NSFG’s 2006–2010, 2011–2013, 2013–2015, and 2015–2017 cycles (U.S. Department of Health and Human Services 2018) and the NLSY97’s annual 2002 to 2011 waves and biennial 2013, 2015, and 2017 waves (Bureau of Labor Statistics nd). Additional details on the two sources are given in Appendix A. In both surveys, we analyzed samples of women who reported having used and then discontinued use of a LARC. A major difference between the NSFG and the NLSY97 is that the NSFG samples over multiple cross sections of U.S. reproductive-age women from 2006 through 2017, whereas the NLSY97 is sampled from the population of young women who were born between January 1, 1980, and December 31, 1984 and resided in the U.S. in 1997, and then followed as they aged through to their early-to-mid-30s in 2017. We discuss differences in sociodemographic composition of LARC-discontinuing women between the two surveys as a result of these different sampling and observation schemes when presenting descriptive results from Table 1, and we discuss potential implications of these survey differences for the estimation of reproductive event risks when presenting the results of the multivariate estimation of those event risks in Table 2. A second major difference between the NSFG and the NLSY97 is that the NSFG identified contraceptive method from a monthly contraceptive calendar involving just over three years of recall, whereas the NLSY97 identified contraceptive method from an annual contraceptive-use report involving recall only in the last 12 months. The implications of these offsetting advantages and disadvantages in terms of detail of period reported on (months versus a year) and length of recall period (3+ years versus only 1 year) for estimation bias are unclear. We present comparisons of event hazard risks by single year of duration since LARC discontinuation between the NSFG and NLSY97 in Appendix B, and find that overall patterns of annual demographic event risks between the two surveys are more similar than they are different.

Table 1:

Percent distribution of characteristics of women at LARC discontinuation, United States 2002 to 2017.

	NSFGa	NSFGa (women ages 20–34 at LARC discontinuation)	NLSY97b
N	718	550	514
Age at LARC discontinuation	pdiff < .001
15–19	6.0	NA	1.4
20–24	24.3	33.6	15.4
25–29	30.1	41.7	44.8
30–34	17.9	24.7	36.2
35+	21.7	NA	2.1
Calendar year of LARC discontinuation	pdiff = .724
2002–2006	8.6	8.6	8.6
2007–2011	36.5	38.0	40.9
2012–2017	54.8	53.4	50.4
Race/ethnicity	pdiff < .001
White	52.1	50.5	72.0
Hispanic	29.6	28.1	17.0
Black	9.4	10.8	10.0
Other	8.9	10.6	1.0
Education at LARC discontinuation	pdiff < .010
<High School	13.7	12.1	16.7
High School	28.4	29.1	25.0
Some College	35.7	38.9	28.3
BA+	22.3	19.9	30.0
Union status at LARC discontinuation	pdiff = .393
Married	49.3	46.9	46.9
Cohabiting	20.0	21.6	25.6
Single	30.8	31.5	27.6
Parity at LARC discontinuation	pdiff = .769
0	16.2	15.4	17.6
1	29.4	32.8	32.5
2+	54.4	51.8	49.9
Duration of LARC use before discontinuation	pdiff < .001
1 year	23.3	21.5	24.3
2 years	20.4	20.6	35.5
3 years	15.9	17.8	12.5
4+ years	40.4	40.2	27.7

Notes: All percentages are weighted. p_diff. indicates results of F-test of null hypothesis that there is no difference between the NSFG (women ages 20–34 at LARC discontinuation) and NLSY97 estimates.

Sources:

^aNational Survey of Family Growth (NSFG), 2006–2010, 2011–2013, 2013–2015, and 2015–2017 cycles. LARC discontinuation occurred no more than three calendar years before the calendar year of the woman’s survey interview.

^bNational Longitudinal Survey of Youth 1997 (NLSY97) Cohort.

Table 2:

Multinomial logistic regression of birth, new LARC insertion, and sterilization after LARC discontinuation: Women who discontinued LARC use, 2002–2017, Relative Risk Ratios.

	Live birth	New LARC insertion	Sterilization
Age at LARC discontinuation		***
15–19	1.17	0.18*	0.21
20–24	1.97*	1.04	0.57
25–29	1.70*	0.56	0.52+
30–34	1.00	1.00	1.00
35+	1.31	1.47	3.51*
Education at LARC discontinuation
<High School	1.21	0.90	0.62
High School	1.00	1.00	1.00
Some College	1.13	0.89	0.53+
BA+	1.43	0.58	0.60
Race/ethnicity		***
White	1.00	1.00	1.00
Hispanic	1.08	0.98	0.85
Black	0.69+	1.57	0.42+
Other	4.79***	8.44***	3.34+
Parity at LARC discontinuation		***
0	0.43***	0.72	0.75
1	1.00	1.00	1.00
2+	0.62**	0.69	2.40*
Union status at LARC discontinuation		**
Married	1.00	1.00	1.00
Cohabiting	0.67*	0.76	0.53
Single	0.47***	1.09	0.60
Duration of LARC use before discontinuation		***
1 year	2.43***	7.96***	1.38
2 years	1.00	1.00	1.00
3 years	1.46+	1.65	1.51
4+ years	1.53*	0.66	1.51
Time since the year of LARC discontinuation		***
1 years	0.61**	1.01	2.98*
2 years	1.00	1.00	1.00
3 years	0.76	1.73	1.28
4 or more years	0.25***	0.68	1.81
Calendar year of LARC discontinuation
2002–2006	0.78	0.84	0.65
2007–2011	1.04	1.18	1.19
2012–2017	1.00	1.00	1.00

Notes:

⁺p < .10,

^*p < .05,

^**p < .01,

^***p < .001.

Symbols opposite of variable names indicate results of Wald test of null hypothesis that the full set of variable coefficients are jointly equal to one.

Sources: 2006–2010, 2011–2013, 2013–2015 and 2015–2017 NSFG cycles, N = 598 person-years, and NLSY97 2002 to 2017 waves, N = 1,189 person-years.

Our main outcome of interest is whether or not the woman experienced a live birth. We define the cumulative fertility rate for up to nine post-LARC years of fertility exposure, where either one of two competing events may end her spell of fertility exposure: a new LARC insertion; or sterilization. Cumulative risks of a new LARC insertion and of sterilization are similarly defined for up to nine years following LARC discontinuation. We also consider whether or not the woman experienced a live birth that resulted from an intended pregnancy. We define the intended fertility rate as the proportion of all women who discontinue LARC use who then subsequently have a next birth that results from an intended pregnancy. We specify the intended fertility rate as the product of a cumulative fertility rate and the conditional probability that, if a birth does occur, it results from an intended pregnancy.

The cumulative fertility rate, the cumulative new LARC insertion risk, and the cumulative sterilization risk are calculated from an annual competing hazard of birth, new LARC insertion, and sterilization by number of years since LARC discontinuation (Model (1) below). Both NSFG and NLSY97 data are used in the estimation of this hazard. Combining data from the two surveys requires that we simplify the sequences of contraceptive method use and discontinuation to those occurring in annual periods. We coded events in the NSFG by the calendar year in which they occurred, and in the NLSY97 by the approximately 12-month period between annual survey waves, imputing those events in the 2012, 2014, and 2016 years between the biennial survey waves 2011 to 2017 (see Appendix A for more detail). For the NSFG, year 0 indicates the calendar year that includes the month of last reported LARC use before the month of first reported non-use of LARC. For the NLSY97, year 0 indicates the approximately 12-month period between the waves of last reported LARC use (retrospectively reported LARC use over the 12 months before the more recent wave). For the NSFG, most women can be followed in the year of LARC discontinuation (year 0), and for only one (year 1) or two (year 2) years after the year of LARC discontinuation. This is due to the limited time period covered by the NSFG contraceptive calendar (i.e., a maximum of three to four years). For the NLSY97, women can be followed for up to 11 years after the year of LARC discontinuation (year 0). However, because of the very sparse data at long durations, we limit our simulated period to nine years, from year 0 to year 8.

Our estimation of the intended fertility rate combines data on the cumulative fertility rate (a simplified version of Model (1) below) with an estimate of the conditional probability that, if a birth does occur, it results from an intended pregnancy. Direct information on the intendedness status of a pregnancy resulting in live birth after LARC discontinuation comes from NSFG data only. Intendedness was determined based on a series of retrospective NSFG questions asking the woman to report her pregnancy intentions “Right before you became pregnant …” Following previous work (Mosher et al. 2012), we coded an intended pregnancy as one reported as having occurred at the right time or later than desired. Responses of “didn’t care, indifferent” (N = 1) or “don’t know, not sure” (N = 1) were also classified as intended. Information used in the prediction of intendedness status, specifically duration since LARC discontinuation and whether the woman switched to a non-LARC method following LARC discontinuation, comes from both the NSFG and NLSY97.

STATISTICAL ANALYSIS

We first use data from the NSFG and NLSY97 to estimate a multinomial logit (MNL) model of the annual competing hazard of birth, new LARC insertion, and sterilization in the years following the year of LARC discontinuation:

$$\text{Pr}\left\{\text{birth; new LARC; sterilization}\right\}\text{ = MNL}\left(\text{sociodemographic variables; years of LARC use; years since LARC discontinuation}\right)$$ (1)

We coded woman’s characteristics, including her age, education, union status, parity, and race/ethnicity, at the time of LARC discontinuation. In addition, we include a predictor of the number of calendar years (NSFG) or survey years (NLSY97) during which the LARC device was used before it was discontinued. Number of years (“duration”) since LARC discontinuation, finally, is included as it is a key predictor variable in our study. In the NLSY97, we are able to code this from 12-month retrospective contraceptive use questions at each panel wave. In the NSFG, only 0, 1, and 2 years duration since LARC discontinuation can be coded from the contraceptive calendar. Therefore, for our analyses of reproductive events following LARC discontinuation, we only use these recently-observed discontinuations in the NSFG.

To estimate the annual competing hazard with the full set of regressors (Model (1) above), observations of LARC discontinuers in the NSFG and NLSY97 were pooled. We exclude the year of LARC discontinuation here, year 0, because of the very low hazard of any event occurring in year 0 (see Appendix Table A1 for the hazard including year 0). Ours is not the first study to conduct pooled estimation using National Longitudinal Survey (NLSY) cohort data and one other survey. Hellerstein and Imbens (1999) combined NLSY 1979 cohort data with Current Population Survey data for labor market outcomes. Weden et al. (2012) combined NLSY 1979 cohort data with Early Childhood Longitudinal Study, Birth Cohort data for a child obesity outcome. The Weden et al. (2012) study is closer to ours in its using the same regressor variables in both surveys. They evaluated the appropriateness of pooling observations across surveys using model-fit assessment methodology described in Rendall et al. (2013), and we apply that same methodology in the present study (see Appendix B, presentation and discussion of Appendix Table A3).

All our estimates are weighted, using the NLSY97 and NSFG sample weights. Before pooling observations across the four NSFG cycles (2006–2010, 2011–2013, 2013–2015, and 2015–2017), we normalize the weight of each cycle to have a mean of 1, and we normalize the weight of the NLSY97 similarly to have a mean of 1. This ensures that equal weight is given to observations sampled in these different periods and in different surveys. This procedure follows the methodology for pooled-survey analysis recommended in Rendall et al. (2008). We additionally adjust for clustering and stratification in the NSFG-only analyses of Model (1). Because the clustering and stratification of observations occurs differently between the NSFG and NLSY97, solutions for which have yet to be developed in pooled-survey analysis, we do not adjust for these in our pooled-survey estimation.

In addition to the full competing-hazard model above, we conduct a synthetic cohort simulation that joins the predicted probabilities from a simplified version of Model (1) with the conditional probability that, if a birth does occur, it results from an intended pregnancy (see Appendix B for details on the estimation of this conditional probability). We simulated nine years of post-LARC exposure to a birth (i.e., the fertility rate), to a birth resulting from an intended pregnancy (i.e., the intended fertility rate), to a LARC re-insertion, or to sterilization. This simulation includes as predictors only time since LARC discontinuation and whether she used a non-LARC method in the year before a birth. We use this as a proxy for whether she switched to a non-LARC method between LARC discontinuation and having a birth. We use this proxy for its availability in both the NLSY97 and the NSFG. A full description of this simulation method is given in Appendix B.

RESULTS

CHARACTERISTICS OF WOMEN WHO DISCONTINUE LARC

In Table 1, we describe U.S. women who discontinued LARC use between 2002 and 2017, as identified in the NLSY97 and NSFG. Previous studies have described current LARC users (e.g., Kramer et al. 2018; Xu et al. 2011) and the characteristics associated with early LARC discontinuation (e.g., Grunloh et al. 2013; Phillips et al. 2017). However, no nationally-representative study to our knowledge has described women who have discontinued LARC use. Although both the NSFG and NLSY97 provide such nationally-representative estimates of women who discontinue LARC use in the observed, 2002–2017 period, we prioritize description from the NSFG in our presentation of findings on the characteristics of LARC-discontinuing women. This is because of both its coverage of women at all reproductive ages (15 to 44) and because its sampling over multiple cross sections of U.S. reproductive-age women from 2006 through 2017 allows for ongoing period representation of the sociodemographic characteristics of women discontinuing LARC use, including representation of recent immigrants. In contrast, the NLSY97 is sampled from the population of young women who were resident in the U.S. in 1997, thus representing a smaller fraction of Hispanic women than in the NSFG, and the 1997 NLSY youth cohort is then followed as they aged through to their early to mid-30s. In our comparisons immediately below of estimates of characteristics of LARC-discontinuing women between the NSFG and NLSY97, we compare to the NSFG subpopulation of LARC discontinuers at ages 20 to 34. Our primary substantive interest in this study is of women who discontinued LARC use up to age 34, as for the up to nine years following LARC discontinuation of our simulation, those women are still of reproductive ages, and therefore exposed to reproductive events.

The NSFG estimates are derived from a sample size of 718 women age 15–44 and 550 women age 20–34 (at the time of LARC discontinuation). The modal group of LARC discontinuers is aged 25–29 for both age ranges. The timing of the re-introduction of hormonal IUDs and implants back onto the U.S. market also means that the majority of LARC discontinuations have been recent. More than half of discontinuations occurred in the 2012–2017 period, over a third in 2007–2011, and fewer than one in ten in 2002–2006.

Hispanic women account for relatively large shares of LARC discontinuers (almost 30%), consistent with estimates of current LARC use among U.S.-born and especially foreign-born Hispanics (Tapales et al. 2018). Non-Hispanic White women account for half of women discontinuing LARC, followed by Black and Other race/ethnicity at around 10% each. Women at all education levels are represented among LARC discontinuers, with high school graduates and “some college” being the largest two educational attainment groups. A relatively high fraction of LARC discontinuers were partnered at discontinuation: half of all discontinuers were married and another fifth cohabiting at the time of LARC discontinuation. The large majority of women were parous at the time of LARC discontinuation, just over half having already given birth to at least two children, and around 30% having given birth to one child only at the time of LARC discontinuation. The modal duration of women’s LARC use before discontinuation is 4 or more years (possibly coincident with the end of the device’s life cycle), followed by durations of only 1 or 2 years’ use. This result is broadly consistent with previous estimates of mean completed durations of IUD use of 47 months in the 2006–10 and 31 months in the 2011–13 NSFG data (Gomez et al. 2016). Short durations may reflect method dissatisfaction or unintended expulsion of IUDs (Goldthwaite et al. 2017; Grunloh et al. 2013), or planning for a timely next birth.

Ages 20 to 34 correspond to the approximate overlap in ages at LARC discontinuation between the NLSY97 and the NSFG. We are thereby able to examine the LARC discontinuation distributions and subsequent dynamics from as many as 1,064 women in this age range. We present statistics separately for the NSFG and NLSY97 samples because the structure of cohort observation in the NLSY97 introduces some differences based on the different time of sample selection and based on different matches between age and period. In particular, by the time LARCs were again increasing in popularity after the earlier IUD problems and the gap in implant availability following Norplant discontinuation, the NLSY97 cohort was mostly in their mid-20s. This means that 30–34 year olds account for a larger fraction of LARC discontinuers in the NLSY97 (36.2%) than do 30–34 year olds represented in the 20–34 year old NSFG sample (24.7%). College graduates are similarly more highly represented among NLSY97 LARC discontinuers (30.0%) than among 20–34 year olds from NSFG-sampled LARC discontinuers (19.9%). The NLSY97 sample was selected from those resident in the U.S. in 1997, and this is apparent in the race/ethnic composition which has less Hispanic women (and more non-Hispanic White women) than in the NSFG estimates of LARC discontinuers. Distributions of women by calendar year, union status, and parity at LARC discontinuation, however, are similar between the estimates based on the 20–34 year old NSFG sample and the NLSY97 sample of LARC discontinuers.

SOCIO-DEMOGRAPHIC AND LARC-HISTORY PREDICTORS OF LIVE BIRTHS, NEW LARC INSERTIONS, AND STERILIZATIONS AFTER LARC DISCONTINUATION

Multivariate estimates of the relative risks ratios (RRRs) for a live birth, new LARC insertion, and sterilization after LARC discontinuation are shown in Table 2. These estimates are based on the pooled NSFG (all women ages 15–44) and NLSY97 samples. Model-fit statistics supporting the pooling of observations across the two surveys are shown in Appendix Table A3.

Compared to the reference age group 30–34, the risk of giving birth was higher for women discontinuing LARC in their 20s (RRR = 1.97 and 1.70 respectively for 20–24 and 25–29 year olds at LARC discontinuation). Compared to non-Hispanic White LARC discontinuers, Black LARC discontinuers have a lower risk of birth (RRR = 0.69, p < .10) and Other race/ethnicity discontinuers have a higher risk of birth (RRR = 4.79). Being at parity one at LARC discontinuation is associated with a higher likelihood of birth than either being nulliparous (RRR = 0.43) or parity two or above (RRR = 0.62). Being married at the time of discontinuing LARC use is also associated with a higher birth risk compared to either cohabiting (RRR = 0.67) or being single (RRR = 0.47). Very short (one year) duration of the use of LARC before discontinuing is associated with an elevated risk of birth (RRR = 2.43), as are longer (four or more years) duration of LARC use before discontinuing (RRR = 1.53). Two years since discontinuation is the peak duration for birth risk. Having been one year since the year of LARC discontinuation (RRR = 0.61) or four or more years since the year of LARC discontinuation (RRR = 0.25) is associated with substantially lower birth risk.

The other two spell-ending reproductive events are new LARC insertion and sterilization. New LARC insertion is mostly dependent on how long they had used their previous LARC device. Women were much more likely to have a new LARC inserted if they had used the previous LARC device for only one year (RRR = 7.96). Women of Other race/ethnicity (RRR = 8.44) were much more likely to have a new LARC inserted compared to non-Hispanic White women, and age 15–19 at discontinuation was associated with a lower risk of having a new LARC inserted (RRR = 0.18).

Regarding factors associated with sterilization, age has the expected positive gradient, sterilization risk being higher when over age 35 (RRR = 3.51), and lower at ages 25–29 at LARC discontinuation (RRR = 0.52, p < .10) than for women aged 30–34 at LARC discontinuation. Higher parity was also positively associated with sterilization, those with two or more live births more likely (RRR = 2.40) to transition to sterilization following LARC discontinuation relative to those at parity one. These findings are consistent with general reproductive life-course expectations of who is most at risk of sterilization (Chandra 1998; Eeckhaut and Sweeney 2016). Finally, women are at higher risk of sterilization one year after the year of discontinuing LARC (RRR = 2.98), compared to the reference duration of two years since the year of LARC discontinuation. In these cases, we infer that LARC might have been used by the woman to defer a decision to stop childbearing, and when she is more certain about the decision, she replaces her LARC device with a permanent method.

Before describing the results from our simulations of reproductive events subsequent to LARC discontinuation, for which we control only for years since discontinuation, it is useful to note how the similarities and differences in sociodemographic characteristics of LARC-discontinuing women between the NSFG (age 20 to 34) and NSLY97 samples might impact our pooled-survey estimates of risks of those events. We first note that distributions for two of the strongest predictors of risks of reproductive events, union status and parity at LARC discontinuation, are substantively and statistically indistinguishable between the NSFG and NSLY97 (see again Table 1). Being married and of parity 1 are the characteristics associated with the highest risk of birth following LARC discontinuation, and account respectively for 47% and 33% of LARC discontinuers in both the NSFG age 20–34 and NLSY97 samples. Being of parity 2+ is associated with easily the highest risk of sterilization following LARC discontinuation, and parity 2+ women accounted respectively for 52% and 50% of LARC discontinuers in the NSFG age 20–34 and the NLSY97 samples. Also reassuring is that one of the sociodemographic characteristic that differs most between the NSFG and NLSY97, that of Hispanic versus non-Hispanic White ethnicity (Hispanics represent 11 percentage points more LARC discontinuers in the NSFG than the NLSY97, see again Table 1), does not have a substantively or statistically significant multivariate association with birth, new LARC insertion, or sterilization event risk. However, two other characteristics that have different distributions between LARC discontinuers in the NSFG age 20–34 and the NLSY97 samples, being age at discontinuation and duration of LARC use before discontinuation, do have substantively and statistically significant multivariate associations with birth and sterilization events. Being aged 20–24 is associated with almost twice the risk of giving birth relative to being age 30–34 at LARC discontinuation, and 20–24 year olds account for only 15% of LARC discontinuers represented by the NLSY97 sample versus 34% of LARC discontinuers represented by the NSFG age 20–34 sample. This may downwardly bias the annual birth hazards at 3 or more years following LARC discontinuation that are used in the simulation, as only NLSY97 sample members can be used for those 3+ year durations. Countering this downward bias, however, is the lower fraction of NLSY97 than NSFG age 20–34 sample members who had been using a LARC device for 4 or more years before discontinuation (28% versus 40%). This characteristic is associated with a much lower annual birth hazard than being of the reference category of 2 years duration of LARC use (represented by 36% of NLSY97 versus 21% of NSFG age 20–34 women). This higher fraction at 2 years’ than 4+ years’ duration of LARC use before discontinuation in the NLSY97 will have an upward bias on annual birth hazards at 3 or more years following LARC discontinuation. Overall, then, the differences in sociodemographic characteristics between the NLSY97 and NSFG samples result in no clear direction of estimation bias implied for the simulation results, to which we now turn.

SIMULATION OF REPRODUCTIVE EVENTS AFTER LARC DISCONTINUATION

Our simulation of births, new LARC insertions, and female or male sterilizations up to nine years after LARC discontinuation is summarized in Table 3. Only 7.4% of LARC discontinuations result in none of these three reproductive events within nine years; 61.9% of women are projected to have experienced a birth, 17.6% a new LARC insertion, and 13.1% female or male sterilization. The 61.9% of women experiencing a birth by nine years following LARC discontinuation constitutes the cumulative fertility rate. These include 45.2% of all LARC-discontinuing women who experience a birth resulting from an intended pregnancy. This constitutes the cumulative intended fertility rate.

Table 3:

Summary statistics of births and other reproductive events, by time since LARC discontinuation and whether switched to a non-LARC method after LARC discontinuation.

	Years after LARC discontinuation			Contraceptive method in year before birth
	0–8	0–5	0–2	None	Non-LARC method
Distribution of LARC discontinuation spell-ending events
Birth	61.9	58.0	44.5	46.1	15.8
resulting from an intended pregnancy (intended fertility rate)	45.2	42.8	34.6	36.6	8.6
resulting from an unintended pregnancy	16.7	15.2	9.9	9.5	7.2
New LARC insertion	17.6	15.7	9.6
Sterilization	13.1	11.6	8.8
No event	7.4	14.8	37.0
Percentage from intended pregnancy among births following no method	79.4	79.3	79.7
Percentage from intended pregnancy among births following switch to a non-LARC method	54.5	55.6	65.8

Source: authors’ simulations using NSFG 2006–2010, 2011–2013, 2011–2015, and 2015–2017 and NLSY97 2002–2017 data.

We divide the occurrence of each of the reproductive events into three cumulative periods, 0–2 years, 0–5 years, and 0–8 years, and we divide births into those occurring after switching to a non-LARC contraceptive method between the LARC discontinuation and the birth versus those births not preceded by contraceptive switching. Recall that the duration since LARC discontinuation and whether the woman switches to a non-LARC contraceptive method between the LARC discontinuation and the birth are the two factors in our simulation model that predict whether the woman reports the birth as resulting from an intended pregnancy. As shown in the last two rows of Table 3, the fraction of births that follow an intended pregnancy among women who give birth without switching to a non-LARC method is 79.4% (36.6% out of the 46.1% of women who give birth without switching), versus 54.5% intended among women who give birth after switching to a non-LARC method (8.6% of the 15.8% of all LARC-discontinuers who switch methods and then give birth).

Regarding duration since LARC discontinuation, by two years after the year of discontinuation (year 2), 71.9% of all births that will eventually occur within nine years post-discontinuation will have already occurred (44.5% out of a total of 61.9% women having a post-LARC birth). An even higher fraction, 76.5%, of all births resulting from an intended pregnancy will have already occurred (34.6% out of a total of 45.2% women having a post-LARC intended birth within nine years). By five years after the year of LARC discontinuation, almost all of the LARC-discontinuing women who will experience a birth (58.0%) will have already done so. Moreover, only 14.8% of women discontinuing LARC will not have experienced any of a birth, new LARC insertion, or sterilization within five years. This follows from an especially high fraction of eventual sterilization events occurring by years 0–2 (8.8% of the cumulative sterilization rate of 13.1%), and a majority too of eventual new LARC insertions occurring within years 0–2 (9.6% of the cumulative new LARC insertion rate of 17.6%). In years 6–8, another 3.9% of women give birth, 1.9% experience new LARC insertions, and 1.5% experience transitions to sterilization.

DISCUSSION

This study aimed to broaden investigation of LARC beyond the focus of existing quantitative research on LARC’s potential to reduce unintended pregnancy to consider LARC’s potential to increase intended pregnancies resulting in births. We estimated that 62% of women who discontinue LARC use between the ages of 20 and 34 have a live birth within nine years of discontinuation. This is a relatively high proportion given that about half of LARC-discontinuing women in this age group already had two or more children. However, LARC discontinuers also had characteristics facilitative of going on to a next birth, including the more than two thirds of women who were either married or cohabiting at the time of LARC discontinuation. We estimated a cumulative intended fertility rate of 45%. That is, just under half of LARC discontinuers go on to have a birth that results from an intended pregnancy inside a spell of nine years after discontinuing LARC use. This is broadly consistent with Kramer et al.’s (2018) findings from analysis of the birth intentions of current LARC users in 2011–2015, in which just under half of non-Hispanic White (45%) and Hispanic (38%), and just over half (57%) of Black women expressed an intention to consider a future birth. We infer from this high intended fertility rate that many women were using LARC with the intention to successfully transition to an intended pregnancy following LARC discontinuation.

We also considered two other reproductive behaviors following LARC use and subsequent discontinuation of use. Most women who do not have a birth within nine years after discontinuing LARC either have a new LARC inserted (18%), or initiate (female or male) contraceptive sterilization (13%). We estimate that only 7% of women experience none of these three reproductive events within nine years of LARC discontinuation.

Reproductive life-course characteristics are not generally associated with the event of a new LARC insertion. Only having recently (within a year) had their LARC removed (or expelled) is strongly (and positively) associated with having a new LARC insertion, suggesting that LARC discontinuation may not have been due to readiness for pregnancy but instead because of problems with the previously-inserted LARC device, such as IUD expulsion or dissatisfaction due to adverse side effects. By subsequently having a new LARC inserted—possibly of a different type—these women opted to continue highly effective, but reversible, pregnancy protection. This we infer to be either to further postpone a (next) birth or to further postpone a decision about whether or not to have a (next) birth.

Reproductive life-course characteristics are important in predicting which women are most likely to proceed to sterilization in their years following LARC discontinuation. Notably, women who are older and of higher parity are more likely to have their post-LARC-discontinuation fertility exposure end with sterilization. Women who had their LARC device inserted for three or more years before its removal were also more at risk of transitioning to sterilization. These may be women who have reached the end of their LARC device’s lifecycle (which is three to ten years, depending on LARC type) and, rather than having a new LARC inserted, decided to switch to sterilization—possibly because their decision to permanently end childbearing had crystallized.

The high predictiveness of reproductive life-course readiness characteristics for either giving birth or for transitioning to sterilization following LARC discontinuation, and mostly lower predictiveness of socio-economic characteristics such as race/ethnicity and educational attainment after taking reproductive life-course factors into account, suggests that women across social groups typically use LARC in ways that match their reproductive life-course stage. Although necessarily speculative with respect to women’s contraceptive motives, the profiles we estimated of women most at risk of birth, of new LARC insertion, and of sterilization following LARC discontinuation were consistent respectively with: (1) using LARC to successfully transition to a birth resulting from an intended pregnancy; (2) using LARC to further postpone a birth when the initial LARC device use was discontinued after a brief time, and when the woman may not yet have been ready to have a next birth or to make a decision on ceasing further childbearing; and (3) using LARC to transition to a permanent method of preventing further childbearing after first deferring that decision. Taken together, our results suggest that women not only use LARC to reduce the likelihood of an unintended pregnancy, but also to better time their first or next birth, or as part of a trajectory towards definitively completing childbearing. Another way of putting this is that we find substantial evidence of LARC use for family planning purposes, and not only for immediate pregnancy prevention.

Looking ahead, the composition of women who will start and then subsequently discontinue use of LARC while still of reproductive ages will clearly affect LARC-using women’s future fertility, and their future intended fertility in particular. This composition may change as LARC use in the U.S. continues to increase (Foster et al. 2015). If future increases in LARC use continue among nulliparous women, as is suggested in recent analyses by Kavanaugh and Jerman (2018), this will imply greater use of LARC to optimize timing of births, thereby potentially both decreasing unintended fertility and increasing intended fertility. In contrast, if future increases in LARC use were to be concentrated among higher parity women—for example, by LARC acting as a substitute to, or a precursor to use of permanent sterilization for women who plan no future births—this would likely mean that effects on intended fertility will be limited primarily to that resulting from women changing their childbearing plans. However, the effects of LARC on intended fertility could still be sizeable given the high current incidence of sterilization regret (Eeckhaut et al. 2018; Grimes and Mishell 2008).

Our analyses advance understanding of how continued increases in LARC use may shape U.S. (intended) fertility by helping women to intentionally time the arrival of their first or next birth, or to prevent a next birth. These findings build on previous research showing LARCs’ ability to reduce contraceptive failure and unintended fertility (e.g., Winner et al. 2012), and are essential to improving our understanding of the potential broader impact of increasing LARC use on U.S. women’s fertility and reproductive life courses. Considerable efforts have been undertaken in recent years to increase access to LARCs, including efforts to improve LARC supply, provider training, reimbursement, practice guidelines, and product labeling (Biggs et al. 2014; CMS 2016; Pace et al. 2016; Society of Family Planning Guideline 20092 2010). Such efforts may advance women’s reproductive autonomy by reducing common barriers to LARC use. Further research that informs about the different purposes for which LARC is used by women to shape their reproductive life courses towards achieving their family-planning goals will also be helpful in informing the above-mentioned efforts by providers and policymakers towards ways that align well with women’s contraceptive and reproductive decision-making.

APPENDIX A: Data and Coding

Data for this study were drawn from the 2006–2010, 2011–2013, 2013–2015, and 2015–2017 cycles of the National Survey of Family Growth (NSFG), and from the 1997 cohort of the National Longitudinal Survey of Youth (NLSY97) observed through 2017.

The NSFG is a series of cross-sectional surveys of fertility and contraceptive behavior, designed and administered by the National Center for Health Statistics (NCHS), that has been conducted since 1973. The NSFG data are representative of the U.S. non-institutionalized population ages 15–44 when properly weighted,¹ and include oversamples of teens, Blacks, and Hispanics. For the 2006–2010, 2011–2013, 2013–2015 and 2015–2017 surveys, a total of 12,279, 5,601, 5,699, and 5,554 women were interviewed, respectively, resulting in response rates of 78%, 73%, 71%, and 67% (U.S. Department of Health and Human Services 2018).

The NLSY97 is a nationally representative probability sample of 8,984 youths aged 12 to 16 years on December 31, 1996, including 4,385 females (Bureau of Labor Statistics nd). The data are representative of people living in the United States during the initial survey round and born between January 1, 1980, and December 31, 1984, and include oversamples of Blacks and Hispanics. Participants in the NLSY97 were followed annually through 2011 and biennially thereafter. We use data from the annual waves from 1997 through 2011 and from the biennial 2013, 2015, and 2017 waves. We use LARC discontinuations from 2002 onward to best match the timeline of LARC discontinuers in the NSFG. The baseline wave response rate was 90.7%. Of the original wave 1 (1997) sample, 75% were again interviewed in 2017 (Bureau of Labor Statistics nd).

SAMPLES AND CODING

Our analytic samples were limited to women who reported having used and then discontinued LARC. We focus on births, which may differ from pregnancies due to miscarriages and abortions, because the latter are poorly reported in surveys (Lindberg and Scott 2018). This latter concern also motivates our decision to focus on the intended fertility rate, whose denominator is all women who discontinued LARC use. In contrast, the unintended fertility rate and the intended fertility ratio measures both include unintended pregnancies in either their numerator or denominator. Recent estimates indicate that as many as two fifths of unintended pregnancies result in abortions (Finer and Zolna 2016), and our survey data will miss most of these.

To estimate the annual hazards of experiencing any of the three spell-ending outcomes (birth; new LARC insertion; and sterilization), observations of LARC discontinuers in the NSFG and NLSY97 were pooled (see Table 2 and Appendix Table A1). For the NSFG, we included only those women who reported having discontinued LARC during the contraceptive calendar period. This calendar measures monthly contraceptive use (up to four methods each month) for a minimum of three, and a maximum of four years (depending on month of interview) prior to the survey interview. We considered women to have discontinued LARC use if they reported using an IUD or implant in a given month of the contraceptive calendar, but not in the following month. For the NLSY97, LARC discontinuation can be coded from contraceptive questions asked from the 2001 wave onwards. Our coding was based on the answer to the following question, asked in 2001 and in every subsequent wave: “Still thinking about birth control you used within the past 12 months, which one of these methods did you or your partner use most often…?” The NLSY97 provides several options as a response, including IUDs and implants.² Women were considered as having discontinued LARC use if they reported using an IUD or implant in a given survey year, but not in the following survey year. We relied on annual data through 2011 and, following the switch to a biennial survey calendar after 2011, we applied rules to impute contraceptive method use in the even numbered “off” years 2012, 2014, and 2016. First, for all survey years, if a removal is reported on the same year as a birth, we assume that the removal precedes a birth. Similarly, if we see an insertion on the same year as a birth, we assume the insertion follows the birth. These assumptions are carried through to the post-2011 imputation. For example, if a woman is not using LARC in 2011, is using LARC in 2013, and experienced a birth in 2013, then we assume that the 2013 LARC use started after experiencing the 2013 birth. Therefore, assuming non-use of LARC in 2012. Second, and along those same lines, if a woman is using LARC in 2011 and using LARC in 2013, but has a birth in 2013 then we assume that she was not using LARC in 2012 and thus experienced a LARC removal in 2011. The rationale here is that the very high contraceptive effectiveness of LARC makes it highly unlikely that she was still using a LARC in 2012, and that any LARC use in 2013 would then be a new insertion. Likewise, we checked that if a woman is using LARC in two consecutive odd numbered years, but has a birth in the most recent odd numbered year then we would assume that she was not using LARC in the previous even year and thus experienced a LARC removal in the first of these odd numbered years. Finally, if a respondent had a removal in 2011 or before (and is therefore not using LARC in 2011) and they are not using a LARC in 2013, then we assume they are still not using a LARC in 2012. Similarly, if a respondent is not using a LARC in the odd years post-2011, then we assume no LARC use in the even numbered year prior, thus continuing the exposure period. In both our NSFG and NLSY97 sample, some LARC discontinuing women, particularly those with short durations of LARC use, may be missing, though likely for different reasons. For the NSFG, the long recall period of the contraceptive calendar (i.e., up to 3–4 years prior to survey interview) may mean that some women fail to report earlier episodes of LARC use. For the NLSY97, some LARC use may go unreported because of the limited level of detail of the single contraceptive use question or the switch to a biennial survey calendar after 2011.

NEW LARC INSERTIONS AND STERILIZATION

Both the NSFG and NLSY97 allow for the coding of contraceptive method in the years observed after LARC discontinuation. We use contraceptive information in the years following LARC discontinuation to code a new LARC insertion and either female or male sterilization³ as spell-ending events. To increase comparability with the yearly reporting of contraceptive use in the NLSY97, in the NSFG, women who reported reliance on sterilization in the calendar year of LARC discontinuation were considered as starting sterilization in year 1 (rather than year 0). Women reporting having a new LARC inserted in the year of LARC discontinuation (i.e., in year 0) were omitted from the analysis, as we would not be able to detect these short LARC discontinuations based on the yearly reporting of contraceptive use in the NLSY97.

We are also interested in whether the woman switched to a non-LARC contraceptive method following LARC discontinuation. We code this in both the NSFG and NLSY97 using information on contraceptive method use in the year before a birth. Specifically, we code whether the woman used a non-LARC contraceptive method versus no contraceptive method in the year preceding the year of a birth that followed LARC discontinuation. We rely on this broad categorization of post-LARC contraceptive use because non-LARC method use is often transient with the potential for (multiple) switches between methods. If a respondent had a birth in the year of or the year after LARC discontinuation (i.e., “0–1 years”), they were coded as not having used any non-LARC contraception (i.e., value “no method”) in the year before the live birth. In the NLSY97, if a respondent had a birth in 2013, 2015, or 2017, we considered the method used two years prior to the year of the birth, instead of one year, as we do not know the birth control method in the “off” years 2012, 2014, or 2016.

APPENDIX B: Simulation of Cumulative Births, Intended Births, new LARC Insertions, and Sterilizations following LARC Discontinuation

OVERVIEW

We describe here the structure of, and estimated parameter values for, our simulation model that produces the results presented in Table 3 described in the main text. These quantities are all derived for exposure of up to nine years after discontinuing LARC use. The years are enumerated from year 0 to 8. Year 0 is the year in which the woman last used LARC, otherwise described as the year in which LARC was discontinued. Because this discontinuation may have been early in the year, we allow for a positive probability that she will give birth in year 0. Year 8 is the farthest year out from LARC discontinuation because of the limited observed data beyond this duration. This also means that no spells to women discontinuing LARC by age 34 will extend beyond her reproductive years. Competing with the birth event in these nine years are the event of new LARC insertion and the event of sterilization. We derive the annual and cumulative rates of giving birth, of new LARC insertion, and of sterilization in these nine years after discontinuing LARC use.

Although not all women will have experienced one of the three reproductive events within nine years, we estimate from our simulation that this will account for only 7% of LARC-discontinuing women. Discontinuing LARC use may in the simulation model result in a birth, a birth resulting from an intended pregnancy, a new LARC insertion, sterilization, or in no event. We derive annual and cumulative (intended) fertility rates, and annual and cumulative rates of new LARC insertion and of sterilization.

Data on LARC discontinuers after the first 3 years are relatively sparse, both because the NSFG’s contraceptive calendar only asks about contraception in the year of the survey and the 3 years before that, and because relatively few women in the population have been using LARCs until recent years, resulting in relatively small numbers who have both used and then discontinued use of LARC devices for more than a few years. Fortunately, we find that most reproductive events following LARC discontinuation, including births, occur soon after that discontinuation. We handle the relative sparseness of data more than a few years after LARC discontinuation by assuming constant probabilities after that. Specifically, from Year 4 onwards, we assume constant hazards of birth, new LARC insertion, and sterilization (see Appendix Table A1), and from Year 3 onwards, we assume constant probabilities of using a non-LARC method in the year before the birth, and constant probabilities of pregnancy intendedness of births (see Appendix Table A2).

MODEL STRUCTURE

Our simulation of (intended) fertility after LARC discontinuation, taking into account the competing events of new LARC insertion and sterilization, has two components:

1. The unconditional probability of a birth b, a new LARC insertion r, and sterilization s at each annual duration x since the year that LARC was discontinued. These probabilities are unconditional in the sense that they are among all women who discontinued LARC use. They are generated from a triple-decrement life table in which birth b, new LARC insertion r, and sterilization s, are competing decrements. Following life table notation, we denote by d_b(x) the probability of birth x years after the year that LARC was discontinued among all women who discontinued LARC use. We denote respectively by d_r(x) and by d_s(x) the probabilities of new LARC insertion and sterilization x years after the year that LARC was discontinued among all women who discontinued LARC use. The sequence of unconditional probabilities {d_b(x), d_r(x), d_s(x)} is generated from a sequence of annual competing hazard probabilities. Following standard life table notation again, these are denoted by {q_b(x), q_r(x), q_s(x)}. The cumulative probability of giving birth, $\text{Pr}\left(\text{B}\right)={\sum }_{x=0}^{x=8}\left[d_b\left(x\right)\right]$, of new LARC insertion, $\text{Pr}\left(\text{R}\right)={\sum }_{x=0}^{x=8}\left[d_r\left(x\right)\right]$, and of sterilization $\text{Pr}\left(\text{S}\right)={\sum }_{x=0}^{x=8}\left[d_s\left(x\right)\right]$ nine years after discontinuing LARC are then simple sums.

2. The intended probability of birth at duration x years since LARC discontinuation. Denoting by Pr(B_i(x)|B(x) = 1) the conditional probability that a birth at duration x resulted from an intended pregnancy, we model this probability as depending on both duration and the fraction of LARC-discontinuing women switching to another, non-LARC method of contraception before giving birth at duration x. We expect there to be a higher probability that a birth resulted from an intended pregnancy if the woman uses no contraceptive method after LARC discontinuation than if she switches to another contraceptive method between LARC discontinuation and the birth. Denoting births that occur after no method by c = 0, and after a non-LARC method by c = 1, we first estimate Pr(B_i(x, c)), the intendedness probability at duration x conditional on whether she switched to another, non-LARC method before the birth. These probabilities are then weighted by the probability that a birth at duration x occurs after no method versus after a non-LARC method, Pr(c|x, B = 1) We thereby obtain the intendedness probability for all births at duration x years since LARC discontinuation:

   $$\text{Pr}\left(B_i\left(x\right)|B\left(x\right)=1\right)={\sum }_{c=0}^1\left[\text{Pr}\left(c|x,B=1\right)\text{Pr}\left(B_i\left(x,c\right)\right)\right].$$

Putting together the two components to compute intended fertility rates

The (unconditional) probability that a woman discontinuing LARC will then go on to have a birth that resulted from an intended pregnancy x years later is given by the product of the above two components:

$$d_{b,i}\left(x\right)=d_b\left(x\right)\text{Pr}\left(B_i\left(x\right)|B\left(x\right)=1\right)$$

The intended fertility rate within nine years of LARC discontinuation is for all women who discontinue LARC, f_i, and is therefore given by:

$$f_i={\sum }_{x=0}^{x=8}\left[d_{b,i}\left(x\right)\right]$$ (A1)

MODEL PARAMETER ESTIMATION

The quantities that need to be estimated from data in the above equations are the annual hazard probabilities q_b(x), q_r(x), and q_s(x), the probability that a birth at duration x occurs after no method versus after a non-LARC method is used in the year before the birth, Pr(c|x, B = 1), and the probability that a birth is intended conditional on both duration x and whether no method versus a non-LARC method is used in the year before the birth Pr(B_i(x, c)). In this section we describe the data used to estimate each of these quantities, and present in tables the estimates themselves.

Neither the NSFG nor NLSY97 alone is sufficient to estimate the parameters of this model. Only in the NSFG is birth intendedness asked. Only in the NLSY97, however, are there sufficient longitudinal data to analyze exposure to births, together with any permanent or non-permanent contraceptive method use, for more than two or three years following LARC discontinuation. Moreover, given the low prevalence of LARC use in the U.S. until very recently, in both surveys there are relatively few observed LARC discontinuations.

ESTIMATION OF THE ANNUAL COMPETING HAZARD PARAMETERS OF THE TRIPLE-DECREMENT LIFE TABLE

The set of parameters needed to estimate the triple-decrement life table are the three annual hazards, {q_b(x), q_r(x), q_s(x)}. They can be estimated either from the NSFG or NLSY97 or from observations pooled across the NSFG and NLSY97. In the simulation, we include only those NSFG sample members who discontinued LARC use at ages 20 to 34, to match the NLSY97 age distribution of women discontinuing LARC use. Pooled estimation effectively averages estimates across the two data sources, thus disregarding variation between them. Model-fit statistics are used to assess the appropriateness of pooling observations across the NLSY97 and NSFG in the same way as for Table 2. Those model-fit statistics are presented and discussed below in Appendix Table A3, and show that this approach is appropriate also when including duration since LARC discontinuation only (in Appendix Table A1). Pooled estimation is optimal because of the increases in sample sizes thereby obtained at short durations after LARC discontinuation and because in the NLSY97 the period over which exposure to a birth following LARC discontinuation is measured is longer (in the NSFG it is no more than three years).

The pooled and separate NSFG and NLSY97 estimates are presented in Appendix Table A1. Note that the samples of women for these estimates are exactly those presented in Table 1, hence the match of sample sizes at Year 0 between Appendix Table A1 and Table 1 (N=514 NLSY97 and N=550 NSFG ages 20–34).

We first discuss results for women ages 20 to 34 using the NLSY97+NSFG pooled. We indicate results of tests for statistically-significant differences between the hazards as estimated from the two different data sources for those similar age ranges. In the pooled estimates, the hazard of birth is seen to be highest at durations 1, 2, and 3 years following LARC discontinuation, at 0.238, 0.314, and 0.239 respectively. Thereafter, the annual birth hazard is 0.109, at each year 4 to 8. The probability of new LARC insertion at durations 1, 2, 3, and 4 to 8 are respectively .059, .060, .102, and .053. For sterilization at durations 1, 2, 3, and 4 to 8, the probabilities are respectively .073, .026, .026, and .043. Thus the birth hazard exceeds the combined hazard of new LARC insertion and sterilization at every duration, though only by a large amount in years 1, 2, and 3.

The estimates separately for the NLSY97 and NSFG can be compared at durations 0, 1, and 2, and show that the pooled estimates average over different patterns by duration at years 1 and 2 in the two data sources, although these differences attain statistical significance between the two data sources only in year 1. The pooled birth hazard that is somewhat higher at duration year 2 than at year 1 (.314 versus .238) is produced by an average of a much higher birth hazard at year 2 than at year 1 in the NLSY97 (.329 versus .173), offset by a slightly lower birth hazard at year 2 than at year 1 in the NSFG (.263 versus .330). New LARC insertion probabilities are similarly higher in year 2 than in year 1 in the NLSY97 (.065 versus .039) but lower in year 2 than in year 1 in the NSFG (.044 versus .086). Probabilities of sterilization by duration are very similar across the NLSY97 and NSFG, being highest at Year 1 (.070 and .077 respectively) in both data sources.

USE OF NO METHOD OR SWITCHING TO A NON-LARC CONTRACEPTIVE METHOD AFTER LARC DISCONTINUATION

We are only able to estimate intendedness of a pregnancy resulting in a birth from NSFG data (see below), as only in the NSFG is the intendedness of the birth observed. However, we are able to incorporate NLSY97 data indirectly into the estimation of the intendedness probability by including two predictor variables generated from both the NSFG and NLSY97 data that we know from the NSFG are strongly associated with the intendedness of a pregnancy resulting in a birth. These are duration since LARC discontinuation, x, and switching to another contraceptive method, c, after LARC discontinuation. The probability of a birth being from an intended pregnancy is first expressed as the product of two components: Pr(c|x, B = 1), which is estimated from both NSFG and NLSY97 women who have discontinued LARC and subsequently had a live birth; and the intendedness probability of births by duration x and by whether the birth occurred after no method or after a non-LARC method, Pr(B_i(x, c)), which is estimated from the NSFG alone.

We code switching to a non-LARC method in both the NSFG and NLSY97 using information on contraceptive method use in the year before a birth. Because non-LARC method use is often transient with the potential for (multiple) switches between methods, we rely on a broad categorization of post-LARC contraceptive use that measures only whether the woman used any non-LARC contraception in the year before birth. Specifically, we code whether the woman used a non-LARC contraceptive method versus no contraceptive method in the year preceding the year of a birth that followed LARC discontinuation. Because we rely on the “ever use” question of the NSFG for periods of LARC discontinuation of more than three years (see below), we are not able to code a more complete measure of contraceptive switching following LARC discontinuation.

For births at durations 0 and 1, the Pr(c|x, B = 1) component is assigned to “no contraceptive method” (c = 0) for all women. This is necessary because of the annual question about contraceptive method in the NLSY97. For births at duration 0, we therefore do not know if a reported non-LARC contraceptive method was used before or after the birth. For births at duration 1, we do not know if any reported non-LARC contraceptive method was used in the previous year, because the previous year is that in which LARC use was last reported as the main contraceptive method. For births at duration 2, Pr(c|x = 2, B = 1), we are able to observe any non-LARC contraceptive method used one year before in both the NSFG and the NLSY97, although there are some differences in the operational definitions. In the NSFG, c = 1 is defined for any use of a non-LARC method in the year before the year of the birth at duration 2 years. In the NLSY, c = 1 is defined by the main contraceptive method used in the year before the birth at duration 2 years. Consequently, the estimated probability of a non-LARC method in the year before the birth is higher in the NSFG than in the NLSY97 (results not shown). We effectively average over the two in using a pooled-survey estimation. For the years 2012, 2014, and 2016 in the NLSY, when it had moved to biennial surveying, no contraceptive method is asked. The contraceptive method before the birth (none versus non-LARC) is then assigned to be the method used two years before the year of the birth. For durations 3 and above, we estimate Pr(c|x, B = 1) from an NLSY-only sample. For these durations, we assume a constant probability of no contraception and non-LARC contraceptive method. This constant probability assumption is needed because the conditional probability of intendedness is estimated from an NSFG sample (see below) that does not identify duration year above three years post-LARC-discontinuation.

We show in Panel A of Table A2 our estimates of the probability of no method and the probability of switching to another contraceptive method between discontinuing LARC and giving birth, Pr(c|x, B = 1) As noted above, in most cases this probability is approximated by the contraceptive method, if any, used in the year before the year of giving birth following LARC discontinuation. Also as noted above, only for years of birth at durations 2 and above do we allow for a positive probability of having switched to a non-LARC contraceptive method. At duration 2, this probability of use of a non-LARC contraceptive method in the year after the year of last LARC use is estimated by pooling across the NLSY97 and NSFG samples. We find that 68.2% of births occurring two years after the year of last LARC use followed a year in which no contraception was used, with the other 31.8% of births occurring two years after the year of last LARC use following a year in which a non-LARC contraceptive method was used. We find that 45.0% of births occurring three or more years after the year of last LARC use at durations 3 to 8 followed a year in which no contraception was used, and the other 55.0% of births followed a year in which a non-LARC contraceptive method was used.

INTENDEDNESS PROBABILITY BY DURATION SINCE LARC DISCONTINUATION AND WHETHER THE WOMAN SWITCHED TO ANOTHER CONTRACEPTIVE METHOD AFTER DISCONTINUATION

Estimates of the probability that a birth following LARC discontinuation is from an intended pregnancy come from NSFG data only, as only in the NSFG is the intendedness of the birth observed. We use a sample of women who had a live birth in the year of or the year before survey interview after having discontinued LARC use, where the fact of their having discontinued LARC use was either determined based on the NSFG’s contraceptive calendar (for women who discontinued LARC 0, 1, or 2 years before the year of the birth), or determined based on the NSFG question on whether the woman has “ever used” LARC (for longer durations since LARC discontinuation).

The NSFG’s contraceptive calendar allows a short period of post-LARC-discontinuation exposure to a birth (i.e., a maximum of two years). Most births (167 out of our sample of 267 births overall), however, occur within this period (years 0, 1, 2). For births that occur in years 3 to 8, we use the “ever use” questions for IUD or implants. Those questions unfortunately do not specify when the woman started and stopped LARC use. We therefore only know (from her responses for the contraceptive calendar) that she discontinued LARC use 3 or more years before the birth (which occurred always in the year of or the year before survey interview survey). We do not know from this question when she last used LARC, nor what intervening events, including any birth, occurred between her last LARC use and the year of the recent birth (for which intendedness is reported). To reduce the likelihood that any of these births occurred following another birth that occurred after LARC discontinuation, we exclude the births of women whose LARC discontinuation was identified by the “ever used LARC” question and who had a prior birth within the contraceptive calendar period.

Estimates of the probability that a birth resulted from an intended pregnancy as a function of both duration (number of years x) since LARC discontinuation and whether in the year before the year of the birth the woman used another (non-LARC) form of contraception, Pr(B_i(x, c)), are presented in Panel B of Table A2. We find that births with the highest probability of resulting from an intended pregnancy are those occurring within two years of discontinuing LARC and after using no contraception in the year before the year of the birth. Births in the year of or after LARC discontinuation have a 0.784 probability of resulting from an intended pregnancy. Births two years after the year of LARC discontinuation and following no contraceptive method use in the previous year have a 0.822 probability of resulting from an intended pregnancy. Births occurring three or more years after the year of LARC discontinuation and following no contraceptive method use in the previous year have a 0.775 probability of resulting from an intended pregnancy. Births occurring two years after the year of LARC discontinuation and following a year in which a non-LARC contraceptive method was used have a 0.658 probability of resulting from an intended pregnancy. Births occurring three or more years after the year of LARC discontinuation and following a year in which a non-LARC contraceptive method was used have the lowest probability of resulting from an intended pregnancy, at 0.471.

MODEL-FIT ASSESSMENTS FOR POOLING NLSY97 AND NSFG DATA TO ESTIMATE THE COMPETING HAZARD OF BIRTH AND OTHER REPRODUCTIVE EVENTS

For both main text Table 2 and Appendix Table A1, we present estimates of the competing hazard of birth, new LARC insertion, and sterilization for pooled NLSY97 and NSFG data. The theoretical statistical justification for this pooled estimation is that the two samples are drawn probabilistically from the same universe, being women who discontinued LARC use in any year between 2002 and 2017 when aged between 20 and 34, and therefore should generalize to the same population. The statistical rationale for pooled estimation is that sample size is doubled by the pooling of the NLSY97 and NSFG samples, and that especially at the durations in the first 2 years following the year of LARC discontinuation, where the hazards of birth are especially high, the gains in precision of the estimates will therefore be high. For broader consideration of the often large statistical gains to combined-survey estimation, see Ridder and Moffitt (2007).

Because our estimation combines observations from two nationally representative surveys of approximately the same ages and years, we begin by assuming that they sample from a common social process except for a potential difference in levels of the outcome variable (the hazard of birth, new LARC insertion, and sterilization). We test the validity of this assumption by conducting diagnostics under a model-fitting framework, following Rendall et al. (2013). This procedure consists of including as a regressor “survey” (the woman’s being observed in the NSFG rather than in the NLSY97), and then “survey” and its interaction with all covariates, and assessing change in AIC and BIC model-fit statistics with the addition first of the “survey” intercept shifter and second with the addition of “survey” interactions with the covariates. Only variables common to both surveys are included in this model-fit test. As recommended by Rendall et al. (2013, p. 498, following Weakliem 2004), we emphasize the BIC over the AIC as the preferred model-fit statistic. A finding of model-fit improvement when adding a “survey” intercept shift variable for overall hazard of birth, new LARC insertion, and sterilization differences between the surveys would not be problematic. However, a finding of model-fit improvement when adding a full set of covariate interactions with “survey” would be evidence against the assumption that the surveys sample from a common social process, calling into question the appropriateness of a pooled-survey method. This model-fit assessment needs to be conducted separately for the two regression specifications, respectively in Table 2 and (with fewer regressors) in Appendix Table A1. As we show immediately below, we find no model-fit improvement using the BIC either when adding an intercept shift variable for survey or when adding survey by covariate interactions. Our conclusion is therefore that pooled estimation is appropriate both when including duration since LARC discontinuation only (Appendix Table A1) and when including duration since LARC discontinuation and socio-demographic regressors (Table 2).

Appendix Table A3 summarizes the AIC and BIC statistics for the three versions of the model (no ‘survey’ regressor, ‘survey’ main effect regressor only, and ‘survey’-by-covariate interactions) in each specification. Lower values of the fit statistics indicate a better-fitting model, after penalizing the addition of parameters for ‘survey’ and covariate interactions with ‘survey’ (Weakliem 2004). We would expect pooled survey analysis to be aided by the inclusion of more regressors that would account for structural and realized differences in population representation between the two surveys. These differences are discussed in the main text, in the context of distributions of LARC-discontinuing women presented in Table 1, such as the larger fraction of Hispanic women represented in the NSFG and the larger fraction of women discontinuing LARC after age 30 in the NLSY97. As expected, the BIC statistics are indeed more clearly indicative of the pooled-survey model without including ‘survey’ main effect or ‘survey’-by-covariate interactions being the best-fitting model for the Table 2 specification (see Panel A). The BIC fit statistic is 3,555.8 for that model, and increases to 3,561.3 when including the NSFG ‘survey’ intercept, and increases still further to 3,798.7 when additionally including the ‘survey’-by-covariate interactions. For the simpler model specification of Appendix Table A1, in which duration is the only regressor, the BIC fit statistic is 2,666.3 for the model without ‘survey’ main effect or ‘survey’-by-covariate interactions, and does not improve when including the NSFG ‘survey’ intercept (2,666.4), and gets substantially worse (2,675.4) when additionally including the ‘survey’-by-covariate interactions (see Panel B). Thus our conclusion from conducting this model-fit assessment is that pooled estimation is appropriate for the competing hazard regressions both with and without socio-demographic predictor variables.

Appendix Table A1:

Annual hazard of experiencing a live birth, new LARC insertion, or sterilization among women who discontinued LARC, 2002–2017, ages 20–34.

Year of follow-up	N at riska	No event	Live birth	New LARC insertion	Sterilization	p diff. c
	NLSY97
Year 0	514	99.2	0.8	0.0	0.0	NA
Year 1	502	71.7	17.3	3.9	7.0	NA
Year 2	291	58.4	32.9	6.5	2.2	NA
Year 3	174	63.3	23.9	10.2	2.6	NA
Years 4–8	222	79.6	10.9	5.3	4.3	NA
	NSFG (women ages 20–34)
Year 0	550	97.6	2.4	0.0b	0.0b	NA
Year 1	366	50.7	33.0	8.6	7.7b	NA
Year 2	96	65.1	26.3	4.4	4.2	NA
	Pooled NSFG (women ages 20–34) and NLSY97
Year 0	1,064	98.4	1.6	0.0b	0.0b	.080
Year 1	868	63.0	23.8	5.9	7.3b	<.001
Year 2	387	59.9	31.4	6.0	2.6	.596
Year 3	174	63.3	23.9	10.2	2.6	NA
Years 4–8	222	79.6	10.9	5.3	4.3	NA

Notes: All percentages in this table are weighted.

^aWomen who experienced an event (i.e., live birth, new LARC insertion, or sterilization) in a previous follow-up period were excluded from the risk set but, for the NLSY97 women only, could reenter the risk set (i.e., be included in N at risk in year 0) after experiencing another LARC discontinuation.

^bTo increase comparability with the yearly reporting of contraceptive use in the NLSY97, in the NSFG, women who reported reliance on sterilization in the calendar year of LARC discontinuation were considered as starting sterilization in year 1 (rather than year 0). Women reporting new LARC insertion in the year of LARC discontinuation (i.e., in year 0) were omitted from the analysis, as we would not be able to detect these short LARC discontinuations based on the yearly reporting of contraceptive use in the NLSY97. To increase comparability with the ages of women in the NLSY97, we only included NSFG women ages 20–34 at LARC discontinuation. The age distribution of LARC discontinuation in the NLSY97 data is as shown in main text Table 1, including small numbers of women discontinuing LARC use before age 20 and after age 34.

^cp_diff. indicates results of F-test of null hypothesis that there is no difference between the NSFG and NLSY97 estimates for that year.

Sources: 2006–2010, 2011–2013, 2013–2015 and 2015–2017 NSFG and NLSY97 2002–2017.

Appendix Table A2:

Intended Fertility Simulation Model Parameters.

A. Proportion Using No Contraceptive Method versus a non-LARC Contraceptive Method in the Year Before the Birth for Births by Years Since LARC Discontinuation
Years since LARC discontinuation	No method (t-1)	non-LARC method (t-1)
0 or 1	1	0
2	.682	.318
3 to 8	.450	.550
Sample n (Years 2 to 8)		196

B. Proportion Intended for Births by Years Since LARC Discontinuation and Whether No Method or a non-LARC contraceptive Method was Used in the Year Before the Birth
Years since LARC discontinuation	No method (t-1)	non-LARC method (t-1)
0 or 1	.784	n.a.
2	.822	.658
3 to 8	.775	.471
Sample n	175	101

Sources: Panel A, National Longitudinal Survey of Youth 1997 Cohort, and National Survey of Family Growth, 2006–2010, 2011–2013, 2013–2015, and 2015–2017 (NSFG); Panel B, NSFG, 2006–2010, 2011–2013, 2013–2015, and 2015–2017 only.

Appendix Table A3:

Model-fit statistics.

A. (For Table 2)
Pooled NLSY97 and NSFG Model	No NSFG intercept or regressor interaction	NSFG intercept, no regressor interaction	NSFG intercept and regressor interaction
Fit statistics
AIC	3177.2	3166.5	3090.7*
BIC	3555.8*	3561.3	3798.7
* best fitting model (lower = better fit)
Sources: NLSY97 and 2006–10, 2011–13, 2013–15, and 2015–17 NSFG.
Notes: “NSFG intercept” indicates that the pooled-survey (NLSY97 and NSFG) regression model specification includes a dummy variable for the observation’s coming from the NSFG sample. “NSFG intercept and regression interaction” indicates that the pooled-survey (NLSY97 and NSFG) regression model specification includes a dummy variable for the observation’s coming from the NSFG sample plus an interaction variable for NSFG*<regressor> for each of the regressors in the model, with the exception of ‘time since LARC discontinuation’ (because there are no NSFG cases beyond duration years 1 and 2).

B. (For Appendix Table A1)
Pooled NLSY97 and NSFG Model	No NSFG intercept or regressor interaction	NSFG intercept, no regressor interaction	NSFG intercept and regressor interaction
Fit statistics
AIC	2620.3	2597.4	2583.4*
BIC	2666.3*	2666.4	2675.4
* best fitting model (lower = better fit)
Sources: NLSY97 and 2006–10, 2011–13, 2013–15, and 2015–17 NSFG.
Notes: “NSFG intercept” indicates that the pooled-survey (NLSY97 and NSFG) regression model specification includes a dummy variable for the observation’s coming from the NSFG sample. “NSFG intercept and regression interaction” indicates that the pooled-survey (NLSY97 and NSFG) regression model specification includes a dummy variable for the observation’s coming from the NSFG sample plus an interaction variable for NSFG*Year of follow-up in the model.