data
We estimate contraceptive use using quarterly sales data for emergency contraceptives (levonorgestrel and ulipristal acetate) and regular hormonal contraceptives (birth control pills, hormonal IUDs, injections, and patches) from the second quarter of 2014 to the first quarter of 2020 obtained from the IQVIA MIDAS (35) database.footnote 1 IQVIA collects data from wholesalers, retail/hospital pharmacies, or manufacturers and, if necessary, extrapolates that data to ensure it is representative of the entire country. Both Rx and OTC markets are included unless otherwise specified. Non-hormonal contraceptive methods (e.g. copper intrauterine device or condoms) were excluded due to incomplete reporting in many countries, and for copper intrauterine devices it was not possible to distinguish between routine and emergency use. Emergency contraceptive use was measured in IQVIA MIDAS standard units (SU), which reflect the number of counting units (tablets, grams, etc.) divided by the lowest common dose. In the case of conventional contraceptives, the SU of regular contraceptives accounts for a 4-week period of effective contraception, while the SU of hormonal intrauterine devices accounts for the 3-5 years of protection provided by a single device. We addressed this issue by multiplying the SU of a generic contraceptive pill by the number of quarters effectively covered, called quarterly standard units (QSU) (see Table A3 in the SI for conversion details).
With respect to potential demographic and health outcomes, the availability of contraceptives can affect birth rates, abortion rates, and the incidence of various sexually transmitted infections (STIs). Data on the number of births (both annual and monthly) from 1998 to 2021 were collected from Eurostat ( 36 ). Data on the number of abortions per year were taken from WHO, with gaps filled in by data from national offices and the Johnston’s Archive (37, 38). As discussed in the background section, easier access to emergency contraceptives may affect STI incidence rates due to changes in risk-taking behavior. Annual data on STI incidence for HIV, gonorrhea, hepatitis B, and syphilis were obtained primarily from the European Center for Disease Prevention and Control ( 39 ) and supplemented with data from WHO whenever possible ( 39 ).40,41,42,43) and National Statistics Office. Data missing from the previously mentioned sources for Russia, Kazakhstan, and Bosnia and Herzegovina were obtained from the World Bank (44). If there were multiple data sources, two people independently checked for consistency. To reduce bias that may arise from the annual structure of the data, we considered the year following the switch as the first year of active treatment for countries that switched in the second half of the year (e.g., 2010 for Austria, which switched to OTC in December 2009) (see Tables A5 and A6 in the SI for country-specific details).
We correlated data on sales, births, and abortions with the size of the reproductive female population (i.e., women aged 15–49).45,46,47)), the data were derived from Eurostat (48, 49) and the World Bank (44). STI incidence rates represent the total population (incidence rate per 100,000 people) as they directly affect both men and women (see Table A5 in the SI for country-specific details).
OTC transition dates were sourced from the scientific literature or, if unavailable or conflicting, by contacting relevant national authorities directly (see Table A1 in the SI for country-specific details). When analyzing fertility rates, we applied a one-year lag to OTC and Rx transition dates due to the length of pregnancy and the annual data structure.
Our analysis is based on a sample of 32 (European) countries (Austria, Belgium, Bosnia and Herzegovina, Bulgaria, Croatia, Czech Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Iceland, Ireland, Italy, Kazakhstan, Latvia, Lithuania, Luxembourg, Malta, Netherlands, Norway, Poland, Portugal, Russia, Slovakia, Slovenia, Spain, Sweden, Switzerland and United Kingdom). I did it.footnote 2
For sales/usage analysis, data availability was limited to the period from Q2 2014 to Q1 2020, during which only a few countries experienced OTC transitions: Germany, Poland, Italy, and Croatia. For Rx transitions, Poland was the only intervention country compared to 31 control countries with always treated or never treated. However, for demographic and health outcomes, the data included OTC switches for all countries in the dataset. We therefore restricted the analysis to a subset of these countries based on outcome, intervention and statistical approach. The following chapters describe each statistical method along with additional methodological reasons. Table A7 in the SI provides a comprehensive overview of the country samples used in each analysis.
empirical strategy
OTC switch
We used a staggered generalized two-way fixed effects (TWFE) difference-in-differences (DiD) approach as follows.
$$\begin{sort}\:{Y}_{\text{c}\text{t}}=&{{\upbeta\:}}_{0}+{{\upbeta\:}}_{1}{OTC}_{\text{c}\text{t}}\\&+{{\upbeta\:}}_{2,\text{t}}{ \:\text{
(1)
\(\:{Y}_{ct}\) Outcomes (EC/RC use, live birth rate, abortion rate, and STI incidence) are described. \(\:{\gamma\:}_{c}\:\)and \(\:{\theta\:}_{t}\) While constructing country and time fixed effects, OTCConnecticut Indicates a dummy variable for country status. aspirate At the time, it promoted OTC access to emergency contraceptives. tea (Time is defined as quarters for usage data and years for demographic and health outcomes.) Therefore, \(\:{\beta\:}_{1}\) Represents the average difference in outcomes due to OTC conversion. vector \(\:{\:X}_{ct}\) Indicates confused people.
To account for within-correlation error terms in clustered data, we adopted constrained wild cluster bootstrapping using Rademacher distribution for the lagged approach. This is because the literature shows that rejection rates for cases with few clusters are superior to rejection rates derived from large sample theory due to better finite sample properties (50,51,52,53).
Importantly, poor control of (time-varying) confounding can violate causal assumptions such as parallel trends, resulting in biased average treatment effects for the therapist (54, 55). To minimize bias, we assessed potential confounders following the approach of Zeldow and Hatfield (55) . Adjustments for life expectancy, population density, or unemployment were not necessary because they were independent of the intervention and had time-invariant marginal effects on the outcomes. However, we included dummy variables for gross domestic product (GDP) per capita, the proportion of women in the total population, and whether strict abortion laws were in place, and interacted with time because the results showed time-varying marginal effects. For a complete list, including reasons for including or excluding confounding variables, see Table A8, Table A9, and Figure A2 in the SI. See Table A10 in the SI for data sources of confounders.
However, an imbalance in the number of treated and untreated countries can lead to significant bias and loss of power in TWFE DID analyzes (56). To mitigate this, we restricted the control countries in the TWFE DID analysis of sales/usage data.footnote 3 The first “always-on” countries to move emergency contraception to OTC status (Belgium, France, Norway, Portugal, Sweden, and the United Kingdom). These countries are where the potential long-term effects of OTC transition are most likely to be present at the time of analysis, resulting in minimal impact on DiD results. Hungary, Russia, Kazakhstan, and Bosnia and Herzegovina served as no-treatment controls. Results for the full country sample are presented in Table A11 of the SI.
Robustness checks included several approaches. First, we added a country-specific linear time trend (CSLTT) to the TWFE DID to address potential confounds due to the underlying linear trend. Second, we utilized a new estimator from Chaisemartin and D’Haultfoeuille ( 57 , 58 ) to address the ongoing debate in the literature about the bias of the standard TWFE estimator in the presence of dynamic/heterogeneous intervention effects and the adoption of staggered interventions due to negative group weights and undesirable late versus early 2 × 2 comparisons ( 54 , 59 ).footnote 4 Finally, we generated dynamic event plots using a classic TWFE event study in the manner of Schmidheiny and Siegloch (60) and the estimator of Chaisemartin and D’Haultfoeuille (58) to gain better insight into the (parallel) prior trends and dynamics of treatment effects around the intervention period.
Rx switch
Polish Rx switch analysis differs from traditional OTC analysis due to data limitations. First, Poland is the only country receiving treatment. Second, to avoid bias due to the early OTC transition in Poland in 2015, the pre-intervention period should be limited to the time between OTC and Rx transition in Poland (e.g. 2015Q2 to 2017Q2). Although we had access to quarterly usage/sales data and monthly birth data, other demographic and health variables were available annually, providing only two pretreatment observations per country (see Table A4 in the SI). Lacking a single method that addresses both problems simultaneously, we used two separate approaches, each addressing one of the limitations.
First, we used the same TWFE difference-in-differences approach described in Eq. (1). Although this approach can handle a small number of preprocessing periods, it runs the risk of underrejecting the null hypothesis when there is only one treated cluster and many untreated clusters, even when using limited wild cluster bootstrapping or random sampling inference as previously described (56, 61). As a result, we used a reduced sample of countries as in the analysis of OTC transitions in sales data, except that the control intervention resulted in a rotation of “always treated” and “never treated” countries. (See Table A7 in SI for details)
Second, since a composite control approach is intended when there is only one processed unit, we used the “synthetic difference-in-difference” (SDID) estimator ( 62 ). As opposed to classical composite control (63), which does not allow for control over unit fixed effects, and classical difference-in-differences, which is constrained by the underlying parallel trends assumption, the SDID estimator uses units (\(\:{\widehat{\omega\:}}_{c}^{sdid}\)) and time (\(\:{\widehat{{\uplambda\:}}}_{\text{t}}^{\text{s}\text{d}\text{i}\text{d}})\) The weights were adjusted to build a composite control group with optimized parallel chucks. Therefore, the algorithm for calculating ATT (based on two-way fixed effects regression) is:
$$\begin{aligned}{l}\left(\widehat\tau^{sdid},\:\widehat{\mu\:},\widehat{\alpha\:},\widehat\beta\right)&=arg\;\underset{\tau,\;\mu,\;\alpha,\beta}{min}\\&\lef t\{\:\sum\:_{c=1}^N\sum\:_{t=1}^T(Y_{ct}-\:\mu\:-\alpha_c-\beta_t-{Rx}_{ct}\tau )^2\:\widehat{\omega\:}_c^{sdid}\widehat\lambda_t^{sdid}\right\}\end{aligned}$$
(2)
where \(\:{Y}_{ct}\) indicate the result, receptionConnecticutIndicates a dummy variable for country status. aspirate At the time, it promoted Rx access to emergency contraception. tea (Year, quarter or month depending on results) no wayaspirate and raintea Configure unit and time fixed effects. We adjusted SDID for the same time-varying covariates as the previous TWFE DID to avoid bias due to potentially different evolution of control variables between treatment and control groups (64).
The properties of the SDID theoretically allowed the inclusion of all countries in the Rx analysis, except for “always treat” countries, such as Hungary, which never allowed OTC access. However, in practice, some countries had to be excluded due to missing values because the SDID estimator requires strictly balanced panel data.footnote 5 (See Table A7 in SI for details)
We address in our discussion the limitations of this method, as for some demographic and health outcomes there are few pretreatment observations.
