This study consisted of five stages: (1) identifying the universe drawn from multiple platforms (2) removing sellers advertising on more than one site (3) draw a roughly probability proportional to size sample (4) contacting sellers as mystery clients through text/chat and (5) purchasing drugs offered by the seller. We provide below the protocol we followed for each of these steps.
Patient and Public Involvement: Patients and the public were not explicitly involved, although as abortion is not a visible experience, it’s possible that members of the study team had previously bought medication abortion drugs online and used them to carry out a pregnancy termination.
Identify The Universe Drawn from Multiple Platforms
To construct our universe, it was necessary to identify as many online misoprostol sellers as we could find. We identified sellers through three platforms: 1) websites, identified through searching on Google; 2) Instagram profiles; and 3) online marketplaces (sites similar to Amazon). We included the following marketplace sites because they are the most popular e-commerce sites in Indonesia: Tokopedia (www.tokopedia.com), Bukalapak (www.bukalapak.com), Shopee (www.shopee.co.id), Lazada (www.lazada.com), and Blibli (www.blibli.com). This universe is fluid; we identified the universe at one point in time.
Selection of search terms
For conducting our web searches to find sellers, we chose the most popular relevant search terms on Google Trends as of August 5, 2019 (Table 1). We used the word jual (“sell” in Bahasa) in addition to the drug name. We also added phrases lacking specific drug names to our search term list, which were also popular according to Google Trends, and which may be employed by women who do not know the names of abortion drugs. On Instagram, we conducted searches for the terms listed in Table 1 using Instagram’s regular search function.
Identifying sellers through marketplace sites required a slightly different strategy. Marketplace sites use algorithms to detect advertisements for illegal products; we found that many sellers add characters or deliberately misspell words to evade detection, e.g. “Cy-to-tec” or “Cyt0tec.” Each marketplace platform had slightly different sets of commonly used misspellings, which we believe is because each site uses different algorithms to flag profiles that do not comply with site regulations. The data collectors identified and searched for misspellings at their discretion, discussing their process with members of the study team and recording all search terms used.
Listing procedures
Six data collectors conducted the searches after receiving a one-day training. Data collectors searched for each term only once, keeping those results open for the duration of the listing. Marketplace sites were continuously refreshing such that a vendor that came up during a search no longer appeared when the fieldworker tried to list it. All searches were conducted on Google Indonesia (www.google.co.id) using the Chrome browser’s Incognito mode. Each search term yielded thousands of results. We began with the hits that appeared first, as these are the most relevant according to the Google search algorithm. The listing of websites continued until the hits were no longer relevant: either they stopped linking to misoprostol sellers or linked to the same pages as before. The Google search results turned up misoprostol advertisements buried in blog posts on sites such as www.medium.com, www.asus.com, www.weebly.com, and www.kompasiana.com; websites meant for other purposes such as www.goodreads.com; and sales sites, for example those selling colour vision sensors: www.cmucam.org.
During the listing, we excluded the following types of results: pages that were no longer active or had been taken down (i.e. “404-page not found”), sites that did not provide any means of contacting a seller, and irrelevant pages which did not advertise the sale of an abortion drug, such as news stories. The website listing also captured a big drug store chain; big drug store chains are legally allowed to sell online with a prescription. We excluded this site since it was advertising legal sales. The Google search listers also ignored any hits that led to Instagram profiles and online marketplace vendors, which we had another process for identifying.
To conduct searches in Instagram, data collectors logged in to their Instagram accounts in a Google Chrome incognito window. Searches were filtered in order of most recent posting. Most of the accounts that came up in searches were private, meaning that unless the account holder accepted to be “followed” by the buyer, the information of that account holder was not visible. The lister created a fake Instagram profile to follow these sellers; if they accepted her “follow” request by the end of the listing period, they were included in the sample. Some Instagram sellers directed prospective buyers to a website link in their profile; in these cases, we counted the website as part of the Instagram page.
Listing took place over two weeks in August 2019 and resulted in 727 hits: 441 websites, 153 marketplace sellers, and 133 Instagram profiles. For each search result, listers completed a web-based data capture form created using Open Data Kit (ODK). In addition to the information needed to identify the seller such as the URL and contact information, the data form also captured information on the site: specific brand names and types of pills advertised, asking price, the maximum gestational ages for which sellers offered abortion drugs, and any additional information about misoprostol or abortion given on each site. Once completed, each form was uploaded to a secure server hosted by SurveyCTO. In addition to filling out the listing form, data collectors also took screen captures of the landing page, and any testimonials or pictures of pills.
Removing Sellers Advertising on More Than One Site
We identified duplicate listings using the telephone number listed, usually presented as a WhatsApp number. There was high variability in the number of separate sites a seller had. For example, eight sellers had 18 or more separate sites; the biggest seller had 51.
After removing duplicate listings, we identified 281 unique sellers. Some sellers had listings on multiple platforms, and we randomly chose one listing to “represent” each of them for the purpose of drawing a sample. Our sampling frame consisted of 147 websites, 49 Instagram accounts, and 85 marketplace sellers (Table 2).
Drawing A Roughly Probability Proportional To Size Sample
Based on budgetary constraints, we aimed to have a final sample of 100 sellers with whom we were able to make contact. Without having any priori information about how likely it was going to be to establish contact with online vendors, we oversampled by 25%. We aimed to conduct probability sampling proportional to size of online presence. We wanted to be sure to include those sellers who had a larger presence and so we divided sellers into strata by size, each with approximately even measure of size totals as follows: sellers with between 4 and 12 listings; sellers with 2 or 3 listings; and sellers with only one listing. We selected all sellers with greater than 12 listings, 60% of sellers with 4-12 listings, 50% of sellers with 2-3 listings, and 40% of sellers with only one listing. We did not stratify by seller type (web, Instagram, or marketplace) since we do not know the relative market share among the three types, and some sellers were linked to listings across multiple platforms (e.g. both a website and an Instagram account).
This resulted in a sample of 128 sellers: 73 (57%) websites, 19 (15%) Instagram accounts, and 36 (28%) marketplace sellers. By size, 66 sellers (51%) had only one listing, 38 sellers (30%) had 2-3 listings, and the remaining 24 sellers (19%) had 4 or more listings, including the 8 sellers we designated as certainty units because of their exceptionally high number of listings per seller. For comparison, in the sampling frame, 60% of sellers have a single listing, 28% have 2-3 listings, and 12% have 4 or more listings.
In case the sellers were well-connected to one another and our mystery client study was received with suspicion among a broad swath of sellers, we wanted to start contacting the sellers with the largest online presence first. We therefore organized the sample list in descending order by measure of size (number of listings linked to that seller), and instructed the mystery clients to contact the sellers in order of size. When sellers had the same bank account number and quoted the same price, or gave near-identical responses to questions, or engaged in non-simultaneous texting with fieldworkers, we treated them as duplicates and stopped engaging with them (n=12). In response to losing additional sellers from our sample, we selected an additional four sellers to contact, which resulted in a total sample of 132 sellers: 73 websites, 20 Instagram accounts, and 39 marketplace sellers (Table 2).
Contacting sellers as mystery clients through text/chat
The study team selected three of the fieldworkers who had conducted the listing to stay on and pose as mystery clients. We first conducted a pilot test of five sellers. The five sellers we selected were those with only one listing.
Mystery clients contacted all selected sellers using WhatsApp using SIM cards purchased exclusively for use in the study, except for 18 marketplace sellers who we were required to contact through the instant chat window hosted on the marketplace site. If no contact information was available and the seller’s URL no longer existed, that seller was marked as having invalid or missing contact information (n=22).
Mystery clients were able to send initial messages to 110 sellers, of which 16 never responded. If the seller did not respond for 24 hours after the last message, the mystery client should move on to the next assigned seller on their list. After 48 hours with no response, the mystery client marked that seller as unresponsive. Some sellers responded after this period, in which case mystery clients proceeded with the interaction. Mystery clients did not have set “office hours,” texting with sellers during evenings, early mornings, and weekends. Some sellers requested video calls before engaging further; mystery clients refused these requests and in each situation, the seller nevertheless continued to engage with the mystery client. After beginning exchanges with our mystery clients, 2 sellers stopped responding. Mystery clients declined to pursue purchases from four sellers for various reasons: one made threatening and sexually explicit comments which made the mystery client feel unsafe, one asked for a prescription, one quoted prices that were far above the amount expected, and one said they would only sell in-person with cash payment.
Before contacting a seller, mystery clients reviewed the seller’s listing data and website or profile. Some websites gave instructions for what information clients should provide in their first message, such as approximate gestational age and location where they lived. In order for mystery clients to convincingly provide information requested by sellers, we developed profiles, or “scripts” for them to use. All mystery clients were to say that they did not know how far along in their pregnancy they were, only that their last menstrual period was six weeks before the date on which they contacted the seller and that they had not taken a pregnancy test; mystery clients were supposed to ask the sellers how pregnant they were. We had created profiles marital status, age, and number of children, no sellers asked the mystery clients for their age or marital status. A small number asked whether the mystery client had ever given birth previously, but not the number of children they had.
Mystery clients filled in a self-administered structured questionnaire on their interaction with each seller, including information given, price paid, and drugs received. Once completed, these forms were uploaded and stored in SurveyCTO. Mystery clients also took screen shots of their text conversations with sellers, which were stored in a secure folder with the name of the seller’s ID in order to link data from the completed survey form to the logs of the interactions. To ensure all texts would be searchable, mystery clients also selected the text messages on a screen and pasted it into a Word document.
Purchasing drugs offered by the seller
Most sellers offered “packets” of drugs, rather than misoprostol alone. When sellers offered a choice between various packets, we instructed mystery clients to ask for the seller’s recommendation. Mystery clients were instructed to ask what the other pills were for and what they do; they asked the seller to advise them on how many misoprostol pills to get. If any of the indicated packets exceeded the allotted budget of 1,500,000 Indonesian rupiah (about 106 USD) per purchase, the mystery client bought the packet closest to the recommended option that fit the budget. Mystery clients were encouraged to attempt to negotiate the price; they were instructed to ask for a lower price once. Some sellers offered to sell fewer pills for a lower price; in these cases, the mystery clients asked whether this would still cause an abortion and if the seller said yes, then agree on that packet. Mystery clients only ordered using standard shipping, and, once they made the purchase, asked sellers to send the tracking number for the shipment.
Mystery clients reached out to the seller again as soon as the pills arrived, typically texting them something like “What do I do now?” Of the 76 sellers from whom we purchased misoprostol, 12 stopped responding to mystery clients as soon as payment was sent, with some blocking the mystery client’s number. In all, we received drugs from 64 sellers in our sample. We describe the drugs received and information given by sellers in a separate paper.[7]