Our first objective was to detect mentions of SN as well as other specific substances and suicide methods. As is common over social media, many lexical variants are used to discuss SN and other substances or methods. Not including commonly used lexical variants leads to low-sensitivity data collection [19]. We took a data-driven approach to identify all the relevant lexical variants for SN. We trained a Word2Vec model using n-grams (n=1, 2, and 3) from the entire data set, which enabled us to automatically identify lexical expressions that are the most semantically similar to SN. We manually reviewed these terms and identified, in addition to alternative names for SN, keywords associated with other suicide mechanisms. In consultation with our subject matter experts, we manually grouped these keywords into 14 categories: “sodium nitrite,” “cyanides,” “firearms,” “hanging,” “acid regulators,” “ricin,” “plant-based poisons,” “antiemetics,” “other preservatives,” “nitric oxide,” “household chemicals,” “barbiturates,” “benzodiazepines,” and “opioids.” The complete list of suicide mechanism–related terms identified is given in Table S1 in Multimedia Appendix 1.

We performed automatic searches over the whole data set to compute the frequencies of posts mentioning each method of suicide over time. A post that mentioned any number of lexical variants associated with a suicide method counted toward that method. For example, a post mentioning only “NaNO₂” is assigned the group “sodium nitrite,” whereas a post mentioning “gun” and “full suspension” is assigned the group labels “firearms” and “hanging.” We computed the monthly and normalized frequencies of posts mentioning each of these categories to analyze their temporal trends.

We compared the temporal trends of SN mentions we discovered from the above analysis with relevant metrics reported in two more traditional sources: (1) intentional exposures to SN in the US National Poison Data System (NPDS) and (2) death counts in the CDC Wide-Ranging Online Data for Epidemiologic Research (WONDER) database. For the first comparison, we compared our data against the quarterly intentional exposures to SN from the NPDS reported by McCann et al [10]. For the second comparison, we compared against the reported deaths under the underlying cause of death codes U03, X60–X84, and Y87.0 and multiple causes of death code T50.6 from the CDC WONDER database. For the latter, to make the comparison better aligned with the traditional data source, we combined our keyword mention counts pertaining to “sodium nitrite” and “other preservatives.” We performed a Spearman rank correlation test to assess possible associations between the pairs of statistics.

We performed a co-occurrence analysis to compute the number of times different suicide methods we already identified were mentioned together. The intuition behind this analysis was that suicide methods that are considered together or substances that are taken together (eg, substances taken alongside SN) are likely to be mentioned more frequently in the same posts.

To obtain further insights about the topics associated with SN chatter, we conducted a topic modeling experiment. Topic modeling is a class of unsupervised algorithms that attempt to identify clusters of lexical elements that belong to latent topics from large sets of texts. Since the process is purely unsupervised, the topic clusters are not known a priori. We applied the BERTopic model, an unsupervised topic modeling approach that automatically clusters content from posts into a mathematically optimized number of topics [20].

Purchasing, sourcing, and procurement of SN were identified as common topics of discussion during the analyses mentioned above (see the Results section), so we used a 2-fold approach to identify information about both the retail venues and the geographic locations where SN was being sought for purchase. Specifically, we leveraged named entity recognition (NER) models trained on large web-based data sets to identify potential sources of SN: locations, organizations, and keywords used to look for SN on online marketplaces. NER is an information extraction technique used to discover named entities (such as organizations, locations, etc) in a textual corpus (Figure 2).

We used an NER algorithm available in the Python programming language (spaCy NER [21]) to detect possible locations or sources from which people seek or obtain SN. Apart from using spaCy’s location entity label, we created a custom entity called “Suicide_Method” that would identify the substance in the text and highlight it. We used the rule-based pattern recognizer in spaCy to detect mentions of SN. In particular, we used the pattern “sn” to detect mentions of SN (the code snippet is shown in Multimedia Appendix 1). Posts where SN was mentioned in conjunction with location names were used for further analysis. The location entity recognizer of spaCy has a reported F₁-score of 0.916 [22].

The locations identified by the NER methods represented both sources of SN and consumer locations, and thus, the process required further disambiguation. We modeled this disambiguation as a supervised classification task and trained a transformer-based location classifier to classify the locations obtained from NER as “consumer location” and “purchase location.” Transformer models leverage large amounts of “pre-trained” language data, which can then be fine-tuned on a specific task—such as location type identification, in the case of this study [23]. Sentence-level annotation of location mentions was carried out by 2 of the authors to create a gold-standard training data set for fine-tuning the model. The annotation process was carried out iteratively, with the annotation guidelines being refined after each round of annotation. Disagreements were resolved after a detailed discussion between 3 of the authors to reach a consensus after each round. Two rounds of annotation were performed, and interannotator agreement was computed based on Cohen κ, revealing good agreement for both source location (κ=0.80) and consumer location (κ=0.84) [24]. A total of 722 samples were annotated, of which 577 (80%) were used for training and the remaining 145 (20%) were used for testing.

From the many transformer-based models that are publicly available, we chose the RoBERTa model [25], which is based on Bidirectional Encoder Representations From Transformers [26] and has been shown to achieve state-of-the-art results on several language processing tasks similar to those by the base model, including for health-related, social media–based text classification tasks [27]. Our use of “state-of-the-art” refers to the best-performing machine or deep learning models currently available for each task, in the rest of the paper. We used the embeddings obtained from RoBERTa to fine-tune our model using the training data. We computed the distributions of the sources and locations identified automatically for analysis. Classification results and all outputs are provided in the Results section.

A total of 1,337,982 posts were collected. Of these, 1,302,620 were posted under the “Discussion” threads; 28,666 under the “Poll” thread; and 6696 under the “Question” threads. Preprocessing and removal of duplicate posts resulted in 1,329,042 total posts. There was a steady increase in the total number of posts from 2018 to 2020, followed by a slight decrease in 2021. The highest number of posts on the website was made in 2020. In the months leading up to September 2022, the final full month of data collection, the total number of monthly posts was generally higher than in the corresponding months in 2021 (Figure S1 in Multimedia Appendix 1).

Figure 3 shows the relative monthly frequencies for posts mentioning SN, as well as other substances and methods associated with suicides. From the figure, it can be observed that SN was the most popular means of suicide discussed in this online community, and the frequency of mentions of SN increased over time. A sharp rise in SN mentions can be seen towards the end of 2019, with the frequency of mentions remaining elevated thereafter.

SN-related deaths due to intentional consumption have increased since 2019 [28]. Our analyses align with this uptick of SN ingestions and suicides: a sharp increase in the mentions of purchases of SN was seen toward the end of 2019 (Figure 3). In our comparison against the NPDS data, we obtained a Spearman ⍴ of 0.866 (P=.001), revealing a statistically significant association between the 2 data sources. On visualizing the normalized frequencies of intentional exposures in the NPDS and purchases made as obtained from our data set (see the Sources and Consumer Locations section), we found that both the noncumulative and cumulative frequencies showed similar trends (Figure 4) for the 10 quarters spanning from 2018 to 2020 for which data from both sources were available (quarters 1 to 4 for 2018 and 2019, and quarters 1 and 2 for 2020). Although the mentions of purchases on Sanctioned Suicide are not exclusive to the United States, the similar trends are a strong indicator of online suicide community content reflecting real-life suicide incidences.

In the comparison against metrics from the CDC WONDER database, we obtained a Spearman ⍴ of 0.727 (P<.001) for month-by-month frequencies during the time period from March 2018 to December 2021 for SN purchases versus actual deaths and a Spearman ⍴ of 0.775 (P<.001) for keyword mentions versus actual deaths, revealing statistically significant correlation in both cases (Figure 5).

The heatmap in Figure 6 presents the co-occurrence frequencies between different substances and methods. As illustrated in the figure, co-occurring mentions of SN and antiemetics are common. Manual qualitative inspection of these posts revealed that antiemetics are often mentioned when discussing SN, as forum users recommended these substances to ensure that individuals do not feel nauseous after consuming SN.

Topic modeling revealed further insights and some key differences in the content of posts. First, the topics discovered reinforced some of the insights revealed in Figure 3. For the SN group, unigram and bigram topic clusters represented the following:

Figures S2 and S3 in Multimedia Appendix 1 present all the topics.

Based on the aforementioned topic modeling experiment results and supplemented with manual qualitative inspection of SN-related posts, we curated a list of phrases related to sourcing (eg, “seller,” “bought,” and “purchase”). The complete list is given in Table S2 in Multimedia Appendix 1. Since we collected data from Sanctioned Suicide on October 7, 2022, we extrapolated the frequency of posts from 279 days before (January 1, 2022, to October 7, 2022) by multiplying the per-day frequency with the total number of days in the year. Frequency analysis of sourcing-related posts pertaining to SN shows a sharp rise in “purchase” toward the end of 2019, with the highest raw annual “purchase” frequency observed in 2020 (Figure 7). Figure 8 shows the sourcing frequency (purchase frequency) of SN normalized by the posting frequency of SN-related topics. We found that discussions about the sourcing of SN gradually increased over the years.

The NER approach detected locations that included countries, states, cities, counties, etc. We manually analyzed the detected locations to create a location-mapping dictionary, which was used to map cities to countries. The United States, the United Kingdom, Canada, China, and Germany were found to be the most popular potential locations for obtaining or using SN (Table 1). Since the locations detected were primarily from the United States, we also mapped locations detected from within the United States to the state level. The states of California, New York, Florida, Texas, and Oregon were found to be mentioned the most often in association with SN.

Although NER-based geographical sources are informative, we found that consumers often tend to circumvent naming the geographical and online sources from where they obtained SN. Consider the following post: “I ordered it off of the big river in Brazil website.” NER identifies “Brazil” as a potential geographical source of SN in this example. However, the user is referring to the online source “Amazon” rather than the geographical source “Brazil.” Furthermore, locations mentioned in posts may also refer to the consumer’s location rather than the location where SN was sourced. This necessitated building a classifier to identify the intent of mentioning the location. Our proposed location mention–intent classifier achieved an accuracy of 72.92% on the unseen test data set, outperforming traditional machine learning–based baselines (Table 2).

Based on the NER and subsequent classification process, the United States, the United Kingdom, Canada, Australia, and China were found to be the most popular source locations for obtaining SN (Table 3). The states of California, New Mexico, Florida, Rhode Island, and Oklahoma were reported to be the most popular sourcing locations for SN within the United States.

The consumers who were interested in or attempted to obtain SN primarily were from the United States, the United Kingdom, Canada, Mexico, and Australia (Table 4). Within the United States, the states of California, New York, Texas, Florida, and Pennsylvania were found to be the most common locations of consumers attempting to obtain SN. For user-level post frequency, the mean was 72.63 (median 13, IQR 4-46; range 1-14,795).

Our NER-based approach also revealed potential online sources of obtaining SN or information about obtaining SN. Online marketplaces were the most commonly mentioned potential sources of SN (Table 5).