Download the data

• trends.csv: one row per item and year, with the item name, group, share of facilities offering it, and the number of facilities that answered.
• sample.csv: the number of facilities in the sample each year.

Please cite this page if you use the data. The code, codebook and data are on GitHub at jdada360/mh-facility-trends and archived on Zenodo: doi:10.5281/zenodo.23574161.

Sources

The 2015 to 2020 data come from the National Mental Health Services Survey (N-MHSS). In 2021, SAMHSA merged N-MHSS with the National Survey of Substance Abuse Treatment Services (N-SSATS) into one survey, the National Substance Use and Mental Health Services Survey (N-SUMHSS). I use the public-use files for every year, all from SAMHSA:

• N-MHSS: 2015, 2016, 2017, 2018, 2019, 2020
• N-SUMHSS: 2021, 2022, 2023, 2024

The series starts in 2015 because of a change in question format. The 2014 survey asked about each service as its own yes-or-no question. From 2015 on, the survey lists the services and asks facilities to mark all that apply. The two formats give quite different answers for the same item: 66% of facilities reported suicide prevention services in 2014, against 48% in 2015 and 54% in 2016. I left 2014 out so that every year in the charts uses the same format.

Which facilities count

For 2015 to 2020, I include every facility in the N-MHSS file. For 2021 to 2024, I include the facilities that completed the mental health section of N-SUMHSS, since these items are blank for facilities that only treat substance use. That is the closest match to the old N-MHSS group, though the two groups differ somewhat. Keep this switch in mind when comparing 2020 with 2021. In 2024 the mental health section grew to about 14,300 facilities, up from about 10,100 in 2023, so the 2024 point reflects a larger group with more outpatient providers.

The table shows how many facilities are in the sample each year. The total counts facility-years: most facilities answer in several years, so it is not a count of distinct facilities.

sample = FileAttachment("sample.csv").csv({ typed: true })
html`<table class="table mh-sample">
  <caption>Facilities in the sample by year, 2015–2024</caption>
  <thead><tr><th>Year</th><th>Survey</th><th class="num">Facilities</th></tr></thead>
  <tbody>${sample.map(d => html`<tr><td>${d.year}</td><td>${d.year >= 2021 ? "N-SUMHSS" : "N-MHSS"}</td><td class="num">${d3.format(",")(d.facilities)}</td></tr>`)}</tbody>
  <tfoot><tr><th>Total</th><th></th><th class="num">${d3.format(",")(d3.sum(sample, d => d.facilities))}</th></tr></tfoot>
</table>`

Missing answers

Each share counts only the facilities that answered that question in that year. The numerator is facilities answering “yes” and the denominator is facilities answering “yes” or “no.” If a facility skipped a question, left it blank, or has a missing code (−1 or −2), I drop it from that item’s share for that year. It still counts toward every other item it answered, so denominators differ across items and years. Hover over any point to see how many facilities answered.

Matching items across years

I kept only items asked in all ten years with the same yes/no coding: 54 in total. Most kept the same variable name throughout. I left out items first asked after 2015, such as transcranial magnetic stimulation, ketamine, and EMDR, because they don’t have a full series. The substance use side of the survey isn’t covered here either.

Caveats

Service offerings are self-reported by facility staff. Response rates differ by year (see each year’s codebook).

Use of AI

I built this page with Claude Opus 5.5, an AI model made by Anthropic. Claude matched variables across the survey years, wrote the code that cleans the data and draws the charts, and drafted the text. I chose the questions, checked the results against the survey files, and edited the writing.