Working Paper No. 04

Revision without disclosure in California involuntary detention reporting

Three documents on the state health department's own domain all present themselves as the FY2019-20 involuntary detention report. The statewide count is 17,736 in one, 29,117 in the next, and 117,620 in the third.

1. Why three versions is the finding

Welfare and Institutions Code section 5402 requires the California Department of Health Care Services to collect involuntary detention data from every county mental health program and treatment facility and to publish it. The statute exists so that the public, the legislature, courts, and researchers can see how often the state's power to detain people against their will is used, and where.

That purpose depends on the published number being the number. A reader consulting the official record for how many Californians were held for 72-hour psychiatric evaluation in fiscal year 2019-20 can currently retrieve three different official answers, depending entirely on which file a search engine returns.

The fiscal year in question ended June 30, 2020. Every detention it counts had already happened before the first version was published. Nothing in the world changed between February 2021 and August 2022 to make more people have been detained. Only the record moved.

This is the same argument as Working Paper No. 01 and Working Paper No. 03, which examined published measures that failed to capture what they claimed to measure. The failure here is one step earlier in the chain. The measure is unstable, and the instability is undisclosed.

2. Data

Three PDF documents retrieved from dhcs.ca.gov on July 28, 2026.

All three are titled "California Involuntary Detentions Data Report, Fiscal Year (FY) 2019/2020." All three carry a DHCS division attribution, though the division named changes across versions, from Mental Health and Substance Use Disorder Services Division in v1 and v2 to Licensing and Certification Division in v3.

Rate values in these documents always print with one decimal place and raw counts always print as integers, which provides a reliable structural test for distinguishing the two tables during extraction. Values suppressed by DHCS appear as *** or * and are recorded as missing rather than zero. Reports for six neighboring fiscal years were also parsed, to establish whether the values in each version are plausible against adjacent years.

Known limitations

Version 3 contains fewer county rows. This extraction recovers 57 county rows from v1 and v2 but only 30 from v3. Several counties present in the earlier versions produce no value in v3. This paper does not claim those counties were deleted. Later DHCS reports adopt a convention of leaving a county blank when it contains no LPS-designated facility, and v3 may have adopted that convention early. The distinction is unresolved and is reported as an observation.

The department's current link points to v3. As of July 28, 2026, the DHCS data page lists the FY2019-20 report at a path resolving to FY19-20-IDR-Report.pdf. A reader who navigates the official page today receives the corrected file. The two earlier versions remain live at their original addresses and are still reachable by search, which is the channel through which a stale figure would most plausibly be retrieved.

3. Method

For each county, the adult 72-hour value was captured from the raw count table and from the rate table separately, then aligned across the three versions on county name. Two alternative explanations were tested directly. A denominator revision would move rates while leaving counts fixed, which comparing raw counts rules out. Redistribution between counties would move individual counties while holding the statewide total roughly constant, which comparing statewide totals rules out. Finally, the full text of all three documents was searched for the terms a revising agency would ordinarily use: revised, corrected, supersedes, erratum, amended, updated version, previously published, replaces.

4. The second version corrects exactly one county

The difference between v1 and v2 is 11,381 adult detentions. San Diego County appears as 0.0 per 10,000 in v1 and 44.4 in v2, and no other county changes. San Diego's adult population places its v2 count at approximately 11,100, which accounts for essentially the entire statewide difference between the two versions.

So v2 is not a general revision. It is v1 with one county added, published five months later, with no indication that it differs from its predecessor.

5. Individual county revisions reach two orders of magnitude

Los Angeles County is the largest county in the United States. For roughly eighteen months, the official public record of how many of its adult residents were detained for involuntary psychiatric evaluation in fiscal year 2019-20 was 168.

6. Neither alternative explanation survives

Rates and counts move together and by comparable magnitudes. A denominator revision cannot produce a 231-fold increase in a raw headcount. Back-solving the implied adult population from each version's count and rate pair gives 30.06 million, 29.71 million, and 29.85 million. Those agree within one percent across all three publications. The count changed. The denominator did not.

The statewide total does not hold constant either. It rises from 17,736 to 117,620. Information entered the record that was previously absent, rather than moving between counties within a fixed pool.

A full-text search of all three documents for revised, corrected, supersedes, erratum, amended, updated version, previously published, and replaces returns nothing in any of the three. Each document presents itself as the report for fiscal year 2019-20, without qualification.

7. The plausible values are the late ones

Against neighboring years, the v3 figures are the credible ones and the earlier versions are the outliers.

The v3 value of 50.5 sits inside the range set by adjacent years. The v1 and v2 value of 0.2 does not. The early versions were wrong by amounts large enough to invert any conclusion drawn from them.

8. The republication pattern may not be specific to this fiscal year

The currently published FY2015-16, FY2016-17, and FY2018-19 reports each carry a date stamp of October 2021, several years after their respective fiscal years closed. Multiple historical years appear to have been republished in a single month.

This is an observation, not a finding. It does not establish that those years had earlier versions with different values, only that the current files were issued long after the periods they describe. Testing whether the pattern documented here extends to other years would require locating the earlier files, which has not been done.

9. Some published values cannot be correct in any version

Extracting the raw count tables across every available year surfaces arithmetic that cannot be correct regardless of any revision history.

Of 199 county-year observations where both figures are published, 19 report more 14-day certifications than 72-hour holds. A 14-day certification follows a 72-hour hold. Under facility-based reporting a modest excess is possible, since a person may be held in one county and certified in another, but the magnitudes here are not modest.

A separate check runs the same test from the other direction. Dividing a published count by its published rate recovers the population denominator the department used. For most counties this returns a sensible figure: Los Angeles 7.7 million adults, Kern 626,000, Humboldt 108,000. For Ventura in fiscal year 2018-19 it returns 39,139, against an actual county population of roughly 840,000. Ventura also returns two incompatible populations in adjacent years.

These are not revision artifacts. They appear inside single published documents and are internally inconsistent within them.

10. What this does and does not support

11. Reproduction

Scripts, dataset snapshot, and the three source PDFs under their original filenames, so the comparison can be re-run against the files as they existed on the retrieval date:

zatara-research-data/04-three-versions

Cite this paper: