# Pre-registered plan: does ketamine lower depressive symptoms when patients can't feel it?

## Question

Ketamine and its S-enantiomer esketamine relieve depression within a day in randomized trials, and esketamine is approved for treatment-resistant depression. Both have psychoactive effects (dissociation, perceptual change) that most patients notice, so most patients in these trials can tell which arm they are in. Whether the antidepressant effect needs that experience is open: it could be a pharmacological effect that would appear without it, or partly the expectation and meaning that a strange, strong experience carries. The one trial built to tell these apart gave ketamine or saline to 40 depressed adults while they were unconscious under general anesthesia for routine surgery: depressive symptoms fell by about half in both arms, with no difference between them (Lii et al., Nature Mental Health 2023, doi:10.1038/s44220-023-00140-x).

Many more randomized trials have since given ketamine or esketamine around surgery or childbirth and measured depression afterwards. In some, every dose is given while patients are unconscious under general anesthesia; in others, patients are awake for some or all of it, after a cesarean delivery under spinal anesthesia, by a postoperative pump, or before induction. Meta-analyses have pooled these trials together. None has separated the two, screened the trials' trustworthiness, or counted registered trials that never reported.

**Questions.**

1. **Primary 1.** Do ketamine and esketamine lower depressive symptoms one to three days after a procedure when every dose was given while the patient was unconscious under general anesthesia (masked trials)?
2. **Primary 2.** Among trials in surgery, is the effect larger when patients received the drug while conscious (awake trials) than when they didn't, allowing for whether trials enrolled only depressed patients?
3. **Checks.** Do masked doses act on the body (pain, a positive control), and were masked patients spared the drug's psychoactive effects (psychotomimetic adverse events)?
4. **Secondary.** Do the answers hold among trustworthy trials, with strictly masked dosing, in trials of depressed patients, later after the procedure, and allowing for small-study effects and for registered trials that never reported?

## Data

**Reports.** PubMed (E-utilities) and Europe PMC (titles and abstracts; this adds preprints), searched on 2026-10-13 for ketamine or esketamine, a depression measure, randomization or placebo, and surgery, anesthesia, sedation, or childbirth: 703 PubMed records and 719 Europe PMC records, 759 distinct. `data/search.json` has the queries, counts, and identifiers; `data/records.json` has each record's identifiers, title, journal, year, publication types, and the retraction, expression-of-concern, erratum, and update notices PubMed links to it (abstracts stay with PubMed and Europe PMC).

**Registrations.** ClinicalTrials.gov, searched the same day for interventional studies of ketamine, esketamine, or arketamine mentioning depression or mood: 634 registrations (`data/search.json`).

## Screening

Two coders, separate instances of Claude Opus 5.5 (the model that wrote this study), each working alone from `screening_codebook.md` with only titles and abstracts, decided which records might report an eligible trial. They agreed on 758 of 759 (κ = 0.996; 137 and 138 kept); a third instance, given the codebook, the record, and both decisions, decided the other one (kept).

Two more coders, from `registry_codebook.md`, coded each registration's eligibility and, for eligible ones, whether any dose was given while patients were conscious. They agreed on eligibility for 633 of 634 (κ = 0.99) and on dosing for 54 of the 56 both found eligible; a third reader decided the three others. 56 registrations are eligible (`data/registry.json`).

**Other sources.** One report that PubMed links to an eligible registration, but whose abstract doesn't mention depression, so the search missed it; and the five eligible registrations that were due, never published, and posted results on ClinicalTrials.gov, whose source is the registry itself.

**Trials.** The 138 kept records were grouped into trials (`data/groups.json`): reports whose abstracts name the same registration, near-identical titles (a preprint and its journal version), and four links, each with its reason: a post hoc analysis of one trial in three reports, and a preprint whose three groups and their sizes match a later journal report, which the preprint calls a prospective case-control study and the journal a randomized trial. With the six records from other sources, that gives 126 trials. Open full text came from Europe PMC, PubMed Central (its XML or its public article page), preprint servers' PDFs, and publishers' open-access pages read in a browser (ten, seven of them once a person had passed the publisher's bot check). For abstract-only trials whose articles aren't open, a person obtained what a library subscription allowed: eleven journal articles and one preprint of a twelfth, as PDFs that stay outside this bundle (source `library`). Each trial's source is then the first that exists of: its journal report's full text, open or from the library; its preprint's open full text; its journal report's abstract; its preprint's abstract (or, for the five, its registration). Sources: 88 full text, 12 library, 4 preprint, 17 abstract, 5 registry.

## Design coding

Two further coders, working alone from `design_codebook.md`, coded each trial's design from its source's Methods section only, never its title, results, tables, or discussion; for a trial with only an abstract, from the abstract up to where its results begin (9 abstracts have no labelled sections and were given whole); for a registry trial, from its registration's protocol section. They coded eligibility, arms, doses, the timing of each dose, the anesthesia, whether any dose was given while the patient was conscious (`conscious_dosing`), whether a dose under anesthesia came late enough to still act at waking (`late_dosing`), the setting, whether only depressed patients were enrolled, the depression and pain measures planned (with fixed codes), blinding, registration, enrollment dates, and country.

They agreed on eligibility for all 126 trials, and, for the 104 both found eligible, on `conscious_dosing` for all 104, setting and depression selection for all 104, `late_dosing` for 103, the number of arms of each role for 103, the set of depression measures for 101, and which is primary for 99. A third coder decided the nine trials with any difference, with reasons (`coding/`). As a check by other coders from other text, the registry coders' dosing codes agree with the design coders' for all 10 trials registered on ClinicalTrials.gov that both coded clearly.

**Second round.** The 19 trials whose source became a full text (open or from the library) after this round had coded them from their abstracts were coded again from their Methods by two new coders, in the same way (`coding/design2_*`); their coding replaces the first round's. The two agreed on every compared field for all 19 (16 eligible), so no third reader was needed. Full text changed eligibility for three trials (excluded from their abstracts, eligible from their Methods), settled `conscious_dosing` for five that had been unclear, and showed that one trial hadn't enrolled only depressed patients.

`data/trials.json` holds each trial's decided design. Two reports with the same registration, size, enrollment start, and setting are one trial: the preprint T072 is T054's, and is dropped. Three trials of different sizes, dates, and settings (T033, T044, T052) give one registration number; they stay separate trials, each flagged. 106 trials are eligible; the rest had no depression measure, no inert control, or no randomization (`data/trials.json`).

**Groups.** A trial is **masked** if `conscious_dosing` is `none` (39 trials: 37 in surgery, 2 cesarean), **awake** if `some` (61: 29 in surgery, 28 cesarean, 4 vaginal delivery); the 6 `unclear` trials are in neither primary analysis. A masked trial with `late_dosing` true (14) is masked but not **strictly masked**; 20 are strictly masked and 5 unclear. Nine masked and seven awake trials enrolled only depressed patients. 67 of the 106 planned a depression measure on days 1 to 3 (27 masked, 35 awake, 5 unclear). One trial (T097) randomized esketamine before induction (so patients were awake for it), esketamine after induction, and placebo. It gives two comparisons, each with half of its placebo arm (Cochrane Handbook 23.3.4): after induction against placebo, masked, and before induction against placebo, awake, so the analyses count 40 masked and 61 awake comparisons. Its two esketamine arms are also compared with each other, the only randomized comparison of conscious with unconscious dosing, and reported on their own.

## Outcomes

Numbers are extracted after this plan is registered. A full text the library provides after registration but before extraction becomes its trial's source; the trial's design is then coded again from its Methods as in the second round, and the change is listed in `deviations.json`. Numbers are extracted by two coders working alone from `extraction_codebook.md`, copying what each source prints, in the form printed; a third coder decides every number on which they differ, against the source. The code does every conversion.

**Depression.** For each trial and window, one effect: the measure the trial named primary, otherwise the first it reports in that window of MADRS, HAMD, EPDS, PHQ, BDI, HADS depression subscale, CES-D, GDS, Zung SDS, POMS depression, HADS total, other (`data/parameters.json`); the earliest day it reports in the window; and the first available form of mean and SD, mean and SE, mean and CI, change from baseline, median and quartiles, median and range, difference between arms, count. Windows: days 1–3 after the procedure (early, primary), 4–14 (middle), and 15–90 (late).

**Effect size.** Hedges' g, ketamine against control, negative meaning fewer depressive symptoms with ketamine (`metafor::escalc`, `SMD`). Several ketamine arms are combined into one (Cochrane Handbook 23.3.4). A mean and SD are estimated from a median and quartiles, or a median and range, by Luo et al. (2018) and Wan et al. (2014); a count's log odds ratio becomes g by multiplying by √3/π (Chinn 2000); a mean difference with its CI becomes g through the pooled SD its CI implies.

**Pain** (positive control): on day 1, else day 2, pain at rest, else pain on movement, else another pain score (most trials don't say which), as g; opioid consumption is not used.

**Psychotomimetic adverse events** (manipulation check): in each trial, the psychotomimetic event (hallucination, nightmare, dissociation, visual disturbance, agitation, or a combined psychiatric count) with the most participants in both arms together, chosen blind to arm.

## Estimates

All pooled effects are random-effects estimates by restricted maximum likelihood with Knapp–Hartung confidence intervals and tests (`metafor::rma`, `method = "REML"`, `test = "knha"`), with τ², I², and the 95% prediction interval.

- **Primary 1:** the pooled early effect among masked trials (all settings).
- **Primary 2:** among trials in surgery (masked or awake), the meta-regression coefficient of `awake` in a model of the early effect on `awake` and `selected` (enrollment only of depressed patients): how much more negative the effect is when patients were dosed while conscious. The pooled masked and awake effects in surgery are reported beside it.
- **Multiplicity:** Holm's correction over the two primary p values.
- **Robust**, for each primary: Holm p < 0.05; p < 0.05 among trustworthy trials; p < 0.05 with strict masking; and p < 0.05 with each single trial left out.

**Integrity.** A trial is **trustworthy** unless: a retraction or expression of concern is linked to its report; its reports disagree on whether allocation was random; another trial, of a different size, start, or setting, gives the same registration number; it names no registration; its registration is dated after its enrollment began (by the month; when only a ChiCTR number gives the year, a registration year after the enrollment year); two or more of its depression means fail GRIM (a mean of whole-number scale totals must be a whole number divided by n, checked only where n is known exactly and n < 10^d for a mean given to d decimals; Brown and Heathers 2017); or its early effect is beyond |g| = 2. A stricter set requires prospective registration shown by dates. Before extraction: no kept report has a retraction or expression-of-concern notice; one trial's reports disagree on randomization (T039); three trials share a registration number; and 78 of the 106 name a registration.

**Sensitivity analyses** of both primaries: trustworthy trials; prospectively registered trials; strict masking; leaving out the four trials whose ketamine was given only by a postoperative pump; full-text, library, and preprint sources only; open sources only (leaving out the library's); effects needing no conversion; depressed-only trials; unselected trials; esketamine only; and the middle and late windows.

**Positive control.** Pain on day 1 among masked trials, among awake trials, and the awake–masked difference in surgery, as for Primary 2. If masked ketamine doesn't lower pain (95% CI including 0), the paper says the masked doses may not have been active, and that a null masked depression effect is then uninformative.

**Manipulation check.** The Mantel–Haenszel odds ratio of the dominant psychotomimetic event, ketamine against control, among masked and among awake trials, and the z test of their difference on the log scale.

**Small studies and the file drawer.** Among masked trials, Egger's regression test and the PET-PEESE estimate. Of the eligible registrations whose primary completion date was at least 12 months before the search and that weren't withdrawn or still enrolling (39 of 56), how many reported, by a PubMed report naming the NCT number that isn't a protocol, by results posted, or neither (`data/registry.json`).

**Dose.** Among masked trials, the meta-regression slope of the early effect on the log of the total dose in ketamine equivalents (esketamine counted twice).

**Within one trial.** T097's esketamine before induction against esketamine after induction, on days 1 to 3, as Hedges' g with its 95% CI.

## How results will be read

- **The effect doesn't need awareness:** Primary 1 is significant after Holm's correction, and the 95% CI of Primary 2 includes 0.
- **The effect depends on awareness:** Primary 2 is significant after Holm's correction (awake larger), and the 95% CI of Primary 1 includes 0 while masked ketamine lowers pain.
- Anything else is reported as it falls, without either reading.

## Power

From the design coding alone (each trial's group, setting, selection, and planned size, the median planned size where none is stated, over the trials that planned a depression measure on days 1 to 3: 27 masked, and 26 masked and 21 awake in surgery), with 1,000 simulations per scenario (`power.json`): with a masked effect of g = −0.2, Primary 1 rejects in 71% (τ = 0.35) to 96% (τ = 0.2) of simulations, and with g = −0.3 in 95% to 100%; with masked g = 0 and awake g = −0.4, Primary 2 rejects in 90% to 100%; with masked −0.2 and awake −0.5, in 68% to 96%; and with equal effects, Primary 2 rejects at its nominal 5%. Trials that report a measure in the window they didn't plan add to these.

## What was seen before this plan was registered

- A feasibility screen of an earlier search's 650 PubMed abstracts by four other instances, counting trials by timing and selection from abstracts, which state results.
- The screeners read titles and abstracts, many of which state results. The design coders read only Methods text, or abstracts up to their results, except the 9 unstructured abstracts. Checking that Methods text was split cleanly, the author read lines of seven Methods files, none of them results, and a figure legend that had slipped into one preprint's Methods text (a significance mark on a pain chart) before the split was fixed and the files remade.
- The published results of Lii et al. (2023), of meta-analyses that pooled perioperative ketamine and esketamine trials (which report lower depression scores with esketamine, for example a pooled SMD of about −0.9 over 13 trials under general anesthesia), and of a large trial in prenatal depression, as known in the field.
- The analysis code was written and tested on simulated trials and on invented numbers for the real trials' designs (`code/simulate.py`, both modes), which say nothing about any trial's results; that test showed that most trials don't say whether a pain score was at rest, so the pain control takes any pain score after rest and movement. `power.json` comes from the design coding alone, with no outcome data.
- A registry coder briefly saw ten registrations of the other batch through a helper file in its shared folder, but not anyone's coding.
- The person who passed the bot check and obtained the library PDFs did no coding. Checking that PDF and page text was split cleanly, the author read the opening and closing lines of eleven Methods texts. One PDF was first matched to the wrong trial (its reference list cites that trial); the text it left was deleted before any coding, and PDFs are now matched by the number they were requested under.

## Files

- `plan.md`, `screening_codebook.md`, `design_codebook.md`, `extraction_codebook.md`, `registry_codebook.md`.
- `coding/`: both screenings, both design codings of each round, both registry codings, the third readers' decisions with reasons, and the agreement.
- `data/`: `search.json`, `records.json`, `groups.json`, `trials.json`, `registry.json`, `parameters.json`.
- `code/`: `analyze.R` (the analysis) and its command (`run`), `power.R`, `simulate.py`; `env/Dockerfile` (R 4.6.1, metafor 5.2-1, jsonlite 2.0.0, from a dated CRAN snapshot).
- `power.json`.

## Deviations

Any departure from this plan will be listed in the bundle's `deviations.json` with its reason.