Antidepressant withdrawal as its people record it
Follows people coming off antidepressants through what they write about themselves over months and years: how severe their symptoms are, where they are in stopping (tapering, acute withdrawal, protracted, holding, reinstated, recovered), which of sixteen symptoms they report, how they are tapering and how their prescriber responded. Each post is read alone, with no history, platform or date shown to the model; the journey is assembled in code. It reports what a support forum's record can honestly say: change within each person, which symptoms persist, whether severity moves in waves, and who stops writing, beside the snapshot by time since the first post. Reviews of a drug (WebMD) are read with the same questions, and report how often stopping it is described, by drug.
The program
2 steps- 1
postsjev Jev · measuring postsTwenty-six questions about one forum or Reddit post, as the author tells it: a gate (is it a report on themselves), severity, phase, trajectory, outlook, windows, sixteen symptoms, taper method, prescriber, drug.
→ writesreadings - 2
reviewsjev Jev · measuring reviewsThe same questions about a drug review, and whether it describes what happened on stopping the drug.
→ writesreadings
Stages
52 questionsposts · 26 questions
jev-1.13.0Twenty-six questions about one forum or Reddit post, as the author tells it: a gate (is it a report on themselves), severity, phase, trajectory, outlook, windows, sixteen symptoms, taper method, prescriber, drug.
lens withdrawal_posts· the whole text is given as post· census genre sa· reads works where {"platform":["sa","reddit"]}
As measured, the yes/no criteria of 1 questions were not sent to Jev (is_self_report): the card reproduces that so its track record holds.
reviews · 26 questions
jev-1.13.0The same questions about a drug review, and whether it describes what happened on stopping the drug.
lens withdrawal_reviews· the whole text is given as review· census genre webmd· reads works where {"platform":["webmd"]}
As measured, the yes/no criteria of 2 questions were not sent to Jev (is_self_report, withdrawal_described): the card reproduces that so its track record holds.
The questions
as Jev is asked themyes / no is_self_report is_self_report
[object Object]
- yes
- —
- no
- —
score Severity described in a review (0 = none, 100 = crisis) severity
[object Object]
- 1 No symptoms reported, or the author says they feel well or recovered — I'm doing really well, sleeping fine, back at work
- 2 Mild: symptoms mentioned but described as tolerable; daily life largely unaffected — Some zaps in the morning but nothing I can't handle
- 3 Moderate: symptoms interfere with some sleep, work, social life, or activities — Insomnia most nights and I had to cut back my hours
- 4 Severe: symptoms dominate the day; the author struggles to work, care for themselves, or leave the house — I can barely make it through the day, haven't left the house in a week
- 5 Very severe or crisis: bedridden, unable to function at all, or in acute danger — I am bedbound and terrified, I don't know how much longer I can do this
choice phase phase
[object Object]
pre_taper- Still taking the medication at a steady or full dose and has not started reducing; may be considering stopping
tapering- Says they are currently reducing the dose step by step
acute_withdrawal- Stopped or cut the dose within the last few weeks and is describing symptoms since
protracted- Symptoms continuing months or years after the last dose or after a large cut
holding- Paused the taper and staying at the current reduced dose to let symptoms settle
reinstated- Went back on the medication or raised the dose to relieve withdrawal symptoms
recovered- Says withdrawal symptoms have resolved and they feel normal again
unclear- The post does not say where the author is with the medication: no dose, no stopping, no taper is mentioned, or the text is about something else
choice trajectory trajectory
[object Object]
improving- Says things are better than before; symptoms easing
worsening- Says things are worse than before; symptoms returning or intensifying
fluctuating- Describes alternating better and worse periods, windows and waves
stable- Describes no change, or a steady state
not_stated- No comparison with an earlier state is made
score hope hope
What is the emotional outlook of the author of post?
- 1 Despairing: hopeless, sees no way out, cannot go on
- 2 Uncertain or mixed: worried but not without hope, or no outlook expressed
- 3 Hopeful or reassured: expects to improve, encouraged by progress or by others
yes / no reports_window reports_window
Does the author of post describe a recent period of relief from their symptoms, feeling noticeably better for a while, even if symptoms later returned?
yes / no seeks_advice seeks_advice
Does the author of post ask others for advice, information, or reassurance?
yes / no sym_brain_zaps sym_brain_zaps
[object Object]
yes / no sym_insomnia sym_insomnia
[object Object]
yes / no sym_anxiety_panic sym_anxiety_panic
[object Object]
yes / no sym_low_mood sym_low_mood
[object Object]
yes / no sym_akathisia sym_akathisia
[object Object]
yes / no sym_derealization sym_derealization
[object Object]
yes / no sym_dizziness sym_dizziness
[object Object]
yes / no sym_gi sym_gi
[object Object]
yes / no sym_headache sym_headache
[object Object]
yes / no sym_fatigue sym_fatigue
[object Object]
yes / no sym_emotional_blunting sym_emotional_blunting
[object Object]
yes / no sym_irritability sym_irritability
[object Object]
yes / no sym_suicidal sym_suicidal
[object Object]
yes / no sym_cognitive sym_cognitive
[object Object]
yes / no sym_sensory sym_sensory
[object Object]
yes / no sym_sexual sym_sexual
[object Object]
choice taper_method taper_method
[object Object]
cold_turkey- Says they stopped abruptly, all at once, or 'cold turkey', with no taper at all
rapid_taper- Reduced over days or a few weeks, or in large cuts such as halving the dose
slow_taper- Reduced in small steps over months, such as 10% cuts, liquid dosing, bead counting, or micro-tapering
not_stated- How the dose was reduced is not described, including posts that only say the author stopped or is off the drug
choice clinician_response clinician_response
[object Object]
supportive- Acknowledged withdrawal and helped plan, slow, or adjust the taper
dismissive- Attributed the symptoms to relapse or the original illness, or said the drug does not cause withdrawal
unaware_or_unhelpful- Did not know how to taper, gave no plan, or advised stopping quickly
no_clinician- No doctor, psychiatrist, or prescriber interaction is described
choice primary_drug primary_drug
[object Object]
sertraline_zoloftescitalopram_lexaprocitalopram_celexafluoxetine_prozacparoxetine_paxil_seroxatvenlafaxine_effexorduloxetine_cymbaltadesvenlafaxine_pristiqmirtazapine_remeronbupropion_wellbutrinquetiapine_seroquelbenzodiazepine- Any benzodiazepine: diazepam, clonazepam, lorazepam, alprazolam, or a Z-drug such as zopiclone
other- A medication not listed
none_named- No medication is named
yes / no is_self_report is_self_report
[object Object]
- yes
- —
- no
- —
score Severity described in a review (0 = none, 100 = crisis) severity
[object Object]
- 1 No symptoms reported, or the author says they feel well or recovered — I'm doing really well, sleeping fine, back at work
- 2 Mild: symptoms mentioned but described as tolerable; daily life largely unaffected — Some zaps in the morning but nothing I can't handle
- 3 Moderate: symptoms interfere with some sleep, work, social life, or activities — Insomnia most nights and I had to cut back my hours
- 4 Severe: symptoms dominate the day; the author struggles to work, care for themselves, or leave the house — I can barely make it through the day, haven't left the house in a week
- 5 Very severe or crisis: bedridden, unable to function at all, or in acute danger — I am bedbound and terrified, I don't know how much longer I can do this
choice phase phase
[object Object]
pre_taper- Still taking the medication at a steady or full dose and has not started reducing; may be considering stopping
tapering- Says they are currently reducing the dose step by step
acute_withdrawal- Stopped or cut the dose within the last few weeks and is describing symptoms since
protracted- Symptoms continuing months or years after the last dose or after a large cut
holding- Paused the taper and staying at the current reduced dose to let symptoms settle
reinstated- Went back on the medication or raised the dose to relieve withdrawal symptoms
recovered- Says withdrawal symptoms have resolved and they feel normal again
unclear- The post does not say where the author is with the medication: no dose, no stopping, no taper is mentioned, or the text is about something else
choice trajectory trajectory
[object Object]
improving- Says things are better than before; symptoms easing
worsening- Says things are worse than before; symptoms returning or intensifying
fluctuating- Describes alternating better and worse periods, windows and waves
stable- Describes no change, or a steady state
not_stated- No comparison with an earlier state is made
score hope hope
What is the emotional outlook of the author of review?
- 1 Despairing: hopeless, sees no way out, cannot go on
- 2 Uncertain or mixed: worried but not without hope, or no outlook expressed
- 3 Hopeful or reassured: expects to improve, encouraged by progress or by others
yes / no seeks_advice seeks_advice
Does the author of review ask others for advice, information, or reassurance?
yes / no sym_brain_zaps sym_brain_zaps
[object Object]
yes / no sym_insomnia sym_insomnia
[object Object]
yes / no sym_anxiety_panic sym_anxiety_panic
[object Object]
yes / no sym_low_mood sym_low_mood
[object Object]
yes / no sym_akathisia sym_akathisia
[object Object]
yes / no sym_derealization sym_derealization
[object Object]
yes / no sym_dizziness sym_dizziness
[object Object]
yes / no sym_gi sym_gi
[object Object]
yes / no sym_headache sym_headache
[object Object]
yes / no sym_fatigue sym_fatigue
[object Object]
yes / no sym_emotional_blunting sym_emotional_blunting
[object Object]
yes / no sym_irritability sym_irritability
[object Object]
yes / no sym_suicidal sym_suicidal
[object Object]
yes / no sym_cognitive sym_cognitive
[object Object]
yes / no sym_sensory sym_sensory
[object Object]
yes / no sym_sexual sym_sexual
[object Object]
choice taper_method taper_method
[object Object]
cold_turkey- Says they stopped abruptly, all at once, or 'cold turkey', with no taper at all
rapid_taper- Reduced over days or a few weeks, or in large cuts such as halving the dose
slow_taper- Reduced in small steps over months, such as 10% cuts, liquid dosing, bead counting, or micro-tapering
not_stated- How the dose was reduced is not described, including posts that only say the author stopped or is off the drug
choice clinician_response clinician_response
[object Object]
supportive- Acknowledged withdrawal and helped plan, slow, or adjust the taper
dismissive- Attributed the symptoms to relapse or the original illness, or said the drug does not cause withdrawal
unaware_or_unhelpful- Did not know how to taper, gave no plan, or advised stopping quickly
no_clinician- No doctor, psychiatrist, or prescriber interaction is described
choice primary_drug primary_drug
[object Object]
sertraline_zoloftescitalopram_lexaprocitalopram_celexafluoxetine_prozacparoxetine_paxil_seroxatvenlafaxine_effexorduloxetine_cymbaltadesvenlafaxine_pristiqmirtazapine_remeronbupropion_wellbutrinquetiapine_seroquelbenzodiazepine- Any benzodiazepine: diazepam, clonazepam, lorazepam, alprazolam, or a Z-drug such as zopiclone
other- A medication not listed
none_named- No medication is named
yes / no withdrawal_described withdrawal_described
[object Object]
- yes
- —
- no
- —
What it reports
2 indicators| Indicator | Kind | From | Needs |
|---|---|---|---|
| Reviews describing symptoms on stopping, missing or cutting the drugheadline | frame | withdrawal_described | 1 of 1 yes |
| Severity described in a review (0 = none, 100 = crisis) | share | severity | – |
What it follows
7 reportsA journey is a member's own posts in their thread; the clock is days since its first post, read or not; only posts in which the author reports on themselves are observations.
- subject
- one work: a journey: one member's own posts in their introduction-and-updates thread, in order, as one work
- clock
aton each unit, counted from the work’s first- gate
- observations where
is_self_reportis above 0.5 - the end of a record
- unexplained: people leave when they recover and when they give up
- follows works where
{"platform":["sa"]}
| Report | Kind | What it says |
|---|---|---|
change_withinheadline | first_last | Severity at a journey's first and last self-report (journeys with five or more); improved or worsened by more than an eighth of the scale |
by_month | time_bins | The snapshot: means over every self-report by months since the journey's first post a census of who is still writing, not of how people fare: read it beside change_within and who_stays |
persistence | persistence | Of journeys reporting a symptom in their first three months and still writing at 6-18 months, the share reporting it again |
waves | series | Do windows and waves show from post to post? Severity's order within journeys of ten or more self-reports |
waves_by_gap | gaps | Consecutive self-reports' severity (each from its journey's mean), correlated, by the days between them |
who_stays | who_stays | Who is still writing at 9-18 months, by severity in the first three self-reports; protracted (severity at least moderate, or read as protracted) among those who are |
phases | transitions | From one self-report's phase to the next one's, within journeys (posts whose phase is unclear are passed over) |
Track record
against human coders and external anchorsNo human coding. External anchors: the share of reviews describing withdrawal follows the drugs' half-lives, from paroxetine (30%) to fluoxetine (7%) and vortioxetine (3%), with no drug named in the questions and the drug never shown. Severity read from a review correlates with the reviewer's own satisfaction rating at Spearman -0.56, and outlook at +0.72.
source: project_research/experiments/withdrawal/RESULTS.md
Phase agrees exactly with the earlier labels on 64% of posts across eight phases, and severity correlates at +0.74; disagreements sit at boundaries (unknown vs pre-taper, tapering vs acute withdrawal). Model against model: evidence of consistency, not validation.
source: project_research/experiments/withdrawal/RESULTS.md
No human coding (a blind 300-post coding sheet is ready in project_research/data/validation/). Paraphrase disagreement on the 310-post pilot: 3.5% of posts for severity, 2.9% for the gate, 0.3% for suicidal thoughts. Within journeys of five or more self-reports severity falls from 0.50 to 0.38, while the snapshot by month since the first post stays flat (0.36-0.40): the forum loses the recovered and the most severe first. Psychological symptoms persist (anxiety 71%, low mood 62%) and brain zaps least (22%). No windows and waves from post to post (lag-1 autocorrelation +0.02; reversals 0.66 against 0.67 at random).
source: project_research/experiments/withdrawal/COHORT_RESULTS.md, as census.longitudinal computes it now (engine/scripts/check_withdrawal.py)
Rules and limitations
- The model sees one post at a time and nothing else: no earlier post, no platform, no date, no author. The journey is put together in code, afterwards.
- Only self-reports are observations: posts where the author describes their own condition (the gate, above 0.5). A journey's clock counts every post, read or not.
- Days, months, cut sizes and changes are arithmetic in code. The model is never asked to count, compute or compare dates.
- Anonymity: no handle, no verbatim quotation of a forum or Reddit post in any report or product. A person is a pseudonym. Keep outputs aggregate and paraphrase.
- Journeys come from the research instance only (data/research.duckdb), never the main one or a deployment.
- Not validated against human coders yet. The anchors (reviewers' own ratings, an earlier model's labels, the half-life ordering) are consistency, not truth.
- Attrition: people stop writing when they recover and when things get worse, and do not say which. Any snapshot by time is of those still writing; read change within people and who_stays beside it. Survival curves would assume leaving says nothing about how someone is, which is false here, so the card does not draw them.
- A post is what its writer chose to say that day. Post-to-post statistics (waves) describe the record, not lived experience.
- Severity and outlook are positions on described levels, not measurements; the model is weakly calibrated between levels.
- A symptom counts only when the author says they have it. Symptoms implied but not stated are missed.
- Yes/no criteria did not reach the model in this version: YAML read their unquoted true:/false: keys as booleans, so the gate (is_self_report) and withdrawal_described were asked with their question and focus but without the examples of each side. The track record describes the questions as they were sent; a version that sends them must be measured again.
- Posts longer than 12,000 characters are cut there (24 of 22,566 in the large run). English only.
- Observational and self-selected: associations with taper method or cut size are not effects.