Design of studies

The majority of the 62 studies or cohorts included in the incidence analysis were acute efficacy studies, extension studies, or relapse prevention studies, in which withdrawal effects were an incidental concern and not reliably measured (see below). Only 16 (26%) were designed primarily to study withdrawal, and these were mostly small (See Table 1).

Table 1. Characteristics of studies used in Henssler et al.’s analysis of the incidence of any withdrawal symptom (eFigure 1 in Henssler’s supplementary appendix)

Abbreviations: GAD, generalized anxiety disorder; OCD, obsessive compulsive disorder; SSRI, selective serotonin reuptake inhibitor; SNRI, serotonin and noradrenaline reuptake inhibitor; OL, Open Label; DESS, Discontinuation-Emergent Signs and Symptoms.

^a We have classified studies according to the aim of the original study from which the data were gathered. Some papers focus on data on withdrawal from studies that were set up with a different aim, hence we have classified them as ‘secondary’ whereas Henssler et al. classified them as ‘primary’.

^b ‘High’ = high potential for misclassification because withdrawal and relapse/psychiatric symptoms measured concurrently with evidence suggestive of misclassification (such as the noninclusion of typical emotional symptoms among reported effects of withdrawal); ‘Medium’ = medium potential for misclassification where misclassification may have influenced ratings of withdrawal; ‘Low’ = misclassification not likely to have been a significant problem. See Supplementary Table S1 for more detailed rationale for individual studies.

^c Some or all participants did not stop their antidepressant.

Forty-six (74%) of the 62 studies had definite or probable funding from a pharmaceutical company (Table 1). Since funded studies were larger than non-funded studies, they accounted for 96.2% (12,119/12,603) of the participants included in the analysis.

Assessment of withdrawal effects

A fundamental problem with the review is the manner of assessment of withdrawal symptoms (Table 1). In 52 of the 62 study cohorts, figures were derived from data on adverse events, responses to open questions, clinician judgment, or no method was specified. Where details were provided, adverse events and symptoms were ‘spontaneously reported’ in all but one study. In this study, specific withdrawal symptoms were enquired about during the measurement of adverse events (Tourian et al., Reference Tourian, Padmanabhan, Groark, Brisard and Farrington2009). Whether adverse events counted as withdrawal effects was further determined by the subjective judgment of the researchers who decided ‘if they occurred for the first time or worsened following discontinuation of treatment ’(p. 208) (Perahia, Kajdasz, Desaiah, & Haddad, Reference Perahia, Kajdasz, Desaiah and Haddad2005).

It is known that the detection of adverse events in studies designed to evaluate efficacy is unreliable, inconsistent, and likely to underestimate effects (Chrysant, Reference Chrysant2008; Hammad, Pinheiro, & Neyarapally, Reference Hammad, Pinheiro and Neyarapally2011; Phillips, Hazell, Sauzet, & Cornelius, Reference Phillips, Hazell, Sauzet and Cornelius2019). Ratings show poor reliability even for physical symptoms (Forster, Taljaard, Bennett, & van Walraven, Reference Forster, Taljaard, Bennett and van Walraven2012) and when raters are guided by a list of specific symptoms (Atkinson et al., Reference Atkinson, Li, Coffey, Sit, Shaw, Lavene and Basch2012). In a trial of a chiropractic intervention, 88 times more adverse events were identified using proactive monitoring than when relying on spontaneous reports (Pohlman et al., Reference Pohlman, Carroll, Tsuyuki, Hartling and Vohra2020). Subjective adverse effects, including symptoms such as fatigue and emotional changes, are more likely to be under-detected than objective physical signs such as oedema (Chrysant, Reference Chrysant2008). The reporting, as well as the detection of adverse effects in such studies, is also unreliable (Mayo-Wilson et al., Reference Mayo-Wilson, Fusco, Li, Hong, Canner, Dickersin and investigators2019; Phillips et al., Reference Phillips, Hazell, Sauzet and Cornelius2019).

Underestimation of adverse effects following antidepressant withdrawal is particularly likely because the most common symptoms include anxiety, fatigue, impaired concentration, and worsened mood, as documented in a study of over 1000 participants (Moncrieff, Read, & Horowitz, Reference Moncrieff, Read and Horowitz2024), which overlap with symptoms of the disorders for which antidepressants are most commonly prescribed. Therefore, withdrawal symptoms can be overlooked or misclassified as symptoms of the underlying condition.

Eleven studies that rated withdrawal symptoms and mental disorder symptoms concurrently were rated as showing a ‘high potential’ for misclassification (See Table 1 and Supplementary Table S1).

In several of these studies, the authors acknowledged the problem by conducting sensitivity analyses of their efficacy measure, excluding data from the first few weeks after randomization (Allgulander, Florea, & Huusom, Reference Allgulander, Florea and Huusom2006; Kornstein et al., Reference Kornstein, Bose, Li, Saikali and Gandhi2006; Rosenthal et al., Reference Rosenthal, Boyer, Vialet, Hwang and Tourian2013). However, they did not consider how the potential misclassification might have impacted the detection of withdrawal symptoms.

Only 13 of the 62 studies were rated as showing a low potential for misclassification.

In addition to these problems, five studies in the incidence analysis were retrospective case note reviews identifying reports of withdrawal symptoms entered by clinicians during routine clinical care (Table 1). Such studies are likely to miss all but the most distinctive and severe symptoms of withdrawal due to the lack of awareness of the range of effects (Guy et al., Reference Guy, Brown, Lewis and Horowitz2020).

Use of structured instruments

Although 18 of the 62 studies included in Henssler’s incidence analysis used a structured instrument to assess withdrawal symptoms, in 10 of these, Henssler et al.’s analysis was based on adverse events because data from the instrument were not available in the required form (Table 2). In three studies, the instrument was developed for other purposes and did not include common antidepressant withdrawal symptoms. In one, data did not reflect the proportion of people experiencing ‘any’ symptom as per Henssler et al.’s criteria (Table 2) (Yasui-Furukori et al., Reference Yasui-Furukori, Hashimoto, Tsuchimine, Tomita, Sugawara, Ishioka and Nakamura2016). In only four studies were withdrawal symptoms measured using a relevant instrument and reported in such a way as to be eligible for Henssler et al.’s analysis. Two of these used the Discontinuation-Emergent Signs and Symptoms (DESS) (Cohen et al., Reference Cohen, Soares, Lyster, Cassano, Brandes and Leblanc2004; Stein et al., Reference Stein, Chartier, Hazen, Kroft, Chale, Coté and Walker1996) and two used similar instruments or sets of questions (Mourad, Lejoyeux, & Adès, Reference Mourad, Lejoyeux and Adès1998; Murata et al., Reference Murata, Kobayashi, Imuta, Haraguchi, Ieiri, Nishimura and Mine2010). One further study presented data on ‘specific’ adverse events that were elicited alongside the DESS questionnaire (Tourian et al., Reference Tourian, Padmanabhan, Groark, Brisard and Farrington2009).

Table 2. Studies included in Henssler et al.’s incidence analysis that used a structured instrument

References for articles not cited in-text can be found in the supplementary material.

Abbreviations: AE, adverse event; DESS, Discontinuation-Emergent Signs and Symptoms.

^a The figures used by Henssler et al. for these studies were based on any withdrawal-related symptom (criteria for the incidence analysis) measured by a structured instrument or specific questions relevant to antidepressant withdrawal.

Therefore, only 8.1% (5/62) of the studies included in Henssler et al.’s meta-analysis, involving 4.8% (601/12,603) of total participants, presented data derived from a systematic and relevant assessment of withdrawal symptoms (Figure 1).

Figure 1. Flow diagram of studies included in Henssler et al.’s incidence calculation that used a structured instrument or method of assessment of antidepressant withdrawal symptoms.

Discrepancies and uncertainties

Minor discrepancies in data extraction are common in systematic reviews, but some of those in the Henssler et al.’s review are likely to have impacted the results of the analysis, given the size of the studies involved (see Table 1).

For example, participants in a large, pooled analysis of studies of desvenlafaxine by Montgomery et al. (Reference Montgomery, Fava, Padmanabhan, Guico-Pabia and Tourian2009) were double counted, so that 947 participants from these studies were included in the meta-analysis twice.

Figures for several further studies are unreliable due to minimal reporting of adverse events, leading to uncertainty about the total number of people who were followed up after discontinuing their antidepressant (details in Table 1 and Supplementary Table S1). Henssler et al.’s use of the number randomized as the denominator in these cases would tend to reduce the rate of reported withdrawal effects, unless there were no dropouts (which is unlikely).

In two studies, all or some participants only reduced the dose of their antidepressant and did not stop (Bakish et al., Reference Bakish, Bose, Gommoll, Chen, Nunez, Greenberg and Khan2014; Murata et al., Reference Murata, Kobayashi, Imuta, Haraguchi, Ieiri, Nishimura and Mine2010). The reductions made may not have had a large enough impact on receptor occupancy to trigger a withdrawal reaction (Horowitz & Taylor, Reference Horowitz and Taylor2019).

The study by Rapoport et al. should not have been included because the number of participants who experienced a withdrawal symptom is not reported. Henssler et al. used the number of people who withdrew from the trial due to a discontinuation-emergent adverse effect (Rapaport et al., Reference Rapaport, Wolkow, Rubin, Hackett, Pollack and Ota2001).

Observation periods

Observation periods in the studies included in Henssler et al.’s review were generally short – the mode was two weeks. Short follow-up periods are likely to miss some withdrawal effects, which may not necessarily start immediately (Stockmann, Odegbaro, Timimi, & Moncrieff, Reference Stockmann, Odegbaro, Timimi and Moncrieff2018) due to receptor occupancy taking weeks to fall for many drugs (not just fluoxetine)(Sørensen, Ruhé, & Munkholm, Reference Sørensen, Ruhé and Munkholm2022b) or the accumulation of downstream effects that are not well understood (Horowitz & Taylor, Reference Horowitz and Taylor2024).

Duration of treatment

Previous research has shown that the incidence and severity of antidepressant withdrawal effects are greater following long-term use (Horowitz et al., Reference Horowitz, Framer, Hengartner, Sørensen and Taylor2023; NICE, 2022a).

The weighted average duration of exposure to antidepressants in the 58 studies included in the incidence analysis, which reported this data, was less than six months (23.4 weeks). In 30 of these, participants had used antidepressants for less than three months, and only nine involved a majority of participants who had taken antidepressants for a year or more (Table 1). Moreover, in two of these, figures that underestimate withdrawal events were inadvertently used in Henssler et al.’s analysis (Rapaport et al., Reference Rapaport, Wolkow, Rubin, Hackett, Pollack and Ota2001; Tourian, Pitrosky, Padmanabhan, & Rosas, Reference Tourian, Pitrosky, Padmanabhan and Rosas2011).

Placebo withdrawal

Henssler’s final estimates were computed by subtracting the incidence of withdrawal effects reported following the discontinuation of a placebo (nocebo effects), in trials that reported this data, from the incidence rate among people who had withdrawn from an antidepressant, derived from a larger group of studies. Although the occurrence of nocebo effects, or the misclassification of non-specific symptoms as withdrawal-related effects, is an important consideration, the use of different groups of studies to estimate antidepressant and placebo withdrawal contravenes recommendations because of likely differences between the groups (Glenny et al., Reference Glenny, Altman, Song, Sakarovitch, Deeks and D’Amico2005).

Henssler et al.’s strategy also assumes that the adverse effects reported by people withdrawing from a placebo and an antidepressant are the same. However, it is unlikely that these are ‘like for like’.

Antidepressant withdrawal is associated with common, nonspecific symptoms such as dizziness, headache, and anxiety (as well as more specific symptoms, such as electric ‘zaps’). These will occur to some extent as ‘background noise’ in the placebo group, as highlighted by Baldwin, Montgomery, Nil, and Lader (Reference Baldwin, Montgomery, Nil and Lader2007). However, antidepressant withdrawal symptoms are likely to be more severe and occur more frequently. It has been reported that people can become so dizzy they have physical accidents (Moncrieff et al., Reference Moncrieff, Read and Horowitz2024) or be referred for neurological workups (Haddad, Devarajan, & Dursun, Reference Haddad, Devarajan and Dursun2001), for example. Therefore, incidental or background symptoms can only be distinguished from genuine withdrawal symptoms by measuring their severity and frequency, in the same way that symptoms of anxiety and depression are usually rated. Only one study included in Henssler et al. reported the severity of individual symptoms, but there was no placebo group in this study (Khan et al., Reference Khan, Musgnung, Ramey, Messig, Buckley and Ninan2014).

This point is supported by the fact that although withdrawal symptoms in general were only about twice as frequent among people taking an antidepressant compared to those taking a placebo in Henssler et al.’s analysis, the limited indicators of severe withdrawal used (see below) were almost five times more common in antidepressant users.

Severity

Table 3 shows the data used by Henssler et al. to calculate the proportion of people experiencing severe withdrawal effects. This was not presented in the paper, and readers might assume the figures referred to withdrawal symptoms whose severity had been measured using an instrument, or at least to adverse effects that had been judged to be severe. However, in 11 of the 19 studies, the analysis was based on figures for adverse events that led to study discontinuation or on Serious Adverse Events (SAEs). The basis for the selection of these particular studies is unclear since others presented such data.

Table 3. Studies used in Henssler et al.’s analysis of the incidence of severe withdrawal (eFigure 2 in Henssler’s supplementary appendix)

References for articles not cited in-text can be found in the supplementary material.

Abbreviation: SAE, serious adverse event.

^a The SAEs in this study are described as occurring either during treatment with the antidepressants or in the 14-day period after discontinuation, so they are not necessarily related to withdrawal.

^b Participants did not necessarily stop the drug completely in this study, and there were high rates of use of concomitant medications, including benzodiazepines (see Supplementary Table S1).

^c According to Henssler et al. (personal communication), this was based on the authors’ conclusions. In the paper, the authors justify this on the basis that there was no excess risk of early relapse in the placebo group during the relapse prevention trial, and that the mean number of DESS symptoms following discontinuation of agomelatine and switch to placebo at the end of the trial was similar to the mean among those who continued agomelatine.

In any case, neither is a valid indicator of the severity of withdrawal symptoms. Decisions to discontinue from a trial involve many considerations. Researchers usually make concerted efforts to retain participants so as not to lose data and power, thereby making it likely that only unusually severe events culminate in someone leaving a trial. SAEs are a formal category of events with a precise definition, which includes events that lead to death, are life-threatening, lead to hospital admission, cause persistent or significant disability or incapacity, or a congenital abnormality (Health Research Authority, 2024). Therefore, there is a high threshold for categorizing an event as an SAE, and severe symptoms, even if painful, uncomfortable or debilitating, would rarely qualify, especially after short-term exposure.

The few studies that reported authors’ qualitative assessment of the severity of withdrawal symptoms yielded varied results. Some suggested that symptoms were generally mild (Kragh-Sorensen et al., Reference Kragh-Sorensen, Hansen, Larsen, Nasestoft and Hvidberg1974; Rickels, Schweizer, Weiss, & Zavodnick, Reference Rickels, Schweizer, Weiss and Zavodnick1993), and some suggested that they were commonly or not infrequently severe (Dallal & Chouinard, Reference Dallal and Chouinard1998; Kramer, Klein, & Fink, Reference Kramer, Klein and Fink1961; Murata et al., Reference Murata, Kobayashi, Imuta, Haraguchi, Ieiri, Nishimura and Mine2010). The authors of one noted that the symptoms of withdrawal in general were ‘fairly distressing and uncomfortable’ and that people who had severe withdrawal had ‘very distressing symptoms’ (p. 16) (Murata et al., Reference Murata, Kobayashi, Imuta, Haraguchi, Ieiri, Nishimura and Mine2010).

In another small study identified by Henssler et al. but not included in their analysis of severity, 12 of 14 participants who abruptly stopped fluvoxamine after 7–8 months experienced withdrawal symptoms, and of these five had to take time off work, six contacted researchers for help, three sought medical attention, one was re-medicated because of panic, and one became suicidal (Black, Wesner, & Gabel, Reference Black, Wesner and Gabel1993). Incidentally, it is also interesting to note that several studies documented rare cases of hospitalization and other serious events that were considered likely or possible complications of withdrawal (see Supplementary Table S1).

Meta-analysis of studies using a systematic and relevant assessment of withdrawal symptoms

We identified five studies that conducted a systematic assessment of withdrawal symptoms using an appropriate structured instrument or method (Figure 1; Table 2). In all but one of these trials, participants had used antidepressants for 12 weeks or less. In one trial of paroxetine withdrawal, the mean duration of prior use was 106 weeks, but 59% of participants in this trial underwent a very slow withdrawal over a period of up to four years and not all participants discontinued their antidepressant. The majority were also using concomitant benzodiazepines and other drugs prescribed for depression and anxiety (Murata et al., Reference Murata, Kobayashi, Imuta, Haraguchi, Ieiri, Nishimura and Mine2010). All five studies were rated as having a low probability of the misclassification of withdrawal and relapse (Table 1).

The pooled rate of withdrawal symptoms in all five trials was 0.55 (95% confidence interval CI, 0.36–0.72, N = 601) using a random effects model, without subtracting nocebo effects (Figure 2). Heterogeneity was high (I ² = 77%; τ ² = 0.59; Q = 17.1, df = 4, p = 0.002). Excluding the trial by Murata, in which not all participants stopped their antidepressant, yielded a pooled estimate of 0.61 (CI 0.38–0.80; N = 545; I ² = 74.4%; τ ² = 0.67; Q = 11.7, df = 3, p = 0.009). Removing the trial of agomelatine (Stein, Ahokas, & de Bodinat, Reference Stein, Ahokas and de Bodinat2008), which has a different mechanism of action from other antidepressants and has consistently been found to have a low potential for dependence (Goodwin, Emsley, Rembry, & Rouillon, Reference Goodwin, Emsley, Rembry and Rouillon2009; Montgomery et al., Reference Montgomery, Kennedy, Burrows, Lejoyeux and Hindmarch2004), produced an estimate of 0.69 (0.43–0.87; N = 482; I ² = 72.7%; τ ² = 0.63; Q = 7.3, df = 2, p = 0.02) (see Supplementary material Figures S1–S2).

Figure 2. Meta-analysis of studies using a structured instrument or method of assessment of antidepressant withdrawal symptoms: Forest Plot.

These figures are likely to include nocebo withdrawal or incidental symptoms. Although these were not reliably measured in the original review, for illustration, we deducted Henssler et al.’s estimate of placebo withdrawal in trials using a structured instrument (30%) from our estimates. This resulted in a range of 25%–39% of people experiencing withdrawal symptoms.