Dose-Response Relationship in Selective Serotonin and Norepinephrine Reuptake Inhibitors in the Treatment of Major Depressive Disorder: A Meta-Analysis and Network Meta-Analysis of Randomized Controlled Trials

Selective serotonin and norepinephrine reuptake inhibitors (SNRI), such as venlafaxine, and duloxetine, are among the most often prescribed antidepressants [1‒3]. Antidepressant dose escalation is common in the clinical routine if there is insufficient response to initial treatment [4]. This practice has been critically discussed, for example by Dold et al. [5], who emphasized a lack of evidence. Also, the guideline of the World Federation of Societies of Biological Psychiatry is cautious with regard to dose increase in case of non-response, but it states that venlafaxine seemed to be more effective at higher doses [6]. Moreover, the authors of the 2020 guideline of the Royal Australian and New Zealand College of Psychiatrists firmly recommend a dose increase of antidepressants in the event of insufficient response to normal recommended doses [7]. This advice is based on the assumption that there is a dose-response relationship of antidepressants, including SNRIs.

However, there is a dearth of research on dose-response relationships in SNRIs. An early meta-analysis on antidepressants included the first studies on milnacipran and venlafaxine, and the authors found an imipramine equivalent of 100–200 mg most effective but did not specify results in regard to single substances [8].

In one of the more recent meta-analyses, Furukawa et al. [9] focused on SSRI, venlafaxine, and mirtazapine. They reported no clear evidence of clinically meaningful effect gradients along different doses. This study largely rested on indirect evidence through a comparison via placebo.

To our knowledge, there is no meta-analysis investigating direct comparisons of various SNRI doses. We therefore investigated the efficacy of different doses of SNRIs as a group and specific SNRIs in direct comparisons in patients with major depressive disorder. As a sensitivity analysis, we employed a network meta-analysis (NMA).

This systematic literature review and meta-analysis is part of a larger project on dose escalation and dose-response relationships in pharmacologic antidepressant treatment, registered in Prospero (https://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD42018081031) [10, 11]. Our literature search was based on the Cochrane Handbook for Systematic Reviews of Interventions [12]. No language or date restrictions applied; grey literature was not excluded. We focused our search on the CENTRAL database and complemented our results with systematic searches in PubMed, EMBASE, and PsycINFO for studies published after December 31, 2011 (last updates: Embase, February 2017; CENTRAL, PubMed, and PsycINFO: March 2020). We used generic terms for unipolar depression, dose increase, and randomized controlled trials. The following search terms serve as an example and reflect that the present study is embedded in a larger project (CENTRAL):

MeSH descriptor: [Dose-Response Relationship, Drug] explode all trees

#2 ((dose (tw) or dosage (tw)) and (increase (tw) or escalat* or elevat* or raise))

#3 ((dose (tw) or dosage (tw)) and ((maxim* (tw)) or (upward (tw) and titrat* (tw))))

#4 dose-effect or high-dose

#5 “dose-response relationship”

#6 #2 or #3 or ((#1 or #4) or #5)

#7 depress* or dysthymi* or adjustment disorder* or mood disorder* or "affective disorder" or "affective symptoms"

#8 antidepressant* or agomelatin* or amineptin* or amitriptylin* or amoxapin* or bupropion* or butriptylin* or chlorimipramin* or citalopram* or clomipramin* or desipramin* or desvenlafaxin* or dibenzepin* or dosulepin* or dothiepin* or doxepin* or duloxetin* or escitalopram* or fluoxetin* or fluvoxamin* or imipramin* or isocarboxazid* or lofepramin* or levomilnacipran* or MAOI* or "monoamine oxidase inhibitors" or maprotilin* or mianserin* or milnacipran* or mirtazapin* or moclobemid* or nefazodon* or nortriptylin* or paroxetin* or phenelzin* or protriptylin* or reboxetin* or selegilin* or sertralin* or setiptilin* or SSRI or SSNRI* or SNRI* or tca or "selective serotonin reuptake inhibitors" or "serotonin-norepinephrine reuptake inhibitors" or tetracyclic* or tianeptin* or tranylcypromin* or trazodon* or trimipramin* or tricyclic* or venlafaxin* or viloxazin* or vortioxetin*

#9 #7 and #8 and #6

Online supplementary Table 1 (for all online suppl. material, see www.karger.com/doi/10.1159/000520554) provides details of the search algorithm for EMBASE, PsycInfo, and PubMed.

In addition, reference lists of all review articles retrieved and of all studies included were hand searched. Finally, we carried out a forward search in Web of Science, screening all articles that cited the papers included.

All search results, titles, and abstracts were screened by two authors (L.R., C.Br.) independently. Studies that seemed eligible were read in full text. If eligibility remained unclear, consent was found through discussion among others (L.R., C.Br., C.Ba.).

• RCTs randomizing patients to at least two different doses of SNRI monotherapy;

• Patients diagnosed with major depressive disorder according to established diagnostic criteria, such as ICD [13], or DSM [14];

• Documentation of treatment success or failure based on established scales, e.g., Hamilton Depression. Rating Scale (HDRS-17 or higher) [15], or Montgomery-Asberg Scale for Depression (MADRS) [16].

• Different doses that were not investigated in parallel groups of different patients but consecutively in the same group of patients, as in schemes with one dose as the second step, another as the third step of antidepressant treatment algorithms;

• Dose increase after non-response if resulting in different durations of exposure to a dosage: for example, one arm dose increase, one arm continuation;

• Studies on secondary depressive syndromes, e.g., post-stroke depression;

• Parallel treatment with another antidepressant or augmentation therapy in the same study arm;

• Less than 3 weeks of active treatment.

We used a standardized Excel-form to extract data, as employed in earlier studies by our group [17]. Data extraction was carried out by two authors independently (L.R., C.Br.). If data were missing, trial authors or manufacturers were contacted. Data only available from graphs were read out by two authors independently (L.R., C.Br.) using Plot digitizer software (SourceForge Project; sourceforge.net) [18].

To assess each study’s risk of bias we applied the Cochrane Collaboration tool for assessing risk of bias [19]. Studies were evaluated by 2 authors independently (L.R., C.Br.). At least five out of seven domains of the tool needed to be graded “low” to consider a study as carrying a low risk of bias.

The primary outcome of this study was the difference in efficacy between the doses under study as represented in depression rating scores. If more than one scale was used, we included results according to the following ranking defined a priori: (1) HDRS (Hamilton Depression Rating Scale) [15], (2) MADRS (Montgomery-Asberg Depression Rating Scale) [16], (3) BRMS (Bech-Rafaelsen Melancholia Rating Scale) [20], (4) other scales. Change from baseline was preferred over follow-up value and other endpoints. Intention-to-treat data were preferred, and we followed the method applied by the study authors, e.g., LOCF (last observation carried forward).

As the secondary outcome we carried out analyses of dropouts due to any reason and of dropouts due to adverse events. To keep analyses at a reasonable number we decided a priori to not analyze responder and remitter rates. In one case [21] we requested individual patient data from the manufacturer to calculate necessary outcomes.

We measured efficacy using standardized mean differences (SMD) and 95% confidence intervals (CI). We followed Cohen in considering 0.2 the lower limit of a small effect and, in accordance with our analysis of the SSRI dose-response relationship, classified an SMD of 0.2 as the lower bound of a clinically meaningful effect [22]. If change from baseline was not available, we computed it from baseline and final scores. In case of missing standard deviations (SD) we derived it from CI or p values. If neither SD, SE, CI, or p values were reported, we imputed SDs from similar studies according to a validated imputation method or from baseline and final scores assuming a correlation coefficient of 0.5 [23].

We investigated dose-response relationships with two approaches. Firstly, we rated SNRI doses as low, medium, and high following manufactures’ drug information such as product monographs (Table 1) to analyze dose-response across all SNRIs. We defined a “low dose” as all doses below the recommended daily dosage and “high dose” as the highest daily dose recommended by the product monographs and all doses above. All doses recommended for clinical use lower than the “high dose” definition were rated as medium dose. This was the primary analysis of the present study.

Secondly, we performed meta-analyses regarding specific SNRIs including all direct dose comparisons if at least two comparisons were available (e.g., desvenlafaxine 50 vs. 100 mg), in order to investigate the dose-effects across and within the three dosage levels.

We applied random-effects meta-analyses throughout and the I² statistic as a measure of heterogeneity [24]. All calculations were carried out with Comprehensive Meta-Analysis, version 2, SPSS, and R [25, 26].

An NMA was carried out as an a priori planned sensitivity analysis of the comparison of low, medium, and high levels of SNRIs doses. The NMA aimed at utilizing all published comparisons of different doses of SNRIs, directly and indirectly, and included placebo arms as the third comparator. Combining sufficiently similar data, we used the frequentist weighted least squares approach and random effects meta-analyses.

We tested the robustness of our results by deleting studies that appeared to increase heterogeneity. Finally, we estimated risk of bias with funnel plots and Egger’s test contingent on a sufficient number of studies.

This meta-analysis investigated whether, in patients diagnosed with major depressive disorder (P), different doses of SNRIs (I, C) exert different antidepressant efficacy (O).

Our literature search retrieved 2,558 articles out of which we screened 2,070 article titles and abstracts, evaluated 224 full texts, and included 26 studies for quantitative analysis (shown in online suppl. Fig. 1 and online suppl. Table 2) [21, 27‒51]. Desvenlafaxine was used in 10 studies, duloxetine in 5, levomilnacipran in 2, milnacipran in 4, and venlafaxine in 5.

In 47 direct comparisons, studies compared 25 SNRI doses in 30 different dose pairs. Across all studies, the ratio of the highest and the lowest dose investigated averaged 3.3 (95% CI 1.94–4.73). Two studies were open-label, all other trials were at least double-blinded (online suppl. Table 2). Six studies did not publish SD or SE and we imputed measures of dispersion. Overall, 7,695 patients received active drugs, and 2,547 placebo. Female patients represented 65% of all subjects. Across the studies, the median age was 42 years.

We included 14 comparisons of different dose levels in pairwise meta-analyses of SNRIs as a group, 6 contrasting low to medium and 8 medium to high doses. There was no investigation of low versus high doses.

The summary SMD of high versus medium doses was – 0.06 (–0.16 to 0.03) and –0.06 (–0.16 to 0.04) in medium versus low doses, favoring higher doses. I² was 0% (medium vs. low) and 19% (high vs. medium; Table 2; Fig. 1) [27, 31, 33‒35, 37‒39, 43, 44, 46‒49].

Within the 47 direct comparisons, 11 dose pairs were represented at least twice and thus eligible for the meta-analysis. We found no statistically significant differences between different doses of specific SNRIs (shown in online suppl. Table 4). Some SMDs were substantial, for example, –0.56 in one RCT comparing 75 versus 150 mg of milnacipran, and, generally, the CIs were wide and did not preclude considerable effect sizes [40].

One meta-analytic result bordered on statistical significance: desvenlafaxine 100 mg may be more effective than 200 mg (p = 0.054), with an SMD of 0.26 (–0.004 to 0.522; 3 studies, I² 50%) [33, 45, 46]. In the best-studied dose comparison, desvenlafaxine 50 versus 100 mg, no substantial difference emerged: SMD –0.054 (–0.151 to 0.042; 6 studies, I² 11%) [30, 32, 42, 45, 46, 50].

Relative to medium doses, dropouts due to any reason were not statistically significantly more frequent in high doses, with a relative risk of 1.05 (95% CI 0.91 to 1.20), as were dropouts due to adverse events (RR 1.16; 95% CI 0.92 to 1.48) [33, 34, 37, 39, 47‒49]. Comparing medium and low doses, patients receiving medium doses of SNRI were less likely to drop out for any reason (RR 0.84; 95% CI 0.68 to 1.05) [27, 35, 38, 43, 44, 46]. However, the relation was reversed regarding dropouts due to adverse events: RR 1.31 (95% CI 0.83 to 2.08) [35, 38, 43, 44, 46].

Six studies carried a low risk of bias, 18 trials were considered unclear in this regard, and 2 had a high risk of bias. Owing to the small number of low-risk studies, we included studies with low or unclear risk (shown in online suppl. Table 3), and thus excluded NCT 00619619 from the analysis of medium versus low doses: SMD –0.06 (–0.17 to 0.05; I² 4%) [46]. Analyzed in the same way, the difference between high and medium doses remained similar: SMD –0.06 (–0.16 to 0.03; I² 19%). Additionally, we calculated estimates based on studies with low risk of bias especially, without change in SMDs of medium versus low (–0.05; –0.25 to 0.15) and high versus medium (0.02; –0.19 to 0.23) comparisons [31, 38, 43].

Funnel plots and Egger’s tests revealed no indication of small study effects in medium versus high and low versus medium comparisons (p = 0.95 and 0.83, respectively). Owing to the lack of studies, no reporting bias analysis was carried out for the comparisons of low versus high doses.

After excluding Emslie et al. [35] and NCT00619619 [46] because they included only children and adolescents, the contrast between medium versus low dosage levels remained unchanged (data not shown).

We carried out a NMA as a sensitivity analysis. With 14 studies in the network and little inconsistency, high versus medium and medium versus low dose level contrasts stayed in the same range [27, 31, 33‒35, 37‒39, 43, 44, 46‒49]. An effect of –0.12 (–0.26 to 0.02) SMD was estimated for high relative to low dose levels (Table 2). When we added placebo arms – resulting in 23 studies in the network – high, medium, and low dosages of SNRIs were statistically significantly superior to placebo (shown in online suppl. Table 5) [27‒39, 41‒44, 46‒51]. High versus medium and medium versus low dose effect comparisons were similar to our primary outcome analysis and in the NMA without placebo (data not shown). The estimate of high versus low level dosages resulted in an SMD of –0.24 (–0.48 to 0.00), favoring high doses and with substantial inconsistency (I² 76%).

In the NMA including placebo arms we found substantially higher dropout rates in active than in placebo arms. We estimated the relative risk of high versus low doses, where direct comparisons are not available, as 1.02 (0.81 to 1.30) in favor of low doses. Risk differentials from direct comparisons were confirmed (Table 3).

This study yielded three important results. Firstly, there is no high-level evidence in favor of a dose-response relationship for SNRIs as a group or for specific SNRIs. Secondly, higher doses of SNRIs are associated with nominally more dropouts related to adverse events. Thirdly, confidence intervals for several compared doses are wide.

At the level of different doses of SNRIs as a group, effects of both comparisons are far below our definition of a clinically meaningful effect (SMD 0.2). With p values invariably well above 0.05, none of the comparisons in our primary analysis became statistically significant.

Our main result appears to be robust and has been corroborated in additional calculations: risk of bias analyses returned similar findings, as did the exclusion of single studies. There is no indication of reporting or publication bias, even though the number of studies is small and preclude definitive judgment. However, no values needed to be added in a trim and fill procedure. Summary estimates of two network meta-analyses support what we saw in primary analyses. Heterogeneity is low and CIs are reasonably small.

On the other hand, we lack direct evidence comparing high versus low doses. Our NMA of indirect comparisons suggests a small but clinically relevant effect of high doses (SMD –0.24). This finding is uncertain because it is based on indirect evidence only, heterogeneity is considerable, the CI includes a negligible effect, and it is not adjusted for multiplicity.

Broken down to different doses of specific substances, we did not find a single statistically significant comparison, although some comparisons clearly were underpowered with n in five studies at or below 15 per arm. Assuming a mid-range effect size of 0.5 SMD, two arms of 64 patients each are necessary to show a positive signal with a likelihood of 80%. By this standard, 7 of 30 comparisons in online supplementary Table 4 were insufficiently powered. For a small effect, such as an SMD of 0.2, the lower bound of our definition of a clinically meaningful effect, almost all sample sizes are below the roughly 800 participants needed. Accordingly, for several comparisons, CIs are wide and substantial effects cannot be ruled out. Also, extreme comparisons, at the range of a dose ratio of 10 or more, have only rarely been studied. Still, it is possible to state with reasonable certainty that some higher doses will not result in more efficacy: for example, 100 mg as opposed to 50 mg desvenlafaxine. Consequently, while analyses at the level of single substances remain inconclusive at this point, no positive evidence for a dose-response relationship of SNRIs emerged from this line of inquiry.

When compared to placebo in a sensitivity NMA, we found high, medium, and low doses of SNRI effective, a finding that points to a general problem in interpretation: indirect evidence can suggest a dose-response relationship while direct evidence is not indicative of such an effect. This may explain the differences in results between the NMA by Furukawa et al. [9] and our approach. In principle, NMA is a powerful tool. However, by using indirect comparisons it adds another layer of variance to conventional meta-analysis, a method already at risk of over simplification. For example, placebos are considered a particularly reliable common comparator, but Breilmann et al. [52] showed that results from placebo arms can vary widely across RCTs. Therefore, to privilege results from direct comparisons is the more conservative way of reasoning. Still, it is a very reasonable future research agenda to compare the results of the present investigation with findings from NMA.

We view our largely negative results as preliminary and as less secure than our findings on SSRIs. In general, however, while dose-response relationships may be a possibility in tricyclic antidepressants or tranylcypromine [53, 54], the findings accord well with our earlier meta-analysis of SSRI dose-response relationships [11]: comparing different dose levels of SSRIs at the group level, we saw no improvements in efficacy with increasing, or decreasing, dose level. No clear-cut advantages of higher doses of specific SSRIs were detected – quite the opposite: the only robust finding showed 60 mg fluoxetine to be statistically significantly inferior to 20 mg. The results are also in line with earlier meta-analyses on dose escalation in treatment-resistant depression [5, 10], not suggesting dose increase as a promising second step strategy. Non-response, for the purposes of the present discussion is defined by less than 50% improvement on a depression severity scale, for example HAM-D-21, after at least 4 weeks of treatment with an adequate dose of an antidepressant. However, a universally agreed upon definition of treatment resistance is yet to be found, as Bartova et al. [53] have discussed.

Beyond its definition, our view of treatment resistance may generally need refinement. Moreover, Fava et al. [55] proposed the consideration of antidepressant pharmacotherapy not only as helpful but also as potentially associated with paradoxical, rebound, and withdrawal phenomena and thereby contributing to the problem of treatment failure.

In contrast to the present investigation, higher doses of SSRIs came with statistically significantly more dropouts linked to adverse events. As for dropouts due to adverse effects among patients treated with SNRIs, the confidence interval ranges up to a 2-fold risk for medium relative to low doses and, thus, we cannot exclude a relevant effect. Our NMA points to the possibility of 15–40% more side effect-associated dropouts with higher doses compared to lower ones, and yet this has to be viewed as a plausible but preliminary finding because it is entirely based on indirect comparisons.

In addition to side effects during dose escalation treatment, recent publications point to withdrawal syndromes after SNRI treatment discontinuation that may be particularly relevant if doses have been high. In recent systematic reviews, withdrawal symptoms were reported after discontinuation of any type of SNRI but seemed most prevalent with venlafaxine [56, 57]. In the same vein, Cosci and Chouinard [58] concluded in their comprehensive review that among antidepressants, upon discontinuation or dose decrease, especially SSRIs and SNRIs can induce: (1) new withdrawal symptoms (often unspecific, e.g., headache, fatigue), (2) rebound syndromes (e.g., depression, anxiety), and (3) persistent post-withdrawal symptoms, such as mood and anxiety psychopathology, but also tinnitus, nausea, and dizziness. Again, venlafaxine figured prominently in the reports reviewed [58]. In general, the risk of withdrawal or discontinuation syndromes appears to increase when antidepressants have been taken in high doses, such as 120 mg duloxetine [57].

Among the studies in this meta-analysis, only 3 desvenlafaxine papers [30, 42, 50] provided detailed data on withdrawal symptoms as measured by the discontinuation emergent signs and symptoms instrument (DESS). In all studies, DESS scores were statistically significantly higher in verum than in placebo arms at some point during tapering.

To our knowledge, this meta-analysis is the first focussing on SNRI as well as the first analysing SNRIs other than milnacipran and venlafaxine. Compared to an earlier meta-analysis by Bollini et al. [8], we included 24 new studies.

In light of the common practice of dose escalation, the number of comparisons testing different SNRI doses is relatively small. Several dose gradients have not been investigated, especially high versus low ones, and almost all single studies that have been carried out are underpowered for detecting small effects. In such a situation, meta-analytical approaches seem particularly apt. Still, it is possible we missed studies. To complement our literature search in CENTRAL, we searched Embase, PsycINFO, and all NIH databases via PubMed for studies published in the last decade. Apart from screening narrative review articles, we also approached authors and, in one case, a manufacturer. We are therefore confident that the current investigation is reasonably based on the available evidence.

Due to its short time of active treatment and in consistence with our earlier meta-analyses, we excluded 2 studies by Dunner et al. [59] and Whitmyer et al. [60]. Both studies compared early dosing strategies of duloxetine with doses of 30 and 60 mg. Taken together, they reach an SMD of –0.048 (–0.195 to 0.099). With a p value of 0.52, those findings are in line with our findings.

By the standards of the Cochrane collaboration, most trial reports in this meta-analysis carried a substantial risk of bias, as is common in antidepressant research [11, 61, 62]. However, the available RCTs represent the best evidence to hand, and we do not see better studies to rely on.

Subdividing dosage levels into low, medium, and high along the recommendations by pharmaceutical manufacturers is only one of several conceivable categorizations. However, the categorization chosen is not arbitrary but uses boundaries that are investigator independent. Of note, manufacturer recommendations may differ slightly from country to country.

Doses used for oral treatment do not necessarily reflect serum concentrations of the compounds under study. Individual patients with doses usually sufficient for therapeutic serum concentrations quite possibly may show lower concentrations on testing. A comprehensive review and consensus guideline recommends the use of TDM (therapeutic drug monitoring) at least for the SNRIs duloxetine, milnacipran, and venlafaxine [63]. In non-responders to these drugs, it may be wise to measure serum levels and escalate the dose in case of a serum level below the recommended range – a situation that may arise among the small patient group of rapid metabolizers [5, 53]. Other reasons for a subtherapeutic serum level need to be excluded before raising the dose.

Of note, we did not include studies on patients with treatment-resistant depression in the analysis because dose increase after non-response is the topic of an earlier meta-analysis by our group [10]. In that systematic review, we did not find any studies on SNRI dose increase in treatment resistant-depression. As a result, we cannot rule out that SNRI may be effective in that indication. However, there is no positive evidence for SNRI dose increase as a second-step strategy, neither directly, from studies among patients with treatment resistant depression, nor indirectly, from the present analysis on dose-response relationships in SNRIs. Initially, we had planned to use regression slopes for various doses and effects – a method that assumes a linear log-dose-effect relationship. Unfortunately, we did not find such a relationship and were forced to change our method. While it was inevitable in our investigation, and, to us, seems hardly unusual in meta-analyses, it deserves mention as a deviation from the protocol published on Prospero.

Although there are gaps in SNRI dosage studies, little evidence suggests new dose-effect trials. If anything, comparisons between low and high doses are under-researched, and possibly also studies on medium versus high doses of venlafaxine, an antidepressant prescribed particularly often. If results from the present investigation and from SSRIs are any guide, a dose increase study in treatment refractory patients seems not to hold much promise.

For clinical practice we advocate, in uncomplicated cases, to stay within medium recommended doses. In light of other, more evidence-based second step strategies in antidepressant pharmacotherapy, such as second-generation antipsychotic augmentation [5, 6, 64], lithium augmentation [65, 66], or combination of antidepressants [67], we do not advise colleagues to increase the dose of an SNRI after a first treatment attempt has failed ‒ except in the case of subtherapeutic serum levels. While we cannot rule-out increased efficacy with increased doses regarding single substances, it is fair to say that we lack positive evidence for this often-employed treatment strategy.

This publication is based on research using data from Eli Lilly and Company that has been made available through Vivli Inc. Vivli has not contributed to or approved, and is not in any way responsible for, the content of this publication.

There was no funding source for this study.

Lena Rink and Anne Adams contributed equally to this work.