Systematic Review and Meta-Analysis: Dose-Response Relationship of Selective Serotonin Reuptake Inhibitors in Major Depressive Disorder

While the efficacy of selective serotonin reuptake inhibitor (SSRI) medications and their widespread use is generally accepted in major depressive disorder (especially in more severe cases), there remains some uncertainty as to the optimal dose for SSRI pharmacotherapy of major depressive disorder (1–9). Current APA Practice Guidelines state that “optimizing the medication dose is a reasonable first step if the side effect burden is tolerable and the upper limit of a medication dose has not been reached” (3). This recommendation is based on level-II evidence indicating that escalating antidepressant doses was “recommended based on moderate clinical confidence” (3).

Based on currently available evidence, the APA Practice Guidelines regarding antidepressant dosing appear reasonable. A previous meta-analysis examining dosing of antidepressant medications in major depressive disorder demonstrated a flat dose-response curve within the therapeutic range for antidepressant medications (≥100-mg imipramine equivalents) (10). Furthermore, the meta-analysis demonstrated a greater side effect burden at higher doses as evidenced by an escalating adverse events rate with increasing dosage of antidepressants (10). Although this meta-analysis employed quite advanced methodology for the time, the findings may be somewhat antiquated for use in clinical practice for several reasons. First, the authors grouped other classes of antidepressants (monoamine oxidase inhibitors, tricyclic antidepressants, and atypical antidepressants) alongside SSRIs. Other antidepressants likely have a different dose-response relationship and tolerability profile with dose when compared with SSRIs. Second, the authors examined dose as a categorical rather than continuous outcome, which may reduce overall power to detect a dosing effect in the meta-analysis. In contrast, another meta-analysis, which was quite stringent in its inclusion criteria, examined the dose-response of SSRI medication in only four fixed-dose and four dose-escalation trials in major depressive disorder. This meta-analysis demonstrated a weak positive association between higher doses and treatment response (11). This meta-analysis examining the dose-response curve in SSRI suggests the possibility that SSRIs may behave differently than other antidepressants.

The goal of the present meta-analysis is to improve the existing evidence-base regarding the dose-response relationship of SSRIs in major depressive disorder. Specifically, our goal is to determine whether there exists any evidence in SSRI trials of major depressive disorder to suggest that higher doses are associated with improved outcome. We conducted a meta-analysis and used meta-regression to examine the relationship between target SSRI dose in trials and the measured efficacy (and tolerability) of SSRI treatment compared with placebo.

Method

Search Strategy and Study Selection

A literature search was conducted on October 10, 2013 on PubMed and CENTRAL, The Cochrane Collaboration database of controlled trials (in the Cochrane Library). Published randomized controlled trials comparing all SSRIs versus placebo in short-term treatment of unipolar depression were sought by two reviewers (A.L.V. and M.H.B.), using the search term: (“SSRI”[MESH] OR “fluoxetine”[MESH] OR “fluvoxamine”[MESH] OR “citalopram”[MESH] OR “escitalopram”[MESH] OR “sertraline”[MESH] OR “paroxetine”[MESH]) AND “placebo”[MESH] AND “depression”[MESH]. Trials were included if 1) efficacy data were available for both SSRI- and placebo-treated participants for at least one time point other than baseline and endpoint and 2) they utilized standardized, validated outcome measurements of depression. Trials were excluded if 1) the age of participants was <19 years or >60 years; 2) a cross-over design was used; 3) psychiatric diagnosis other than major depressive disorder or a dual diagnosis was studied; 4) an SSRI was not studied; 5) not randomized; 6) not placebo-controlled; and 7) adjunctive psychotherapy was provided to the active or control group.

Data Extraction

Included trials provided depression ratings on the Hamilton Depression Rating Scale or the Montgomery-Åsberg Depression Rating Scale for at least three time points (baseline, endpoint, and at least one intermediate time point). If trials reported outcomes in a figure rather than in a table, a computer program (Dexter, German Astrophysical Virtual Observatory, University of Heidelberg, Heidelberg, Germany) was used to extract weekly data points from the figures (software available at http://dc.zah.uni-heidelberg.de/sdexter). Additionally, the number of treatment responders (as defined by study criteria) and the number of participants who discontinued during the course of the study (all-cause dropouts and dropouts due to side effects) were recorded. Additional data were collected on the type of SSRI, maximum dosage of medication, duration of the trial, and year of the trial. All SSRI doses were transformed into imipramine equivalent doses using previously described methodology (10).

Data Analysis

Data collection and preparation were conducted in Microsoft Excel 2007, and the effect of dose on time course of SSRI response was analyzed in SAS 9.3 (SAS Institute, Cary N.C.). We used generalized estimating equations to examine the effects of trial, treatment, modeling different forms of treatment effect, accounting for different periods within trials as repeated measures, and defining a new covariance structure for each trial by defining these as random effects. For each trial and week, the standardized mean difference in outcome scores between SSRI and placebo groups was calculated and weighted by the number of randomly assigned patients in the trial. Previous research has demonstrated that a logarithmic model provided the best fit for the time course of SSRI response compared with placebo (12, 13). The effects of SSRI were modeled using an autoregressive variance function, and the model with the lowest values on the Akaike Information Criterion was selected (14). Further details on this technique can be found elsewhere (13). We then examined the moderating effects of SSRI dosage using similar methodology. A mixed model was conducted that included the main effect of time and an interaction between SSRI dosage (in imipramine equivalents) and time. The main effect of SSRI dose was not included in the model, since this effect should be trivial. There should be no differences in depression severity (compared with placebo) at baseline. Differences should only be seen later when different SSRI doses start taking effect. Dose of SSRI was converted into imipramine equivalents based on previously defined methodology based on the therapeutic dose range of each medication (10, 15). Imipramine dose equivalents were chosen as the standard for antidepressants, since it was the first medication introduced in the class. For SSRI analysis, the dose equivalents were as follows: 100 mg of imipramine=120 mg of sertraline=100 mg of fluvoxamine=20 mg of paroxetine or fluoxetine=33.3 mg of citalopram=16.7 mg of escitalopram. We additionally tested the SSRI dose model with an additional term to account for a delayed effect of SSRI dosing. We examined models in which the dosing effect of SSRI was only included after a lag of 2, 3, and 4 weeks and the initial model with no lag to see whether a modified log model can fit the data better. For this, we coded each week as a dummy variable and ran four models adding in a three-way-interaction between week, dose, and time. The goal of analysis was to determine whether the interaction between dose and time was significant, which would indicate that there exists a delayed effect of dose response relationship for SSRI in major depressive disorder.

As an alternative method of analysis, we also examined endpoint data from included trials. We examined 1) the standardized mean difference between endpoint depression scores and 2) the odds ratio of treatment response between SSRI and placebo using Comprehensive Meta Analysis, Version 3 (Biostat, Englewood, N.J.). We conducted a meta-regression in Comprehensive Meta Analysis, Version 3, using a fixed-effects model that plotted the standardized mean difference (or odds ratio) for each trial against SSRI dose (in imipramine equivalents). A statistically significant meta-regression result would indicate an association between SSRI dose (in imipramine equivalents) and reported effect size of SSRI treatment compared with placebo. Additionally, in order to examine how our data replicated previous analysis in the area, we conducted an analysis examining previously utilized categories of SSRI dose. A stratified subgroup analysis was conducted using endpoint data with studies stratified by SSRI dosing (dose range categories: <100 mg, 100–199 mg, 200–250 mg, and >250 mg). This analysis examined the possibility that there might not be a linear association between SSRI dose and therapeutic response or that a linear relationship might exist but only up to or after a certain dose threshold. For clinician-friendly interpretation of the resultant data, we additionally converted all standardized mean difference outcomes to odds ratios in Comprehensive Meta Analysis, Version 3. We also calculated the number needed to treat or number needed to harm for each outcome based on the odds ratio and control event rate using the Center for Evidence-Based Medicine odds ratio to number needed to treat converter (16).

The analyses described thus far used all available data. The following sensitivity analyses were added to examine time and dose effects specific to intent-to-treat studies. The treatment effect was compared in intent-to-treat and completer studies both in the logarithmic model (by including an intent-to-treat status-by-time interaction) and in the endpoint data meta-analysis (via subgroup analysis). Furthermore, in the logarithmic model, the robustness of effect of dose (time-by-dose interaction) was tested by controlling for the intent-to-treat/completer study-by-time interaction. Finally, endpoint data from only intent-to-treat studies were used to conduct a meta-regression testing the effect of dose, as well as to conduct a subgroup analysis comparing the above-mentioned dose ranges.

We additionally examined the relationship between tolerability and SSRI dose in major depressive disorder trials using fixed-effects meta-regression in Comprehensive Meta Analysis, Version 2.2. Specifically, we examined the association between all-cause dropout (and dropouts due to side effects) as expressed in pooled odds ratios and SSRI dosage (in imipramine equivalents). A statistically significant meta-regression result would indicate an association between SSRI dose (in imipramine equivalents) and likelihood of participant dropout compared with placebo. Subgroup analysis was also performed between the four imipramine-equivalent SSRI dose ranges.

For all analyses, we conducted an additional sensitivity analysis excluding trials involving fluvoxamine. We chose to include fluvoxamine in our primary analysis, since fluvoxamine is an SSRI with an indication for major depressive disorder in many countries (e.g., the United Kingdom, Australia, and Russia). However, fluvoxamine does not possess a Food and Drug Administration indication for major depressive disorder and could have a different dose-response relationship compared with other SSRIs, and thus we decided to present our findings without fluvoxamine trials as a sensitivity analysis (see the figures in the data supplement accompanying the online version of this article).

Results

Included Studies

A flowchart describing the selection of eligible trials is presented in Figure 1. Our search identified 1,707 studies, and an additional four studies were identified in references of other included trials and meta-analyses in the area. Forty studies met our inclusion criteria (17–55). The included studies reported 49 active treatment arms involving 10,039 adult patients with major depressive disorder. The characteristics of the included studies are summarized in Table S1 of the online data supplement. Six different SSRIs were studied in placebo-controlled trials with major depressive disorder: fluoxetine (k=9, N=2,386) (17, 21, 29, 31, 46, 47, 49, 51, 56), fluvoxamine (k=8, N=910) (18, 24, 27, 38, 39, 42, 45, 54), paroxetine (k=16, N=3,424) (19, 20, 22, 23, 25, 26, 28, 30, 33, 34, 36, 44, 48, 52, 55), sertraline (k=3, N=865) (43, 50, 57), citalopram (k=4, N=1,349) (32, 37, 40, 50), and escitalopram (k=3, N=1,105) (37, 41, 53).

SSRI Efficacy

Best-fitting model of SSRI response.

The natural logarithmic (log_e) model of SSRI treatment response had the best model fit. Based on Akaike information criterion, the logarithmic treatment model was significantly better than a model using the square root of week (χ²=4.9, df=1, p=0.03). The estimate of treatment effect by log (week +1) from the final model was 0.32 (95% confidence interval [CI]=0.27–0.37; p<0.001). A log_e response curve indicates that the incremental SSRI benefit compared with placebo was greatest in the first week and gradually declined in magnitude as time persisted in short-term treatment trials. Models that introduced delayed treatment effects all produced similar (but worse or equivalent) model fits to when the dosing effect was introduced at baseline (week 2: χ²=0, df=1, p=1; week 3: χ²=0.6, df=1, p=1; week 4: χ²=3.7, df=1, p=0.054).

Dose-response curve in continuous model of SSRI response.

The logarithmic models at different imipramine equivalent dose isoquants are presented in Figure 2. There was significant effect of time (log [week +1]=0.23 [95% CI=0.13–0.33; p<0.0001]) and a significant interaction between dose and time (interaction=0.0007 [95% CI=0.0001–0.0013; p=0.0196]). This result indicates that higher doses of SSRIs were associated with a greater therapeutic response. In sensitivity analysis, the dose-by-time interaction remained significant when use of non-intent-to-treat analysis was adjusted for in the model (interaction=0.0007 [95% CI=0.0000–0.0014; p=0.00480]). Similarly, when fluvoxamine trials were excluded from the analysis, there remained a significant dose-by-time interaction (interaction=0.0008 [95% CI=0.0002–0.0014; p<0.001]) (see Figure S2 in the online data supplement).

Traditional meta-analysis examining depression severity.

Meta-regression described a significant association between SSRI dose (in imipramine equivalents) and measured efficacy of SSRIs in reducing depression severity (β=0.00053, 95% CI=0.00018–0.00088, z=2.98, p=0.0029). A scatterplot that depicts the relationship between imipramine equivalent dose of SSRIs and measured efficacy of SSRIs compared with placebo in terms of standardized mean difference is presented in Figure 3A. In sensitivity analysis, this result remained significant when restricted to trials using intent-to-treat analysis (β=0.00062, 95% CI=0.00025–0.00098, z=3.32, p=0.00090) but not when trials involving fluvoxamine were excluded (β=0.00029, 95% CI= –0.00010 to 0.00066, z=1.44, p=0.15) (see Figure S3A in the data supplement).

When SSRI dose was examined as a specific dosing category rather than as a continuous variable, there remained a significant effect of dose (test for subgroup differences: χ²=54.4, df=3, p<0.001). The estimated efficacy of each SSRI dose category compared with placebo is described in Figure 3B. The greatest measured efficacy of SSRIs was observed in the dosing range of 200–250 imipramine equivalents. In sensitivity analysis, the differences between groups remained significant when restricted exclusively to trials employing intent-to-treat analysis (χ²=56.2, df=3, p<0.001) or when fluvoxamine trials were excluded (χ²=42.4, df=3, p<0.001) (see Figure S3B in the data supplement).

Traditional meta-analysis examining treatment response.

Meta-regression demonstrated a significant association between SSRI dose (in imipramine equivalents) and measured efficacy of SSRIs with regard to the odds ratio of treatment response (β=0.0016, 95% CI=0.0005–0.0027, z=2.86, p=0.004). A scatterplot that depicts the relationship between imipramine equivalent dose of SSRIs and measured efficacy of SSRIs compared with placebo in terms of the odds ratio of treatment response is presented in Figure 3C. In sensitivity analysis, this result remained significant when restricted to trials using intent-to-treat analysis (β=0.00062, 95% CI=0.00025–0.00098, z=3.32, p=0.00090) and when trials involving fluvoxamine were excluded (β=0.0015, 95% CI=0.0003–0.0026, z=2.47, p=0.013) (see Figure S3C in the data supplement).

When SSRI dose was examined as a specific dosing category rather than as a continuous variable, there remained a significant effect of dose (test for subgroup differences: χ²=14.5, df=3, p=0.002). The odds ratio of each SSRI dose category compared with placebo is presented in Figure 3D. The greatest measured efficacy of SSRIs was again observed in the dosing range of 200–250 imipramine equivalents. In sensitivity analysis, the differences between groups remained significant when restricted exclusively to trials employing intent-to-treat analysis (χ²=56.2, df=3, p<0.001) or when fluvoxamine trials were excluded (χ²=11.4, df=3, p=0.01) (see Figure S3D in the data supplement).

SSRI Tolerability

Higher SSRI dose was slightly, but significantly, associated with a lower likelihood of all-cause dropout (β=–0.00093, 95% CI=–0.00165 to −0.00021, z=−2.54, p=0.0110) in meta-regression analysis. The association between SSRI dose (in imipramine equivalents) and likelihood of all-cause dropout compared with placebo is presented in Figure 4A. However, when SSRI dose was divided into previously defined categories, there was no significant association between SSRI dose and likelihood of all-cause dropout (test for subgroup differences: χ²=4.8, df=3, p=0.19). The likelihood of all-cause dropout compared with placebo was highest in the 100–200 imipramine-equivalent group and slightly, but not significantly, lower if the dose was lowered or raised from this dose. The association between SSRI dose and likelihood of all-cause dropout for each of the dosing categories is shown in Figure 4C. In the sensitivity analysis, excluding fluvoxamine trials, results remained similar for likelihood of all-cause dropout in the meta-regression (β=–0.00092, 95% CI=–0.00174 to −0.00010, z=−2.20, p=0.03), but became nonsignificant in the subgroup analysis (χ²=3.7, df=3, p=0.29) (see Figure S4 in the data supplement).

Meta-regression described a significant association between higher SSRI dose and increased likelihood of dropout due to side effects (β=0.00207, 95% CI=0.00071−0.00342, z=2.98, p=0.0028). A scatterplot demonstrating the relationship between SSRI dose and the likelihood of dropout due to side effects is presented in Figure 4B. Stratified subgroup analysis by SSRI dose category also demonstrated a significant association between SSRI dose and likelihood of dropout due to side effects. All dosing categories of SSRIs were associated with a greater likelihood of dropout as a result of side effects compared with placebo. Higher dosing categories of SSRIs were associated with a greater likelihood of dropout as a result of side effects. The association between SSRI dose and likelihood of dropout due to side effects for each of the dosing categories is shown in Figure 4B. In the sensitivity analysis, excluding fluvoxamine trials, results remained mostly unchanged for the likelihood of dropout due to side effects in the meta-regression (β=0.00249, 95% CI=0.00073−0.00425, z=2.77, p=0.006), but became nonsignificant in the subgroup analysis (χ²=7.7, df=3, p=0.052) (see Figure S4 in the data supplement).

Discussion

Meta-analysis demonstrated a significant association between higher SSRI doses and greater measured efficacy of SSRIs in placebo-controlled trials. This significant association between SSRI dose and measured efficacy was demonstrated in 1) longitudinal mixed-model meta-analysis, 2) endpoint meta-regression, and 3) stratified subgroup analysis by dose. These findings remained significant if the analysis was restricted to only data from studies employing intent-to-treat analysis and when fluvoxamine trials were excluded. Meta-analysis suggests that there may also be a consequence associated with escalating the dose of SSRI associated with reduced tolerability, as evidenced by a greater likelihood of dropout due to side effects with higher SSRI dose.

The results of this meta-analysis both extend upon and contradict a previous meta-analysis in this area (10). We replicated previous evidence suggesting a reduced tolerability of SSRIs at higher doses as evidenced by a higher likelihood of dropouts as a result of side effects. However, we demonstrated a significant positive association between SSRI dose and measured efficacy that flattened out only at the higher end of the recommended dosing range (greater than 250-mg imipramine equivalents). Specifically, meta-analysis demonstrates that using a higher dose of SSRI for major depressive disorder is associated with increased likelihood of response. Table 1 depicts odds ratio and number-needed-to-treat comparisons for different initial dosing strategies of SSRIs. Our results suggest a modest improvement in efficacy of high doses (200 mg–250 mg or >250-mg imipramine equivalents) compared with low doses (100 mg–200 mg imipramine equivalents), with odds ratios of approximately 1.3 and numbers needed to treat in the 14–16 range.

Previous meta-analysis and fixed-dose trials in this area have provided no evidence for escalating dose beyond the minimum recommended therapeutic dose (10, 57). Our meta-analysis differed in methodology in several important ways from this previous meta-analysis that likely explains the difference in results. First, we restricted our analysis to SSRI trials and did not include other antidepressants that likely have different dose-response and dose-tolerability curves. Second, we examined symptom improvement as a continuous measure rather than examining clinical improvement (yes/no) as the primary outcome of the meta-analysis. This decision likely increased power of the meta-analysis by increasing sensitivity of the primary outcome measure and reducing heterogeneity by eliminating differences in definition of therapeutic response. Third, we additionally examined the dosing effects of SSRIs not only with treatment response as a dichotomous outcome but also as a continuous measure. Meta-regression with a continuous measure is more sensitive to a change in SSRI benefit with dose. Fourth, we also included several trials published after the first meta-analysis. The additional trials provided more power to conduct this analysis.

Our meta-analysis demonstrates that there is substantial evidence for a modest increase in efficacy with higher doses of SSRIs starting from the point of initial titration. We also demonstrate that this benefit is at the cost of reduced tolerability. Given this tradeoff between the risks and benefits, another potential prudent clinical strategy is to raise SSRI doses in nonresponders to low-dose treatment. Other systematic reviews have previously examined whether dose-escalation strategies are effective in nonresponders to low-dose antidepressant treatment. A systematic review that examined dose-escalation studies in low-dose SSRI nonresponders suggested that SSRIs have a flat dose-response relationship within the therapeutic range and that higher SSRI doses were only associated with a greater side-effect burden (58). By contrast, a later systematic review that examined the efficacy of dose escalation strategies in SSRI nonresponders suggested a modest benefit (number-needed-to-treat range: 12–82 in trials) of increasing to a higher-dose SSRI if subjects had received previous low-dose SSRI treatment for at least 4 weeks (59). This systematic review suggested that when the dose-escalation strategy was initiated before 4 weeks of SSRI treatment, there was no evidence of benefit to raising SSRI dose on likelihood of treatment response (59). Our results extend upon these previous dose-escalation studies and systematic reviews by demonstrating that the dose-response relationship of SSRI is mildly positive and not flat within the SSRI therapeutic range even when started from the initial point of treatment. Further research is needed to extend upon our results in order to 1) better gauge the risk/benefit of SSRI dose-escalation in low-dose SSRI nonresponders and 2) determine the ideal time point for starting SSRI dose escalation.

Given the potential clinical implications of the results of our meta-analysis, it is important to be clear in its limitations. Publication bias is a well-identified problem in trials involving antidepressant agents (6). We employed a comprehensive search strategy to try to identify all available published and unpublished trials of SSRIs. Given that our meta-analysis examined the difference in efficacy of different doses of SSRIs rather than the overall efficacy of the underlying therapeutic class, it is not clear how publication bias could have influenced the relationship between SSRI dosing and measured efficacy. Assuming that the positive association between SSRI dose and measured efficacy is true, then publication bias, if present, would likely have dampened our measured association. Publication bias would have likely caused the suppression of negative trials, which based on the findings of this meta-analysis would be more likely to occur at lower SSRI doses, and potentially lead to a reduced measured association between dose and efficacy in the meta-analysis. Other limitations were present in our meta-analysis examining tolerability of SSRI agents at different doses. We would have liked to have analyzed the frequency of different side effects (e.g., .sexual dysfunction, nausea, sedation, etc.) at different SSRI doses. However, measurement and reporting of side effects have changed dramatically over the three decades during which these trials were published. Selective reporting of side effects in earlier manuscripts and changes in how side effects are screened for over time made this analysis not feasible. We would have also liked to examine how timing of dose titration affected likelihood of subject dropout in the trials employing higher SSRI doses, but titration schedules are variably reported in trials. Another general limitation is the generalizability to the community population. Most SSRI trials included in this meta-analysis had strict inclusion criteria. Therefore, many patients seen in typical clinical practice with depression, such as those with significant comorbid medical or psychiatric conditions or taking adjunctive medications, would be specifically excluded from these trials. Clinical patients with additional comorbid illness or concomitant medication use may respond differently to SSRI dose escalation both in terms of efficacy and side effects compared with clinical trials samples (11, 58, 59).

Our meta-analysis provides evidence to support clinical guidelines that recommend raising SSRI dose in adults with major depressive disorder who do not respond to SSRI medications at or below the lower end of the therapeutic dose range. Higher doses of SSRIs are associated with increased efficacy (number needed to treat for treatment response range: 14–16) but also reduced tolerability as evidenced by a higher likelihood of dropouts due to side effects in trials (number needed to harm range: 22–24). However, overall dropout rates were reduced at higher doses of SSRIs, which is likely attributable to their greater efficacy. Further research needs to be performed to examine the ideal timing of dose escalation of SSRIs in major depressive disorder in order to maximize benefit while reducing unnecessary additional side effects caused by higher-dose SSRI treatment.