Nonsteroidal anti-inflammatory drugs (NSAIDs) have been among the most commonly prescribed medications worldwide. NSAIDs act by inhibiting cyclooxygenase (COX) enzymes, which convert the fatty acid arachidonic acid to prostaglandins (particularly prostaglandin H2), eicosanoids, and thromboxanes, which in turn increase pain, fever, and both local and systemic inflammation. Most COX inhibitors are nonselective inhibitors of COX-1 and -2 enzymes (e.g., aspirin and ibuprofen). Selective COX-2 inhibitors, such as celecoxib and rofecoxib, were developed to reduce side effects, particularly gastric ulceration and bleeding (1). By the year 2000, there were more than 75 million prescriptions for NSAIDs in the United States alone and over $3 billion in sales for celecoxib (2). However, by the mid-2000s, data emerged indicating that COX inhibitors (both selective and nonselective agents) increase risk for cardiovascular disease, stroke, and heart failure, which has reduced sales of at least the COX-2 selective inhibitors. Celecoxib is the only COX-2 inhibitor that remains on the market in the United States, and its use has declined considerably since 2005. At the same time, the use of naproxen has held steady, and the use of both ibuprofen and meloxicam have increased (3). Any significant interaction between antidepressants and COX inhibitors would be very important given the high degree of comorbidity of depression with the inflammatory diseases treated with these drugs.

A concern about a possible attenuation of the effects of selective serotonin reuptake inhibitors (SSRIs) by COX inhibitors was raised initially by Warner-Schmidt et al. (4). They demonstrated that in the frontal cortex of mice, the SSRI citalopram increased the cytokines tumor necrosis factor alpha (TNF-α) and interferon-gamma (IFN-γ) and a protein called p11, which had previously been linked to both depressive-like states and antidepressant response in animal models. The NSAID ibuprofen reduced TNF-α, IFN-γ, and p11 and attenuated the antidepressant-like actions of SSRIs but not of other antidepressants, such as tricyclics and monoamine oxidase inhibitors, in mice in an apparent cytokine- and p11-dependent manner.

This is in contrast to a compelling literature suggesting that depressed patients are more likely to show increased peripheral and central levels of cytokines, including interleukin-6 (IL-6), TNF-α, and IFN-γ (5), and that administration of exogenous inflammatory mediators such as interferon increase the risk for depression (6). Moreover, blockade of cytokines such as TNF-α appears to have antidepressant actions in the face of inflammatory disease (7), and some antidepressants, including SSRIs, appear to reduce cytokines in depressed patients (8). However, Warner-Schmidt et al. went on to evaluate the effect of concomitant administration of NSAIDs on response to initial treatment with citalopram in depressed patients using data from the Sequenced Treatment Alternatives to Relieve Depression (STAR*D) study and found that NSAID treatment was associated with reduced response to citalopram (4). The authors concluded with the following recommendation: “Analysis of our clinical data strongly suggests that remission rates among depressed individuals may be improved by avoiding certain common over-the-counter medications such as ibuprofen, aspirin, and acetaminophen.” However, is this conclusion warranted?

Sampling bias may occur in any analysis of outcomes data in which the study group was not randomly assigned to the variable of interest (in this case, NSAID use) (9). Retrospective case-control studies of the type described earlier often suffer from fundamental differences between groups. Stated simply, depressed people taking NSAIDs regularly are likely to be different in many ways from those who do not; they are much more likely to have medical illnesses, including inflammatory diseases, that reduce responsiveness to treatment. Notably, Warner-Schmidt et al. (4) went on to state that “it is possible that underlying condition(s) [i.e., concomitant medical illnesses] contribute to treatment resistance rather than any one particular mechanism of action of concomitant medication.”

There are several ways to control for these kinds of problems. In a prospective study with multiple treatment conditions (e.g., active treatments or placebo), randomization can be stratified based on the presence or absence of concomitant medical conditions, that is, assuring that equal numbers of people with specific conditions are included in each condition. However, this is often difficult to accomplish because of the necessity of controlling for a large number of illnesses and treatments. A second approach in a prospective study is to conduct tests of moderation of the outcomes based on specific covariates, such as concomitant diseases or drug exposures. In a retrospective case-control study, a common approach to controlling for extraneous variables that may influence outcomes is propensity score matching, first proposed by Rosenbaum and Rubin in 1983 (10). This approach adjusts the analyses based on covariates that may influence the outcomes. Rather than matching on single characteristics of interest such as NSAID exposure, propensity score matching adjusts for selection bias in studies by summarizing the effects of all known covariates. Adjusting for concomitant conditions that might influence outcome helps to offset the lack of randomization.

This approach is similar to the methods used by Gallagher et al. (11), presented in this issue of the Journal, in an attempt to resolve the question, “Is it the NSAID or what the NSAID is used to treat that reduces response to antidepressants?” They first analyzed data from 1,528 outpatients with depression treated with antidepressants at 45 academic health centers, using an automated method to assess treatment outcome. They then entered a variety of possible factors that could have affected outcome, including age, sex, ethnicity, and comorbidities. The authors used a validated tool to summarize the health burden of concomitant illnesses, referred to as the Charlson comorbidity index, as well as health care utilization data. Prescription history data were employed to determine whether an NSAID was used. Logistic regression was used to compare outcomes in depressed patients with or without NSAID exposure. Binary logistic regression is an analytic approach in which there are only two possible outcomes, in this case antidepressant response and nonresponse, which can be adjusted for multiple covariates.

The second part of the study involved a reanalysis of the STAR*D citalopram treatment data, as well as level-2 treatment with cognitive-behavioral therapy (CBT) alone using similar methods. Medical comorbidity was recorded systematically in STAR*D using the 14-item Cumulative Illness Rating Scale, which assesses disability on a scale of 0–4 for 14 different categories of illnesses (e.g., cardiac, musculoskeletal, or vascular disease).

People who were taking NSAIDs were more likely to be older, to be female, to be non-White, to have public insurance, and to have greater medical comorbidity and health care utilization. In the unadjusted analyses of antidepressant response in both cohorts, concomitant treatment with nonselective NSAIDs (alone or in combination with other COX inhibitors), but not with COX-2 selective inhibitors or salicylates, was associated with reduced response to antidepressant therapy; the odds ratios for treatment nonresponse in the NSAIDs-only group were 1.74 in the first cohort and 1.44 in the second. The odds ratios were reduced in the fully adjusted models to 1.31 and 1.26, which remained statistically significant. Therefore, there appear to be significant effects of NSAIDs alone even after adjustment for confounds. Additionally, chronic but not intermittent NSAID use was more likely to be associated with treatment resistance.

To control for the effects of pain on nonresponse, the group went on to test for significant effects of concomitant narcotic use without NSAID treatment in the STAR*D data set. Narcotic use in the absence of NSAID therapy was associated with an effect that was equivalent to that for NSAID use in both the unadjusted and adjusted models (odds ratio, 1.54 and 1.26, respectively). This indicated that the effect was not specific to NSAIDs.

As a final step, the investigators tested the effects of NSAID use on CBT at level 2. Participants who were not adequately responsive to citalopram at level 1 moved on to several different treatment options at level 2. For some, citalopram was discontinued and CBT alone was initiated. The effect of NSAID therapy on outcome with level-2 CBT was even more robust than the effect on antidepressant response in both the unadjusted and adjusted models (odds ratio, 2.42 and 2.14, respectively); this did not achieve statistical significance, probably because of the relatively small sample size. This effect is striking and suggests that NSAID therapy or the diseases that NSAIDs are used to treat also reduce responsiveness to CBT. Therefore, the “NSAID effect” on antidepressant response does not seem to be specific to either NSAIDs or antidepressants.

So what can we make of the fact that the effect of NSAIDs on outcome of treatment with both antidepressants and CBT did not go away completely even after adjusting for comorbid medical conditions? It is possible that NSAIDs (and narcotics) adversely affect outcome in a variety of different antidepressant treatments and psychotherapies. However, it is more likely that this is a result of residual confounding, which occurs when the confounding variables (here, concomitant medical conditions) are measured imperfectly or when the covariate adjustment does not completely remove the effect of the confounders. This explanation seems likely in this case for at least three reasons. First, the effect was seen with treatments that are likely to act through different mechanisms (antidepressant drugs and probably CBT). Second, the effect was seen with narcotics used for some of the comorbid medical conditions, although narcotics do not act through the same mechanism as NSAIDs. Third, the measurement of comorbid conditions in both cohorts was imperfect at best and was likely to miss a lot of the variations in those conditions. Based on the current state of the data from both Warner-Schmidt et al. (4) and Gallagher et al., it is not possible to completely exclude at least some effect of NSAIDs on outcome. However, a more likely explanation is that the effects are accounted for by the concomitant medical conditions for which NSAIDs are used.

But what about the original finding that in an animal model of depression and antidepressant-like response, NSAIDs seem to reduce the effects of SSRIs? This may be related to the differing roles of cytokines in brain; under some circumstances, cytokines such as IL-6, TNF-α, and IFN-γ have a neuroprotective role (12). In other situations, the opposite may be true. The original study in mice used acute stress paradigms (i.e., the forced swim test, the tail suspension test, and novelty suppressed feeding), which are considered animal equivalents to human depression or anxiety. However, there may be important differences between the physiological state represented by these animal paradigms and human depression. Alternatively, there may still be important implications for human applications. Are there human states that are more analogous to the animal paradigms used in the original study in which NSAID suppression of central cytokine response interferes with normal functions of cytokines that could produce adverse effects long term? Whether NSAIDs influence antidepressant response is by no means settled, and the findings by Warner-Schmidt et al. (4) raise important concerns that need to be addressed in future research.