Exploring psychotherapeutic issues and agents in clinical practice
The headline, “Antidepressant Use Linked to Increased Risk of Microbleeds” (Janeczko, 2016), caught my eye. This news story described potentially alarming findings from a recently published observational study (Akoudad et al., 2016). Because of the widespread use of antidepressant drugs and the potential public health significance of this finding, it is instructive to critically review the methodology and findings from this study, illustrating the limitations inherent to observational studies.
Methodology of Observational Studies
In experimental studies, an exposure (e.g., a treatment) is manipulated randomly to assess its effect on an outcome. A randomized controlled trial is an example of an experimental study. By contrast, observational studies only allow observation of the effect of an exposure on an outcome; the exposure is not deliberately manipulated.
The main types of observational studies are cohort (follow-up), case-control, and cross-sectional studies (Jepsen, Johnsen, Gillman, & Sorensen, 2004). In cohort studies, groups of patients with an exposure(s) of interest are followed prospectively to determine the incidence of an outcome(s) of interest. In case-control studies, patients are selected by an outcome of interest (cases). A non-case control group is selected from the same source population. The level of an exposure(s) of interest is compared between cases and controls. In cross-sectional studies, exposures and outcomes are assessed simultaneously at a single point in time within a source population.
Interpreting the Findings from Observational Studies
Calculation of an odds ratio (OR) is performed to interpret the findings from cohort and case-control studies. For case-control studies, the OR is the odds of an exposure among cases compared to the odds of an exposure among controls. For a cohort study, the OR is the odds of an outcome among those exposed compared to the odds of an outcome among those unexposed. Calculation of the relative risk (RR; also referred to as risk ratio) can also be used to interpret findings from cohort studies, but not case-control studies. The RR is the ratio of the incidence of an outcome among those exposed to the incidence of the outcome among those unexposed. Because the OR and RR are ratios, their value is >0. An OR or RR of 1 is interpreted as no association between exposure and outcome. The higher the value is above 1, the stronger the association between an exposure and outcome. Values <1 are interpreted to mean that the exposure is associated with a reduced outcome.
The confidence interval (CI) is a statistical measure of the precision of the calculation of the OR or RR. A 95% CI is typically used, meaning that if a study is performed 100 times, the true OR or RR would lie within this interval 95 times and outside of it 5 times. If the CI includes the value of 1, the results are not considered statistically significant.
Antidepressant Drug Use and the Increased Risk of Developing Microbleeds
Participants in the Akoudad et al. (2016) study were drawn from the Rotterdam Study, a prospective, population-based cohort of 14,926 participants followed to investigate factors associated with the development of various diseases in older adults. Initially, 3,054 participants were selected because they were affiliated with pharmacies serving the study area and had baseline and repeat brain magnetic resonance imaging (MRI) scans between 2005 and 2013. Four hundred ninety-five individuals with evidence of microbleeds on their baseline scan were excluded, leaving 2,559 participants for data analysis.
Antidepressant drug exposure (yes or no) was assessed between baseline and follow-up scans, based on pharmacy records. Antidepressant drugs were categorized in two ways: (a) based on their affinity for the serotonin transporter (low, intermediate, and high degree of serotonin reuptake inhibition) and (b) as selective serotonin reuptake inhibitor (SSRI) drugs or non-SSRI drugs.
Baseline data included age, gender, depression score (based on a depression rating scale), smoking history, diabetes diagnosis, cholesterol levels, blood pressure, lipid-lowering medication use (yes or no), blood pressure–lowering medication use (yes or no), and antithrombotic medication use (yes or no). Mean participant age at baseline was 58 years. Mean interval between MRI scans was 3.9 years. Microbleeds incidence was 3.7% among all participants at follow up (regardless of anti-depressant drug exposure).
Two hundred ninety-nine participants had used antidepressant drugs; 2,260 were classified as nonusers. The first statistical analysis (Model 1) adjusted for age, gender, and scan interval. Antidepressant drug use was associated with a statistically significant higher microbleed incidence than nonuse (OR = 2.22; 95% CI [1.31, 3.76]). The odds of microbleeds among those who used antidepressant drugs compared to the odds of microbleeds among nonusers appear to be doubled (2.22).
The second analysis (Model 2) adjusted for age, gender, scan interval, depression score, and a cardiovascular propensity score (comprising the other baseline variables described above). This statistically significant adjusted OR was 2.29 (95% CI [1.31, 4.02]).
When categorized by serotonin transporter affinity, only the intermediate reuptake inhibition was associated with a statistically significant higher microbleed incidence in Model 1 (OR = 3.07; 95% CI [1.53, 6.17]) and Model 2 (OR = 3.29; 95% CI [1.59, 6.79]). SSRI drugs (Model 1, OR = 2.27; Model 2, OR = 2.39) and non-SSRI drugs (Model 1, OR = 2.28; Model 2, OR = 2.37) were both associated with statistically significant higher microbleed incidence.
Bias and Confounding in Observational Studies
In an observational study, an association between exposure and outcome can be due to bias, confounding, chance, or cause (Jepsen et al., 2004). Unlike experimental studies, however, causality between exposure and outcome cannot be definitively established in observational studies, even if an observed association is not attributable to bias, confounding, or chance.
For an observational study, bias means that the measure of association is systematically wrong. Selection bias and information bias are common types of bias in observational studies.
Selection bias relates to how participants are identified and included in the study, and whether the selection process is associated with the exposure or the outcome in such a way that the association between exposure and outcome is influenced, resulting in false conclusions about the association. Participants in the Akoudad et al. (2016) study were conveniently selected because they were in the Rotterdam cohort, were affiliated with certain pharmacies, and had baseline and repeat MRI scans. The study cohorts they compared were classified according to whether they were taking antidepressant medication (yes or no) based on pharmacy records. How representative are these patients of antidepressant drug users and nonusers? What were the reasons for prescribing antidepressant medications and for selecting certain medications (i.e., an SSRI drug versus a non-SSRI drug)? How similar or dissimilar are antidepressant drug users and nonusers? Do the different cohorts they define have similar or dissimilar risk factors for cerebrovascular events? Because of selection bias, it is likely that the cohorts defined by Akoudad et al. (2016) have differential characteristics (other than taking or not taking an antidepressant medication) that will influence the association between antidepressant drug exposure and microbleeds.
Information bias relates to errors in measuring exposure, covariate, or outcome variables, which may lead to differences in the accuracy or validity of information between the comparison groups and lead to false conclusions about the association between exposure and outcome. Potential sources of information bias in the Akoudad et al. (2016) study include measuring medication exposure, the validity of categorizing antidepressant drugs based on serotonin transporter affinity, the reliability and validity of the depression measure, and the reliability of cholesterol and blood pressure determinations. Drug exposure (antidepressant, lipid, blood pressure, and antithrombotic medications) was based on pharmacy records, which is an imprecise and indirect way of measuring medication adherence. Rates of nonadherence to psychotropic and non-psychotropic medications are high, even when patients fill prescriptions (Chapman & Horne, 2013; Ho, Bryson, & Rumsfeld, 2009; Osterberg & Blaschke, 2005).
Confounding variables are factors that may correlate with the exposure or the outcome or both. The primary purpose of randomization is to evenly distribute measured and unmeasured confounding variables between groups. False conclusions about associations can occur in observational studies because of confounding factors that differ between exposure groups. Potential confounding variables may be identified and measured in observational studies, and they can be partly accounted for by matching the groups or statistical corrections. However, many potential confounding factors are unknown or unmeasured. In the Akoudad et al. (2016) study, the limited variables they include in their analysis are measured at one point in time (i.e., at baseline). Past antidepressant drug exposure, non-depression psychiatric disorders, non-antidepressant psychotropic medication use, current smoking, past or current alcohol or other drug use, use of non-steroidal anti-inflammatory drugs (other than aspirin), diabetes medication use, and a history of falls (a risk for head injury) are examples of unmeasured potential confounding factors in the Akoudad et al. (2016) study. Also, there are no measured variables during the interval between MRI scans. Hence, confounding may very well influence the association between antidepressant drug exposure and microbleeds.
As discussed above, the statistical significance of associations can be calculated for OR and RR in observational studies. Statistical significance, however, does not demonstrate that the association is valid or meaningful (Grimes & Schulz, 2012; Nuzzo, 2014).
Association and Causation in Observational Studies
In a classic article on causation and association in observational studies, Hill (1965) described nine factors to consider. The most important factor is the strength (magnitude) of the association. Based on a review of epidemiological guidelines, Grimes and Schulz (2012) suggested that an OR or RR of 3–4 or greater in cohort or case-control studies should be considered potentially meaningful. Weaker associations are likely to be false and attributable to bias or confounding. The small effect found in the Akoudad et al. (2016) study falls short of this threshold, and their findings are likely to be false.
Other factors described by Hill (1965) are (a) consistency (results from study to study that are similar in magnitude and direction); (b) specificity (no other likely explanation for an association between a factor and an effect); (c) temporality (exposure reliably precedes the outcome); (d) biologic gradient (dose–response relationship); (e) plausibility (plausible mechanism between a factor and an effect); (f) coherence (an effect fitting an observed pattern in different populations or in different types of studies); (g) experiment (the effect being demonstrated in a controlled experimental setting); and (h) analogy (similar effects seen with similar factors).
How does the Akoudad et al. (2016) study hold up according to some of Hill's (1965) criteria? The finding that SSRI drugs may be associated with bleeding is plausible because serotonin is important for platelet aggregation and hemostasis, but prospective laboratory studies comparing participants and controls receiving antidepressant drugs had inconsistent findings (Halperin & Reber, 2007). Akoudad et al. (2016) did not demonstrate a biologic gradient; only intermediate serotonin reuptake inhibition was associated with a significantly higher microbleed incidence. They did not demonstrate specificity for an effect; there was no difference between SSRI and non-SSRI drugs. Other observational studies have found that SSRI drug exposure is associated with a small increased risk of intracerebral and intracranial hemorrhage (Hackam & Mrkobrada, 2012). The magnitude of this effect was <2 in a meta-analysis of cohort and case-control studies (Hackam & Mrkobrada, 2012), a finding that is likely to be false and attributable to bias or confounding.
Dykes (1974) advised that published reports should be carefully read and critically evaluated. Dykes (1974, p. 1,276):
These, of course, are time-consuming processes, and time is often limited. Thus, through lack of critical thinking, engendered by lack of time, material that should at best generate no more than a weak hypothesis, unfortunately may soon be offered as proof of a causal relationship.
Observational studies reveal associations, but cannot establish causality. Any claim coming from an observational study is likely to be wrong, especially for small effects, because of bias and confounding (Young & Karr, 2011). Statistical testing does not establish the validity of an association, nor does a meta-analysis of multiple studies (Berlin & Golub, 2014). A valid association may not be clinically meaningful if the magnitude of the effect is small or the outcome is rare. The microbleeds incidence was only 3.7% in the Akoudad et al. (2016) study. Intracerebral and intracranial hemorrhages were rare events in the SSRI meta-analysis conducted by Hackam and Mrkobrada (2012). The prevalence of cerebral microbleeds in community-based populations ranges from 3.1% to 15.3% (Poels et al., 2010; Sveinbjornsdottir et al., 2008; Viswanathan & Chabriat, 2006). Physicians and nurses should strive to be careful readers and critical evaluators. Doing so will lead us to conclude that antidepressant drugs and the risk of intracranial bleeding is likely to be a false alarm (Grimes & Schulz, 2012). When reading observational studies, Caveat lector!
- Akoudad, S., Aarts, N., Noordam, R., Ikram, M.A., Tiemeier, H., Hofman, A. & Visser, L.E. (2016). Antidepressant use is associated with an increased risk of developing microbleeds. Stroke, 47, 251–254. doi:10.1161/STROKEAHA.115.011574 [CrossRef]
- Berlin, J.A. & Golub, R.M. (2014). Meta-analysis as evidence: Building a better pyramid. Journal of the American Medical Association, 312, 603–605. doi:10.1001/jama.2014.8167 [CrossRef]
- Chapman, S.C. & Horne, R. (2013). Medication nonadherence and psychiatry. Current Opinion in Psychiatry, 26, 446–452. doi:10.1097/YCO.0b013e3283642da4 [CrossRef]
- Dykes, M.H.M. (1974). Uncritical thinking in medicine: The confusion between hypothesis and knowledge. Journal of the American Medical Association, 227, 1275–1277. doi:10.1001/jama.1974.03230240033020 [CrossRef]
- Grimes, D.A. & Schulz, K.F. (2012). False alarms and pseudo-epidemics: The limitations of observational epidemiology. Obstetrics and Gynecology, 120, 920–927. doi:10.1097/AOG.0b013e31826af61a [CrossRef]
- Hackam, D.G. & Mrkobrada, M. (2012). Selective serotonin reuptake inhibitors and brain hemorrhage: A meta-analysis. Neurology, 79, 1862–1865. doi:10.1212/WNL.0b013e318271f848 [CrossRef]
- Halperin, D. & Reber, G. (2007). Influence of antidepressants on hemostasis. Dialogues in Clinical Neuroscience, 9, 47–59.
- Hill, A.B. (1965). The environment and disease: Association or causation?Proceedings of the Royal Society of Medicine, 58, 295–300.
- Ho, P.M., Bryson, C.L. & Rumsfeld, J.S. (2009). Medication adherence: Its importance in cardiovascular outcomes. Circulation, 119, 3028–3035. doi:10.1161/CIRCULATIONAHA.108.768986 [CrossRef]
- Janeczko, L.L. (2016). Antidepressant use linked to increased risk of microbleeds. Retrieved from http://www.psychcongress.com/article/antidepressant-use-linked-increased-risk-microbleeds-25851
- Jepsen, P., Johnsen, S.P., Gillman, M.W. & Sorensen, H.T. (2004). Interpretation of observational studies. Heart, 90, 956–960. doi:10.1136/hrt.2003.017269 [CrossRef]
- Nuzzo, R. (2014). Scientific method: Statistical errors. Nature, 506, 150–152. doi:10.1038/506150a [CrossRef]
- Osterberg, L. & Blaschke, T. (2005). Adherence to medication. New England Journal of Medicine, 353, 487–497. doi:10.1056/NEJMra050100 [CrossRef]
- Poels, M.M.F., Vernooij, M.W., Ikram, A., Hofman, A., Krestin, G.P., van der Lugt, A. & Breteler, M.M. (2010). Prevalence and risk factors of cerebral microbleeds: An update of the Rotterdam scan study. Stroke, 41(10 Suppl.), S103–S106. doi:10.1161/STROKEAHA.110.595181 [CrossRef]
- Sveinbjornsdottir, S., Sigurdsson, S., Aspelund, T., Kjartansson, O., Eiriksdottir, G., Valtysdottir, B. & Launer, L.J. (2008). Cerebral microbleeds in the population based AGES-Reykjavik study: Prevalence and location. Journal of Neurology, Neurosurgery, and Psychiatry, 79, 1002–1006. doi:10.1136/jnnp.2007.121913 [CrossRef]
- Viswanathan, A. & Chabriat, H. (2006). Cerebral microhemorrhage. Stroke, 37, 550–555. doi:10.1161/01.STR.0000199847.96188.12 [CrossRef]
- Young, S.S. & Karr, A. (2011). Deming, data, and observational studies: A process out of control and needing fixing. Significance, 8, 116–120. doi:10.1111/j.1740-9713.2011.00506.x [CrossRef]