The perception of emotion from non-verbal cues is crucial to human social interaction. Many psychological disorders are characterized by deficits or biases in facial emotion recognition (ER), including schizophrenia (Addington et al. Reference Addington, Saeedi and Addington2006), alcoholism (Philippot et al. Reference Philippot, Kornreich, Blairy, Baert, Dulk, Bon, Streel, Hess, Pelc and Verbanck1999), autism (Celani et al. Reference Celani, Battacchi and Arcidiacono1999), anxiety (Button et al. Reference Button, Lewis, Penton-Voak and Munafo2013a ), bipolar disorder (Derntl et al. Reference Derntl, Seidel, Kryspin-Exner, Hasmann and Dobmeier2009), and depression (Rubinow & Post, Reference Rubinow and Post1992).
Affective disorders affect 21 million people in Europe alone and account for nearly half of the costs of all mental disorders (Andlin-Sobocki et al. Reference Andlin-Sobocki, Jönsson, Wittchen and Olesen2005). Understanding the role of ER is especially relevant to depression, as the impaired recognition of emotion has been associated with decreased satisfaction, support, and well-being of interpersonal relationships (Carton et al. Reference Carton, Kessler and Pape1999). Critically, poor interpersonal relationships have been proposed as an important factor in both the aetiology and maintenance of depression (Finch & Zautra, Reference Finch and Zautra1992, Platt et al. Reference Platt, Kadosh and Lau2013), and impaired ER may contribute to the interpersonal difficulties and avoidance seen in depression (Persad & Polivy, Reference Persad and Polivy1993). Since deficits in ER may contribute to the maintenance of depressive symptoms, investigating this relationship has important implications for existing cognitive behavioural interventions and the development of novel interventions.
Many studies have attempted to investigate the relationship between ER and depression over the last 30 years (see Bourke et al. Reference Bourke, Douglas and Porter2010 for a review). However, these have used various paradigms and stimulus sets, thus making the comparison of results across studies difficult. Two recent meta-analyses were conducted to investigate the association between major depressive disorder (MDD) and ER. Demenescu et al. (Reference Demenescu, Kortekaas, den Boer and Aleman2010) examined eight studies and found that ER in depressed adults was moderately impaired compared to controls. Given the small number of included studies, analyses stratified by emotion and analyses of study-level design characteristics were not conducted. Similarly, Kohler et al. (Reference Kohler, Hoffman, Eastman, Healey and Moberg2011) identified a moderate deficit in ER in a meta-analysis of 51 studies of emotion identification or discrimination in bipolar (31 studies) or unipolar (20 studies) depressed patients compared to controls. Notably, impairment did not differ between diagnostic groups, and analyses of all six basic emotions revealed small to moderate deficits across both patient groups. However, data on specific emotions were limited, so it was difficult to determine with certainty whether the nature and strength of the deficit differed by emotion. There was also some evidence suggesting that symptom severity was associated with a greater deficit in ER. Furthermore, demographic characteristics such as old age, sex (females), and higher levels of education (in cases) were also shown to be positively associated with ER performance.
The results from these meta-analyses are inconclusive regarding whether the ER deficit in depression is general or specific to the recognition of one or more emotions. The discovery of a specific ER deficit in depression would have important implications for treatment, allowing clinicians to target the treatment of impairments more effectively. Some researchers have proposed that there is a unique relationship between MDD and the recognition of happiness, suggesting that the recognition of happiness is specifically impaired while the recognition of sadness is spared or enhanced (Gur et al. Reference Gur, Erwin, Gur, Zwil, Heimberg and Kraemer1992; Bourke et al. Reference Bourke, Douglas and Porter2010). Similarly, while studies have demonstrated that some antidepressant pharmacotherapies modify the recognition of emotion (Harmer et al. Reference Harmer, de Bodinat, Dawson, Dourish, Waldenmaier, Adams, Cowen and Goodwin2011, Reference Harmer, Dawson, Dourish, Favaron, Parsons, Fiore, Zucchetto, Bifone, Poggesi, Fernandes, Alexander and Goodwin2013), meta-analyses thus far have not considered the effects of current medication on ER in depressed individuals.
The purpose of this meta-analysis was therefore to extend our understanding of the relationship between ER deficits and MDD. We did this by comparing studies across several different methodologies, paradigms, and design-level characteristics, including stimulus sets, presentation times, and response options. This included stratifying our analyses by medication status (i.e. medicated or unmedicated) in order to investigate the effects of antidepressants on this relationship. In addition to investigating a general deficit of ER, we further stratified our analyses by all six basic emotions in order to investigate specific deficits. In the interest of reducing the moderate levels of heterogeneity detected in the previous meta-analyses, we only included studies using human facial emotional expression stimuli. Finally, we also calculated the statistical power of each study included in our analysis to detect the effect size indicated by the meta-analysis, and tested for possible publication bias. This meta-analysis extends our understanding of the relationship between ER abilities, and provides a more accurate estimate of the real magnitude of the effect of depression on ER deficits in studies using photorealistic stimuli.
Study inclusion/exclusion criteria
Eligibility criteria for study inclusion were as follows: (1) studies were required to have both a clinical sample with a diagnosis of MDD and a control sample; (2) studies were required to have assessed the accuracy of ER; and (3) studies were required to have used stimuli comprising of human facial emotional expressions. Studies using schematic or artistically rendered faces, neuroimaging studies and studies that included experimental administration of drug treatments were excluded. Studies that recruited participants with a diagnosis of both MDD and bipolar disorder were retained.
We performed a search on two databases: PubMed and Web of Science. These databases were searched from the first date available in each database up to 1 June 2013, using the inclusion terms ‘depression’, ‘MDD’, ‘emotion*’, ‘recognition’, ‘perception’ and the exclusion term ‘administration’. After articles had been collected, bibliographies were then searched for additional references.
For each study, the following data were extracted: (1) author(s) and year of publication; (2) data [mean and standard deviation (s.d.) of ER accuracy scores, number of participants, mean age and male/female ratio] and (3) study design characteristics. Study design was coded (where possible/applicable) for: stimulus emotion (anger, disgust, fear, happiness, sadness, surprise), case status (MDD no co-morbidity, MDD co-morbidity, MDD+bipolar disorder), control status (matched, unmatched), treatment status (medicated, unmedicated), diagnostic criteria [Diagnostic and Statistical Manual of Mental Disorders (DSM)/Research Diagnostic Criteria (RDC), International Classification of Diseases (ICD)], stimuli (Ekman & Friesen, Other), use of morphed stimuli (no, yes), stimulus type (dynamic, static), presentation time (<500, >500, 500 ms, self-paced), and response option [two alternative forced choice (AFC), six AFC, other]. We also rated the quality of all included studies using eight items adapted from the Newcastle–Ottawa Scale, a measure for assessing the quality of non-randomized studies in meta-analyses (Wells et al. Reference Wells, Shea, O'Connell, Peterson, Welch and Tugwell2000). Studies were rated on the selection of study groups, the comparability of those groups, and the ascertainment of the outcome of interest.
Effect sizes (Hedges' g) were calculated for the comparison of cases v. controls on ER accuracy for each emotion reported within each individual study. Hedges' g is a measure of standardized mean difference, similar to Cohen's d but including a correction for small sample size. Conventionally, a small effect size is defined as 0.20, a medium effect size as 0.50 and a large effect size as 0.80 (Cohen, Reference Cohen1988).
Data were analysed within a random-effects framework, with g values pooled using DerSimonian & Laird (Reference DerSimonian and Laird1986) methods. A random-effects framework assumes that between-study variation is due to both chance or random variation and an individual study effect, and provides an estimate of the range of likely effect sizes across the populations sampled by individual studies. Random-effects models are more conservative than fixed-effects models and generate a wider confidence interval (CI), but give similar results under conditions of low between-study heterogeneity. The significance of the pooled g values was determined using a Z test. Between-study heterogeneity was estimated using the I 2 statistic. Conventionally, values of 25%, 50% and 75% represent the upper thresholds for low, moderate and high heterogeneity, respectively.
Small study bias, which may reflect publication bias against null results, was assessed using Egger's test (Egger et al. Reference Egger, Davey Smith, Schneider and Minder1997). We also conducted a series of stratified analyses and meta-regression analyses to assess the impact of various study design characteristics. The analyses were conducted using the Comprehensive Meta-Analysis v. 2 statistical software package (Biostat, USA). Exact p values are reported throughout.
Description of studies
Our search strategy across both databases initially identified 728 articles. Of these, 66 articles were identified as duplicates and were removed. Of the remaining 662 articles, we were able to exclude 624 articles because they did not meet our inclusion criteria. A further 16 articles were excluded because they did not report the data required to enable inclusion in our meta-analysis, and attempts to contact the study authors to acquire these were unsuccessful.
A total of 22 studies published between 1992 and 2012 met inclusion criteria and were included in our meta-analysis. A flow chart describing this process is shown in Fig. 1. Characteristics of these studies are described in Table 1.
DSM, Diagnostic and Statistical Manual of Mental Disorders; RDC, Research Diagnostic Criteria; ICD, International Classification of Diseases; AFC, alternative forced choice.
Quality of included studies
The eight items we used to assess the quality of our included studies consisted of four items related to study group selection, two items related to the comparability of groups and two items related to how the studies ascertained the outcome of interest. Each study scored 1 point for each item if the criterion was met. Most studies included in our meta-analysis adequately described the selection of study groups as only three studies scored <3 out of a possible 4 points on these items. Most studies failed to offer sufficient information regarding the comparability of study groups as only six studies scored points on both items while three studies earned only 1 point. All studies met criteria regarding the ascertainment of the outcome of interest, scoring a point for both items.
ER in MDD
Meta-analysis (k = 22) indicated strong evidence of a deficit in ER among cases compared to controls (g = −0.16, 95% CI −0.25 to −0.07, p < 0.001) with negligible between-study heterogeneity (I 2 = 0%). Stratified analyses across the six primary emotions indicated a deficit in ER for anger, disgust, fear, happiness and surprise (k's = 7–22, g's = −0.42 to −0.17, p's < 0.08), but not sadness (k = 21, g = −0.09, 95% CI −0.23 to +0.06, p = 0.23). Sensitivity analysis indicated that no single study disproportionately contributed to these results. These results are presented in Table 2.
CI, Confidence interval.
Impact of study-level design characteristics
Stratified analyses indicated no evidence that any study-level design characteristics altered the deficit in ER among cases compared to controls (p's ⩾ 0.11). In all cases, between-study heterogeneity was moderate to negligible (I 2 ⩽ 55%), with the exception of the two studies in the MDD+bipolar disorder stratum (I 2 = 70%). These results are presented in Table 3. Meta-regression indicated a positive association between year of publication and effect-size estimate (p = 0.029).
CI, Confidence interval, DSM, Diagnostic and Statistical Manual of Mental Disorders; RDC, Research Diagnostic Criteria; ICD, International Classification of Diseases; AFC, alternative forced choice.
a One study with cases classified as MDD+BP included participants diagnosed with co-morbid disorders (Douglas & Porter, Reference Douglas and Porter2010).
b Studies classified as ‘medicated’ include those where only a proportion of participants were medicated (Gur et al. Reference Gur, Erwin, Gur, Zwil, Heimberg and Kraemer1992; Kan et al. Reference Kan, Mimura, Kamijima and Kawamura2004; Bediou et al. Reference Bediou, Krolak-Salmon, Saoud, Henaff, Burt, Dalery and D'Amato2005; Mendlewicz et al. Reference Mendlewicz, Linkowski, Bazelmans and Philippot2005; Joormann & Gotlib, Reference Joormann and Gotlib2006; Langenecker et al. Reference Langenecker, Caveney, Giordani, Young, Nielson, Rapport, Biellauskas, Mordhorst, Marcus, Yodkovik, Kerber, Berent and Zubieta2007; Wright et al. Reference Wright, Langenecker, Deldin, Rapport, Nielson, Kade, Own, Akil, Young and Zubieta2009; Douglas & Porter, Reference Douglas and Porter2010; Milders et al. Reference Milders, Bell, Platt, Serrano and Runcie2010; Anderson et al. Reference Anderson, Shippen, Juhasz, Chase, Thomas, Downey, Toth, Lloyd-Williams, Elliott and Deakin2011; Arteche et al. Reference Arteche, Joormann, Harvey, Craske, Gotlib, Lehtonen, Counsell and Stein2011; Derntl et al. Reference Derntl, Seidel, Schneider and Habel2012; Vederman et al. Reference Vederman, Weisenbach, Rapport, Leon, Haase, Franti, Schallmo, Saunders, Kamali, Zubieta, Langenecker and McInnis2012).
c One study used both morphed and unmorphed stimuli in two separate tasks, and contributed to each stratum of this analysis (Sprengelmeyer et al. Reference Sprengelmeyer, Steele, Mwangi, Kumar, Christmas, Milders and Matthews2011).
Impact of medication status on recognition of happiness and sadness
Given evidence from human psychopharmacology studies indicating that antidepressants modify ER (Harmer et al. Reference Harmer, de Bodinat, Dawson, Dourish, Waldenmaier, Adams, Cowen and Goodwin2011, Reference Harmer, Dawson, Dourish, Favaron, Parsons, Fiore, Zucchetto, Bifone, Poggesi, Fernandes, Alexander and Goodwin2013), we examined the impact of medication status on the recognition of happiness and sadness. The pattern of results described did not differ by medication status for either the recognition of happiness (p = 0.84) or sadness (p = 0.65). Notably, only three studies included in our analysis tested unmedicated cases compared to 19 studies assessing recognition of happiness in medicated samples and 18 assessing recognition of sadness. These results are presented in Table 4.
CI, Confidence interval.
a Studies classified as ‘medicated’ include those where only a proportion of participants were medicated (Gur et al. Reference Gur, Erwin, Gur, Zwil, Heimberg and Kraemer1992; Kan et al. Reference Kan, Mimura, Kamijima and Kawamura2004; Bediou et al. Reference Bediou, Krolak-Salmon, Saoud, Henaff, Burt, Dalery and D'Amato2005; Mendlewicz et al. Reference Mendlewicz, Linkowski, Bazelmans and Philippot2005; Joormann & Gotlib, Reference Joormann and Gotlib2006; Langenecker et al. Reference Langenecker, Caveney, Giordani, Young, Nielson, Rapport, Biellauskas, Mordhorst, Marcus, Yodkovik, Kerber, Berent and Zubieta2007; Wright et al. Reference Wright, Langenecker, Deldin, Rapport, Nielson, Kade, Own, Akil, Young and Zubieta2009; Douglas & Porter, Reference Douglas and Porter2010; Milders et al. Reference Milders, Bell, Platt, Serrano and Runcie2010; Anderson et al. Reference Anderson, Shippen, Juhasz, Chase, Thomas, Downey, Toth, Lloyd-Williams, Elliott and Deakin2011; Arteche et al. Reference Arteche, Joormann, Harvey, Craske, Gotlib, Lehtonen, Counsell and Stein2011; Derntl et al. Reference Derntl, Seidel, Schneider and Habel2012; Vederman et al. Reference Vederman, Weisenbach, Rapport, Leon, Haase, Franti, Schallmo, Saunders, Kamali, Zubieta, Langenecker and McInnis2012).
Small study bias
There was evidence of small study bias for the combined analysis (p = 0.003), while for the stratified analyses this was indicated for sadness (p = 0.028) and anger (p = 0.003). Adjusting for possible publication bias against null results using Duval and Tweedie's trim and fill method (Duval & Tweedie, Reference Duval and Tweedie2000) indicated a reduced effect-size estimate in the combined analysis (g = −0.08, 95% CI −0.18 to +0.01), and the sadness (g = +0.04, 95% CI −0.12 to +0.20) and anger (g = 0.01, 95% CI −0.18 to +0.16) stratified analyses.
The effect-size estimate indicated by our combined meta-analysis (g = −0.16) suggests that a sample size of approximately 615 cases and 615 controls would be required to detect a deficit in ER with 80% power at an alpha level of 0.05. The median sample size among studies included in our meta-analysis was 21 cases and 25 controls, which would correspond to 8% power to detect an effect size of this magnitude.
Our findings indicate a general ER deficit associated with MDD. In addition, analyses stratified by emotion indicate that the recognition of sadness is uniquely preserved, while recognition of the other basic emotions is impaired. We also did not find any evidence that study-level characteristics modified these results, suggesting that these effects may be relatively robust to diagnostic criteria, task parameters and other design factors. Given the variability in these factors across studies, this finding was unexpected, and may suggest that despite the small effect size of the ER deficit, this is a robust feature of MDD.
Medication status among cases did not appear to modify the association of depression with ER. However, this analysis included only three studies where depressed patients were unmedicated at time of testing, making it difficult to draw firm conclusions on the effects of medication on ER in this population. Details of current psychological treatment, which may also modify ER, were often unreported and therefore could not be systematically examined. It is noteworthy that studies of medicated patients tended to report a greater deficit in the recognition of sadness than the studies including only unmedicated patients, although there was not sufficient statistical power to evaluate whether this was a consistent effect. Clearly, further research with untreated depressed samples is required in order to better understand how ER deficits are associated with MDD, rather than medication or therapy per se. In particular, a wide body of research suggests that antidepressant medication reduces the recognition of, and neural responses to, negative facial expressions in healthy participants and patients with MDD (see Pringle et al. Reference Pringle, McCabe, Cowen and Harmer2013). In the relative absence of data from unmedicated patients, it is possible that the current results are a marker of medication status as opposed to the disorder itself. Nonetheless, the results of our analyses stratified by emotion suggests that there is no unique relationship between MDD and the accurate recognition of happiness, as previously proposed (Gur et al. Reference Gur, Erwin, Gur, Zwil, Heimberg and Kraemer1992; Bourke et al. Reference Bourke, Douglas and Porter2010). Instead, the impaired recognition of happiness is merely part of a general recognition deficit across the other basic emotions. Response bias (i.e. the tendency to label ambiguous faces as positive v. negative) was not systematically reported in these studies and therefore has not been directly compared.
Contemporary theories of depression emphasize the importance of negative biases in ER as an important causal factor in illness etiology (Disner et al. Reference Disner, Beevers, Haigh and Beck2011; Roiser et al. Reference Roiser, Elliott and Sahakian2012). In particular, attentional, perceptual and interpretative biases towards negative material is believed to fuel negative self-referent schema in depression (Roiser et al. Reference Roiser, Elliott and Sahakian2012). The current results are broadly consistent with this framework, since the recognition of sadness was preserved across a general landscape of ER deficits in depression. In other words, the recognition of sadness may be greater in relative terms, compared to the other emotional inputs (including happiness). However, the current results are not consistent with a more general negativity bias in depression in terms of accuracy of facial expression recognition. Based on the findings of a recent study, a negative bias in the interpretation of neutral faces rather than accuracy deficits in ER may represent a vulnerability factor for major depression in at-risk individuals (Maniglio et al. Reference Maniglio, Gusciglio, Lofrese, Belvederi Murri, Tamburello and Innamorati2014). Given the effects of medication on the detection of negative emotion in facial expressions (Harmer et al. Reference Harmer, Shelley, Cowen and Goodwin2004, Reference Harmer, Mackay, Reid, Cowen and Goodwin2006), this conclusion needs to be qualified by noting the scarcity of research investigating ER in unmedicated patients. Future research should prioritize assessing ER (and associated measures) in patients free of medication. It is also worth noting that psychological treatments may also impact the processing of emotion in facial expressions, which indicates that studies in patients who are receiving neither pharmacological nor psychological treatments may be informative.
While our results indicate that MDD is associated with a general deficit in ER, the size of this effect is small. One consequence of this is the low statistical power of individual studies in our meta-analysis to detect effects of these associations, with the largest study achieving only 34% power to detect the effect size indicated by our meta-analysis. The problems associated with low statistical power have recently been described, and include an increased likelihood that a statistically significant finding reflects a false positive (Ioannidis, Reference Ioannidis2005; Button et al. Reference Button, Ioannidis, Mokrysz, Nosek, Flint, Robinson and Munafo2013b ). Rather than being endemic to a particular domain, the problem of low statistical power appears to be pervasive across several fields in the biomedical sciences. Our results therefore indicate the need for studies of ER deficits on a scale far larger than has been achieved to date. New technologies and data collection methods, such as the use of Internet and smartphone platforms, could help achieve this (Mar et al. Reference Mar, Spreng and Deyoung2013), and recent studies have shown that data collected via the Mechanical Turk are of comparable fidelity to those collected in a traditional laboratory setting (Crump et al. Reference Crump, McDonnell and Gureckis2013). However, one important limitation of this approach is that it may be difficult to obtain data on clinical status except via self-report.
Our positive test for small study bias reveals evidence of possible publication bias against null results in this literature. This arises when researchers decide to not submit negative findings for publication, largely due to the prevailing tendency for journals to reject papers reporting null findings (Thornton & Lee, Reference Thornton and Lee2000). While we sought unpublished studies, as is common practice in conducting meta-analyses, we did not receive any responses. Given the presence of small study bias, we adjusted using Duval and Tweedie's trim-and-fill method (Duval & Tweedie, Reference Duval and Tweedie2000) in order to account for small-study effects, where smaller studies in a meta-analysis tend to show larger treatment effects (Sterne et al. Reference Sterne, Gavaghan and Egger2000). While reduced in strength, evidence of a general deficit in ER in depression remained.
There are a number of limitations to the present study which should be considered when interpreting these results. First, we excluded studies using stimulus sets generated using schematic or artistically rendered faces. This was done due to the lack of perceived ecological validity for schematic or artistically rendered faces, compared to human facial expression stimuli. We therefore cannot say whether our results would apply to tasks using schematic or artistically rendered faces. Second, there was insufficient data among studies included in our analysis to conduct meta-analysis on response bias. The investigation of false alarms from recognition tasks would offer further insight into the nature of the observed deficits, and would allow us to potentially identify biased responding for specific emotions, i.e. the tendency to mislabel ambiguous faces as sad or happy. However, there were minimal false alarm data available for analysis in the present study; future studies should report false alarm data consistently, alongside accuracy data. Third, we were limited in our ability to draw conclusions on the effect of medication status on ER in depressed individuals as few studies in our analysis assessed unmedicated cases. While the data available did not provide strong evidence that the recognition of emotion differs by medication status, contrary to our expectations given the literature on the effects of antidepressants on ER, future studies explicitly designed to test this (i.e. including both medicated and unmedicated cases) are required. Additionally, we were unable to investigate whether symptom severity moderated recognition performance as these data were not uniformly reported. As elevated depressive symptoms appear to predict poorer performance on recognition tasks (Kohler et al. Reference Kohler, Hoffman, Eastman, Healey and Moberg2011) accounting for symptom severity would be helpful when investigating the effect of medication status on performance. Finally, while we have determined a more accurate estimation of the size of the effect of depression on ER performance, it is not clear how these effects may translate to clinical significance. Therefore more research is needed to explore the relationship between ER and symptom severity.
In conclusion, our analyses confirm a general deficit of ER in depressed individuals compared to controls, albeit with a small effect size. Studies thus far have been considerably underpowered to detect this effect, and primary studies with much larger sample sizes will be required to properly investigate this association. Of the six basic emotions, only the recognition of sadness appears to be spared in depression, and there appears to be no specific association with impaired recognition of happiness. Further research comparing both medicated and unmedicated patients would offer new insights into the effects of depression on ER, with possible implications for treatment.
For supplementary material accompanying this paper visit http://dx.doi.org/10.1017/S0033291714002591.
M.R.M. is a member of the UK Centre for Tobacco Control Studies, a UKCRC Public Health Research Centre of Excellence. Funding from British Heart Foundation, Cancer Research UK, Economic and Social Research Council, Medical Research Council, and the National Institute for Health Research, under the auspices of the UK Clinical Research Collaboration, is gratefully acknowledged.
Declaration of Interest
I.S.P-V. and M.R.M. are directors of Jericoe Ltd, which develops software for assessing and modifying emotion perception. M.R.M. has provided consultancy to Servier. C.J.H. is a director of Oxford Psychologists and has received consultancy fees from Lundbeck, Lily, Servier and P1vital in the last 3 years. She holds shares in P1vital.