Introduction

The ability to recognize others’ emotions benefits many aspects of personal and social functioning [1]. Women generally perform better than men on emotion recognition tests, although effect sizes are usually small [2, 3]. Biologically, men and women have different levels of sex hormones (i.e., estrogen, progesterone, and testosterone [4], which likely influence development of some behavioral sex differences [5, 6]. Estrogen and progesterone fluctuate during the phases of the menstrual cycle in women, which may influence perception and behavior [7, 8] as well as mental health outcomes [9]. During the follicular phase (from the onset of menstruation to ovulation) both estrogen and progesterone levels are initially low. Estrogen markedly increases in the late follicular phase and peaks during ovulation, while progesterone remains low. During the luteal phase (following ovulation until the next menstruation) both estrogen and progesterone are generally high (e.g., [10]). These hormonal changes could influence emotion recognition. Notably, neuroimaging studies suggest that female sex hormones modulate activity in the amygdala and prefrontal cortex, which are regions important for the processing of emotions [11, 12]. Thus, researchers have proposed that sex hormone fluctuation during the menstrual cycle could influence emotion recognition accuracy (e.g., [13]). In the current study, we therefore investigated the effects of sex hormones and menstrual cycle phase on emotion recognition rates in a sample of healthy naturally cycling women.

Previous studies suggest that females may be slightly better at recognizing facial expressions during the follicular/ovulatory phase than during the luteal phase (for reviews, see [14, 15]), although these studies find mixed results. Some studies found higher overall recognition accuracy in the follicular vs. luteal phase [16–19]. In line with this group difference, Derntl et al. [16, 18] reported negative correlations between overall accuracy and progesterone levels. These effects may also vary by stimulus characteristics, such as type of emotion. Pearson and Lewis [13] reported that accuracy for fearful faces was highest during the late follicular phase compared to other cycle phases. Guapo et al. [20] reported that accuracy for angry and sad faces was higher during the follicular vs. luteal period. Some studies have also reported positive correlations between recognition accuracy and levels of estrogen, albeit with some qualifications. In Hamstra et al. [21] this effect applied only to happy faces and participants who carried a certain mineralocorticoid haplotype. In Gamsakhurdashvili et al. [22] this effect only appeared for expression stimuli from male actors.

From an evolutionary perspective, enhanced emotion recognition in the follicular phase could reflect the importance of social interaction during the most fertile period to increase mating chances [15]. Alternatively, decreased accuracy in the luteal phase could reflect heightened sensitivity to emotional cues [23] and bias toward negative emotions [24]. This may result in overperceiving potentially threatening stimuli which has a protective function for the fetus during pregnancy [16].

By contrast, other studies have reported no effects on emotion recognition of cycle phase [25–27], progesterone levels, or estrogen levels (e.g., [19, 27]). Results from some studies even go in the opposite direction as those reviewed above, with negative correlations between estrogen and recognition accuracy [20, 25]. Shirazi et al. [28] also reported no evidence that estrogen or progesterone levels predicted results on the Reading-the-Mind-in-the-Eyes test [29]. However, this test does not measure emotion recognition per se, but instead assesses the related ability of ‘theory of mind’ which focuses on reasoning about the mental states of others.

Most past research investigated the recognition of still images of facial expressions (but see [25, 27]). In real social situations, however, emotions are expressed by a combination of dynamic facial, vocal and bodily expressions [30]. Previous studies have also focused on a limited number of emotion categories, including only a single positive emotion (i.e., happiness). Such a design contrasts with the potential to communicate a wide range of both positive and negative emotions [31].

The current study introduces more ecologically valid measures of emotion recognition into the study of sex hormones and cycle phase. We used two emotion recognition tasks, both of which included dynamic expressions. Both tasks also included a wider range of emotions than in past work, including several positive emotions other than happiness. In the first task, emotions were expressed through facial expressions and speech prosody, and stimuli contained both unimodal (video only, audio only) and multimodal (audio-visual) items. The second task contained non-linguistic vocalizations—i.e., non-speech human sounds such as grunts, screams, and laughter—which are especially suited for emotional expression [32]. These two tasks were used to investigate in detail whether emotion recognition accuracy varies across three factors, namely the follicular vs. luteal phases, high vs. low estrogen, and high vs. low progesterone. We hypothesized that emotion recognition will be more accurate in the follicular vs. luteal phase. On an exploratory basis without prediction based on previous research, we did not have grounds for specific hypotheses about associations between accuracy and hormone levels [15].

Method

Results

Cycle phase differences and hormones for the ERAM test

Table 1 shows the results from four linear mixed effects models using emotion recognition accuracy (Hu) on the ERAM test as the outcome measure. Presentation modality (Models 1 and 2) and specific emotions (Models 3 and 4) appear in separate models, as detailed below.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Mixed effect models of emotion recognition accuracy for dynamic multimodal expressions (ERAM).

https://doi.org/10.1371/journal.pone.0312404.t001

Presentation modality.

Model 1 of Table 1 shows that, by contrast with the hypothesis, the coefficient for cycle phase was not significant and there was no difference in recognition accuracy between the follicular (M = .432, 95% CI [.408, .457]) and luteal phases (M = .435, 95% CI [.404, .467], d = .030). Note that all reported mean values in the results section represent the estimated marginal means. However, indicators for the auditory and multimodal modalities were significant compared to the baseline visual modality. Post hoc multiple comparisons (t-tests with Bonferroni corrections) showed that accuracy with multimodal expressions (M = .520, 95% CI [.487, .554]) was significantly higher than with visual expressions (M = .420, 95% CI [.387, .454], d = .950), which in turn was higher than with auditory expressions (M = .361, 95% CI [.328, .395], d = .562), with all of these comparisons significant. There was a significant main effect of test occasion that reflected a training effect between the first test occasion (M = .409, 95% CI [.381, .436]) and the second test occasion (M = .459, 95% CI [.433, .484], d = .473). We observed a significant negative coefficient for within-person levels of estrogen but a positive between-person effect—that is, generally high levels of estrogen increase accuracy, but higher-than-average levels of estrogen at any given time decrease accuracy. There was a moderating effect whereby within-person estrogen interacted with the follicular phase indicator, such that women with higher levels of estrogen were more accurate when tested in the follicular phase (see Fig 1A). A similar interaction between progesterone and the follicular phase indicator did not reach significance.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Interactions between estrogen (within-person), cycle phase, and cycle days in predicting emotion recognition accuracy (ERAM task).

(A) Interaction between estrogen levels and cycle phase shown when accuracy is distinguished via presentation modality. In the follicular phase, the +1 SD line shows that accuracy is higher for women who have higher levels of estrogen and the -1 SD line shows accuracy is lower for women who have lower levels of estrogen. (B) Illustration of the interaction between estrogen levels and the days squared variable when accuracy is distinguished via presentation modality. Accuracy is highest in the middle of cycle for women who have higher levels of estrogen (+1 SD line). (C) Interaction between estrogen levels and cycle phase shown when accuracy is distinguished via specific emotion. In the follicular phase, the +1 SD line shows that accuracy is higher for women who have higher levels of estrogen and the -1 SD line shows accuracy is lower for women who have lower levels of estrogen. (D) Illustration of the interaction between estrogen levels and the days squared variable when accuracy is distinguished via specific emotion. Accuracy is highest in the middle of cycle for women who have higher levels of estrogen (+1 SD line). Accuracy levels are generally higher when distinguished by specific emotion vs. presentation modality because accuracy for the reference category (anger) is higher than for all other emotions.

https://doi.org/10.1371/journal.pone.0312404.g001

As secondary analyses, we tested for a potential quadratic relationship between the days of the cycle and accuracy. In this Model 2, Table 1, neither day of cycle nor the quadratic term were significant. The indicator variables that represented the auditory and multimodal modalities were significant compared to the baseline visual modality, which is similar to the prior set of analyses. The effect of testing occasion was also significant, with greater accuracy in the second occasion than the first. This again represents a training effect for participants engaging in the same task on two occasions. The estimated marginal means for the secondary analyses are reported in S1 Table. Between-person estrogen predicted greater accuracy, which indicates that generally having higher levels of estrogen predicts greater accuracy. We observed a significant interaction between within-person levels of estrogen and the days-squared variable (Fig 1B), which suggests that accuracy is highest in the middle of the cycle for women with higher levels of estrogen. The interaction between progesterone and the days-squared variable was not significant.

Specific emotions.

In addition to the effects of modality, we examined the influence of specific emotions. In Model 3, Table 1, the coefficient for cycle phase was not significant, indicating no difference in recognition accuracy between the follicular (M = .430, 95% CI [.407, .454]) and luteal (M = .437, 95% CI [.407, .467], d = .039) phases. Accuracy varied across the various emotional categories. Table 2 shows a matrix of comparisons between emotions, with Bonferroni adjusted p-values. Anger (which was also the reference category in the analyses) was the most recognized emotion and anxiety the least recognized. There was a significant effect of testing occasion, again representing a training effect, with lower accuracy in the first (M = .409, 95% CI [.383, .436]) vs. the second occasions (M = .458, 95% CI [.433, .483], d = .276). Both within-person and between-person levels of estrogen were significant predictors of accuracy. In this case, within-person levels of estrogen negatively predicted accuracy and between-person levels of estrogen positively predicted accuracy, as in Model 1, Table 1. Also similar to the preceding analyses, within-person estrogen interacted with the follicular phase indicator, such that women with higher levels of estrogen at the follicular phase were more accurate (see Fig 1C). A similar interaction between progesterone and follicular phase indicator was not significant.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Accuracy on the ERAM task.

https://doi.org/10.1371/journal.pone.0312404.t002

We repeated the analyses using the quadratic term in place of the follicular phase indicator. In Model 4, Table 1, neither day of cycle nor days squared were significant. A significant coefficient of test occasion was observed, reflecting a training effect between the first and second occasion, as in the prior analyses (see S1 Table for estimated marginal means). Within-person level of estrogen marginally predicted accuracy and between-person estrogen significantly predicted accuracy. We again observed a significant interaction between estrogen and the quadratic term, as shown in Fig 1D. An interaction between progesterone and the quadratic term was not significant.

Cycle phase differences and hormones for the VENEC test

Table 3 shows the results from four linear mixed effects models using emotion recognition accuracy (Hu) from non-linguistic vocalizations as the outcome measure. Positive emotions (Models 1 and 2) and negative emotions (Models 3 and 4) were included in separate models, as detailed below. Selected estimated marginal means from the models reported below are available in S1 Table.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Mixed effect models of emotion recognition accuracy for non-linguistic vocalizations (VENEC).

https://doi.org/10.1371/journal.pone.0312404.t003

Positive emotions.

Starting with positive emotions, Model 1, Table 3, showed a non-significant coefficient for cycle phase, with no difference between recognition accuracy in the follicular (M = .322, 95% CI [.298, .346]) vs. luteal (M = .317, 95% CI [.288, .345], d = .041) phase. However, there was a significant effect of testing occasion, with lower accuracy in the first (M = .309, 95% CI [.283, .334]) vs. second (M = .330, 95% CI [.306, .353], d = .153) occasion. Neither estrogen nor progesterone predicted accuracy. There was considerable variability in the accuracy of individual emotions, with Bonferroni corrected pairwise comparisons shown in Table 4.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Accuracy on the positive emotions on the VENEC task.

https://doi.org/10.1371/journal.pone.0312404.t004

As with the ERAM analyses, we examined if there was a quadratic relationship between cycle phase days and accuracy. In Model 2, Table 3, the day of cycle, days squared, and estrogen did not predict accuracy. However, within-person levels of progesterone negatively predicted accuracy. There was also a marginally significant coefficient for testing occasion, with lower accuracy in the first occasion compared to the second.

Negative emotions.

In Model 3, Table 3, the coefficient for cycle phase was not significant, with no significant difference between the follicular (M = .441, 95% CI [.417, .465]) and the luteal (M = .430, 95% CI [.401, .459], d = .078) phases. There was no significant difference in accuracy between the first (M = .428, 95% CI [.402, .454]) vs. second (M = .444, 95% CI [.420, .468], d = .115) occasions. Within-person estrogen negatively predicted while between-person positively predicted accuracy. In addition, within-person estrogen significantly interacted with the follicular phase indicator, suggesting higher accuracy in the follicular phase for women with higher levels of estrogen (see Fig 2). Neither progesterone nor interactions between progesterone and cycle phase predicted accuracy. There was considerable variability in the accuracy of individual emotions, with Bonferroni corrected pairwise comparisons shown in Table 5.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Interaction between estrogen and cycle phase in predicting emotion recognition accuracy for negative emotions (VENEC task).

Interaction between estrogen levels and cycle phase shown when accuracy is distinguished via type of negative emotions. In the follicular phase, the +1 SD line shows that accuracy is higher for women who have higher levels of estrogen and the -1 SD line shows accuracy is lower for women who have lower levels of estrogen. As in Fig 1, accuracy levels are high when distinguished by specific emotion because accuracy for the reference category (anger) is higher than for all other emotions.

https://doi.org/10.1371/journal.pone.0312404.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Accuracy on the negative emotions on the VENEC task.

https://doi.org/10.1371/journal.pone.0312404.t005

As with the analyses above, we further examined if there was a quadratic relationship between the cycle phase days and accuracy. In Model 4, Table 3, there were no significant coefficients for the day of cycle term, days squared term, within-person estrogen, nor progesterone. However, between-person levels of estrogen predicted accuracy as in the prior analyses, and the interaction between progesterone and days squared was marginally significant.

Discussion

Our first main finding was the lack of significant associations between menstrual cycle phase and the ability to recognize emotions from dynamic multimodal expressions (ERAM) or non-linguistic vocalizations (VENEC). This was the case in both the main analyses and in additional robustness tests.

Our second main finding was a significant positive association between recognition accuracy and between-person estrogen level, yet a negative association between accuracy and within-person estrogen level. Taken together, this suggests that having generally higher estrogen may increase accuracy, whereas a higher-than-average level of estrogen at any given time may instead decrease it. These results were consistent in analyses involving positive and negative emotions in the ERAM test and for negative emotions for the VENEC test. In terms of robustness, the negative associations between accuracy and within-person estrogen occurred only in the models that controlled for cycle phase, and not in the models that included a continuous measure of days of the cycle, whereas the positive associations between accuracy and between-person estrogen appeared across both types of models.

Our third main finding was consistent interactions between cycle phase and within-person level of estrogen for the ERAM task, which suggests that women with higher levels of estrogen were more accurate when tested in the follicular phase.

Finally, our fourth main finding concerned a quadratic relationship between days of the menstrual cycle and emotion recognition accuracy. While we did not observe significant coefficients for the quadratic term, there were consistent significant interaction coefficients for within-person levels of estrogen across analyses for the ERAM test. This suggests an inverse-U like relationship where higher level of estrogen was associated with greater accuracy during the middle of the menstrual cycle, but with lower accuracy toward the beginning and the end of the cycle.

In addition to the main findings, our analyses detailed how emotion recognition varied as a function of task characteristics such as presentation modality and emotion. Because these findings did not relate to cycle phase or hormone levels, they are not discussed in depth, although they replicate well what has been observed in previous studies using the same emotion recognition tasks (e.g., [37, 46]). We also noted a significant negative association between within-person progesterone level and recognition accuracy in one analysis for the VENEC test. While this is consistent with some previous studies [16, 18], we hesitate to emphasize this finding due to concerns about robustness, given that results for progesterone tended to vary across analyses in contrast with consistency for analyses of estrogen.

Contrary to several previous studies [13, 16–20], but in line with others [25–28], we did not observe menstrual cycle phase differences on emotion recognition accuracy. We suggest that the mixed results regarding cycle phase differences could result, at least in part, from variability in the method used to define and measure cycle phases in different studies (see [28]). Whereas some studies (like ours) have only categorized participants into two broad categories (follicular and luteal), other studies have instead attempted more fine-grained classifications (e.g., early follicular, ovulatory, luteal; see e.g., [20, 48]). If the effects of sex hormones vary continuously across the menstrual cycle, then more detailed categorization would be preferable because it would, for example, allow one to detect effects that occur specifically during ovulation. Luteinizing hormone tests could also be used to validate cycle phases (e.g., [49, 50]). Another limitation of our study, and several others, is the use of the forward method—that is, counting days from the onset of the last menses forward to the day of testing—to categorize cycle phase. It has been suggested that this method is prone to errors, and more reliable ways of obtaining self-reports of cycle phase have been proposed [49, 51]. We suggest that future studies could combine reliable measures of cycle phase with measurement of sex hormones to achieve a deeper understanding of how emotion recognition accuracy varies as a function of cycle phases. Alternatively, studies could also dispense with the categorization of cycle phases, and instead use days of the menstrual cycle as a continuous variable, akin to what we attempted in our analyses on a possible quadratic relationship between cycle day and emotion recognition.

A relatively novel feature of our study was that we assessed associations between both within- and between-person hormone levels and emotion recognition accuracy (see also [27]). Interestingly, the coefficients pointed in different directions with positive associations for between-person levels of estrogen, but negative associations for within-person levels (see e.g., Fig 1). This pattern (higher estrogen levels on the aggregate being beneficial, but higher-than-usual estrogen being detrimental at any one time) is called Simpson’s paradox and has been observed in multiple contexts. In one classic example, Bickel and colleagues [52] discussed the paradox in terms of admission rates at a university. While it may appear on the aggregate that male applicants were admitted at a higher rate than female applicants, the pattern reversed when examining admission rates by department—women were more likely to be admitted than men at the level of the department (see [53] for a detailed discussion). In the case of sex hormones and emotion recognition, some previous studies have reported positive associations between estrogen levels and emotion recognition accuracy (e.g., [21, 22]), whereas others have reported negative associations (e.g., [20, 25]). In light of our findings, we suggest that such mixed findings could result, at least partly, from design choices and analysis methods, where between- and within-person designs could potentially lead to different conclusions.

We also observed significant interactions that suggest the associations between estrogen and recognition accuracy may vary across the phases and days of the menstrual cycle. These interactions appeared most consistently for the ERAM test, and showed that women with higher estrogen levels had higher accuracy in the follicular phase and in the middle of the menstrual cycle. Such interactions may have also potentially contributed to the mixed findings in previous studies. Considering that estrogen levels vary the most during the follicular phase, with a peak during ovulation but relatively stable values during the luteal phase (e.g., [10]), the observed interactions could reflect increased emotion recognition accuracy closer to ovulation (see [13]).

Our findings regarding estrogen should be interpreted with caveats. All findings should be regarded as tentative until replicated, and further research is needed especially to understand better how between-person and within-person estrogen levels may have different associations with emotion recognition. Our observations do not allow us to draw conclusions about causality and further research is needed to establish potential causal roles for estrogen in enhancing or decreasing emotion recognition ability. Such studies could, for example, compare groups of women that use hormonal vs. non-hormonal contraceptives at several occasions across the menstrual cycle, which would allow some experimental control of hormone levels (e.g., [28, 48]). We also note that the validity of the immunoassay method to measure salivary estrogen and progesterone, which has been used in many studies including ours, has recently been criticized [54]. Replication efforts could thus seek additional techniques for hormone measurement such as liquid chromatography-mass spectrometry. Finally, the study was not pre-registered and a pre-registered replication attempt could ensure the robustness of the findings.

Replication efforts could systematically investigate the impact of task characteristics, given that previous work often finds associations with estrogen only in some conditions but not others (e.g., only for some emotions) (e.g., [13, 20, 21]). Previous studies have usually measured emotion recognition based on a small number of emotions appearing in still pictures of facial expression. In contrast, we measured recognition of a wide range of emotions using dynamic multimodal expressions and non-linguistic vocalizations. We argue that the latter approach provides a more ecologically valid measure of emotion recognition ability, as it manifests in daily social interactions (e.g., [30]). Given the large number of emotions included in our tasks, the most reliable measures included all together rather than examining them each separately, and further investigation of possible emotion-specific effects using more ecologically valid stimuli remains an exciting topic for future research.

The small sample sizes (typically around 20–40 participants, but see [27, 28]) and cross-sectional designs employed in previous studies may have caused problems with statistical power [49, 50, 51]. In our study, by contrast, we used a different type of design in which most participants were tested twice at different phases. Although repeated measurement does increase power, it can also lead to learning effects with improved performance between the two test occasions, and these effects can make it more difficult to assess the possible associations between cycle phase and hormones. Learning effects appeared especially for the ERAM test in our study (see also [55]). We suggest that future studies could develop tasks and designs that minimize learning effects while still benefiting from the advantages of within-subjects measurement.

To conclude, the current study expanded upon previous research on sex hormones, cycle phase and emotion recognition by using dynamic multi-modal stimuli of a wide range of positive and negative emotions. The main positive finding was that level of estrogen was associated with emotion recognition accuracy, albeit not in a straight-forward way. Between-person estrogen was positively associated with accuracy, within-person estrogen was negatively associated with accuracy, and associations with estrogen further varied across the cycles and days of the menstrual cycle. Analyses did not demonstrate consistent significant cycle phase differences for emotion recognition across analyses. These findings contribute another piece in the puzzle to the growing literature on the psychological correlates of sex hormones. More research is needed to understand the possible causal mechanisms of how sex hormones may affect psychological functioning factors that include emotion recognition, and such knowledge could have important implications for well-being and social functioning.

Supporting information

Acknowledgments

Data and code are available on the Open Science Framework (https://osf.io/uw6a7/). The present study was not pre-registered.