Sex Differences in Depressive Symptoms and Their Networks in a Treatment-Seeking Population – a Cross Sectional Study

Abstract

Background.

The higher prevalence of major depressive disorder (MDD) in females relative to males is well-established. Some authors have posited this difference arises to divergent symptom profiles in females vs. males. However, empirical tests of this hypothesis have yielded equivocal results. Here, we investigate sex differences in MDD of individual symptoms and symptom networks in a treatment-seeking sample.

Methods.

We assessed depressive symptoms using Hamilton Depression Rating Scale (HDRS-17) in 590 treatment-seeking adults with MDD (300 females). We examined group differences in symptom endorsement. We investigated symptom networks and estimated Gaussian Graphical Models. Finally, we compared the female and male networks using the NetworkComparisonTest.

Results.

Females scored significantly higher in psychological anxiety (p <.001; rB = −0.155), somatic anxiety (p = .001; rB = −0.150) and feelings of guilt (p = .002; rB = −0.139). Male and female patients did not differ in depression sum scores. There were no sex differences in network structure or global strength.

Limitations.

Our study was sufficiently powered to detect only medium sized symptom differences. The generalizability of our study is limited to clinical samples and further studies are needed to investigate if findings also translate to outpatient samples.

Conclusion.

Females reported elevated anxiety symptoms and guilt. Clinicians should assess these symptom differences and tailor treatment to individual symptom profiles. No differences between sexes emerged in MDD network structures, indicating that features may be more similar than previously assumed. Sex differences in psychopathological features of MDD are important for future research and personalized treatment.

Introduction

The increased prevalence of major depressive disorder (MDD) in females compared to males has been reported consistently over time and in different populations (Kessler et al., 2003; Lim et al., 2018). The reason for these differences in prevalence has been attributed to various underlying sex differences, for example in neurobiology (Rubinow & Schmidt, 2019) and the prevalence of subtypes of depression (Silverstein et al., 2017). With regard to the latter, multiple studies found that females are more often affected by “atypical” depression (Angst et al., 2002; Marcus et al., 2008). Atypical depression is defined by a specifier of MDD in different versions of the Diagnostic and Statistical Manual of Mental Disorders (DSM) by the predominance of symptoms such as weight gain or hypersomnia (Blanco et al., 2012; Marcus et al., 2008) compared to symptoms considered more typical such as disturbances of appetite and sleep disturbances. Depressed females tend to report more somatic symptoms (e.g., gastrointestinal symptoms and loss of appetite; (Angst et al., 2002; Silverstein et al., 2013) and further symptoms not included in the DSM definition of MDD (e.g., loss of libido and anxiety; Martin, Neighbors & Griffith, 2013). One meta-analysis included data from 32 studies with a total of 108,260 individuals and found differences in the symptom profiles of females and males. Females reported higher intensity of somatic symptoms (e.g., sleep disturbance or fatigue) but also of core symptoms of depression (e.g., depressed mood or diminished interest; Cavanagh et al., 2017). Males presented more often with impulse control problems and substance use (e.g. nicotine/tobacco; Lamers et al., 2011). These differences, however, were moderate and heterogeneity was generally large. This might be due to the vast variety of measures used to assess depressive symptoms across studies, variances in definitions of depression and symptoms, and differences in participant samples and study designs (Cavanagh et al., 2017). Thus, it is not surprising that numerous different factor structures of depression have been proposed (Pancheri et al., 2002). Consequently, there is also evidence that differential symptom profiles may be associated with distinct risk factors (Fried et al., 2014). Several risk factors (depression history, childhood stress, stressful life events, and sex) all predicted progression of depressive symptoms, they each had a specific association to individual depressive symptoms (Fried et al., 2014). Moreover, there is evidence that cognitive and memory biases are more strongly related to some, but not all, symptoms of depression (Beevers et al., 2019; Marchetti et al., 2018). Finally, treatment of depression does not affect all symptoms uniformly (Boschloo et al., 2019; Mullarkey et al., 2020). Taken together, this adds to the importance of symptom-level approaches in depression research.

Differences on the symptom level are of relevance when depression is investigated from a network approach perspective (Fried, van Borkulo, et al., 2017). This approach conceptualizes mental disorders based on a complex systems theory framework as the interaction of their symptoms (Borsboom & Cramer, 2013). For example, in depression, insomnia may lead to concentration problems. This may then negatively impact one’s performance at work, which again exacerbates insomnia, because one ruminates about the low performance at work (Cramer et al., 2016). According to Borsboom (2017), current conceptualizations of mental disorders presume an underlying latent disease entity to be the common cause of the symptoms that reflect its presence. From this perspective, symptoms are conceptualized to be diagnostically equivalent and interchangeable (Cramer et al., 2010; Lux & Kendler, 2010) and can therefore be summed up to an overall score indicating the severity of a mental disorder (Fried, 2015; Fried & Nesse, 2015b). This seems not only clinically implausible, but also negates the potential relevance of differences in symptom profiles of depression in females and males. While traditional conceptualizations have their shortcomings, novel approaches may be an alternative. The network approach posits that symptom co-occurrence arises not from a common underlying cause, but from symptom-to-symptom interactions (Borsboom, 2017). From this network perspective, differences between groups in the severity of individual symptoms or the interactions among those symptoms are exceedingly important as they indicate the possibility of differences between those groups in the causes of depression. Nevertheless, it has been noted, that the contrast between the two interpretations of correlations among symptoms (reflecting an underlying disorder versus a causal network), is likely less dogmatic than often pictured (Bringmann & Eronen, 2018). Furthermore, there is no general test to investigate if the “true” model is a network or a factor model (Fried, 2020).

Although there are numerous studies that investigate depression from a network perspective (e.g., van Borkulo et al., 2015), including studies about the association of depressive symptoms with environmental and genetic risk factors (van Loo et al., 2018) and co-expression of symptoms of anxiety and depression (Beard et al., 2016), only one study has analysed sex differences (Mullarkey et al., 2018). Mullarkey et al. (2018) found that the depressive symptom networks of 646 male and 744 female adolescents from the general population differed in one relationship between two symptoms, namely that the association between self-hatred and negative body image was stronger in females. In another non-peer-reviewed preprint, van Borkulo et al. (2017) reported no differences between the symptom networks of depressive symptoms of 351 male and 701 female adults of a clinical population. To our knowledge, no published study has analysed both sex differences in depression symptom networks in adults, despite the importance of these analyses for our understanding of depression and the role of sex in its development. The main purpose of the study was an explorative assessment of sex differences in depressive symptom profiles and symptom networks in a clinical sample of adults with a diagnosis of MDD. We investigated the interaction of these symptoms in treatment seeking patients of a psychiatric hospital.

Methods

Participants and Procedure

As part of the regular clinical documentation at admission to the department for affective disorders at the Psychiatric Hospital of the University of Zurich, Switzerland, inpatients and day-clinic patients undergo basic clinical assessment. Data were collected as part of the routine clinical care procedure and completely anonymized. No specific written informed consent was thus obtained.

We analysed data from patients admitted during July 2007 to June 2018. For this study, inclusion criteria were: age between 18 and 70, clinical diagnosis of major depressive disorder (single episode or recurrent; ICD-10 code F32 or F33; World Health Organization, 1992) as the main diagnosis, and data was only included from the first admission. To ensure the same power for detecting differences for all symptoms, we excluded participants if they failed to provide rating for all HDRS-17 items. Among the 611 patients who met the inclusion criteria, this led to the exclusion of additional 21 participants, leaving a final sample of 590 patients (300 female and 290 male patients).

Measures

Sex was determined as a binary variable (female, male). Individual sex was obtained from the patient’s electronic patient record, which is documented in accordance with the sex on the official ID-card or passport.

Depressive symptoms were assessed using the clinician-administered Hamilton Depression Rating Scale (HDRS-17; Hamilton, 1960). Clinically experienced physicians and psychologists completed HDRS questionnaires during the first three days of admission. The intensity of some symptoms (e.g., depressed mood, suicidal ideation, decreased interest, or psychomotor retardation) was rated from 0 to 4 (0 = none/absent to 4 = most severe), whereas other symptoms’ intensity (e.g., decreased appetite or general somatic symptoms) was rated from 0 to 2 (0 = none/absent to 2 = severe). The HDRS-17 has extensively been used in research on depression (Trajković et al., 2011). The reliability of the total depression symptom scale derived in our study (Cronbach’s α = .74) was comparable to other studies (Trajković et al., 2011).

Statistical Analyses

First, we compared females and males on both their overall HDRS score as well as their endorsement of individual HDRS items. Due to evidence of a non-normal distribution of the data, we used non-parametric Mann-Whitney U tests. To control for the family-wise error rate, Bonferroni correction was applied. Given the current sample size and global alpha level of .05, post-hoc power to detect small, medium and large effects (d = 0.2, 0.5, 0.8; Cohen, 1988), was 0.68, 1.00 and 1.00, respectively (calculated with g*power; Faul et al., 2009). Second, we used network analysis to examine relationships among HDRS items. We modelled our network analyses after those completed by Fried et al. (2017), and structured our results using the four components identified by those authors: (a) network estimation; (b) network characterization; (c) network stability, and (d) network comparison. All network analyses were carried out using R version 3.6.2. Networks were visualized using the R-package qgraph (Epskamp et al., 2012). All other statistical analyses were conducted with JASP Version 0.9.2.0 (JASP Team, 2020).

Given that a lack of variability of an item can bias the network structure (Terluin et al., 2016), we checked the variance of all items. One item (Insight) showed very low variability and was therefore excluded from the network analysis. In addition, prior to network estimation, we tested if the included symptoms overlapped, using the suggested default settings of the goldbricker function of the networktools package (Jones, 2019). This analysis suggested that the items Insomnia – Early and Insomnia – Middle had more than 75% (but less than 90%) of topologically overlapping correlations. Thus, we investigated the effect of removing each item individually on our results in two sensitivity analyses. Further details are available in the supplement.

Network estimation

Following the recommendations for network analysis with cross-sectional ordinal data (Costantini et al., 2017; Epskamp et al., 2018), we estimated partial correlation networks. In the resulting network (a Gaussian Graphical Model; GGM), nodes represent symptoms of depression, and edges represent partial correlations between symptoms. The techniques used to estimate the networks are both based on graphical lasso, which is a regularization procedure (Costantini et al., 2017; Epskamp et al., 2018). First, we used the Fused Graphical Lasso (FGL) to estimate the networks for females and males jointly. We used the information criterion based FGL, to facilitate comparison with individually estimated networks (Costantini et al., 2017). Second, we estimated the networks for the subsamples individually for technical reasons (e.g., stability analysis is only available for individually estimated networks, see below). We used the R-packages bootnet (Version 1.2.2) (Epskamp et al., 2018) and EstimateGroupNetwork (Costantini & Epskamp, 2017).

We used the graphical lasso estimation procedure to estimate the networks (Epskamp et al., 2018). First, we used the Fused Graphical Lasso (FGL) to estimate the networks for females and males jointly. We used the information criterion based FGL, to facilitate comparison with individually estimated networks (Costantini et al., 2017). Second, we estimated the networks for the subsamples individually for technical reasons (e.g., stability analysis is only available for individually estimated networks, see below).

Network characterization

Subsequently, we examined the characteristics of individual nodes within the network. First, we computed undirected strength centrality to assess the relative interconnectedness of nodes in a network. Next, we estimated “predictability” using the mgm package (Haslbeck & Waldorp, 2016). “Predictability” is the upper bound of the shared variance of a given node (measured in R²) with all its neighbours, assuming that all connections go towards the given node (Haslbeck & Fried, 2017).

Network stability

Stability and accuracy analyses, as well as the estimation of predictability are not yet available for jointly estimated networks. We used the R-package bootnet Version (1.1) (Epskamp et al., 2018) to assess the stability for the individually estimated networks, reflecting a lower bound for stability for the jointly estimated networks. Bootnet uses bootstrapping procedures to compute 95% confidence intervals for the edge weights, and to calculate the correlation-stability coefficient. We performed edge weights difference test and centrality difference tests. The results of all these analyses are outlined in the Supplementary materials.

Network comparison

We used the R-package NetworkComparisonTest (NCT; van Borkulo et al., 2017) to test for differences in network structure (assuming that the structure of both networks is exactly the same), global strength (assuming that overall connectivity in both networks is exactly the same) and edge strength (assuming that all edges of both networks are of similar strength) between the female and male symptom networks (van Borkulo et al., 2017).

Results

Sample characteristics

Average age (in years) did not differ between females (M = 43.44, SD = 12.80) and males (M = 43.99, SD = 12.26); t(1) = 0.524, p = .601). Female and male patients did not differ in their main diagnosis (F32 or F33), use of medication, or the comorbidity with an additional mental disorder. However, males were more likely to have an additional diagnosis of Mental and behavioural disorders due to psychoactive substance use (F10-F19). Females were more likely to have an additional diagnosis of disorders associated with physiological disturbances and physical factors (F55-F59). Further sample characteristics are presented in Table 1.

Table 1.

Demographic and clinical characteristics (n = 590).

Variables	Females (n = 300)		Males (n = 290)		χ2	p-value
	n	%	n	%
MDD, single episode (F32)	120	40	113	39	.66	.797
MDD, recurrent episode (F33)	180	60	177	61	.66	.797
Medicationa	219	73	213	73.4	.015	.902
Non-psychotropic drugs	88	29.3	109	37.6	4.52	.034
Antidepressants	198	66	190	65.5	.015	.902
Anxiolytics	48	16	35	12.1	1.89	.170
Detoxication and withdrawal	1	.3	3	1	1.08	.299
Hypnotics	28	9.3	24	8.3	.21	.651
Neuroleptics	84	28	65	22.4	2.44	.118
Mood stabilizers	36	12	38	13.1	.164	.686
Stimulants	5	1.7	10	3.4	1.89	.169
Patients with comorbid diagnosesb	160	53.3	165	56.9	.76	.384
F00 - F09	0	0	0	0	-	-
F10 - F19	53	17.7	80	27.6	8.31	.004*
F20 - F29	2	0.7	4	1.4	.74	.388
F30 - F39	12	4	8	2.8	.69	.405
F40 - F48	74	24.7	76	26.2	.19	.668
F50 - F59	21	7		1	13.45	.000*
F60 - F69	44	14.7	22	7.6	7.44	.006
F70 - F79	0	0	1	.3	1.04	.309
F80 - F89	0	0	0	0	-	-
F90 - F98	9	3	14	4.8	1.32	.252

MDD = Major depressive disorder

^aPatients can take more than one drug.

^bPatients can have more than one comorbid diagnosis.

^*p < .05.

^**p < .01. – Adjusted p-value for medication: 0.00625 – Adjusted p-value for comorbidities: 0.005

F00-F09: Mental disorders due to known physiological conditions, F10-F19: Mental and behavioral disorders due to psychoactive substance use, F20-F29: Schizophrenia, schizotypal and delusional disorders, F30-F39: Mood [affective] disorders, F40-F48: Neurotic, stress-related and somatoform disorders, F50-F59: Behavioral syndromes associated with physiological disturbances and physical factors, F60-F69: Disorders of adult personality and behavior, F70-F79: Mental retardation, F80-F89: Disorders of psychological development, F90-F98: Behavioral and emotional disorders with onset usually occurring in childhood and adolescence.

Individual MDD Symptoms

We found no sex difference in the sum score of the HDRS-17. Regarding individual items, female patients had higher ratings in Anxiety Psychic (p <.001; W = 36763; r_B = −0.155), Anxiety Somatic (p = .001; W = 36979; r_B = −0.150) and Feelings of Guilt (p = .002; W = 37468; r_B = −0.139). No item was more frequently endorsed by males. Mean scores and standard deviations for each of the 17 depression symptoms indexed by the HDRS are presented in Figure 1. Two sensitivity analyses, individually excluding patients with an additional F10-F19 or F50-F55 diagnoses showed the same differences (for more details see the Supplementary Materials).

Figure 1.

Mean endorsement for each symptom for female (n = 300) and male patients (n = 290)

Standard error is presented in the figure by the error bars.

DeprMD = Depressed Mood, Guilt = Feelings of Guilt, Suic. = Suicidal ideation, InsEa = Insomnia – Early, InsMid = Insomnia – Middle, InsLa = Insomnia – Late, WorkAct = Work and Activities, Retard = Retardation, Agit = Agitation, AnxPsy = Anxiety Psychic, AnxSom = Anxiety Somatic, SomGast = Somatic Symptoms – Gastro-intestinal, SomGen = Somatic Symptoms – General, Genital = Genital Symptoms, Hypoch = Hypochondriasis, Weight = Loss of Weight; Insight = Insight

MDD Symptom Networks

The networks for males and females are presented in Figure 2. The NCT revealed no differences between sexes in network structure (p = 0.5578), or global strength (p= 0.2626; see R script for more details). Because the network structure was found to be invariant, we did not test individual connection strengths (van Borkulo et al., 2017). Neither removing Insomnia – Middle or Insomnia – Early from the network, as suggested by the goldbricker function, nor excluding patients with an additional F10-F19 or F50-F59 diagnosis altered the results (see Supplementary Materials). In accordance with the NCT, visual inspection of the network suggests that depressive symptom networks for females and males shared many edges and network features. Across both sexes, Somatic Symptoms – Gastro-intestinal and Loss of Weight were strongly connected. There were also strong connections among the three Insomnia items, as well as connections between Anxiety Somatic and Hypochondriasis in both males and females. Agitation showed the weakest edges in both sexes’ networks.

Figure 2.

Network of depressive symptoms for female (left, n = 300) and male (right, n = 290) treatment-seeking adults

Nodes represent symptoms, edges partial correlations. Blue edges indicate positive, red edges negative correlations. The bigger the partial correlation, the ticker the edge. The grey area in the rings around the nodes depicts predictability (the upper bound of variance of a given node explained by all its neighbours), with a full circle corresponding to an R2 of 1.

DeprMD = Depressed Mood, Guilt = Feelings of Guilt, Suic. = Suicidal ideation, InsEa = Insomnia – Early, InsMid = Insomnia – Middle, InsLa = Insomnia – Late, WorkAct = Work and Activities, Retard = Retardation, Agit = Agitation, AnxPsy = Anxiety Psychic, AnxSom = Anxiety Somatic, SomGast = Somatic Symptoms – Gastro-intestinal, SomGen = Somatic Symptoms – General, Genital = Genital Symptoms, Hypoch = Hypochondriasis, Weight = Loss of Weight; Insight = Insight

Individual Networks

Centrality

Node strength estimations are shown in Figure S1. For both sexes, nodes with high strength values were Depressed Mood and Anxiety Somatic. For females, the node with the highest strength was Insomnia – Middle, for males Depressed Mood. Agitation had lowest centrality estimates in both sexes. However, the CS-coefficient for strength for the female networks was 0.440 and the male network 0.438, both not exceeding the recommended threshold of 0.5 (Epskamp et al., 2018). In line with these results, the significance testing revealed that only Agitation had lower centrality than most other items in the female and male network (Figure S2 for females and Figure S3 for males). In the male network, Depressed Mood exhibited greater node strength than the majority of the other symptoms (Figure S3). Details and results of the significance tests for the individual edge weights are available in the Supplementary materials.

Predictability

For both sexes, Agitation had the lowest predictability. Highest predictability was found for Insomnia – Middle (.367) in females and for Depressed Mood (.414) in males. Mean predictability was .214 for females and .224 for males.

Discussion

The current study investigated sex differences in depressive symptom profiles and in the corresponding symptom network structure in an adult, treatment-seeking population suffering from major depressive disorder. Females reported more anxiety and more feelings of guilt than males. We did not find any additional sex differences in individual symptom endorsement or in the overall sum score. Moreover, no differences in symptom networks emerged.

In accordance with a meta-analysis on sex differences in depressive symptoms by Cavanagh et al. (2017), we found that females reported more anxiety than males. Moreover, this was the case for both psychic (e.g., disproportional worries or fear) and somatic symptoms (e.g., tremor or sweating) of anxiety. It is important to note that we based our analyses on a treatment seeking population and our data facilitated an estimation of the effect of comorbidities, while in the aforementioned meta-analysis (Cavanagh et al. 2017) none of the included studies reported any existing comorbidities. Also, in line with the meta-analysis was the small effect size for both differences. One reason females report more anxiety might be a higher prevalence of comorbid anxiety disorders in females, which has been well documented (McLean et al., 2011; Simonds & Whiffen, 2003). In our sample, however, the rate of co-morbidity with a neurotic, stress-related, or somatoform disorders (F4; World Health Organization, 1992), which include anxiety disorders, was equal in both sexes, suggesting that differences in comorbid anxiety cannot account for our findings. Hence, further research is needed to identify and evaluate potential causes for the increased level of anxiety in females with MDD. Independent of the underlying cause, increased levels of anxiety have been shown to negatively impact the efficacy of treatment in MDD (Fava et al., 2008). Therefore, clinicians should carefully assess symptoms of anxiety in MDD patients, especially in females. Additionally, clinicians should be aware that males may tend to underreport anxiety symptoms (Bekker & van Mens-Verhulst, 2007), resulting in a reporting bias (but also see McLean & Hope, 2010).

We also found females to express more feelings of guilt than males. Comparing our result with the meta-analysis by Cavanagh et al. (2017) has limited validity, because they combined worthlessness and guilt in one item. Still, they did not find a significant difference for the composite item. Our results and also the effect sizes, however, are in line with the sex difference in experiencing guilt in general (Else-Quest et al., 2012). Because our sample consisted of treatment-seeking individuals, an additional reason for higher levels of guilt in females might be the increased attention to emotions in females with severe depression (Thayer et al., 2003). Hence, we suggest clinicians to be aware of elevated levels of guilt in this sub-population. Moreover, clinicians should be mindful of the presence of anxiety and guilt in female patients, considering the potential role of these symptoms within the patients’ case conceptualizations and, therefore, assess these symptoms pro-actively. Our results may suggest a focus on anxiety and guilt as possible treatment targets.

No other sex differences in symptom severity emerged. This is also partially consistent with the meta-analysis by Cavanagh et al. (2017). They did not find sex differences with regard to suicidality, psychomotor retardation and hypochondriasis. However, in contrast with our findings, they reported that females suffered more from depressed mood, problems with sleep, somatic difficulties and problems with appetite or weight. There are several potential reasons for these different findings. First, the meta-analysis covered symptoms assessed with different questionnaires. Thus, the HDRS symptoms did not exactly match with the items analysed in the meta-analysis. Second, regarding the results of the individual studies, the heterogeneity was mostly high (Cavanagh et al., 2017). This limits the grade of evidence of the meta-analysis. One of multiple reasons underlying this heterogeneity is the varying method of symptom assessment. Several studies found significant differences in overall self-reported symptom severity. However, no significant sex differences were found in clinical assessments (Dekker et al., 2008; Kornstein et al., 2000; Scheibe et al., 2003). We used the same type of assessment, i.e., clinician-based, and this has to be considered. Third, the effect sizes reported in the meta-analysis by Cavanagh et al. (2017) were small or even very small. Our study achieved a post-hoc calculated power of 0.69 for small effect sizes after the correction for the family-wise error rate. Hence, our limited power may have resulted in false negative findings (see limitations).

We did not find any sex differences regarding the network structure or global network strength. This is in line with the results of Mullarkey et al. (2018) who reported no difference in global strength in depressive symptom networks of female and male adolescents. Although they found the network structure to be different across the two sexes, this was due to one out of dozens of edges. Furthermore, this edge was between two symptoms not commonly attributed to depression, namely self-hatred and negative body image. Our findings are also in line with a preprint by van Borkulo et al. (2017), which found no sex differences in symptom networks in depressed adults. There are several potential reasons for these null findings. First, the NCT is a conservative test (van Borkulo et al., 2017; Williams et al., 2019). Therefore, investigations with larger samples might be able to detect sex differences in network structures. Moreover, additional statistical tools for comparing networks are under current development (e.g., Haslbeck et al., 2019) and at least one will have higher power to find differences (Williams et al., 2019). Applying these methods, once available, might reveal sex differences. Second, statistical power to detect the effect of a variable on network structure depends on the sample and effect size, but also on the model (i.e., the symptoms included). A GGM is estimated on the basis of all individual symptoms’ variance accounted for by all other included symptoms (Epskamp et al., 2018). Consequentially, when networks are compared with each other, only the shares of the variance accounted for by the networks are compared with each other (e.g., using the NCT; van Borkulo et al., 2017). Hence, given an effect and sample size, the statistical power of the NCT directly relates to the total amount of variance accounted for by the models which are compared with each other. Predictability is an estimate for the upper bound of variance (measured in R²) of a given symptom explained for by all other items in the network. In our study, mean predictability was approximately 0.2 for both sexes, which is comparable to the mean predictability in other studies of networks of depressive symptoms (Haslbeck & Fried, 2017). In other words, on average at least 80% of a symptom’s variance was not explained by other symptoms. This additionally limits the power of the NCT. Moreover, such values for mean predictability are not only low absolute values, but also relatively low, when compared to mean predictability in network models of other disorders (Haslbeck & Fried, 2017). This indicates that symptom network models of depression perform badly at capturing the variance in symptoms of depression. This poses a significant challenge for all network models of MDD. One solution is comparing predictability of network models based on different symptoms sampled out of a set of depressive symptoms and selecting a model with high predictability. However, individual symptom profiles of MDD are known to have substantial heterogeneity (Fried et al., 2016; Fried & Nesse, 2015a), which might limit the optimization of predictability in group-level network models of MDD per se. Third, as outlined above, sex differences in symptom profiles have a small effect size and individual findings are heterogeneous (Cavanagh et al., 2017). Furthermore, given the high heterogeneity of individual symptom profiles (Fried & Nesse, 2015a), one could question the utility of group level network models of symptoms of depression at all. Moreover, the network approach perspective is primarily focused on within-subject, rather than between-subject effects (Borsboom, 2017). Thus, investigations of network models based on within-subject effects (e.g., Bringmann et al., 2015) may provide important additional information about the relationships among depression symptoms in males and females.

Limitations

This study has several limitations. First, our analysis focused on individual HDRS items and we did not investigate sex differences on the factor level of the HDRS (e.g., based on the factor structure described by Cleary and Guy (1977). Due to our focus on differences on the level of individual HDRS items, our limited sample size and the fact that the software to conduct latent network analysis (Epskamp et al., 2017) was still under development, we did not estimate a latent network. Nevertheless, further research is needed to explore potential sex differences in latent networks of depression. Furthermore, such an investigation could also address a second limitation of our study, namely that some of the symptoms assessed with the HDRS probably measure overlapping constructs, which artificially inflates edge weights between these items. This was confirmed by the goldbricker function indicating topologically overlap between the items Insomnia – Early and Insomnia – Middle. As noted above, however, networks of females and males did not differ when either one of these items were excluded from the analysis. Third, as outlined above, our post-hoc calculated power to detected differences in symptom profiles with small effect size was 0.69. It is possible that we would have found more sex differences with greater power. However, to detect small effect sizes with sufficient power (0.9), while also accounting for the family-wise error rate, an approximate sample size of 1830 would have been needed. Fourth, although the HDRS is one of the most commonly used rating scales for depression (Santor et al., 2006), it does not reflect the current DSM-5 or ICD-10 definition of MDD. Still, its broad use enables comparability with existing literature. Fifth, the investigation of a clinical compared to a healthy population leads to spurious negative or weaker correlations between symptoms in a network due to Berkson’s bias (de Ron et al., 2019). This limits the generalizability of our findings based on network analysis to non-clinical populations. Sixth, to our knowledge, there is no possibility to calculate or estimate post-hoc power of an NCT-analysis. Thus, effective power of our analyses remains undetermined. In addition, the NCT is not designed to collect evidence for the null hypothesis. Consequently, we could not assess if the null hypothesis (that there are no sex differences in network structures) is true. Nevertheless, with future methodological advances (e.g., Bayesian methods for estimating GGMs; Williams et al., 2019) this will be possible. Lastly, sex was assessed as a binary variable based on participants’ sex in their official identity documents. Transgender people in Switzerland can change their sex in their identity documents but must choose between male or female. Therefore, no third option for sex besides female or male was available.

Future Directions

Based on our results and the limitations of our study, we have several suggestions for future research. With regard to sex differences of depressive symptom profiles, the heterogeneity of the existing literature should be addressed systematically. As an example, there is evidence that females with severe depression are more aware of their emotions (Thayer et al., 2003). Consequentially, the kind of symptom assessment (as a self-report or by a clinician) likely influences symptom reporting and comparison of symptom severity between females and males. We have several additional recommendations for research on sex differences in network models of MDD to the ones outlined above. First, our study needs to be replicated, especially with a bigger sample size to increase power. Second, future studies should investigate network structure in community samples, which avoids the induction of Berkson’s bias. However, these results could only be compared to clinical samples under the assumption that the underlying network structure is independent of the severity of the symptoms. Third, we strongly recommend future studies to estimate predictability to assess the limits of their model in explaining individual symptoms’ variance. Our findings provide evidence, however, that there are no, or only subtle, sex differences in the network structure of depressive symptoms.

Conclusion

This study is amongst the first to investigate sex differences in depressive symptoms and their corresponding network structure in adults. We found that females reported more symptoms of anxiety and guilt, which is in line with the literature. However, in contrast to a recent meta-analysis (Cavanagh et al., 2017), females were not found to have more symptoms of guilt. Similar to the two existing similar studies (Mullarkey et al., 2018; van Borkulo et al., 2017), we found no sex differences in symptom network structures. Given that the sex differences in individual depressive symptoms were of small effect, one potential reason for our null finding is the design of the used network analytic method, which is aimed to optimize specificity at cost of lower sensitivity. Taken together, our results indicate, that sex differences in depressive symptoms in treatment-seeking adults are few and subtle.

Highlights.

• We assessed sex differences in symptom profiles in depressed inpatients.

• Females showed higher psychological anxiety, somatic anxiety and feelings of guilt

• There were no sex differences in symptom networks

• Sex differences in psychopathological features of MDD may affect personalized treatment.