INTRODUCTION
Major depressive disorder (MDD) and bipolar disorder (BD) formalized in the Diagnostic and Statistical Manual of Mental Disorders (DSM), Third Edition, are representative mood disorders [1-3]. Nevertheless, to date, the misdiagnosis rate of MDD and BD has been considerably high in psychological settings [4]. A previous study reported that BD is frequently misdiagnosed as MDD as it has more depressive episodes than manic–hypomania [5]. Additionally, the prodromal features of BD generally arise before onset [6]; sometimes, depression precedes a diagnosis of BD [7-9]. Furthermore, patients with BD with several depressive episodes tend to accompany higher comorbidity with other psychiatric disorders and experienced more psychotic symptoms during illness phases [5].
Despite efforts to characterize features between MDD and BD, findings from studies indicate that clinicians experience difficulties discriminating between them. Several studies have identified features that differentiate patients with MDD from those with BD for early recognition [10-13]. These attempts are crucial for appropriately treating patients on the basis of their diagnosis [14]. Misdiagnosis may worsen outcomes, thereby increasing poor treatment effects [15] and cost burden. Moreover, the risk of mood instability or suicide attempts is clearly deteriorated in BD if the initial intervention is inappropriate [16-18]. In this regard, identifying BD and MDD remains a challenge and should be investigated thoroughly [19].
Conversely, several studies have reported that DSM, Fifth Edition (DSM-5) diagnosis frequently results in the delay of diagnosis and intervention in the early phase of illness duration. Therefore, to resolve this issue, balancing is needed [20,21]. Although several scales measure mood disorders including the Mood Disorder Questionnaire, Hypomania Symptom Checklist-32, Patient Health Questionnaire, and Beck Depression Scale [22], it is most practiced in patients with MDD, and scales that measure depression cannot address this issue [19].
There have been recent attempts to classify MDD and BD as a measure of affective temperament for mood disorders. The Temperament Evaluation of Memphis, Pisa, Paris, and San Diego Autoquenstionnaire (TEMPS-A) is useful for identifying affective temperament in patients with MDD and BD [23]. TEMPS-A is mainly utilized to distinguish BD I from BD II and BD from MDD by identifying the affective temperament [24-29]; specifically, cyclothymic, hyperthymic, and irritable temperaments distinguish between patients with MDD and those with BD, and depressive temperament distinguishes between patients with BD I and those with BD II [26].
Moreover, this scale was used for detecting suicide risk; Tondo et al. [30] conducted a study on patients with BD, MDD, and anxiety disorders to determine if specific affective temperaments were associated with suicide attempts. Furthermore, there is an attempt to develop a short version of the TEMPS-A, consisting of 35 items [31]. However, the consensus on the affective temperament distinguishing MDD and BD has been unclear.
The affective temperaments of the TEMPS-A for the differentiation between MDD and BD vary across studies, while the subscales that distinguish MDD from normal and BD I from BD II also vary across studies. Another limitation is that most previous studies used the full-length version of the TEMPS-A. This full-length version, however, is costly in terms of time and increases participant fatigue [23]. Although a study using the short version of the TEMPS-A consisting of 39 items (TEMPS-A-SV) was conducted and found that cyclothymic and anxiety temperaments distinguish MDD from BD [32,33], other studies supplementing this result is lacking. Therefore, we conducted a latent profile analysis (LPA) that supplements the limitations of cluster analysis conducted in previous studies using the TEMPS-A [34,35]. Latent profiles represent unobserved different latent subgroups in the population and classify participants with similar patterns on the basis of their responses into these subgroups [36,37]. This method has advantages over other traditional clustering techniques, such as k-means, by offering precise classification outcomes, unbiased means for each indicator within profiles, several statistical methods for determining the optimal number of profiles and probability-based classification rather than simple cluster assignment [37]. These features enable a more accurate assignment of participants into subgroups and provide insight towards understanding the characteristics of these subgroups. Therefore, we aimed to explore the utility of the TEMPS-A-SV by determining whether patients in the MDD, BD, and control groups have different response patterns on this scale.
METHODS
Participants
The sample consisted of patients diagnosed with BD (BD I and II) and MDD, as well as a control group, with a total of 2,458 participants. From July 2013 to January 2023, 288 patients with MDD, 538 patients with BD (BD I, n=111; BD II, n=427), and the control group (n=1,632) were included.
All patients were diagnosed with a mood disorder on the basis of the DSM-5. Board-certified psychiatrists (THH and WM) confirmed the diagnoses through structured diagnostic interviews (Mini-International Neuropsychiatric Interview) or a review of case records and other relevant data. On the other hand, participants in the control group were selected according to their sex and age through an anonymous online survey, during which they reported having no psychiatric history. This study was approved by the Institutional Review Board of Seoul National University Bundang Hospital (protocol code B-2205-756-111). Patient consent was waived because the data were gathered through a medical chart review. Comparison consent was also waived, as the researchers did not have direct access to participant personal information and used anonymized survey data for the analyses.
Measures
Five affective temperaments were evaluated using the TEMPS-A-SV, a 39-item version, in the patient (MDD and BD) and control groups. The TEMPS-A is an interview-based self-report scale, and each item received a yes-or-no questionnaire (Yes=1 and No=0) [38]. Specifically, the TEMPS-A has depressive, cyclothymic, hyperthymic, and irritable temperaments as factors [39], consisting of 110 items by adding anxious temperament later based on interview [38]. Subsequently, Akiskal et al. [23] developed the TEMPS-A-SV, which shortened the time to complete the questionnaire. The validity of this shortened version of the scale was noted reliable and valid for use in Korea [40]. The TEMPS-A-SV had the same five subscales as the full-length version, and the subscales belonged to each subscale: 12 items for cyclothymic, 8 items for depressive, 8 items for irritable, 8 items for hyperthymic, and 3 items for anxious. For analysis, we calculated the mean scores for each subscale and showed great internal consistency (ω=0.86).
Statistical analyses
To identify underlying profiles on the basis of the average scores of the five subscales, we conducted LPA using Mplus 8.5 (Muthén & Muthén, Los Angeles, CA, USA). To decide the optimal number of profiles, we fitted a series of competing models with 1–5 latent profiles and subsequently compared their results to select the final model for interpretation. Several statistical criteria including Akaike information criteria (AIC), Bayesian information criteria (BIC), sample size-adjusted BIC (SABIC), Lo–Mendell–Rubin adjusted likelihood ratio test (LMR-LRT) [41], bootstrapped likelihood ratio test (BLRT), and R2-entropy value were considered. Smaller values of information criteria (AIC, BIC, and SABIC), statistical significance in likelihood ratio tests (LMR-LRT and BLRT) (e.g., p<0.05), and higher R2-entropy values indicated better fitting of a model, respectively.
Considering the number of profiles selected in the profile enumeration stage, we conducted multiple-group LPA to evaluate whether the identical profile structure can be retained across participant groups, which consisted of patient groups (BD and MDD) and a control group. Specifically, to ensure their equality across subgroups following the guidelines by Morin et al. [42], we compared fit measures for several nested models that impose different constraints on parameters. Initially, we estimated the configural model, which posits identical number of profiles across subgroups. Subsequently, we fitted the structural similarity model wherein equality constraints were applied to profile means, suggesting that within-profile means were equivalent across subgroups. We then estimated the model with equal profile variability (e.g., the dispersion similarity model), thereby indicating similar within-profile variance across the subgroups, followed by the distributional similarity model by imposing equality constraints on size of the profiles, assuming the similar profile size across groups. To compare nested multiple models forming the sequence of similarity tests, the AIC, BIC, and SBIC were utilized [43,44].
After assessing the similarity in profile structures across subgroups, we conducted additional analyses to explore the relationships between the derived profiles and the covariates. Particularly, we considered age, sex, and participant groups as covariates, which consisted of patient groups (BD and MDD) and a control group, and subsequently investigated how they influence the latent profile memberships of the respondents. In our analysis, age was standardized to enhance interpretability [45], whereas sex and diagnosis were dummy coded, respectively (male=0, female=1; normal=0, patients=1). Among several methods available for assessing covariate effects within the mixture modeling framework, we used the bias-adjusted maximum likelihood three-step method (ML method) proposed by Vermunt [46], which was implemented using the “R3STEP” command in Mplus. A distinguishing characteristic of this approach was its two-stage process; first, it selects the number of profiles; second, it assesses covariate effects in a sequential manner using multinomial logistic regression. During the second stage, this method accounts for the uncertainty (e.g., classification error) associated with assigning individuals to latent profiles. A previous study has shown that the ML method provides unbiased parameter estimates and their standard error estimates for covariate effects, while effectively controlling for the uncertainty associated with profile assignment within the initial profile structure obtained during the profile enumeration stage [47,48]. Covariate effects were assessed using the odds ratio (OR) scale, which represent the likelihood of belonging to a particular profile relative to the reference profile given covariate predicting profile membership. An OR of 1 indicated no covariate effect on profile assignment, whereas an OR of >1 indicated the presence of a substantial covariate effect on profile assignment.
RESULTS
Selection of the number of latent profiles
The log-likelihood and information criteria (IC) values of the estimated models are summarized in Table 1. The results showed that the AIC, BIC, and SABIC values decreased as the number of profiles increased, with the lowest values observed in the model with the maximum number of profiles (e.g., the five-profile model). In such a case, the LPA model with a k profile that exhibits a relatively large decrease in the IC values (i.e., elbow point) [49] compared to k–1 profile model can be considered as optimal. As shown in Table 1, the largest decrease in IC values is observed between the three- and four-profile models, compared with other changes. Moreover, the results indicated that the three-profile model was preferred in the LMR-LRT, as evidenced by the insignificance of the LMR-test statistic (p=0.324). Furthermore, the entropy measure remained the highest among the competing models, 2 profile and 3 profile models (0.790). We selected the 3-profile model as the final model for interpretation on the basis of these results.
Latent profile identification
The latent profile sizes and the average scores of each subscale obtained from the three-profile model are presented in Table 2 and Figure 1, respectively. Among the temperaments of the TEMPS-A-SV, hyperthymic temperament had a uniquely protective effect in most mental disorders [50], whereas other temperaments were considered risk factors; therefore, we labeled the profiles reflecting these aspects of temperaments. Participants who belonged to profile 1 (n=439; 17.86%) showed the highest scores across all subscales compared with other profiles; therefore, profile 1 was labeled “High Affective Temperament Group.” Participants belonging to profile 2 (n=1,014; 41.25%) showed the lowest scores across most subscales but had slightly higher scores on the hyperthymic than profile 3 (n=1,005; 40.89%). Therefore, profile 2 was labeled “Low Affective Temperament Group.” Profile 3 was characterized by “Middle Affective Temperament Group” because the overall scores were relatively higher than those of profile 2 but lower than those of profile 1.
Multiple-group three-profile models with different equality constraints were sequentially estimated. As shown in Table 3, the structural similarity model has resulted in the lowest values on the SABIC, whereas the dispersion similarity model has yielded the lowest values on the BIC. These results supported the similarity of the means and variances within the three-profile solution across subgroups; however, the sizes of the profiles differed across groups.
Demographic characteristics of latent profiles
A summary of the demographic information for each profile is provided in Table 4. The results showed that profile 2 was characterized by the highest average age, education, and employment rate among all the profiles, whereas profile 1 exhibited the highest rate of hospitalization compared with the other profiles.
Covariate effects on latent profile membership
The results of multinomial logistic regressions using the ML three-step method are shown in Table 5. In this study, the Middle Affective Temperament Group was designated as the reference profile. It implies that we examined how changes in age, sex, and the diagnosis influenced the likelihood of belonging to other profiles (e.g., High Affective Temperament Group and Low Affective Temperament Group) compared with the reference profile. When compared to the normal group, respondents diagnosed with MDD were more likely to belong to the Middle Affective Temperament Group than the Low Affective Temperament Group (OR=0.42; 95% confidence interval [CI] 0.28–0.65). However, the diagnosis of MDD had no impact on the likelihood of belonging to either the High Affective Temperament Group or the Medium Affective Temperament Group. Moreover, female respondents had a higher likelihood of belonging to the Low Affective Temperament Group (OR=0.36; 95% CI 0.26–0.51), whereas an increase of 1 year in age was associated with a greater likelihood of belonging to the same profile (OR=2.21; 95% CI 1.87–2.61).
Moreover, Table 5 shows that the respondents diagnosed with BD had higher likelihood of being in the Middle Affective Temperament Group rather than the Low Affective Temperament Group in comparison to the normal group (OR=0.35; 95% CI 0.25–0.49). Additionally, they exhibited a higher probability of belonging to the High Affective Temperament Group relative to the Middle Affective Temperament Group (OR=1.90; 95% CI 1.41–2.57). Similar to patients with MDD, the results indicated that female respondents were more likely to be assigned to the Low Affective Temperament Group in comparison to the Middle Affective Temperament Group (OR=0.31; 95% CI 0.23–0.42), whereas each 1-year increase in age was associated with a higher likelihood of belonging to the Middle Affective Temperament Group (OR=1.99; 95% CI 1.72–2.30).
When comparing between patients with MDD and those with BD, patients with MDD were associated with an increased likelihood of belonging to the Low Affective Temperament Group rather than the Middle Affective Temperament Group (OR=1.59; 95% CI 1.02–2.48). Furthermore, a 1-year increase in age was associated with an increased likelihood of being classified into the Low Affective Temperament Group compared with the Middle Affective Temperament Group (OR=1.99; 95% CI 1.60–2.47), whereas sex had no impact on predicting membership to each profile.
DISCUSSION
This study aimed to identify heterogeneous groups on the basis of response patterns to the TEMPS-A-SV and the possibility of belonging to each profile by analyzing the patient (MDD and BD) and control groups as covariates. Therefore, we conducted the LPA and derived the following three profiles: High Affective Temperament Group, Low Affective Temperament Group, and Middle Affective Temperament Group. The High Affective Temperament Group scored high on all TEMPS-A-SV subscales, indicating that this group was a severe group with problems in their respective affective temperaments. This is consistent with the result that prodromal temperament played a significant role in developing affective disorders [51]. The Low Affective Temperament Group scored lower than the Middle Affective Temperament Group on most TEMPS-A-SV subscales but had a pattern of higher scores than the Middle Affective Temperament Group on hyperthymic temperament. This shows that the Low Affective Temperament Group had weaker features of affective temperaments than the other profiles except for hyperthymic temperament. The intersection of the score in the Low Affective Temperament Group and Middle Affective Temperament Group may be because of the feature of hyperthymic temperament compared with the rest of the temperaments; a previous study has suggested that hyperthymic temperament has a protective effect in most of the mental disorders except for bipolar, separation anxiety, and impulse control disorders [50]. Therefore, the Low Affective Temperament Group had low affective problems with the antidepressant response. The Middle Affective Temperament Group had an intermediate scoring pattern on all subscales, with higher scores than the Low Affective Temperament Group except for hyperthymic temperament. Therefore, hyperthymic temperament was shown as a distinctive feature through the Low Affective Temperament Group and Middle Affective Temperament Group and yielded clinically meaningful results in logistic analyses.
Next, we performed a multinomial logistic analysis with diagnosis as a covariate variable. To summarize the results, the patient group was significantly more likely to be in the High Affective Temperament Group and Middle Affective Temperament Group, whereas the control group was more likely to belong to the Low Affective Temperament Group than the Middle Affective Temperament Group. These results suggest that the patient group was more likely to have several affective problems than the control group in mood regulation and mood instability and indicate that they were distinguishable from the control group at the temperamental dimension. A previous study on the relationship between the TEMPS-A and suicide attempts has suggested that hyperthymic may function as a protective temperament of depression [52]. Furthermore, Akiskal [53] described this temperament as having cheerful, high-energy, and improvident activities. Therefore, the finding that the control group in our study belonged to the Low Affective Temperament Group and scored higher than the patient group on the hyperthymic scale is consistent with previous studies [26] and indicates the presence of an antidepressant response in the control group. Specifically, even when the patient group was divided into BD and MDD and compared to the control group, the control group was significantly more likely to be in the Low Affective Temperament Group in both cases.
Within patient groups, patients with BD had more mood regulation problems and mood instability temperaments than those with MDD. The results show that the probability of each diagnosis belonging to the Low Affective Temperament Group and Middle Affective Temperament Group was significant, indicating that BD has more severe affective problems than MDD. These findings of differences in response patterns for each disorder are supported by previous studies showing differences in affective temperament between MDD and BD using the TEMPS-A-SV [29,33].
When compared to the control group, patients with BD were more likely to belong to both the High Affective Temperament Group and Middle Affective Temperament Group than those in the control group. However, patients with MDD had an increased likelihood of belonging to the Middle Affective Temperament Group relative to the Low Affective Temperament Group, whereas no difference was noted between the likelihood of belonging to either the High Affective Temperament Group or Middle Affective Temperament Group. These results are partially aligned with a previous study that has consistently shown higher cyclothymic and anxious temperament scores among patients with BD and relatively lower scores in those with MDD [29,32,33]. However, the intersection of hyperthymic temperament scores in this case was contrary to the findings of a meta-analysis conducted by Solmi et al. [26] and differed from other studies. In our study, more patients with BD II than those with BD I were observed; therefore, recognizing the features of patients with BD II, who may have co-occurring depressive symptoms during hypomania episodes, which are more chronic and have more severe disabilities than BD I, is needed [54]. Moreover, patients with BD are distinguished by the finding that hyperthymic temperament is not a protective factor, unlike in the control group. [51]. Our result suggests that patients with BD are more vulnerable to stress and suicide than those with MDD, considering that a previous study has linked low hyperthymic scores in patients with BD to suicide [52]. Therefore, in our results, we speculate that the BD group scored lower on the hyperthymic temperament than the MDD group owing to a weaker response to antidepressants.
This study had several limitations. First, owing to the number of patients with BD II, the BD group in our study tended to follow the features of BD II. Additionally, we could not control the episodes that patients with BD experience; therefore, we could not determine which episodes individuals were in when recording their responses on the TEMPS-A. We also might have patients with MDD who could develop BD, which may have affected our results. Furthermore, this was a cross-sectional study, which was limited to clearly explaining the causation between the diagnosis and the response patterns of the TEMPS-A-SV. This is because although temperament is a stable trait, we should not ignore the influence of the acute phase that patients may experience in certain episodes. Lastly, due to the nature of data collection via an online survey, the psychiatric history of the control group could not be identified using diagnostic criteria.
Nevertheless, our study had the following implications. Temperament is a stable trait [55], and recent attempts to differentiate affective disorders through temperament have been noted. These attempts stem from the “bipolar spectrum” concept and assume that affective temperament is strongly associated with mental illness. This is Akiskal’s extension of Kraepelin’s theory that mood variability in BD is a single disorder, with depression, BD I, and BD II as disorders on a spectrum [56]. The self-report based on this theory is the TEMPS-A, and the TEMPS-A-SV has the advantage of being shorter in length, which may reduce time costs and fatigue in clinical individuals. Our findings using this scale showed differences in response patterns between the affective temperaments of individuals with and without disabilities. This supports the perspective that affective temperament can explain the affective characteristics of BD and MDD, which is consistent with previous studies.
Most importantly, we confirmed the potential of the TEMPS-A-SV to discriminate between MDD and BD, demonstrating its utility as an assistive measurement tool for diagnosis in clinical settings. Moreover, to the best of our knowledge, this is the first study to investigate the response patterns of MDD and BD using the LPA method. Previous studies have conducted variable-centered analyses after dividing the groups by disorder; however, our study derived the profiles by focusing on the response patterns of individuals, showing that bottom-up analyses can identify the affective characteristics of each mood disorder. Most of the studies are still conducted using the 110-item TEMPS-A; therefore, studies on the shortened version are lacking. Therefore, confirming the usefulness of the TEMPS-A-SV using various analyses and considering how it can be utilized in clinical practice are needed; more studies should be conducted.
