Deconstructing depression Monden, Rei
IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite from it. Please check the document version below.
Document Version
Publisher's PDF, also known as Version of record
Publication date:
2017
Link to publication in University of Groningen/UMCG research database
Citation for published version (APA):
Monden, R. (2017). Deconstructing depression: A 3D perspective. Rijksuniversiteit Groningen.
Copyright
Other than for strictly personal use, it is not permitted to download or to forward/distribute the text or part of it without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license (like Creative Commons).
Take-down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.
Downloaded from the University of Groningen/UMCG research database (Pure): http://www.rug.nl/research/portal. For technical reasons the number of authors shown on this cover page is limited to 10 maximum.
Winter morning (Painted by Miho Hashimoto) Winter morning (Painted by Miho Hashimoto)
Chapter 5
Decomposition of depression heterogeneity on the Person-, Symptom- and Time-level during the acute phase of antidepressant therapy:
the STAR*D study
Rei Monden
Klaas J. Wardenaar
Alwin Stegeman
Eiko F Fried
Peter de Jonge
Submitted
Abstract
Background: Antidepressants are commonly used to treat depression, but only part of the patients show a treatment response. Unfortunately, it has proven hard to identify the patients for whom antidepressants are truly beneficial. This could be partly due to the fact that personal differences in how different symptoms respond to antidepressants are insufficiently understood. Therefore, this study aimed to capture and investigate these dynamic variations using Three-mode Principal Component Analysis (3MPCA).
Methods: Data came from a group of depressed patients who all received Citalopram as part of phase 1 of STAR*D. The Quick Inventory of Depressive Symptomatology (QIDS) was administered at weeks 0, 2, 4, 6, 9 and 12. The resulting data were decomposed with 3MPCA into person-, symptom- and time- mode components that were associated with patient characteristics, short-term remission and long-term depression severity.
Results: The optimal 3MPCA model (81% explained variance) had two symptom-mode (‘cognitive’ and ‘somatic’), two time-mode (‘first measurements’,
‘last measurements’) and three person-mode components (‘negative/suicidal thoughts’, ‘physical dysfunction’ and ‘general improvement’). ‘Negative/suicidal thoughts’ scores were correlated with suicidal thoughts (r=0.30-0.38) and negatively associated with remission. ‘Physical dysfunction’ was correlated with work/social difficulties (r=0.47), lower quality of life (r=-0.40) and side-effects (r=0.34-0.37), negatively associated with remission and positively associated with long-term severity. ‘General improvement’ scores were negatively correlated with quality of life (r=-0.38) but were positively associated with remission.
Conclusions: 3MPCA offers an insightful longitudinal description of the most important sources of variation in antidepressant treatment response, providing interesting leads for clinical research into empirically-based antidepressant prescription.
Introduction
Antidepressants are commonly used for Major Depressive Disorder (MDD) treatment [1]. Although treatment studies have shown similar efficacy for different antidepressants [2-4], treatment responses are known to differ strongly across patients [5]. Given the large impact of MDD on both patients and society [6,7], optimizing antidepressant treatment is crucial. However, the heterogeneity of the diagnosis and course trajectories of MDD [8,9] has hindered this process.
To obtain more precise prediction of antidepressant treatment response, patient subgroups based on baseline characteristics, such as symptom severity [3, 10, 11], gender [12] and comorbidity [13] have been studied. However, even within such strata, studies have been shown to have diverse treatment outcomes [14, 15]. A reason for this may be that the used outcomes to quantify treatment response are usually based on clinical cut-offs or minimal clinically-relevant change scores on severity scales, whereas it could very well be that the naturally occurring variations in treatment response are not captured very well in that way.
As an alternative, data-driven approaches have been employed to describe the heterogeneity of treatment response. For example, longitudinal Latent Class Analysis has been used to find more homogeneous treatment response trajectories [16], growth mixture models have been used to find patient subgroups with similar treatment responses [17-19] and mixed-effects models have been used to identify the severity trajectories of respondents and non-respondents [20].
Although the latter studies have contributed considerably to our understanding of the existing variations in antidepressant treatment responses, they have not incorporated all relevant sources of heterogeneity (e.g. person- symptom- and time- heterogeneity) into a single model [21]. For instance, traditional methods provide no insight into how patients can vary in terms of their trajectories on different subdomains of depressive symptomatology. To tackle this, more flexible methods, such as two-mode K-spectral centroid analysis [22] and Three-mode Principal Component Analysis (3MPCA [23-29]) can be used. The former approach offer insight into how temporal changes vary as a function of both symptoms and persons, by identification of patient subgroups with different trajectories on distinct symptom clusters [22]. 3MPCA provides an even more flexible approach to explain heterogeneity at the person-, symptom- and time- level in an integrated dimensional model. Provided with a dataset consisting of symptoms that were assessed repeatedly in a group of patients, 3MPCA can decompose symptoms, time points and persons into symptom-, time- and person- mode components, respectively. Moreover, 3MPCA captures the interactions between the different modes’ components. As such, 3MPCA can describe how persons differ in their trajectories on different symptom-domains in response to
Abstract
Background: Antidepressants are commonly used to treat depression, but only part of the patients show a treatment response. Unfortunately, it has proven hard to identify the patients for whom antidepressants are truly beneficial. This could be partly due to the fact that personal differences in how different symptoms respond to antidepressants are insufficiently understood. Therefore, this study aimed to capture and investigate these dynamic variations using Three-mode Principal Component Analysis (3MPCA).
Methods: Data came from a group of depressed patients who all received Citalopram as part of phase 1 of STAR*D. The Quick Inventory of Depressive Symptomatology (QIDS) was administered at weeks 0, 2, 4, 6, 9 and 12. The resulting data were decomposed with 3MPCA into person-, symptom- and time- mode components that were associated with patient characteristics, short-term remission and long-term depression severity.
Results: The optimal 3MPCA model (81% explained variance) had two symptom-mode (‘cognitive’ and ‘somatic’), two time-mode (‘first measurements’,
‘last measurements’) and three person-mode components (‘negative/suicidal thoughts’, ‘physical dysfunction’ and ‘general improvement’). ‘Negative/suicidal thoughts’ scores were correlated with suicidal thoughts (r=0.30-0.38) and negatively associated with remission. ‘Physical dysfunction’ was correlated with work/social difficulties (r=0.47), lower quality of life (r=-0.40) and side-effects (r=0.34-0.37), negatively associated with remission and positively associated with long-term severity. ‘General improvement’ scores were negatively correlated with quality of life (r=-0.38) but were positively associated with remission.
Conclusions: 3MPCA offers an insightful longitudinal description of the most important sources of variation in antidepressant treatment response, providing interesting leads for clinical research into empirically-based antidepressant prescription.
Introduction
Antidepressants are commonly used for Major Depressive Disorder (MDD) treatment [1]. Although treatment studies have shown similar efficacy for different antidepressants [2-4], treatment responses are known to differ strongly across patients [5]. Given the large impact of MDD on both patients and society [6,7], optimizing antidepressant treatment is crucial. However, the heterogeneity of the diagnosis and course trajectories of MDD [8,9] has hindered this process.
To obtain more precise prediction of antidepressant treatment response, patient subgroups based on baseline characteristics, such as symptom severity [3, 10, 11], gender [12] and comorbidity [13] have been studied. However, even within such strata, studies have been shown to have diverse treatment outcomes [14, 15]. A reason for this may be that the used outcomes to quantify treatment response are usually based on clinical cut-offs or minimal clinically-relevant change scores on severity scales, whereas it could very well be that the naturally occurring variations in treatment response are not captured very well in that way.
As an alternative, data-driven approaches have been employed to describe the heterogeneity of treatment response. For example, longitudinal Latent Class Analysis has been used to find more homogeneous treatment response trajectories [16], growth mixture models have been used to find patient subgroups with similar treatment responses [17-19] and mixed-effects models have been used to identify the severity trajectories of respondents and non-respondents [20].
Although the latter studies have contributed considerably to our understanding of the existing variations in antidepressant treatment responses, they have not incorporated all relevant sources of heterogeneity (e.g. person- symptom- and time- heterogeneity) into a single model [21]. For instance, traditional methods provide no insight into how patients can vary in terms of their trajectories on different subdomains of depressive symptomatology. To tackle this, more flexible methods, such as two-mode K-spectral centroid analysis [22] and Three-mode Principal Component Analysis (3MPCA [23-29]) can be used. The former approach offer insight into how temporal changes vary as a function of both symptoms and persons, by identification of patient subgroups with different trajectories on distinct symptom clusters [22]. 3MPCA provides an even more flexible approach to explain heterogeneity at the person-, symptom- and time- level in an integrated dimensional model. Provided with a dataset consisting of symptoms that were assessed repeatedly in a group of patients, 3MPCA can decompose symptoms, time points and persons into symptom-, time- and person- mode components, respectively. Moreover, 3MPCA captures the interactions between the different modes’ components. As such, 3MPCA can describe how persons differ in their trajectories on different symptom-domains in response to
5
treatment, without using any a priori assumptions. The utility of 3MPCA to effectively decompose depression heterogeneity and the predictive ability of the resulting components have previously been demonstrated [30, 31].
Given the limited insight into the heterogeneity of antidepressant treatment responses and the promising results of 3MPCA, the current study aimed to use 3MPCA in a sample of MDD patients (n=2,876) receiving antidepressant treatment during the Level-1 treatment phase of the Sequenced Treatment Alternatives to Relieve Depression (STAR*D [32]). Moreover, the resulting components were associated with clinical characteristics, short-term remission rates and long-term severity scores.
Methods
Participants and data description
The data came from the STAR*D study, of which a detailed description can be found elsewhere [32, 33] and a brief summary is provided here. Four-thousand- forty-one patients were initially enrolled in the study, but 1,165 patients were excluded before entering the first treatment phase (Level-1) because they either refused to participate or had a score on the Hamilton Rating Scale for Depression (HRSD17 [34]) that was too low for inclusion. This resulted in 2,876 (70.9%) outpatients entering the Level-1, where they received the selective serotonin reuptake inhibitor (SSRI) citalopram for 12 weeks. The current study focuses on this Level-1 phase. Patients with an adequate response (HRSD17<8 [32]) in any treatment phase of the STAR*D study entered a 12-months naturalistic follow-up phase. Here, a clinician-scored Quick Inventory of Depressive Symptomatology (QIDS) was administered at the end of the follow-up. The study protocol was reviewed, approved and monitored by the institutional review board at each participating site. All patients signed informed consent at study entry.
Measures
Quick Inventory of Depressive Symptomatology
The QIDS [35, 36] was filled in by the patients themselves (QIDS-SR16) and by clinicians (QIDS-C16) at weeks 0, 2, 4, 6, 9 and 12 during the Level-1 phase.
Several QIDS items were recoded. The four sleep-domain items ‘Sleep onset Insomnia’, ‘Mid-nocturnal insomnia’, ‘Early morning Insomnia’ and
‘Hypersomnia’ were combined into one sleep item by using the highest reported score, which is consistent with the scoring instructions for the QIDS and prevented the sleep-domain symptoms to be over-represented in the dataset. The items ‘appetite increase’, ‘appetite decrease’, ‘weight decrease’ and ‘weight
increase’ were respectively combined into two compound items for ‘appetite change’ and ‘weight change’ because only one of the items for each domain was recorded. This procedure is consistent with a prior publication on this dataset [37].
For ‘psychomotor slowing’ and ‘psychomotor agitation’ items, we did not follow the scoring instruction of QIDS since these items may be associated with different types of depression [38]. Together, recoding resulted in 11 items for the QIDS- SR16 and 11 items for the QIDS-C16, which were used in the current analyses.
Follow-up measures
The HRSD17 and the Inventory of Depressive Symptomatology-Clinician Rated (IDS-C30 [39, 40]) scores were collected together with other secondary outcomes (e.g. side-effect burden) by a telephone-based interview after 12 weeks (end of the Level-1 treatment phase). Additionally, patients who attained at least a response (i.e. ≥50% reduction in baseline symptom severity) at any treatment level entered a 12-month naturalistic follow-up phase where the QIDS-SR16 was administered.
Patient status
Patients visited the clinics at weeks 0, 2, 4, 6, 9 and 12 during the Level-1 phase.
At each visit, a clinician confirmed the patient’s status for the following three categories: ‘remission’, ‘move to the next level’ or ‘exit the study’. The ‘remission’
category included patients that were ‘symptom-free’ (HRSD17 <8 or QIDS-SR16
≤5). These patients stopped the treatment and entered the one-year naturalistic follow-up phase. The ‘move to the next level’ category included patients who received adequate dose-levels but showed either clear intolerance for the treatment or minimal reduction from the baseline severity. These patients were encouraged to move to the next treatment phase (Level-2) before completing the 12 weeks of Level-1 treatment. The ‘exit the study’ category included patient who either withdrew their consent or were lost to follow-up. The status of all patients was also recorded at the end of the Level-1 phase between weeks 12 and 14.
Other patient characteristics
Baseline demographic characteristics and the following questionnaires were administered at study enrollment: the HRSD17 [41, 42], the Psychiatric Diagnostic Screening Questionnaire (PDSQ [43, 44]), the Cumulative Illness Rating Scale (CIRS [45, 46]), psychotropic medication history and psychiatric history.
The following questionnaires were administered at the end of the Level- 1 phase: the Patient Rated Inventory of Side effects (PRISE [35]), the Quality of Life enjoyment and Satisfaction Questionnaire (QLESQ [47]), the Work and
treatment, without using any a priori assumptions. The utility of 3MPCA to effectively decompose depression heterogeneity and the predictive ability of the resulting components have previously been demonstrated [30, 31].
Given the limited insight into the heterogeneity of antidepressant treatment responses and the promising results of 3MPCA, the current study aimed to use 3MPCA in a sample of MDD patients (n=2,876) receiving antidepressant treatment during the Level-1 treatment phase of the Sequenced Treatment Alternatives to Relieve Depression (STAR*D [32]). Moreover, the resulting components were associated with clinical characteristics, short-term remission rates and long-term severity scores.
Methods
Participants and data description
The data came from the STAR*D study, of which a detailed description can be found elsewhere [32, 33] and a brief summary is provided here. Four-thousand- forty-one patients were initially enrolled in the study, but 1,165 patients were excluded before entering the first treatment phase (Level-1) because they either refused to participate or had a score on the Hamilton Rating Scale for Depression (HRSD17 [34]) that was too low for inclusion. This resulted in 2,876 (70.9%) outpatients entering the Level-1, where they received the selective serotonin reuptake inhibitor (SSRI) citalopram for 12 weeks. The current study focuses on this Level-1 phase. Patients with an adequate response (HRSD17<8 [32]) in any treatment phase of the STAR*D study entered a 12-months naturalistic follow-up phase. Here, a clinician-scored Quick Inventory of Depressive Symptomatology (QIDS) was administered at the end of the follow-up. The study protocol was reviewed, approved and monitored by the institutional review board at each participating site. All patients signed informed consent at study entry.
Measures
Quick Inventory of Depressive Symptomatology
The QIDS [35, 36] was filled in by the patients themselves (QIDS-SR16) and by clinicians (QIDS-C16) at weeks 0, 2, 4, 6, 9 and 12 during the Level-1 phase.
Several QIDS items were recoded. The four sleep-domain items ‘Sleep onset Insomnia’, ‘Mid-nocturnal insomnia’, ‘Early morning Insomnia’ and
‘Hypersomnia’ were combined into one sleep item by using the highest reported score, which is consistent with the scoring instructions for the QIDS and prevented the sleep-domain symptoms to be over-represented in the dataset. The items ‘appetite increase’, ‘appetite decrease’, ‘weight decrease’ and ‘weight
increase’ were respectively combined into two compound items for ‘appetite change’ and ‘weight change’ because only one of the items for each domain was recorded. This procedure is consistent with a prior publication on this dataset [37].
For ‘psychomotor slowing’ and ‘psychomotor agitation’ items, we did not follow the scoring instruction of QIDS since these items may be associated with different types of depression [38]. Together, recoding resulted in 11 items for the QIDS- SR16 and 11 items for the QIDS-C16, which were used in the current analyses.
Follow-up measures
The HRSD17 and the Inventory of Depressive Symptomatology-Clinician Rated (IDS-C30 [39, 40]) scores were collected together with other secondary outcomes (e.g. side-effect burden) by a telephone-based interview after 12 weeks (end of the Level-1 treatment phase). Additionally, patients who attained at least a response (i.e. ≥50% reduction in baseline symptom severity) at any treatment level entered a 12-month naturalistic follow-up phase where the QIDS-SR16 was administered.
Patient status
Patients visited the clinics at weeks 0, 2, 4, 6, 9 and 12 during the Level-1 phase.
At each visit, a clinician confirmed the patient’s status for the following three categories: ‘remission’, ‘move to the next level’ or ‘exit the study’. The ‘remission’
category included patients that were ‘symptom-free’ (HRSD17 <8 or QIDS-SR16
≤5). These patients stopped the treatment and entered the one-year naturalistic follow-up phase. The ‘move to the next level’ category included patients who received adequate dose-levels but showed either clear intolerance for the treatment or minimal reduction from the baseline severity. These patients were encouraged to move to the next treatment phase (Level-2) before completing the 12 weeks of Level-1 treatment. The ‘exit the study’ category included patient who either withdrew their consent or were lost to follow-up. The status of all patients was also recorded at the end of the Level-1 phase between weeks 12 and 14.
Other patient characteristics
Baseline demographic characteristics and the following questionnaires were administered at study enrollment: the HRSD17 [41, 42], the Psychiatric Diagnostic Screening Questionnaire (PDSQ [43, 44]), the Cumulative Illness Rating Scale (CIRS [45, 46]), psychotropic medication history and psychiatric history.
The following questionnaires were administered at the end of the Level- 1 phase: the Patient Rated Inventory of Side effects (PRISE [35]), the Quality of Life enjoyment and Satisfaction Questionnaire (QLESQ [47]), the Work and
5
Social Adjustment Scale (WSAS [48]) and the Short-Form Health Survey (SFHS [49]).
Statistical Analyses Missing data
Patients who did not complete the 12-week Level-1 phase were excluded from the study. Moreover, if a patient did not provide any item score on both the QIDS- SR16 and QIDS-C16 at more than one time point, the patient was excluded from the study. This was done to avoid introducing too much bias in the multiple imputation procedure, while retaining as much patients as possible. This resulted in selection of 1,656 (58%) of the initial 2,876 patients for the 3MPCA analysis.
Of these, 875 patients did not have any missing score. In the remainder of the sample, 7.9% of the data was missing. These missing data were imputed 20 times, using the Amelia II package [50] in R. In the multiple imputation model, patients’
characteristics, such as HRSD17 and PDSQ scores, were included besides all the QIDS scores.
Three-mode Principal Component Analysis (3MPCA)
The 3MPCA analytical procedure has been described in detail elsewhere [27, 29, 30] and is briefly summarized below. First, to confirm if the dataset contained a non-negligible three-way interaction of the person-, symptom- and time-mode, and thus warrants the application of 3MPCA, a three-way ANOVA was applied to each of the 20 imputed datasets. The stability of the proportion of the variance explained by the ‘three-way interaction plus error term’ was investigated by averaging the three-way ANOVA results across the 20 imputed datasets and by calculating the standard deviation of the estimated explained variances.
Before applying 3MPCA, each of the imputed datasets was preprocessed by centering across the person-mode and normalizing within the symptom-mode.
Centering is required to ensure that the 3MPCA model only captures the variations of interest around the mean trend (general trend) in the data.
Normalizing is performed to ensure that variations in all 11 QIDS items are treated as equally important. The complexity of the 3MPCA, which reflects the respective numbers of components for the person-, symptom- and time-mode, was first evaluated with the generalized scree test [51, 52]. Also, 3MPCA was applied in each of the 20 imputed datasets and a split-half procedure was performed to evaluate the stability of the resulting components. Once the model complexity was selected, 3MPCA was performed in each of the imputed datasets, using Joint Orthomax rotation [53] with standard weights (with no weight assigned to the person-mode) to obtain an interpretable 3MPCA component structure. This procedure is useful to acquire interpretable results as it helps to simplify the
component structures for the symptom- and time-mode while maintaining the overall model-fit percentage [53].
After the abovementioned steps, the resulting 3MPCA components were averaged for each component and for the component interactions in the core array, using a generalized Procrustes rotation [29, 54, 55]. Following a procedure explained in detail elsewhere [30], two types of fit percentages were calculated for the resulting 3MPCA model: (1) the overall fit percentage including the variance explained by heterogeneity around the general trend (captured by the 3MPCA model) and the general trend itself, and (2) the fit percentage describing only the explained variance around the general trend, i.e. variance captured by the 3MPCA model (see [30] for details about the computation of both types of explained variance).
The complete dataset had a four-way structure: patients, QIDS items, two response types (clinician/self-rated) and time points, but the data were analyzed as three-way data, as inspection of the data (see Appendix A) suggested that QIDS items with different response types could be aggregated in the symptom- mode.
External associations
To gain insight into the characteristics of the 3MPCA components, Spearman or Pearson correlations were calculated between the estimated person-mode components and auxiliary variables at study enrollment (i.e. demographics, HRSD17, PDSQ, CIRS, psychotropic medication history and psychiatric history) to investigate possible predictors of treatment response variations. In addition, correlations with follow-up measures (i.e. HRSD17, PRISE, QLSEQ, SFHS, side effects and WSAS) were calculated to investigate how the variations in treatment- response that were captured by the model are related to eventual treatment outcome.
To further investigate the short- and long-term predictive value of the 3MPCA solution for eventual treatment outcome, two additional regression analyses were performed. Short-term predictive value was evaluated by conducting a logistic regression, using the standardized person-mode components as independent variables and ‘remission’ between week 12 and 14 as dependent variable (1=yes/0=no). The long-term predictive value was evaluated by conducting multivariate linear regression, using the standardized person-mode components as independent variables and the 12-month QIDS-SR16 sum score as dependent variable.
Social Adjustment Scale (WSAS [48]) and the Short-Form Health Survey (SFHS [49]).
Statistical Analyses Missing data
Patients who did not complete the 12-week Level-1 phase were excluded from the study. Moreover, if a patient did not provide any item score on both the QIDS- SR16 and QIDS-C16 at more than one time point, the patient was excluded from the study. This was done to avoid introducing too much bias in the multiple imputation procedure, while retaining as much patients as possible. This resulted in selection of 1,656 (58%) of the initial 2,876 patients for the 3MPCA analysis.
Of these, 875 patients did not have any missing score. In the remainder of the sample, 7.9% of the data was missing. These missing data were imputed 20 times, using the Amelia II package [50] in R. In the multiple imputation model, patients’
characteristics, such as HRSD17 and PDSQ scores, were included besides all the QIDS scores.
Three-mode Principal Component Analysis (3MPCA)
The 3MPCA analytical procedure has been described in detail elsewhere [27, 29, 30] and is briefly summarized below. First, to confirm if the dataset contained a non-negligible three-way interaction of the person-, symptom- and time-mode, and thus warrants the application of 3MPCA, a three-way ANOVA was applied to each of the 20 imputed datasets. The stability of the proportion of the variance explained by the ‘three-way interaction plus error term’ was investigated by averaging the three-way ANOVA results across the 20 imputed datasets and by calculating the standard deviation of the estimated explained variances.
Before applying 3MPCA, each of the imputed datasets was preprocessed by centering across the person-mode and normalizing within the symptom-mode.
Centering is required to ensure that the 3MPCA model only captures the variations of interest around the mean trend (general trend) in the data.
Normalizing is performed to ensure that variations in all 11 QIDS items are treated as equally important. The complexity of the 3MPCA, which reflects the respective numbers of components for the person-, symptom- and time-mode, was first evaluated with the generalized scree test [51, 52]. Also, 3MPCA was applied in each of the 20 imputed datasets and a split-half procedure was performed to evaluate the stability of the resulting components. Once the model complexity was selected, 3MPCA was performed in each of the imputed datasets, using Joint Orthomax rotation [53] with standard weights (with no weight assigned to the person-mode) to obtain an interpretable 3MPCA component structure. This procedure is useful to acquire interpretable results as it helps to simplify the
component structures for the symptom- and time-mode while maintaining the overall model-fit percentage [53].
After the abovementioned steps, the resulting 3MPCA components were averaged for each component and for the component interactions in the core array, using a generalized Procrustes rotation [29, 54, 55]. Following a procedure explained in detail elsewhere [30], two types of fit percentages were calculated for the resulting 3MPCA model: (1) the overall fit percentage including the variance explained by heterogeneity around the general trend (captured by the 3MPCA model) and the general trend itself, and (2) the fit percentage describing only the explained variance around the general trend, i.e. variance captured by the 3MPCA model (see [30] for details about the computation of both types of explained variance).
The complete dataset had a four-way structure: patients, QIDS items, two response types (clinician/self-rated) and time points, but the data were analyzed as three-way data, as inspection of the data (see Appendix A) suggested that QIDS items with different response types could be aggregated in the symptom- mode.
External associations
To gain insight into the characteristics of the 3MPCA components, Spearman or Pearson correlations were calculated between the estimated person-mode components and auxiliary variables at study enrollment (i.e. demographics, HRSD17, PDSQ, CIRS, psychotropic medication history and psychiatric history) to investigate possible predictors of treatment response variations. In addition, correlations with follow-up measures (i.e. HRSD17, PRISE, QLSEQ, SFHS, side effects and WSAS) were calculated to investigate how the variations in treatment- response that were captured by the model are related to eventual treatment outcome.
To further investigate the short- and long-term predictive value of the 3MPCA solution for eventual treatment outcome, two additional regression analyses were performed. Short-term predictive value was evaluated by conducting a logistic regression, using the standardized person-mode components as independent variables and ‘remission’ between week 12 and 14 as dependent variable (1=yes/0=no). The long-term predictive value was evaluated by conducting multivariate linear regression, using the standardized person-mode components as independent variables and the 12-month QIDS-SR16 sum score as dependent variable.
5
Results
Descriptive information
The proportion of women in the sample was 64% and the mean age was 42.0 years (s.d.=13.2). The mean score on the HRSD17 at study enrollment was 39.5 (s.d.=
5.1), the mean sum scores on the QIDS-C16 and QIDS-SR16 in week 0 of the Level-1 phase were 16.2 (s.d. = 3.3) and 15.1 (s.d.= 4.3), respectively. Both the mean QIDS-C16 and QIDS-SR16 decreased considerably during the first part of the study, stabilizing over time (Figure 1 A), but considerable intra-individual variation could be observed between patients’ QIDS mean-score trajectories over time (Figure 1 B). To evaluate if gender-stratification was needed, a two-way repeated measures ANOVA was conducted. This analysis showed that repeated QIDS assessments did not differ significantly between males and females (Appendix B). Therefore, the data were analyzed as a whole.
Three-mode Principal Component Analysis Three-way ANOVA
The results of the fixed-effect three-way ANOVA (Appendix C) showed the largest explained variance for the ‘three-way interaction plus error’ term (30%), followed by the ‘Persons-by-symptoms’ interaction effect (20%). This implied that there was a non-negligible three-way interaction between the different modes of the dataset, and thus, that 3MPCA was likely to be suitable to analyze the variance in the dataset [27].
Model complexity and fit percentages
To perform the generalized scree test, the maximum number of components was set to (5, 5, 3) for the person-, symptom- and time-mode, respectively. This was set to balance the interpretability and the complexity of the model. The test showed that a model with either the structure (3, 3, 2) or the structure (3, 2, 2) best described the data. However, the fit was only 0.4% higher for the (3, 3, 2) structure than for the (3, 2, 2) structure. Moreover, the split-half procedure showed a less stable symptom-component structure for the (3, 3, 2) structure than for the (3, 2, 2) structure. Therefore, the component structure with three person-mode components, two symptom-mode components and two time-mode components was selected.
The total fit percentage of the estimated (3, 2, 2) 3MPCA model, averaged across 20 imputed datasets, was 26% (s.d.=0.03) when only including only the variance explained around the general trend and 81% (s.d.=0.09) when including both the variance of the general trend and the heterogeneity around it.
Small standard deviations indicated high consistency across the 20 imputed
datasets. Note that 3MPCA was applied to the preprocessed data (centered across the person-mode and normalized within the symptom-mode). Specifically, the mean score of each symptom on each time point was subtracted from the data and the equal importance of the symptoms was assured by normalizing. Therefore, a large difference in fit percentage with/without ‘general trend’ indicates that most of the patients followed this ‘general trend’, which involved decreasing severity scores over time, and heterogeneity around this downward ‘general trend’
explains about 26% of the total variance.
Symptom-mode components
The symptom-mode component scores are shown in Table 1. For the first component, high loadings were observed in ‘suicidal ideation’, ‘outlook (self)’
and ‘sad mood’. For the second component, ‘appetite change’, ‘concentration’,
‘fatigability’, ‘psychomotor slowing’, ‘involvement’, ‘sleep problems’ and
‘weight change’ showed high loadings. Based on these patterns, the first and second symptom-mode components were labeled as the ‘Cognitive’ and ‘Somatic’
component, respectively. Table 1 also indicated that the components were similar irrespective of the respondent (patients or clinicians). All the estimated scores had standard deviations equal to/smaller than 0.01, indicating consistency across the 20 imputed datasets.
Time-mode components
Table 2 shows the time-mode components averaged across imputed datasets. The results showed that the last three measurements had high loadings on the first time-mode component and the first two measurements had high loadings on the second time-mode component. Therefore, the first time-mode component was labeled ‘last measurements’ and the second time-mode component was labeled
‘first measurements’. Small standard deviations (<0.01) indicated high consistency across imputed datasets.
Results
Descriptive information
The proportion of women in the sample was 64% and the mean age was 42.0 years (s.d.=13.2). The mean score on the HRSD17 at study enrollment was 39.5 (s.d.=
5.1), the mean sum scores on the QIDS-C16 and QIDS-SR16 in week 0 of the Level-1 phase were 16.2 (s.d. = 3.3) and 15.1 (s.d.= 4.3), respectively. Both the mean QIDS-C16 and QIDS-SR16 decreased considerably during the first part of the study, stabilizing over time (Figure 1 A), but considerable intra-individual variation could be observed between patients’ QIDS mean-score trajectories over time (Figure 1 B). To evaluate if gender-stratification was needed, a two-way repeated measures ANOVA was conducted. This analysis showed that repeated QIDS assessments did not differ significantly between males and females (Appendix B). Therefore, the data were analyzed as a whole.
Three-mode Principal Component Analysis Three-way ANOVA
The results of the fixed-effect three-way ANOVA (Appendix C) showed the largest explained variance for the ‘three-way interaction plus error’ term (30%), followed by the ‘Persons-by-symptoms’ interaction effect (20%). This implied that there was a non-negligible three-way interaction between the different modes of the dataset, and thus, that 3MPCA was likely to be suitable to analyze the variance in the dataset [27].
Model complexity and fit percentages
To perform the generalized scree test, the maximum number of components was set to (5, 5, 3) for the person-, symptom- and time-mode, respectively. This was set to balance the interpretability and the complexity of the model. The test showed that a model with either the structure (3, 3, 2) or the structure (3, 2, 2) best described the data. However, the fit was only 0.4% higher for the (3, 3, 2) structure than for the (3, 2, 2) structure. Moreover, the split-half procedure showed a less stable symptom-component structure for the (3, 3, 2) structure than for the (3, 2, 2) structure. Therefore, the component structure with three person-mode components, two symptom-mode components and two time-mode components was selected.
The total fit percentage of the estimated (3, 2, 2) 3MPCA model, averaged across 20 imputed datasets, was 26% (s.d.=0.03) when only including only the variance explained around the general trend and 81% (s.d.=0.09) when including both the variance of the general trend and the heterogeneity around it.
Small standard deviations indicated high consistency across the 20 imputed
datasets. Note that 3MPCA was applied to the preprocessed data (centered across the person-mode and normalized within the symptom-mode). Specifically, the mean score of each symptom on each time point was subtracted from the data and the equal importance of the symptoms was assured by normalizing. Therefore, a large difference in fit percentage with/without ‘general trend’ indicates that most of the patients followed this ‘general trend’, which involved decreasing severity scores over time, and heterogeneity around this downward ‘general trend’
explains about 26% of the total variance.
Symptom-mode components
The symptom-mode component scores are shown in Table 1. For the first component, high loadings were observed in ‘suicidal ideation’, ‘outlook (self)’
and ‘sad mood’. For the second component, ‘appetite change’, ‘concentration’,
‘fatigability’, ‘psychomotor slowing’, ‘involvement’, ‘sleep problems’ and
‘weight change’ showed high loadings. Based on these patterns, the first and second symptom-mode components were labeled as the ‘Cognitive’ and ‘Somatic’
component, respectively. Table 1 also indicated that the components were similar irrespective of the respondent (patients or clinicians). All the estimated scores had standard deviations equal to/smaller than 0.01, indicating consistency across the 20 imputed datasets.
Time-mode components
Table 2 shows the time-mode components averaged across imputed datasets. The results showed that the last three measurements had high loadings on the first time-mode component and the first two measurements had high loadings on the second time-mode component. Therefore, the first time-mode component was labeled ‘last measurements’ and the second time-mode component was labeled
‘first measurements’. Small standard deviations (<0.01) indicated high consistency across imputed datasets.
5
Table 1 The symptom-mode component scores averaged across 20 imputed datasets.
Item
number 16-item Quick Inventory of
Depressive Symptomatology items Symptom-mode component
Cognitive Somatic
12 Suicidal ideation 0.45 -0.01
12 Suicidal ideation 0.45 -0.01
11 Outlook (self) 0.39 0.07
11 Outlook (self) 0.36 0.09
5 Mood (sad) 0.25 0.19
5 Mood (sad) 0.22 0.18
6,7 Appetite problem -0.19 0.28
6,7 Appetite problem -0.19 0.27
10 Concentration/Decision Making 0.08 0.27 10 Concentration/Decision Making 0.07 0.25
14 Energy/Fatigability 0.02 0.27
14 Energy/Fatigability 0.03 0.25
15 Psychomotor slowing -0.01 0.27
15 Psychomotor slowing 0.00 0.22
13 Involvement 0.07 0.25
13 Involvement 0.13 0.23
1-4 Sleep problem -0.08 0.24
1-4 Sleep problem -0.08 0.23
8,9 Weight problem -0.21 0.22
8,9 Weight problem -0.22 0.21
16 Psychomotor agitation 0.02 0.18
16 Psychomotor agitation 0.01 0.14
Items shaded in grey indicate clinician-rated items. For all loadings, standard deviation across 20 imputed datasets were at most 0.01. Component scores
≥0.20 are printed in bold font.
Table 2 The time-mode component scores averaged across 20 imputed datasets.
Measurement time point Time-mode components
Last measurements First measurements
Baseline -0.03 0.62
2 weeks 0.11 0.62
4 weeks 0.30 0.38
6 weeks 0.43 0.14
9 weeks 0.57 -0.13
12 weeks 0.63 -0.23
Standard deviations across 20 imputed datasets were at most 0.1 for all loadings. The time-mode component scores ≥0.40 are printed in bold.
Core array
The estimated core array and the explained variance for each combination of the person-, symptom- and time-mode components are shown in Table 3. The core array provides an overview of the three-way interactions, with larger core elements indicating stronger three-way interactions. The results showed that most of the total explained variance (12.1%) was associated with the interaction between the second person-mode component, the somatic symptom-mode component and the last measurements time-mode component. The second-most variance (4.7%) was explained by the interaction between the third person-mode component, the somatic symptom-mode component and the first measurements time-mode component.
Table 1 The symptom-mode component scores averaged across 20 imputed datasets.
Item
number 16-item Quick Inventory of
Depressive Symptomatology items Symptom-mode component
Cognitive Somatic
12 Suicidal ideation 0.45 -0.01
12 Suicidal ideation 0.45 -0.01
11 Outlook (self) 0.39 0.07
11 Outlook (self) 0.36 0.09
5 Mood (sad) 0.25 0.19
5 Mood (sad) 0.22 0.18
6,7 Appetite problem -0.19 0.28
6,7 Appetite problem -0.19 0.27
10 Concentration/Decision Making 0.08 0.27 10 Concentration/Decision Making 0.07 0.25
14 Energy/Fatigability 0.02 0.27
14 Energy/Fatigability 0.03 0.25
15 Psychomotor slowing -0.01 0.27
15 Psychomotor slowing 0.00 0.22
13 Involvement 0.07 0.25
13 Involvement 0.13 0.23
1-4 Sleep problem -0.08 0.24
1-4 Sleep problem -0.08 0.23
8,9 Weight problem -0.21 0.22
8,9 Weight problem -0.22 0.21
16 Psychomotor agitation 0.02 0.18
16 Psychomotor agitation 0.01 0.14
Items shaded in grey indicate clinician-rated items. For all loadings, standard deviation across 20 imputed datasets were at most 0.01. Component scores
≥0.20 are printed in bold font.
Table 2 The time-mode component scores averaged across 20 imputed datasets.
Measurement time point Time-mode components
Last measurements First measurements
Baseline -0.03 0.62
2 weeks 0.11 0.62
4 weeks 0.30 0.38
6 weeks 0.43 0.14
9 weeks 0.57 -0.13
12 weeks 0.63 -0.23
Standard deviations across 20 imputed datasets were at most 0.1 for all loadings. The time-mode component scores ≥0.40 are printed in bold.
Core array
The estimated core array and the explained variance for each combination of the person-, symptom- and time-mode components are shown in Table 3. The core array provides an overview of the three-way interactions, with larger core elements indicating stronger three-way interactions. The results showed that most of the total explained variance (12.1%) was associated with the interaction between the second person-mode component, the somatic symptom-mode component and the last measurements time-mode component. The second-most variance (4.7%) was explained by the interaction between the third person-mode component, the somatic symptom-mode component and the first measurements time-mode component.
5
Table 3 The core array of the 3MPCA model and the percentages of explained variance for each combination of person-, symptom- and time-mode components. Symptom-mode components Cognitive symptom-componentSomatic symptom-component
Time- mode components First measurements Last measurements First measurements Last measurements Core elements % EV Core elements % EV Core elements % EV Core elements % EVPerson-mode components Negative/suicidal thoughts 56.92.9460.72.35.190.016.10.04Physical dysfunction13.90.1950.71.6848.61.7413912.1General improvement21.50.32-5 0 82.64.6610.80.1%EV = Explained variance percentages of rotated components for each combination of components. Standard deviations across 20 imputed datasets for the core elements and explained variance were at most 0.56 and 0.07, respectively.
Interpretation of the person-mode components
Figures 1 C and D show the symptom-component scores over time for each of the three person-mode components. Because the plots show change over time relative to the general downward trend in the data, the averaged component scores for the cognitive and somatic symptom-mode components are plotted as reference.
The plots for each of the person-mode components reflect deviance from (i.e.
heterogeneity around) these general trends. Each of the person-mode components was interpreted by combining the person-mode component plots for each symptom domain (Figure 1 C, D) with the observed external correlations with patient characteristics (Table 4).
The plots for the first person-mode component were characterized by consistently higher cognitive symptom-component scores, and by lower somatic symptom-mode component scores, than the respective general trends in these domains. Furthermore, the first person-mode component score was positively correlated with suicidal thoughts and negative thinking (‘negative thoughts about self’) at study enrollment. Given the associated relative persistence of cognitive symptoms and the external correlations, this person-mode component was labeled
‘negative/suicidal thoughts’.
The second person-mode component was characterized by a relative worsening in both domains compared to their respective general trends, but the worsening was most pronounced for the somatic symptom domain. Additionally, this component was correlated with increased work-related and social disability, lower quality of life, and sleep and concentration side-effects, measured during the Level-1 phase. Based on these observations, the second person-mode component was labeled ‘physical dysfunction’. Interestingly, this person-mode component was the only one that was significantly correlated with antidepressant side-effects.
The third person-mode component’s plot was characterized by decreasing severity relative to the general trends in both symptom-domains. This component showed significant correlations with some HRSD items (‘thoughts of dying’,
‘decreased interest’, ‘depressed mood’, and ‘decreased activity’) and was labeled
‘general improvement’.
Table 3 The core array of the 3MPCA model and the percentages of explained variance for each combination of person-, symptom- and time-mode components. Symptom-mode components Cognitive symptom-componentSomatic symptom-component
Time- mode components First measurements Last measurements First measurements Last measurements Core elements % EV Core elements % EV Core elements % EV Core elements % EVPerson-mode components Negative/suicidal thoughts 56.92.9460.72.35.190.016.10.04Physical dysfunction13.90.1950.71.6848.61.7413912.1General improvement21.50.32-5 0 82.64.6610.80.1%EV = Explained variance percentages of rotated components for each combination of components. Standard deviations across 20 imputed datasets for the core elements and explained variance were at most 0.56 and 0.07, respectively.
Interpretation of the person-mode components
Figures 1 C and D show the symptom-component scores over time for each of the three person-mode components. Because the plots show change over time relative to the general downward trend in the data, the averaged component scores for the cognitive and somatic symptom-mode components are plotted as reference.
The plots for each of the person-mode components reflect deviance from (i.e.
heterogeneity around) these general trends. Each of the person-mode components was interpreted by combining the person-mode component plots for each symptom domain (Figure 1 C, D) with the observed external correlations with patient characteristics (Table 4).
The plots for the first person-mode component were characterized by consistently higher cognitive symptom-component scores, and by lower somatic symptom-mode component scores, than the respective general trends in these domains. Furthermore, the first person-mode component score was positively correlated with suicidal thoughts and negative thinking (‘negative thoughts about self’) at study enrollment. Given the associated relative persistence of cognitive symptoms and the external correlations, this person-mode component was labeled
‘negative/suicidal thoughts’.
The second person-mode component was characterized by a relative worsening in both domains compared to their respective general trends, but the worsening was most pronounced for the somatic symptom domain. Additionally, this component was correlated with increased work-related and social disability, lower quality of life, and sleep and concentration side-effects, measured during the Level-1 phase. Based on these observations, the second person-mode component was labeled ‘physical dysfunction’. Interestingly, this person-mode component was the only one that was significantly correlated with antidepressant side-effects.
The third person-mode component’s plot was characterized by decreasing severity relative to the general trends in both symptom-domains. This component showed significant correlations with some HRSD items (‘thoughts of dying’,
‘decreased interest’, ‘depressed mood’, and ‘decreased activity’) and was labeled
‘general improvement’.
5
Fig. 1. A. Examples of the mean QIDS item score trajectories, B. Examples of five patients’ individual QIDS trajectories, C. Cognitive symptom- component plot around the general trend and D. Somatic symptom- component plot around the general trend for the three person-mode components. The two vertical lines in panels C and D indicate the divisions between the two time-components (‘First measurements’ and ‘Last measurements’).
Table 4 Pearson/Spearman correlations between the person-mode component scores and patients’ characteristics Measuremen t time pointQuestion-naireVariablesPerson-mode component Negative self- image/ Suicidal thoughts
Physical dysfunctionGeneral improveme nt Enrollment PDSQ (about past 2 weeks) Thoughts of suicide0.380.070.10 Think you're better off dead0.360.080.16 Wish you were dead0.340.060.15 Think of dying in passive ways 0.330.110.17 Feel like failure 0.320.160.25 Negative thoughts about self 0.300.100.20 Think of dying(0.02)0.130.34 HRSDSuicide0.500.060.16 Work and interests 0.150.160.33 Depressed mood0.180.170.32 Level 1WSAS Sum score of WSAS 0.130.470.25 QLESQQLESQ General activities index-0.18-0.40-0.38 SF-120.11-0.36-0.08 PRISEPoor concentration0.060.370.10 Sleep(0.01)0.340.07 PDSQ = Psychiatric Diagnostic Screening Questionnaire, HRSD = Hamilton Rating Scale for Depression, WSAS = Work & Social adjustment scale, QLESQ = Quality of life enjoyment and satisfaction questionnaire, SF-12 = Short-Form Health Survey, PRISE = Patient Rated Inventory of side effects. Non-significant correlations (p>0.05) are presented in parentheses. The maximum P-value among the significant correlations was P=5.89e-36.
Fig. 1. A. Examples of the mean QIDS item score trajectories, B. Examples of five patients’ individual QIDS trajectories, C. Cognitive symptom- component plot around the general trend and D. Somatic symptom- component plot around the general trend for the three person-mode components. The two vertical lines in panels C and D indicate the divisions between the two time-components (‘First measurements’ and ‘Last measurements’).
Table 4 Pearson/Spearman correlations between the person-mode component scores and patients’ characteristics Measuremen t time pointQuestion-naireVariablesPerson-mode component Negative self- image/ Suicidal thoughts
Physical dysfunctionGeneral improveme nt Enrollment PDSQ (about past 2 weeks) Thoughts of suicide0.380.070.10 Think you're better off dead0.360.080.16 Wish you were dead0.340.060.15 Think of dying in passive ways 0.330.110.17 Feel like failure 0.320.160.25 Negative thoughts about self 0.300.100.20 Think of dying(0.02)0.130.34 HRSDSuicide0.500.060.16 Work and interests 0.150.160.33 Depressed mood0.180.170.32 Level 1WSAS Sum score of WSAS 0.130.470.25 QLESQQLESQ General activities index-0.18-0.40-0.38 SF-120.11-0.36-0.08 PRISEPoor concentration0.060.370.10 Sleep(0.01)0.340.07 PDSQ = Psychiatric Diagnostic Screening Questionnaire, HRSD = Hamilton Rating Scale for Depression, WSAS = Work & Social adjustment scale, QLESQ = Quality of life enjoyment and satisfaction questionnaire, SF-12 = Short-Form Health Survey, PRISE = Patient Rated Inventory of side effects. Non-significant correlations (p>0.05) are presented in parentheses. The maximum P-value among the significant correlations was P=5.89e-36.
5
Prediction of Level-1 treatment outcome
All person-mode components were significantly associated with remission at the end of the Level-1 phase (Appendix D). The ‘general improvement’ person-mode component showed a positive association with remission (OR=1.5), whereas the other two person-mode components were negatively associated with remission (‘negative/suicidal’: OR=0.63; ‘physical dysfunction’: OR=0.15).
Prediction of severity at 12-month follow-up
Of the patients that were included in the 3MPCA analysis (n=1,656), 489 patients (29.5%) had missing QIDS-SR16 scores at 12-month follow-up. Therefore, a subsample of 1,167 patients were analyzed for 12-month severity prediction. The person-mode components explained 17% of the variance in the 12-month follow- up QIDS-SR16 scores. The results (Appendix D) showed that only the ‘physical dysfunction’ component showed a significant association (p<0.05) with the 12- month follow-up QIDS-SR16 (B=2.37).
Discussion
Gaining more insight into the complex inter-personal variations in antidepressant- treatment response is important if we eventually want to identify the characteristics that can be used to differentiate between patients with different treatment responses. To gain these insights, the current study applied 3MPCA to simultaneously capture person-, symptom- and time-level heterogeneity in treatment responses. The optimal model explained 81% of the total variance (including the general trend) and 26% of the variance around the general trend, suggesting that the flexibility of the integrated multi-way approach of 3MPCA helped to account for a large part of the total variance in the treatment-response.
The identified model decomposed the data into two symptom-mode components (‘cognitive’ and ‘somatic’), two time-mode components (‘first measurements’
and ‘last measurements’) and three person-mode components (‘negative/suicidal thoughts’, ‘physical dysfunction’ and ‘general improvement’). The person-mode components were differentiated by their associated interactions between symptom- and time-mode components, and by their distinct patterns of correlations with auxiliary variables. All three person-mode component scores were associated with remission at the end of the Level-1 phase: the ‘general improvement’ component score was positively associated with remission, whereas the ‘negative/suicidal thoughts’ and ‘physical dysfunction’ component scores were both negatively associated with remission. Also, ‘physical dysfunction’ component scores were positively associated with depression severity at 12-month follow-up.
The current results indicate that the variations in patients’ treatment responses are captured very well by making a distinction between symptom- domain specific course-trajectories, rather than only differentiating between patients with course trajectories on a unidimensional severity construct as was done in previous studies (e.g. [16-20]). As such, the results align with previous research showing that variations in the course of depression are best captured by making a distinction between persons’ course-trajectories on mood/cognitive and on somatic symptom domains [30, 31, 56]. These findings suggest that patients can differ in their patterns of domain-specific symptom persistence and monitoring such patterns could help to find out which symptoms do and which symptoms do not respond to antidepressant treatment in which patients. 3MPCA- based person-mode component scores could be a useful tool to distinguish between patients with different types of symptom-specific treatment responses.
From a psychometric perspective, it is interesting that the current study indicated that the QIDS items can be decomposed into a ‘cognitive’ and ‘somatic’
symptom component, since previous factor-analytical work on the QIDS has mostly shown unidimensional structure [57]. Still, the finding does align with previous 3MPCA work [30], although two different instruments were used in the current and the previous studies (QIDS vs BDI). The fact that two components were found in the current study and that only one factor has been found in factor- analytical work could be the result of fundamental differences between the two analytical approaches. Other than factor analysis, which is mostly focused on cross-sectional data, the 3MPCA method is focused on the decomposition of symptom-heterogeneity based on the whole longitudinal dataset (i.e. accounts for time- and person-level heterogeneity when estimating the symptom-mode components). The observed multidimensionality of depressive symptoms in a sample treated for MDD could also be explained by the recent finding that the factor structure of depression questionnaires depends on overall depression severity, with structures becoming less multifactorial with decreasing severity levels in a sample [58]. In sum, the results support the notion that a more symptom-specific focus could be of considerable added value in research on antidepressant (non)response [59].
The current results showed that the time points in the dataset could be decomposed into two components, corresponding to time points early in the study and the time points later in the study. It is likely that these components reflect the time frames within which most (non)response to the medication occurred in the Level-1 phase. The baseline and 2-week time points loaded on the ‘first measurements’ component and fall within the 3-week time interval until a discernible effect of the SSRI treatment would be expected [60]. The 4-week time
Prediction of Level-1 treatment outcome
All person-mode components were significantly associated with remission at the end of the Level-1 phase (Appendix D). The ‘general improvement’ person-mode component showed a positive association with remission (OR=1.5), whereas the other two person-mode components were negatively associated with remission (‘negative/suicidal’: OR=0.63; ‘physical dysfunction’: OR=0.15).
Prediction of severity at 12-month follow-up
Of the patients that were included in the 3MPCA analysis (n=1,656), 489 patients (29.5%) had missing QIDS-SR16 scores at 12-month follow-up. Therefore, a subsample of 1,167 patients were analyzed for 12-month severity prediction. The person-mode components explained 17% of the variance in the 12-month follow- up QIDS-SR16 scores. The results (Appendix D) showed that only the ‘physical dysfunction’ component showed a significant association (p<0.05) with the 12- month follow-up QIDS-SR16 (B=2.37).
Discussion
Gaining more insight into the complex inter-personal variations in antidepressant- treatment response is important if we eventually want to identify the characteristics that can be used to differentiate between patients with different treatment responses. To gain these insights, the current study applied 3MPCA to simultaneously capture person-, symptom- and time-level heterogeneity in treatment responses. The optimal model explained 81% of the total variance (including the general trend) and 26% of the variance around the general trend, suggesting that the flexibility of the integrated multi-way approach of 3MPCA helped to account for a large part of the total variance in the treatment-response.
The identified model decomposed the data into two symptom-mode components (‘cognitive’ and ‘somatic’), two time-mode components (‘first measurements’
and ‘last measurements’) and three person-mode components (‘negative/suicidal thoughts’, ‘physical dysfunction’ and ‘general improvement’). The person-mode components were differentiated by their associated interactions between symptom- and time-mode components, and by their distinct patterns of correlations with auxiliary variables. All three person-mode component scores were associated with remission at the end of the Level-1 phase: the ‘general improvement’ component score was positively associated with remission, whereas the ‘negative/suicidal thoughts’ and ‘physical dysfunction’ component scores were both negatively associated with remission. Also, ‘physical dysfunction’ component scores were positively associated with depression severity at 12-month follow-up.
The current results indicate that the variations in patients’ treatment responses are captured very well by making a distinction between symptom- domain specific course-trajectories, rather than only differentiating between patients with course trajectories on a unidimensional severity construct as was done in previous studies (e.g. [16-20]). As such, the results align with previous research showing that variations in the course of depression are best captured by making a distinction between persons’ course-trajectories on mood/cognitive and on somatic symptom domains [30, 31, 56]. These findings suggest that patients can differ in their patterns of domain-specific symptom persistence and monitoring such patterns could help to find out which symptoms do and which symptoms do not respond to antidepressant treatment in which patients. 3MPCA- based person-mode component scores could be a useful tool to distinguish between patients with different types of symptom-specific treatment responses.
From a psychometric perspective, it is interesting that the current study indicated that the QIDS items can be decomposed into a ‘cognitive’ and ‘somatic’
symptom component, since previous factor-analytical work on the QIDS has mostly shown unidimensional structure [57]. Still, the finding does align with previous 3MPCA work [30], although two different instruments were used in the current and the previous studies (QIDS vs BDI). The fact that two components were found in the current study and that only one factor has been found in factor- analytical work could be the result of fundamental differences between the two analytical approaches. Other than factor analysis, which is mostly focused on cross-sectional data, the 3MPCA method is focused on the decomposition of symptom-heterogeneity based on the whole longitudinal dataset (i.e. accounts for time- and person-level heterogeneity when estimating the symptom-mode components). The observed multidimensionality of depressive symptoms in a sample treated for MDD could also be explained by the recent finding that the factor structure of depression questionnaires depends on overall depression severity, with structures becoming less multifactorial with decreasing severity levels in a sample [58]. In sum, the results support the notion that a more symptom-specific focus could be of considerable added value in research on antidepressant (non)response [59].
The current results showed that the time points in the dataset could be decomposed into two components, corresponding to time points early in the study and the time points later in the study. It is likely that these components reflect the time frames within which most (non)response to the medication occurred in the Level-1 phase. The baseline and 2-week time points loaded on the ‘first measurements’ component and fall within the 3-week time interval until a discernible effect of the SSRI treatment would be expected [60]. The 4-week time
5
point loaded on both components, possibly reflecting the transitory phase in which the SSRI starts to have an effect (or not).
The 3MPCA analyses revealed that variations in response to treatment could be captured by three person-mode components: ‘general improvement’,
‘physical dysfunction’ and ‘negative/suicidal thoughts’ person-mode components.
Scores on these components indeed showed different associations with treatment response at the end of the Level-1 treatment phase. Moreover, the ‘physical dysfunction’ person-mode component scores were associated with severity at the end of the 12-month naturalistic follow-up phase. The explained variance for this latter, long-term outcome was rather low (adjusted R2=0.17), but given the fact that the sample of 12-month follow-up respondents consisted of patients who received different (combinations of) treatment after the Level-1 phase, it is interesting to note that the ‘physical dysfunction’ person-mode component still showed a significant association.
Inspection of the core-array revealed that the interaction of the person- mode component ’physical dysfunction’, the ‘somatic’ symptom-component and the ‘last measurements’ time-mode explained about half of the heterogeneity (around the general trend) captured by the 3MPCA. Also, observations from Figures 1C and 1D showed that there was much more variation in somatic than in cognitive symptom scores between the person-mode components. Together, these findings indicated that variations in somatic symptoms (persistence) forms an important source of overall heterogeneity in MDD patients’ symptom-specific treatment responses.
The current study had several strengths, including the large sample size, the standardized treatment protocol, and the use of both the short- and long-term outcomes. However, there were also some limitations. First, the sample consisted of depressed outpatients treated with citalopram, limiting the study's generalizability of the results to other treatment settings. Second, the 3MPCA model was estimated with uncorrelated components within each mode, which may be unrealistic. Unfortunately, no oblique rotation technique is currently available to obtain simple structure in the core and component matrices simultaneously for all three modes. Third, STAR*D contained only a limited range of auxiliary variables and it would be interesting to investigate a broader range of clinically relevant variables in future research. In line with the Research Domain Criteria (RDoC [61]), other future research could focus on biological factors that may explain the symptom-specific differences in antidepressant treatment responses.
Furthermore, a confirmatory approach (e.g. by fixing some components) could be adopted to investigate how well the currently identified model fits to data collected in other large-scale trials. Finally, 3MPCA models could be run with a
broader range of symptoms including, for instance, anxiety symptoms that often co-occur with depressive symptoms.
Taken together, the current results showed that the heterogeneity antidepressant treatment responses can be captured by means of a 3MPCA model, illustrating the importance of simultaneously considering heterogeneity at the person-, symptom- and time-level. In addition, the results provide several useful leads for future studies that can lead to more empirically-based prescription of antidepressants.
