1
A Watershed Model of Depression
Investigating the Relationship between Affect-States and Depression Symptoms
Watershed model. From Kievit et al. (2016), figure 1, page 189.
By Gaby Lunansky
Supervisors: Claudia van Borkulo & Angélique Cramer1 Research Master Psychology Internship Report
University of Amsterdam Student number: 10206345 July 2017
2 Abstract
Even though one out of twenty people worldwide suffer from depression (Marcus, van Ommeren, Chisholm & Saxena, 2012), it is still unclear which factors contribute to its onset. Traditionally, depression symptoms (e.g. insomnia) were said to have one common, but unobservable cause. However, the diversity in depression patients symptomatology contradicts such a common cause conceptualization of depression. Recent studies have proposed the possibility of multiple, observable underlying causes of depression symptoms, namely affect-states (e.g. not feeling well; Schneiders et al. 2006; Wichers et al. 2009, 2012). We investigate the relationship between affect-states and depression symptoms, by fitting a hierarchical model on empirical data. This model is inspired by the Watershed model (Cannon & Keller, 2006), which has been recently applied to the concept of fluid intelligence (Kievit et al., 2016). In this study, we will use the statistical tests and predictions as discussed by Kievit et al. (2016) and apply them to our dataset. We find reasonable results of our fitted Watershed model of depression, indicating that the relationship between affect-states and depression symptoms could be captured in a hierarchical, multifactorial model, but needs more research to grasp the complexity of the interactions. We give clear suggestions for this future research.
3 Introduction
According to the World Health Organization (2012), depression is the number one cause of total life-years lost due to disability. One out of 20 people worldwide suffer from a depressive episode at least once in their life, and risk of recurrence is high (Marcus, van Ommeren, Chisholm & Saxena, 2012). Despite a plethora of research endeavors into depression and its treatments over the past 40 years, epidemiological surveys suggest that the worldwide prevalence of depression does not seem to decrease over the past decades (Kessler, 2003). This indicates that there still is no consensus with respect to the treatment of depression. In order to know what treatment is best for depression, it is necessary to know what depression actually is and which factors contribute to its onset.
Typically, depression has been conceptualized as a disorder with certain symptoms, such as fatigue, feeling irritated and loss of interest, indicating the absence or presence of the disorder. The covariance between symptoms of depression has been explained by positioning their joint dependence on one latent variable (Kendler, Aggen & Neale, 2004; Borsboom & Cramer, 2013). In such explanations, depression acts as a common cause that determines the depression symptoms (Borsboom, 2008; Cramer et al. 2010). In other words, a person suffers from depression symptoms because he or she is depressed. The reflective latent variable model for depression implies that, because depression symptoms are all caused by one and the same latent variable (i.e., depression), they should have a homogeneous structure. A
homogeneous structure means that, with changing levels of depression, the association structure of the symptoms should remain stable (i.e., symptoms that are more highly correlated in the general population should also be more highly correlated in a depressed sample); this is because the only way in which correlations between symptoms can arise in a latent variable model is through the function that describes their dependence on the latent
4 variable (the so-called Item Characteristic Curve or ICC; Embretson & Reise, 2013), which by assumption does not qualitatively change over the levels of the latent variable.
Figure 1: The reflective, common cause model of depression, where the depression symptoms (for example,
insomnia and loss of interest) are caused by the same, underlying latent variable ‘depression’.
From a latent variable modeling perspective, these properties imply that depression symptoms should behave in a stable manner: they should function in qualitatively similar ways at different stages of depression. However, empirical research shows these assumptions to be problematic: both the dimensionality and the factor loadings of latent variable models for depression are different across populations with different levels of depression (Fried et al., 2016). Also, the symptomatology of patients is not only different between patients, but even within patients: patients can have a different symptomatology for different depressive
episodes (Fried et al., 2016). Thus, the actual development of depression does not appear to be homogeneous, which is inconsistent with the common cause model.
Recent studies have already proposed another perspective on depression, where there is not one underlying common cause for the depression symptoms, but a whole lower level of ‘building blocks’. In this view, the depression symptoms are caused by underlying affect-states (Wichers, 2014). Affect-affect-states are momentary moods and experiences in the flow of
5 daily life, such as enthusiasm, feeling relaxed and feeling insecure. Several studies have shown that having negative affective reactions to daily life stressors predicts the future onset of psychopathology, such as depressive symptoms (Schneiders et al. 2006; Wichers et al. 2009, 2012). Therefore, it has been hypothesized that these momentary experiences are the micro-level building blocks that account for the development of psychopathological disorders such as Major Depressive Disorder (MDD; Wichers, 2014). For example, the combination of the affect-states having a low energy level, feeling agitated and not feeling calm, could lead to having sleeping problems such as the depression symptom insomnia.
Until now, the formal relationship between affect-states and depression symptoms has not been studied. An interesting model to test this relationship is the Watershed model
(Cannon & Keller, 2006). The Watershed model is a hierarchical model, following the analogy of different smaller branches of a river (the affect-states) merging into bigger branches (the depressive symptoms), and eventually all flowing into the same river mouth (MDD). The Watershed model is an interesting alternative for the common cause model of depression, since it makes predictions in line with the multifactorial conceptualization of depression discussed above. Furthermore, this hierarchical model allows to conceptualize affect-states as the lower-order building blocks of depression symptoms. Recently, the Watershed model has been applied in a study by Kievit et al. (2016) about fluid intelligence and the underlying variables ‘white matter integrity’ in the brain and ‘processing speed’. Kievit et al. (2016) translated the theoretical assumptions of the Watershed model into clear statistical predictions.
In this paper, we will propose a theoretical framework for a multifactorial conceptualization of depression. We propose a Watershed model of depression, where depression is caused by lower-order levels of affect-states and depression symptoms. We will apply the Watershed model to empirical data in order to take the first steps into investigating
6 the relationship between affect-states and depression symptoms, as an alternative to the
common cause model of depression. First, we will discuss the assumptions of the Watershed model, explaining why they provide an interesting alternative to the common cause model of depression. Second, we will discuss the statistical predictions that follow these assumptions, as proposed by Kievit et al. (2016). Lastly, we will test the Watershed model by fitting it to empirical data.
The Watershed Model
As stated above, the Watershed model was proposed by Cannon and Keller (2006) to account for the complexity of common mental disorders. Cannon and Keller investigate the genetic factors underlying mental disorders. However, they make a clear distinction between
‘Mendelian disorders’ and ‘common mental disorders’: the former are rare in the population, and caused by only one or two single-gene mutations (for example, sickle-cell anaemia), whereas the latter are thousands of times more common, with potentially very large numbers of genetic and environmental factors affecting each disorder (for example, depression and bipolar disorder). The enthusiastic search for single-gene mutations in patients of common mental disorders is understandable from a historical point of view, looking at the rapid pace of the discovery of molecular bases for Mendelian disorders in the years between 1970 – 1990. However, this focus on single-gene mutations should not be expanded to common mental disorders, since these are inherently different from Mendelian disorders. Classical genetic analyses focus too much attention on finding single-cell abnormalities, making the erroneous assumption that common mental disorders are fundamentally like Mendelian disorders. Cannon and Keller (2006) therefore argue in favour of a new multifactorial framework for genetic analyses of common mental disorders, since these disorders should be viewed as the end products of multiple causes. They propose a model with the analogy of a river, where
7 diverse, smaller branches eventually all flow into the same river mouth: the Watershed model.
The Watershed model comes with theoretical assumptions in line with a multifactorial perspective on mental disorders. These assumptions, and the statistical methods to test them, are discussed in the Cannon and Keller (2006) paper, and more recently in the Kievit et al. (2016) paper. We will explain three of these assumptions and corresponding statistical predictions, within the framework of our investigation of the relationship between affect-states and depression symptoms. These statistical predictions will be tested within a Structural Equation Modeling (SEM) perspective. See figure 2 for an abstract representation – not yet fitted to data – of our proposed Watershed model of depression.
Figure 2: Abstract representation of our proposed Watershed Model of Depression. The lower-order affect-states
all have causal influences on all the depression symptoms. These influences, represented by the directed paths, can be positive (green path) or negative (red path), depending if the affect itself is negative (“I don’t feel well”)
or positive (“I feel enthusiastic”). The depression symptoms have a (positive) causal influence on the development of depression.
8 The first assumption entails the changing dimensionality over the levels of the model. The Watershed model proposes multiple, underlying causal elements for the rise of depression symptoms, instead of one common cause. It accounts for the inherent multiplicity in patients’ symptomatology when they receive a diagnosis of depression. Therefore, the lower-order affect-states should have a high dimensionality, but the higher-order depression latent variable should be unidimensional. This assumption is in line with the former discussed findings of Fried et al. (2016), where both the dimensionality and the factor loadings of latent variable models for depression are different across populations with different levels of depression (Fried et al., 2016). The assumption of changing dimensionality translates into the prediction that both the affect-states and depression symptoms (level 1 and level 2 in figure 2) cannot be captured in a one factor model, however, the manifest indicators of the latent variable
depression (level 4) are captured by a one factor model. Within the SEM framework, this means that both a one factor model of affect-states and a one factor model of depression symptoms should have poor fit, but a one factor model of the manifest indicators of the latent variable depression (level 4) should have acceptable fit.
The second assumption entails multiple-realizability. Multiple-realizability implies a many-to-one relationship between two levels in a hierarchical model. This means that for our model, different clusters of affect-states should lead to the same depression symptoms, just like the different branches of the river merge together into the same river flow. Thus, we have many different possibilities of affect-state combinations, less different combinations of depression symptoms, which all eventually lead to depression. This assumption leads to three statistical predictions. First, the depression symptoms should not be captured by one
underlying (latent) variable, but should instead be captured by many variables (in our case, affect-states). This links to the first assumption of the model on multidimensionality, thus translates into the same statistical prediction: a one factor model of depressive symptoms
9 (level 2) should have poor fit. Second, the different affect-states should all make independent partial contributions to the onset of depressive symptoms. This is important for the possibility of different affect-state combinations leading to a smaller set of depressive symptoms. We will test if a model wherein all affect-state variables have paths to all depressive symptoms fits the data well. Lastly, there should be no direct paths from the lower-levels of the model, the affect-states and depression symptoms (level 1 and 2), to the indicators of depression (level 4). Such direct paths would contradict the many-to-one relationship, and would justify for example searching for a single-gene mutation of depression. The absence of direct paths between the indicators of depression, and the lower-level ‘building blocks’ expresses the notion that depression arises from complex, underlying interactions between affect-states and depression symptoms, thus the ‘essence’ of depression cannot be found in one small element. This assumption can be tested by fitting a model like in figure 2, with no direct paths from the affect-states nor the depression symptoms to the manifest indicators of depression, and check if this model has acceptable fit.
The third and final assumption entails hierarchical dependence between the levels of the Watershed model. Hierarchical dependence is a relationship wherein the levels of the model causally influence their upper level, and no levels can be ‘skipped’ to go from one level to another. Statistically, this is known as d-separation (Pearl, 2000, via Kievit et al., 2016). In a causal, hierarchical chain such as X Y Z, the middle variable Y ‘d-separates’ the relationship between the variables X and Z. This means that X and Z are statistically
independent controlling for Y, i.e. removing the shared variance of X and Z with Y makes X and Z statistically independent. In our model, we hypothesize that the depression symptoms (level 2) d-separates the affect-states (level 1) from the depression latent variable (level 3). This hypothesis represents the idea that a person cannot ‘jump’ from (negative) affect-states into a depression, but should first develop depression symptoms. From a SEM framework,
10 this is tested by having no direct paths between the affect-states and the depression latent variable, and checking if that model fits the data.
In conclusion, the Watershed model of affect-states and depression symptoms
represents the depression disorder as arising from interactions of lower-order affect-state and depression symptoms. It accounts for the multifactorial conceptualization of depression, where differences in the combinations of affect-states and depression symptoms between and within individuals all lead to the depression disorder. The Watershed model assumes changing dimensionality over the levels of the model, multiple-realizability and hierarchical
dependence between the levels of the Watershed model. The five corresponding statistical predictions are: poor fit of a one factor model for affect-states, poor fit of a one factor model for depression symptoms, acceptable fit of a one factor model of the manifest indicators for the latent variable depression, paths from all affect-state variables to all depression
symptoms, no direct paths from the affect-states to the latent variable depression nor its indicators, and no direct paths from the depression symptoms the indicators of the depression latent variable. These predictions will now be tested on empirical data.
Method
The dataset is part of a bigger study and contains 536 variables and 12011 participants2. The affect-states of the participants were measured for five days, and their depression symptoms were measured once during those five days, and four times after. However, many participants stopped the study after one measurement. Selecting only participants with at least one
measurement moment, leaves us with 603 participants which are used in our study.
To measure the affect-states, Experience Sampling Method (ESM) was used. ESM is a recently new research method, where participants fill in a questionnaire for a couple of times
11 a day. The ESM questionnaire used in this study is in Dutch and asks the participants to rate their affect-states on a 7-point Likert scale. Examples of affect-states are: feeling insecure, lonely, relaxed and guilty. The participants in this dataset filled in the ESM questionnaire ten times a day, for five days. Since each participant had repeated measures, we took the mean score on each affect-state variable for each participant. This leaves us with one measurement per affect-state variable, for each participant.
To measure the depression symptoms, the Symptom Checklist – 90 (SCL-90, Derogatis, 1996) was used. The SCL-90 scale has 90 items that check the presence of 90
psychopathological symptoms, divided over 9 scales. The dataset contains the scores on the dimensions: depression, anxiety, psychosis and paranoid thoughts. Their scores on the SCL-90 dimensions are reported in the dataset. Unfortunately, these scales do not fully represent the depression symptoms needed for our study, since they do not present symptoms but scales (clusters of symptoms), and are not restricted to depression but also other psychopathological disorders. Therefore, it is necessary to repeat our study with better measurements of
depression symptoms. However, the main purpose of this study is to present a new theoretical framework for a multifactorial conceptualization of depression, by explaining the Watershed model, its theoretical assumptions and corresponding statistical predictions. This framework, nor the theoretical assumptions or statistical predictions are affected by using these
psychopathological scales instead of depression symptoms.
Different selection criteria were used to select the relevant participants and variables from the large dataset. The whole analysis can be found as R-script at
https://github.com/glunansky/internship/. To select the correct affect-state variables, we made a first selection by removing the variables with more than 95% missing values. Furthermore, we removed the affect-state variables with a variance lower than 1. Lastly, we tried to balance
12 the number of positive and negative affect-states. We selected 11 affect-state variables, which can be found in table 1.
Table 1: Selected affect-state variables
"Worrying" : Im worrying. “Satisfied" : Im satisfied. "Cheerful" : I am cheerful. "Energy" : I feel full of energy. "Unsure" : I feel unsure. "NWell" : I dont feel well. "Relaxed" : I am relaxed. "Agitated" : I feel agitated. "Angirr" : I am angry or irritated. "Quiet" : I feel quiet.
"Enthus" : I feel enthusiastic.
As explained in the former section, the Watershed model has a fourth level of manifest indicators of the latent variable. The Watershed model for fluid intelligence (Kievit et al., 2016) takes the scores on IQ tests as manifest indicators of the latent variable fluid intelligence. However, for our study, it is more problematic to have manifest indicators caused by one latent variable ‘depression’. Even more so, in the former sections we have presented argument against a common cause model of depression. Nevertheless, only having formative variables in a (co-variance based) SEM framework can lead to problems regarding the error of the estimated parameters (Cadogan & Lee, 2013). For this reason, it is not
possible to fit the Watershed model without the fourth level in lavaan. As a solution, we chose to take further measurement moments of the SCL-90 scales as the manifest indicators. The Watershed model states that all lower-level variables eventually all lead to the same construct. In our model, this construct is depression. Thus, we can interpret the presence of this
13 We have four follow-up measurement moments of the SCL-90 scales in our dataset, thus we can test which moment (moment 2, 3, 4 or 5) has the best model fit.
All analyses will be done within the programming language R and the package “lavaan” (Rosseel, 2012). The models are fitted using Maximum Likelihood Estimation (ML), with robust standard errors. Model fit indicators that are taken into account are the chi-square test (p-value should be > 0.05), RMSEA (good fit at RMSEA < 0.05, acceptable fit at 0.05 – 0.08) and the Comparative Fit Index (good fit at CFI > .97, acceptable fit at 0.95-0.97).
Results
The first step in investigating the relationship between affect-states and depression symptoms within the framework of the Watershed model, is testing if our measurement model has good fit. The measurement model indicates that our latent variable ‘depression’ can be explained by its manifest indicators. By testing our measurement model, we also check the hypothesized unidimensionality of our latent variable depression. We fit a one factor model of depression for the four follow-up measurement moments of the depression symptoms. The third measurement moment fits the data best: 𝜒2(2)= 10.89 , p =0.004, RMSEA = 0.092, CFI =
0.99. The model fit indicators are good; CFI has very good fit, RMSEA is close to acceptable. Thus, the third measurements of depression symptoms are the best manifest indicators of the latent variable depression with acceptable fit. Also, by fitting this one factor model of
depression, we tested the unidimensionality of our latent variable depression. This assumption has been met.
The second step in building the Watershed model for depression is testing if a one factor model for the affect-state variables has poor fit. The SEM analysis indicates that this one factor model for affect-states has indeed poor fit: 𝜒2(44)= 1494.86, p < 0.001, RMSEA =
14 hypothesized.
The third step is testing if a one factor model for the depression symptoms has poor fit. The SEM analyses actually indicate an extremely good fit of the depression symptoms: 𝜒2(2)= 4.53, p = 0.11, RMSEA = 0.046 and CFI = 0.995. Thus, we cannot reject the one
factor model of depression symptoms, which is not in line with our hypothesis. This also means that the depression symptoms can be said to be unidimensional, instead of the hypothesized multidimensionality.
Even though our hypothesis of a rejected one factor model for depression symptoms should be rejected, we still investigate the fit of the full Watershed model for the
completeness of this study. The fitted model, with no direct paths from the affect-states to the depression symptoms, nor the latent variable and manifest indicators, and no direct paths from the depression symptoms to the manifest indicators, has poor fit: 𝜒2(64)= 588.64, p < 0.001,
RMSEA = 0.128 and CFI = 0.699.
However, as a post hoc analysis, we fitted the same model again, but added residual correlations of the affect-states (level 4), and residual correlations of the depression symptoms (level 3). Looking at the code of the paper by Kievit et al. (2016), they also added residual correlations for the third level of the model (processing speed). This makes sense
theoretically: if we conceptualize depression as a complex, multifactorial construct, it could be that there is residual variance that cannot be explained by the variables we include in our model. There could be other causal factors playing a role in the onset of depression, for example environmental factors. Figure 3 shows the post-hoc Watershed model where we add residual correlations for the depression symptoms and the affect-state variables. This model fits the data good: : 𝜒2(58)= 203.79, p < 0.001, RMSEA = 0.071 and CFI = 0.974.
15 Figure 3: The fitted post-hoc Watershed model. The names of the affect-state variables can be found in table 1.
Lastly, notable results of the model are the factor loading values from the affect-states to the depression symptoms: the positive affect-states (“Satisfied”, “Quiet”, “Enthusiastic” and “Cheerful”) have negative factor loadings on the depression symptoms, while the negative affect-states (“Worrying”, “Unsure” and “Angry or Irritated”) have positive factor loadings on the depression symptoms. These results are intuitive from our perspective of affect-states being the building blocks of depression symptoms: having negative affect-states for a longer period of time would eventually cause depression symptoms, while the positive
16 affect-states might block the development of depression symptoms. We did not propose any hypothesis on the value of the factor loadings of the model, hence we are making purely post-hoc observations of the results. Still, it would be interesting to see if these results hold in future research.
Discussion
To our knowledge, this study is the first investigation using an empirical dataset containing both measurements of affect-states and depression symptoms of the same participants to investigate their relationship. Various studies have suggested that affect-states are underlying factors of depression symptoms (Schneiders et al. 2006; Wichers et al. 2009, 2012). We presented a theoretical framework for this suggestion, together with the statistical methods to test this proposed relationship between affect-states and depression symptoms on empirical data. This theoretical framework accounts for a multifactorial conceptualization of depression, in line with findings on the changing dimensionality of different levels of depression (Fried et al., 2016) and the multiplicity in symptom combinations of MDD diagnosed patients. We tested if depression has underlying, multiple causes, where affect-states are the building blocks of depression symptoms, which eventually lead to depression. The presented
Watershed model is a promising model to test the hypothesis of affect-states being of causal influence on the development of depression symptoms, in a hierarchical, multi-realizable structure. The model has three important theoretical assumptions: changing dimensionality over the levels of the model, multiple-realizability of affect-states and depression symptoms, and hierarchical dependence between affect-states and the onset of depression. These
assumptions were translated into five statistical predictions, which were tested on empirical data.
17 The results of the conducted analyses support the measurement model and reject the one factor model of affect-states, as hypothesized. However, the one factor model of
depression symptoms cannot be rejected. This goes against our proposed multifactorial conceptualization of depression. Despite this finding, we continued to fit the full Watershed model. Even though the one factor model for depression symptoms could not be rejected statistically, the theoretical consequences of a one factor model for depression symptoms are simply not in line with the findings in clinical practice. The symptomatology of depressed patients is too diverse to assume all symptoms have only one underlying, latent common cause. Thus, although enough of the variance of depression symptoms can statistically be captured in one underlying factor, we argue that the findings on the changing dimensionality over the depression levels and differences in patients’ symptomatology give reason to continue the investigations on multiple underlying causes of depression.
The full Watershed model had good fit, when adding post-hoc residual correlations of the states and the depression symptoms. The residual correlations mean that the affect-states and depression symptoms have enough shared variance, but that this shared variance cannot be explained by the variables in our model. Thus, there may be other factors
contributing to the onset of depression. This supports our theoretical framework of a multifactorial, complex conceptualization of depression. However, it is not yet clear which factors should be taken into account for future research. One possibility is including more affect-states in the model, instead of our proposed selection of 11 affect-states. This could easily be done in future investigations. Another possibility is investigating if there are other environmental factors that account for the onset of depression, like stressful life events (Cramer, Borsboom, Aggen & Kendler, 2012). These factors could be included in further developments of the model, trying to grasp more of its complexity. In conclusion, the results of our post-hoc fitted Watershed model are in line with the proposed multifactorial
18 conceptualization of depression, and ask for more research on the possible causal factors that attribute to the onset of depression. Our study provides the theoretical and statistical tools for conducting such future research.
We have two suggestions for future research, based on the problems we encountered conducting our study. The first suggestion entails the measurement of depression symptoms. The dataset we used did not contain the optimal measurements of depression symptoms for this study, since the symptom values were actually the summed item values of the SCL-90 scales, instead of the individual item values. Future research should include item scores of the SCL-90 (or another questionnaire), not the values on a summed scale. It is possible that the summed scores are too clustered representations of the depression symptoms, when we are in fact interested in discovering the specific interactions between affect-states and individual depression symptoms. If the symptom values in our dataset actually represent many clusters of symptoms, we are not able to investigate the multiple-realizability structure we are interested in. Investigating this multiple-realizability structure entails investigating as many possible interactions, to see if different clusters of affect-states lead to the same depression symptoms. If these depression symptoms are already clustered, a part of this multiple-realizability structure would be inherently hidden. In conclusion, it would be interesting to have all the scores on the depression items of the SCL-90, not only the total scores of the scales for future investigations.
The second suggestion is a point of discussion on the number of layers that are included in the model. Kievit et al. (2016) included manifest indicators of the latent variable ‘fluid intelligence’, namely scores on a IQ test, as the fourth layer of the Watershed model. By doing this, the latent variable is both reflective (see figure 1) and formative (see figure 4): the latent variable is both constituted by manifest indicators, as also causing other manifest indicators. Without the fourth layer, the latent variable of depression would only be
19 constituted by its underlying symptoms, making it a purely formative latent variable. A
formative latent variable has no causal influences on other variables.
We followed the statistical structure as proposed by Kievit et al. (2016) by including a fourth level, using a follow-up measurement of depression symptoms as manifest indicators of the latent variable depression. However, it is debatable if the latent variables ‘depression’ and ‘fluid intelligence’ are comparable in nature, and consequently, if the latent variable
depression should be modeled as a reflective latent variable with manifest indicators. This debate links to the difference in ontological status between a reflective latent variable and a formative latent variable.
Figure 4: A formative model. The latent variable depression is composed by its underlying symptoms.
When we fit a reflective latent variable with manifest indicators, the conducted statistical analysis is only meaningful if we assume ‘entity realism’ of the latent variable (Borsboom, Mellenbergh & Van Heerden, 2003). In other words, we have to assume that the latent variable exists in the real world independently of our measurements, and that it is the unobservable cause of our measurements. Only with that assumption, it is meaningful to interpret the results of a fitted, reflective model with factor analysis. This is not true for a
20 formative model: the formative latent variable can be seen as a label or summary of our
measurements, like in the measurement of Social Economic Status (SES; Borsboom, Mellenbergh & Van Heerden, 2003). We will now shorty explain why it is debatable to assume that the latent variable ‘fluid intelligence’ and the latent variable ‘depression’ are the same in nature, and consequently, if ‘depression’ should be modeled as a reflective latent variable.
The model by Kievit et al. (2016) proposes two aspects of the brain, processing speed and white matter integrity, as underlying attributes of fluid intelligence. The level of fluid intelligence in an individual causes his/her scores on a IQ test. It is assumed that the causal processes that attribute for one’s level of fluid intelligence are similar for every individual, however, that the values of white matter integrity and processing speed are different over individuals. Therefore, individuals have different levels of fluid intelligence and different scores on a IQ test. In other words, we assume one scale of fluid intelligence, but position every individual on a different level of this scale. Thus, everybody has some level of the same ‘unit’ of fluid intelligence. However, this is not the case for our latent variable of depression. First, we cannot say that every individual has some level of depression, since not all people are diagnosed with depressive disorder, nor suffer depression symptoms. There is not a presence of ‘depression’ in all individuals, like there is a presence of “fluid intelligence” in all individuals. Second, we have been arguing against a homogeneous conceptualization of depression, presenting evidence of former studies about the multiplicity and diversity of depression and depression symptomatology. It is at least debatable to assume that the latent variable depression has ‘entity realism’. Therefore, we do not think that the latent variable of fluid intelligence in the model by Kievit et al. (2016) and the latent variable of depression in our model are comparable in nature. Instead of attributing a realist ontological status to the depression latent variable in our model, we could think of this latent variable as formative
21 instead of reflective. This means that we could view the latent variable of depression in our model as a label for the collection of depression symptomatology of patients. Therefore, it is debatable if we need a fourth level of manifest indicators of depression in our model.
Unfortunately, it is not without problems to simply leave out the fourth level of
manifest indicators in the proposed Watershed model of depression. One problem of modeling the latent variable of depression as strictly formative - by leaving out the fourth layer of manifest indicators - is that estimating a hierarchical model with only formative (latent) variables within the SEM framework is problematic in terms of estimating the standard errors (Cadogan & Lee, 2013). A possible solution for this problem is using partial-least squares SEM (PLS-SEM). Instead of doing covariance analysis, PLS-SEM uses partial-least square regression and is able to estimate higher-order, formative models (Suoniemi, Terho & Olkkonen, 2012). Thus, future investigations of the Watershed model of depression without manifest indicators, conceptualizing depression as a formative latent variable, use PLS-SEM. This could be done in R by using the package “semPLS” (Monecke & Leisch, 2012).
Despite the limitations of this study and dataset, namely the measurements of depression symptoms and the debatable choice of including manifest indicators for depression, the first steps into a theoretical framework on a hierarchical, multifactorial conceptualization of depression have been taken. The framework proposed in this study should be seen as a first step into further discovering the complex nature of depression, in order to understand which factors account for its onset and to develop adequate prevention and treatment strategies. We hope this study will encourage other researchers to help unravel the complexity of depression and other mental disorders, by leaving the common cause model and start looking for observable, underlying multiple causes.
22 References
Bringmann, L. F., Vissers, N., Wichers, M., Geschwind, N., Kuppens, P., Peeters, F., & Tuerlinckx, F. (2013). A network approach to psychopathology: new insights into clinical longitudinal data. PloS one,8(4), e60188.
Borsboom, D., & Cramer, A. O. (2013). Network analysis: an integrative approach to the structure of psychopathology. Annual review of clinical psychology, 9, 91-121.
Borsboom, D., Mellenbergh, G. J., & Van Heerden, J. (2003). The theoretical status of latent variables. Psychological review, 110(2), 203.
Cadogan, J. W., & Lee, N. (2013). Improper use of endogenous formative variables. Journal of Business Research, 66(2), 233-241.
Cramer, A. O., Borsboom, D., Aggen, S. H., & Kendler, K. S. (2012). The pathoplasticity of dysphoric episodes: differential impact of stressful life events on the pattern of depressive symptom inter-correlations. Psychological medicine, 42(5), 957-965.
Cannon, T. D., & Keller, M. C. (2006). Endophenotypes in the genetic analyses of mental disorders. Annual Review of Clinical Psychology, 2, 267-290.
Derogatis, L. R. (1996). SCL-90-R: Symptom Checklist-90-R: administration, scoring, and
23
Judea, P. (2000). Causality: models, reasoning, and inference. Cambridge University Press.
521(77362).
Kievit, R. A., Romeijn, J. W., Waldorp, L. J., Wicherts, J. M., Scholte, H. S., & Borsboom, D. (2011). Mind the gap: a psychometric approach to the reduction
problem.Psychological Inquiry,22(2), 67-87.
Kievit, R. A., Davis, S. W., Griffiths, J., Correia, M. M., & Henson, R. N. (2016). A
Watershed model of individual differences in fluid intelligence. Neuropsychologia,91, 186-198.
Kendler, K. S., Aggen, S. H., Prescott, C. A., Jacobson, K. C., & Neale, M. C. (2004). Level of family dysfunction and genetic influences on smoking in women. Psychological
medicine, 34(07), 1263-1269.
Monecke, A., & Leisch, F. (2012). semPLS: structural equation modeling using partial least squares.
Rosseel, Y. (2012). lavaan: An R Package for Structural Equation Modeling. Journal of Statistical Software, 48(2), 1-36.
Schneiders J, Nicolson NA, Berkhof J, Feron FJ, van Os J, deVries MW (2006). Mood reactivity to daily negative events in early adolescence: relationship to risk for psychopathology. Developmental Psychology 42, 543–554.
24 Suoniemi, S., Terho, H., & Olkkonen, R. (2012). The Measurement of Endogenous
Higher-Order Formative Composite Variables in PLS-SEM: An Empirical Application from CRM System Development. In Proceedings of World Academy of Science,
Engineering and Technology (No. 72, p. 571). World Academy of Science, Engineering and Technology (WASET).
Tian, J., & Pearl, J. (2002). A general identification condition for causal effects. In AAAI/IAAI, 567-573.
Wichers M, Geschwind N, Jacobs N, Kenis G, Peeters F, Derom C, Thiery E, Delespaul P, van Os J (2009). Transition from stress sensitivity to a depressive state: longitudinal twin study. British Journal of Psychiatry 195, 498–503.
Wichers, M., Lothmann, C., Simons, C. J., Nicolson, N. A., & Peeters, F. (2012). The dynamic interplay between negative and positive emotions in daily life predicts
response to treatment in depression: a momentary assessment study.British Journal of Clinical Psychology,51(2), 206-222.
Wichers, M. (2014). The dynamic nature of depression: a new micro-level perspective of mental disorder that meets current challenges.Psychological medicine, 44(07), 1349-1360.
Wilson, D., & Carston, R. (2006). Metaphor, relevance and the ‘emergent property’ issue. Mind & Language, 21(3), 404-433.
