Major depressive disorder (MDD) is the second leading cause of disability worldwide.Reference James, Abate, Abate, Abay, Abbafati and Abbasi1 Affecting over 350 million individuals globally, MDD is the most commonly diagnosed mental disorder in clinical practice.Reference James, Abate, Abate, Abay, Abbafati and Abbasi1,Reference Herrman, Patel, Kieling, Berk, Buchweitz and Cuijpers2 Although accurate assessment is an integral part of treatment, differences have been found across patients in the number of ways symptoms can be combined to satisfy the MDD diagnostic criteria.Reference Fried and Nesse3,Reference Zimmerman, Ellison, Young, Chelminski and Dalrymple4 Accordingly, it has been argued that additional clinical characterisation is needed to more precisely understand the psychopathology of the individual patient and to personalise treatments.Reference Maj5,Reference Maj, Stein, Parker, Zimmerman, Fava and De Hert6 International clinical guidelines focus on overall symptom severity as the key criterion for treatment selection.Reference Maj, Stein, Parker, Zimmerman, Fava and De Hert6–Reference Zimmerman9 This is problematic as additional layers of individual differences in psychopathological expression may be present within the same level of severity.Reference Fried and Nesse10 We propose that this layer does not concern the number or severity of symptoms, but rather how symptoms interact with each other. This relates to the concept of symptom dynamics (or symptom relationship patterns), denoting how symptoms interact with one another over time in a given patient.Reference Borsboom and Cramer11,Reference Ebrahimi, Burger, Hoffart and Johnson12 This concept has been largely underrecognised in the mental health field, even though it has been shown that symptom relationship patterns may actively contribute to the expression of and any changes in the psychiatric condition.Reference Borsboom and Cramer11–Reference Cramer, van Borkulo, Giltay, van der Maas, Kendler and Scheffer16 Two individuals with MDD of the same symptom severity may therefore exhibit substantially different patterns of interaction between their symptoms. Information about these patterns can help identify specific features and differentiate patients displaying the same symptom severity from each other.Reference Fried and Nesse10 However, the extent to which individual differences are present along this dimension of psychopathology in patients with the same diagnosis and severity remains uninvestigated. In this study, we aimed to examine the proportion of patients with MDD that display individual differences in their symptom dynamics while sharing the same overall symptom severity.
Method
This study used data collected at four secondary care facilities in The Netherlands during the ZELF-i study, a randomised controlled trial investigating self-monitoring using ecological momentary assessment in MDD patients.Reference Bastiaansen, Meurs, Stelwagen, Wunderink, Schoevers and Wichers17 The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008. All procedures involving patients were approved by the Medical Ethical Committee of the University Medical Center Groningen (reference: 2015/530).
Participants and procedure
Participants were recruited between 2016 and 2018 by clinicians who screened them for eligibility using DSM-IV.18 All participants: (a) were out-patients aged between 18–65 years; (b) had a current primary diagnosis of major depressive disorder (MDD) and indication for treatment of depression; (c) spoke Dutch; and (d) provided written informed consent. Exclusion criteria were: (a) presence of psychotic or (hypo)manic symptoms (i.e. any form of bipolar disorder); (b) need for urgent care (i.e. acute suicidality); and (c) inability to follow research procedures owing to the presence of intellectual disability or significant visual or hearing impairments. The participants received travel reimbursements and an additional €10 as compensation for internet usage if they used their own smartphones during the study. No additional compensation was provided for participation. Among the 110 participants who provided daily measurements in the ZELF-i study, 74 (67.3%) met the criteria of sharing a symptom severity score with at least one other participant and were waiting to start treatment, rendering them eligible for the present study.
Measurements
Depressive symptom severity
Overall symptom severity at baseline was assessed using the self-reported Inventory of Depressive Symptomatology (IDS-SR).Reference Rush, Gullion, Basco, Jarrett and Trivedi19 The IDS-SR is a 30-item instrument querying about the past 7 days, scored on a 4-point scale (0–3), with higher scores reflecting greater severity; total score 0–13: no depression; 14–21: mild depression; 22–30: moderate depression; 31–38: severe depression; 39 or more: very severe depression.Reference Schulte-van Maaren, Carlier, Zitman, van Hemert, de Waal and van der Does20 The internal consistency was good for this scale, with a Cronbach's α of 0.84.
Momentary assessment
Following an ecological momentary assessment (EMA) design, intensive repeated measurements were planned with a fixed 3 h sampling frequency 5 times per day over a 28-day period for each participant, yielding up to a maximum of 140 possible assessments per person. Momentary assessments were conducted via links to a questionnaire on a secure website sent to each participant's smartphone; those without smartphones were lent one for the study.Reference Bastiaansen, Meurs, Stelwagen, Wunderink, Schoevers and Wichers17 The links were sent via text messages, with the participants instructed to respond as soon as possible on receipt. The questionnaires were open for 30 min before the links expired.
All momentary assessment items mapping onto depressive symptomatology were included in this study. Specifically, (a) depressed mood was measured by the items feeling down and feeling cheerful (reversed); (b) anhedonia by feeling indifferent, listless and enthusiastic (reversed); (c) appetite change as reported deviations in hunger from one's stable average (i.e. being more or less hungry than usual); (d) restlessness by feeling calm (reversed), stressed and relaxed (reversed); (e) irritability by feeling irritated; and (f) lethargy with reports of being tired and energetic (reversed) (the items are listed in Supplementary Material 3, available at https://dx.doi.org/10.1192/bjp.2024.19). Items were rated on a visual analogue scale with scores ranging from 0 (not at all) to 100 (very much). Mean scores were used for symptoms with multiple affective indicators to retain all variables on a 0–100 scale. The items adhered to guidelines for dynamic assessment in EMA studies, being brief and unambiguous (measuring one state per item), present in the ESM Item Repository and worded to assess momentary states, to avoid recall bias.Reference Eisele, Kasanova, Houben, Myin-Germeys and Kuppens21–Reference Kirtley, Hiekkaranta, Kunkels, Eisele, Verhoeven and Van Nierop23
Statistical analyses
Statistical analyses were performed in R (version 4.2.2 on macOS).
Dynamic network analysis
The graphical vector autoregressive model (GVAR) in the ‘graphicalVAR’ (version: 0.3.3) and ‘psychonetrics’ (version: 0.11.15) R packages were used to estimate regularised and unregularised network models respectively, yielding person-specific dynamic network models that identify symptom relationship patterns in each participant.Reference Epskamp, Waldorp, Mõttus and Borsboom24,Reference Epskamp25 The regularised network models incorporate the least absolute shrinkage and selection operator (LASSO) and use the Bayesian information criterion (BIC) to identify the best fitting model.Reference Epskamp, Waldorp, Mõttus and Borsboom24 Visualised networks are based on regularised models to facilitate sparsity and more easily portray the most important edges.Reference Epskamp, Waldorp, Mõttus and Borsboom24 In estimating individual differences in symptom dynamics, raw edge weights from unregularised network models were obtained to yield unbiased estimates (detailed below). To facilitate sensitivity and to address missing data, we followed recommendations using the Kalman filter with LASSO regularisation in the regularised models and full information maximum likelihood (FIML) in unregularised models.Reference Mansueto, Wiers, van Weert, Schouten and Epskamp26 We further followed these simulation recommendations advising the inclusion of a maximum of six nodes with sample sizes around 75 to 100 time points per person to optimally recover the person-specific network structures.Reference Mansueto, Wiers, van Weert, Schouten and Epskamp26 Following the stationarity assumption of GVAR models, all variables were investigated for linear, quadratic and cubic trends. Simulation studies show no difference between detrending all variables versus a specific variable if any trend is present.Reference Epskamp, van Borkulo, van der Veen, Servaas, Isvoranu and Riese27 Accordingly, all variables were detrended for consistency.
The GVAR model produces two types of network per individual: a temporal and a contemporaneous.Reference Epskamp, Waldorp, Mõttus and Borsboom24,Reference Epskamp25 Each of these person-specific networks includes the depressive symptoms as variables in the network (nodes), with the lines between the nodes (edges) revealing the statistical relationship between them. These relationships may be positive or negative. They encode deviations from person-specific means, reflecting how higher levels on a variable compared with the person's average are associated with increases or decreases in other variables compared with the person's average.Reference Ebrahimi, Burger, Hoffart and Johnson12,Reference Epskamp, Waldorp, Mõttus and Borsboom24,Reference Epskamp25 The temporal network encodes effects forward in time, with its directed edges portrayed as arrows. These temporal edges are regression coefficients representing a node's associated strength of impact on another node at a consecutive time point, while controlling for all other variables in the network.Reference Epskamp, Waldorp, Mõttus and Borsboom24,Reference Epskamp25 These time-lagged edges are referred to as Granger causal,Reference Granger28 reflecting satisfaction of the temporal criterion of causality, yielding information about which node temporally precedes another in the system. In this study employing a lag-1 GVAR model with 3 h between assessments, edges in the temporal network reflect a symptom's associated strength of impact on another 3 h later. The contemporaneous network embodies partial correlations to display the unique relationships between all nodes in the network after accounting for the temporal effects.Reference Epskamp, Waldorp, Mõttus and Borsboom24,Reference Epskamp25 In the dynamic network literature, these edges may be interpreted as dynamics that are faster than those captured in the lag-1 temporal model,Reference Epskamp, van Borkulo, van der Veen, Servaas, Isvoranu and Riese27 reflecting the relationship between symptoms inside a 3 h time window in the present study.
Individual differences in depressive symptom dynamics
We used the individual network invariance test (INIT)Reference Hoekstra, Borsboom, McNally and Epskamp29 to investigate our research question about differential symptom relationship patterns (i.e. differences in the edge weights) in groups of MDD participants sharing the same overall symptom severity (IDS-SR scores). INIT is a specialised test developed to inspect differences between networks.Reference Hoekstra, Borsboom, McNally and Epskamp29 This test compares two models within the groups of participants matched on symptom severity: one model in which all edges in each of the person-specific network models are freely estimated (where symptom relations are different across participants) and a contrasting model in which all edges in these networks are constrained to be equal (where there are no differences in symptom relations). Simulation studies have identified the Akaike information criterion (AIC) to be the most sensitive criterion to identify the presence (or absence) of such differences between network structures, with lower AIC values reflecting the best fitting among the contrasting ‘no difference’ (homogeneous) versus difference (heterogeneous) models.Reference Hoekstra, Borsboom, McNally and Epskamp29 In sample sizes similar to those in the present study (i.e. approximately 100 responses per individual) and with six nodes present, the INIT test has revealed optimal sensitivity to detect individual differences when constraints are placed on both the temporal and contemporaneous edges.Reference Hoekstra, Borsboom, McNally and Epskamp29 The INIT weighs differences in temporal versus contemporaneous edges equally, and further takes into account differences in the strength versus the absence/presence of edges. Additional mathematical details of the INIT method are available elsewhere.Reference Hoekstra, Borsboom, McNally and Epskamp29 We chose the test as it is conservative with respect to difference detection, minimising the risk of overestimation, as it more easily detects no difference than differences in symptom dynamics (i.e. greater chance of a type II than a type I error). In summary, in a comparison of, for example, three individuals with MDD matching on severity levels (e.g. all three with a symptom severity score of 47), the INIT test reveals whether these three individuals display no difference in their symptom dynamics and thus whether they all can be represented by one (i.e. the same) network model, or whether they display differential symptom relations and therefore three models (one for each) must be estimated to represent their unique symptom dynamics. In a sensitivity analysis using a logistic regression, we checked whether individual difference detection was related to symptom severity levels (IDS-SR total scores range: 15–51) and severity group size (range: 2–6).
To obtain the proportion of MDD participants who shared the same overall symptom severity and displayed individual differences in their symptom dynamics, the total number of participants matched on symptom severity who were identified as displaying differences in their symptom dynamics was divided by the total number of participants in the sample. To construct a 95% confidence interval around the estimated proportion, this calculation procedure was repeated 10 000 times following random draws with replacement of the severity groups.Reference DiCiccio and Efron30
Simulation inspecting the robustness of the INIT method in our sample
To investigate the extent to which the INIT method was susceptible to false detection of individual differences given the specific conditions of our study (e.g. number of individuals in each severity group, precise number of responses and missing values per participant), we conducted a supplementary simulation (full details in Supplementary Material 1). This procedure reflects obtaining the expected proportion of individual differences under a simulated null scenario where no such individual differences should be present.Reference Hoekstra, Borsboom, McNally and Epskamp29
Results
Demographic information for the sample is presented in Table 1. Among the 74 eligible participants sharing a depression symptom severity (IDS-SR) score with at least one other participant, the time series of 1 of the 74 (participant 11) violated the assumption of stationarity, yielding a final sample of 73 individuals. We observed 23 severity levels (i.e. 23 IDS-SR severity values) shared by a minimum of 2 participants, forming 23 groups of participants with MDD matched on overall symptom severity. Each group included between 2 and 6 matched participants. The most frequent IDS-SR severity category in the sample was ‘severe’ depression, consisting of 7 severity levels (IDS-SR scores of 31–35 and 37–38). All eligible participants provided enough data (i.e. a minimum of 75 completed assessments) to be included in the analysis.Reference Mansueto, Wiers, van Weert, Schouten and Epskamp26 The mean response rate to the dynamic momentary assessments was 115 of 140 (82.1%; s.d. = 16.8) completed assessments per person.
Table 1 Demographic characteristics of the sample (n = 73)
The INIT identified individual differences in symptom dynamics across 46 of the 73 participants, revealing differential symptom relationship patterns in 63.0% (95% bootstrapped CI 41.0–82.1) of MDD participants matched on overall symptom severity.
In the simulation study, we found the average false detection of individual differences with our method to be 2.2%. In the sensitivity analysis, we found both IDS-SR severity scores (logit = 0.61, s.e. = 0.40, p = 0.13) and size of severity group (logit = 0.004, s.e. = 0.43, p = 0.93) to be unrelated to individual difference detection in symptom dynamics.
The symptom relationship patterns of all 73 participants in this study are provided in Supplementary Material 2, illustrating differences and similarities in symptom interactions across the matched participants in each of the 23 severity groups.
To demonstrate these individual differences in symptom dynamics, we present the results for the first available symptom severity level in the most frequently observed severity category in our sample (i.e. an IDS-SR score of 31, for ‘severe depression’). We further describe participants matching on demographic characteristics. Five participants had an IDS-SR severity score of 31 (Fig. 1). Participants 19, 20 and 23 were female, in a relationship, aged 23–24 years and had had a high-school education. The temporal networks for these three show that although lethargy temporally precedes an increase in anhedonia for participant 19 (i.e. reflecting that greater lethargy is related to anhedonia 3 h later for this individual), this is opposite for participant 20, where anhedonia (along with depressed mood) is related to greater lethargy. Similarly, among other dynamics, the temporal pattern of restlessness is the opposite for these participants. Deviating even more strongly from the two aforementioned participants, participant 23's temporal symptom dynamics show it is an increase in irritability that is related to increases in anhedonia at the next time point. Moreover, a vicious cycle between depressed mood and anhedonia, where these symptoms increase and amplify one another, is present for some participants (e.g. participant 19), but not participant 23.
Fig. 1 Symptom dynamics of the five participants with a score of 31 on the Inventory for Depressive Symptomatology Self-Report. Solid lines show positive relationships and the dashed line shows the only negative relationship. Node descriptions: 1, Anhedonia; 2, Depressed mood; 3, Appetite change; 4, Restlessness; 5, Irritability; 6, Lethargy.
An inspection of the contemporaneous networks in Fig. 1 also reveals differential symptom relations for these participants, among which we can see that for participants 20 and 23, while these individuals experience anhedonia, this is associated with an increase in the experience of depressed mood, whereas this deleterious relationship between the core two of symptoms of depression is absent for participant 19. Similarly, in contrast to participants 19 and 20, where this relationship was absent, appetite change was associated with decreases in restlessness for participant 23.
Discussion
In this paper, we presented the concept of symptom dynamics as a clinical characterisation dimension and showed that these symptom interactions vary in the majority of individuals with depression who share the same symptom severity. These results are consistent with previous studies identifying variability in symptom relationship patterns.Reference Fisher, Reeves, Lawyer, Medaglia and Rubel31,Reference Hebbrecht, Stuivenga, Birkenhäger, Morrens, Fried and Sabbe32 We found such individual differences also to be present when individuals share the same overall symptom severity, highlighting that symptom dynamics can serve as a characteristic to differentiate between the psychopathological expression of patients with the same diagnosis and severity classification.
In the MDD participants who displayed the same depressive symptom severity in the present sample, we identified differences in how symptoms temporally precede and are associated with the elevation of other symptoms. This is consistent with research finding differential symptom interactions in individuals with schizophrenia.Reference Hays, Keshavan, Wisniewski and Torous33 These identified differences suggest that general psychoeducation on the level of the diagnosis presented to patients is limited in precisely describing the key symptoms that play a dominant role in worsening the condition for a specific individual. Capturing symptom dynamics therefore gives patients an opportunity to obtain a person-specific understanding of their experience together with their clinician.Reference Hays, Keshavan, Wisniewski and Torous33,Reference Burger, Ralph-Nearman and Levinson34
Our findings also highlight the importance of temporal monitoring of depressive symptoms from the perspective that an understanding of how symptoms fluctuate and interact over time can provide information about the formation of the disorder.Reference Fried35 Previous studies have highlighted the active role of symptom dynamics in worsening mental health conditions, where it has been found that not all symptoms of depression should be considered equal with respect to their reach and strength of impact on the aggravation of additional depressive symptomatology over time.Reference Ebrahimi, Burger, Hoffart and Johnson12,Reference Wolfe, Nakonezny, Owen, Rial, Moorehead and Kennard36 Moreover, symptoms identified to be more strongly connected with other symptoms were found to be more likely to predict the onset of future depression compared with less strongly connected symptoms.Reference Boschloo, van Borkulo, Borsboom and Schoevers37
Studies have also found differential treatment effects on specific symptoms of depressionReference Chekroud, Gueorguieva, Krumholz, Trivedi, Krystal and McCarthy38–Reference Bekhuis, Schoevers, de Boer, Peen, Dekker and Van40 and that change in one symptom during the course of psychotherapy is highly dependent on change in other symptoms.Reference O'Driscoll, Epskamp, Fried, Saunders, Cardoso and Stott41 Accordingly, the presence of differences in symptom relationship patterns across patients is likely to have implications for treatment. This is because recent studies have found indirect (i.e. secondary) effects on specific symptoms via changes first occurring in other symptoms directly influenced by treatments.Reference O'Driscoll, Buckman, Saunders, Ellard, Naqvi and Singh42,Reference Boschloo, Bekhuis, Weitz, Reijnders, DeRubeis and Dimidjian43 This reflects that an understanding of the symptom relationship patterns in a particular patient presents opportunities for personalised treatment. A recent trial investigated this by obtaining the symptom relationship patterns of different individuals with eating disorders through ecological momentary assessments, where clinicians thereafter used this information to match patients to specific evidence-based psychotherapy modules aimed at targeting symptoms with the strongest and highest number of connections to other symptoms in each patient.Reference Levinson, Williams, Christian, Hunt, Keshishian and Brosof44 This personalised therapy trial presented promising preliminary results related to the acceptability of dynamic assessment and the personalisation procedure, the feasibility with respect to low drop-out rates, and found treatment effects to be retained up to a year.Reference Levinson, Williams, Christian, Hunt, Keshishian and Brosof44
Moreover, recent studies have shown that individuals are able to describe the dynamic patterns between their experienced symptoms, and that a greater number of feedback loops (i.e. mutually reinforcing relations) between their symptoms is associated with increased symptom frequencies.Reference Klintwall, Bellander and Cervin45,Reference Frewen, Schmittmann, Bringmann and Borsboom46 This highlights the utility of symptom dynamics in clinical assessment settings and its potential for obtaining greater precision in the psychopathological characterisation and treatment monitoring of patients.
Limitations and strengths
This study has several limitations. The proportion of individual differences in symptom dynamics is likely to have been underestimated, given the method's conservativeness.Reference Hoekstra, Borsboom, McNally and Epskamp29 We chose this conservative method to present the minimum likely proportion of individual differences in symptom dynamics. As the underlying GVAR model captures linear dynamics at a lag-1 level, a general limitation concerns this model's ability to capture non-linear dynamics and possible effects operating on different timescales and lags,Reference Ebrahimi, Burger, Hoffart and Johnson12 prompting future investigations to examine these topics through complementary modelling frameworks.Reference Hebbrecht, Stuivenga, Birkenhäger, Morrens, Fried and Sabbe32,Reference Fisher47 The degree to which momentary states can be accurately related to symptoms necessitates further investigation. Four symptoms of depression (concentration difficulties, worthlessness, sleep disturbance and suicidal ideation) were unavailable in our data and thus not investigated. The participants received feedback on provided data during the EMA period. Although this was shown to not be associated with symptom changes,Reference Bastiaansen, Ornée, Meurs and Oldehinkel48 feedback during the assessment period may have influenced the findings, and thus be a limitation. As DSM-5 introduced new specifiers, the screening of patients for MDD using DSM-IV may have led to a broader subgroup of patients. Furthermore, evidence suggests that unipolar depression may in some individuals convert to bipolar depression over time.Reference Baryshnikov, Sund, Marttunen, Svirskis, Partonen and Pirkola49 The short duration of investigation (i.e. 28 days) and the exclusion of individuals with bipolar disorder, however, makes this less likely to have influenced our findings. Although this study includes a large number of observations per person, the study's sample size for between-participant comparison is small, resulting in the availability of 23 of the 84 possible severity levels on the IDS-SR, and thus reduced precision (i.e. wider confidence intervals) in the individual differences estimate in this study.
This study also has strengths. Having obtained 115 completed measurements on average per participant (i.e. 8395 observations), this is a large EMA study with respect to information on person-specific dynamics. The high measurement adherence rate (average of 82%) adds to the quality of the data.Reference Wrzus and Neubauer50 The study recruited individuals diagnosed by mental health professionals in real-world clinical settings, adding to its proximity and generalisability to the clinic.Reference Blonde, Khunti, Harris, Meizinger and Skolnik51 Further adding to its clinical proximity, this study followed recommended guidelines in assessing participants’ overall symptom severity.7 As a next step, investigating individual differences in symptom dynamics when individuals share the same item-level scores could present added opportunities for clinical characterisation and treatment personalisation.
Future directions
Findings from this study highlight areas for future research. One question concerns how differential symptom dynamics may relate to the evolution of the disorder, referring to how a disorder will progress on its own over time. Different types of symptom connections were identified across individuals with MDD in the present study. A naturally ensuing question is whether certain patterns of symptom interactions relate differently to changes in the disorder state, such as clinical profiles in which spontaneous recovery occurs and critical cases where the disorder worsens over time. This would translate into the clinical question of whether it is possible to know, for a specific patient, whether it is best to prioritise a watchful waiting approach (e.g. when the individual is optimally recovering) or a more rapid initiation of treatment to avoid worsening.
Systematic investigations of symptom relationship patterns may also provide information on whether different clusters of diagnostic profiles deriving from symptom dynamics exist within diagnostic domains.Reference Fried and Nesse3,Reference Hebbrecht, Stuivenga, Birkenhäger, Morrens, Fried and Sabbe32 This may be a promising line of research towards improving precision in the psychopathological assessment of patients.Reference Fisher47 In the present study, we also found evidence for sizeable proportions of patients (i.e. 37%) that displayed similar symptom relationship patterns. The presence of shared profiles of symptom dynamics across patients could enable the identification of different risk factors, prognoses and treatment response patterns tied to the profiles.
Moreover, individual differences in symptom relationship patterns may occur not only in major depression, but also in other disorder domains. Findings from the eating disorders domain suggest that this may be the case.Reference Levinson, Hunt, Keshishian, Brown, Vanzhula and Christian52 Accordingly, the extent to which individual differences in symptom dynamics are present across other disorder domains remains an open question. To facilitate the investigation of this topic, we openly provide all materials (code for person-specific network models, individual difference analysis and simulation setup) needed to research this topic in other substantive areas, found at the online repository of the Center for Open Science (https://osf.io/ygk84/). Finally, a next step concerns investigating whether an examination of changes in symptom relationship patterns during treatments can offer additional insight about the symptom-specific and indirect effects of psychological and psychopharmacological interventions.
Supplementary material
Supplementary material is available online at https://doi.org/10.1192/bjp.2024.19.
Data availability
Access to the data that support the findings of this study may be requested from the ZELF-i team (Dr Jojanneke Bastiaansen or Dr Harriëtte Riese). All code for the study has been uploaded to the online repository of the Center for Open Science (https://osf.io/ygk84/), where we also provide all materials (code for person-specific network models, individual difference analysis and simulation set-up) needed to research this topic in other substantive areas.
Author contributions
O.V.E.: conceptualisation, methodology, formal analysis, data curation, visualisation, writing – original draft, writing – review and editing. D.B.: methodology, supervision, writing – review and editing. R.H.A.H.: validation, methodology, writing – review and editing. S.E.: validation, methodology, writing – review and editing. E.G.O.: writing – review and editing. J.A.B.: investigation, writing – review and editing. A.C.: supervision, writing – review and editing.
Funding
O.V.E. is supported by the University of Oxford. D.B. is supported by a Dutch Research Council (NWO) Vici grant (no. VI.C.181.029). R.H.A.H. is supported by an NWO Research Talent GRANT (no. 406-18-532). J.A.B. is supported by the charitable foundation Stichting tot Steun VCVGZ (grant 239). E.G.O. and A.C. are supported by the National Institute for Health Research (NIHR) Research Professorship (grant RP-2017-08-ST2-006), the NIHR Applied Research Collaboration Oxford and Thames Valley (grant NIHR200172), the NIHR Oxford Cognitive Health Clinical Research Facility (award CRF-2016-10014), the NIHR Oxford Health Biomedical Research Centre (grant NIHR203316) and Wellcome Trust (GALENOS project). The views expressed are those of the authors and not necessarily those of the UK National Health Service, the NIHR or the Department of Health and Social Care. The Zelf-i data collection was supported by the charitable foundation Stichting tot Steun VCVGZ (grant no. 239, 2017) awarded to Dr Harriëtte Riese and Dr Jojanneke Bastiaansen.
Declaration of interest
E.G.O. has received research and consultancy fees from Angelini Pharma. A.C. reports research, educational and consultancy fees from INCiPiT (Italian Network for Paediatric Trials), CARIPLO Foundation, Lundbeck and Angelini Pharma.
Open access
eLetters
No eLetters have been published for this article.