Hostname: page-component-745bb68f8f-kw2vx Total loading time: 0 Render date: 2025-02-05T01:40:47.074Z Has data issue: false hasContentIssue false

Triangulated evidence provides no support for bidirectional causal pathways between diet/physical activity and depression/anxiety

Published online by Cambridge University Press:  04 February 2025

Kirsten J. M. van Hooijdonk*
Affiliation:
Behavioural Science Institute, Radboud University, Nijmegen, The Netherlands
Zoe E. Reed
Affiliation:
School of Psychological Science, University of Bristol, Bristol, UK MRC Integrative Epidemiology Unit, University of Bristol, Bristol, UK
Nina van den Broek
Affiliation:
Behavioural Science Institute, Radboud University, Nijmegen, The Netherlands
Madhurbain Singh
Affiliation:
Virginia Institute for Psychiatric and Behavioral Genetics, Department of Psychiatry, Virginia Commonwealth University, Richmond VA, USA Department of Human and Molecular Genetics, Virginia Commonwealth University, Richmond VA, USA
Hannah M. Sallis
Affiliation:
MRC Integrative Epidemiology Unit, University of Bristol, Bristol, UK Centre for Academic Mental Health, Population Health Sciences, Bristol Medical School, University of Bristol, UK
Nathan A. Gillespie
Affiliation:
Virginia Institute for Psychiatric and Behavioral Genetics, Department of Psychiatry, Virginia Commonwealth University, Richmond VA, USA QIMR Berghofer Medical Research Institute, Brisbane, Queensland, Australia
Marcus R. Munafò
Affiliation:
School of Psychological Science, University of Bristol, Bristol, UK MRC Integrative Epidemiology Unit, University of Bristol, Bristol, UK
Jacqueline M. Vink
Affiliation:
Behavioural Science Institute, Radboud University, Nijmegen, The Netherlands
*
Corresponding author: Kirsten van Hooijdonk; Email: kirsten.vanhooijdonk@ru.nl
Rights & Permissions [Opens in a new window]

Abstract

Background

Previous studies (various designs) present contradicting insights on the potential causal effects of diet/physical activity on depression/anxiety (and vice versa). To clarify this, we employed a triangulation framework including three methods with unique strengths/limitations/potential biases to examine possible bidirectional causal effects of diet/physical activity on depression/anxiety.

Methods

Study 1: 3-wave longitudinal study (n = 9,276 Dutch University students). Using random intercept cross-lagged panel models to study temporal associations. Study 2: cross-sectional study (n = 341 monozygotic and n = 415 dizygotic Australian adult twin pairs). Using a co-twin control design to separate genetic/environmental confounding. Study 3: Mendelian randomization utilizing data (European ancestry) from genome-wide association studies (n varied between 17,310 and 447,401). Using genetic variants as instrumental variables to study causal inference.

Results

Study 1 did not provide support for bidirectional causal effects between diet/physical activity and symptoms of depression/anxiety. Study 2 did provide support for causal effects between fruit/vegetable intake and symptoms of depression/anxiety, mixed support for causal effects between physical activity and symptoms of depression/anxiety, and no support for causal effects between sweet/savoury snack intake and symptoms of depression/anxiety. Study 3 provides support for a causal effect from increased fruit intake to the increased likelihood of anxiety. No support was found for other pathways. Adjusting the analyses including diet for physical activity (and vice versa) did not change the conclusions in any study.

Conclusions

Triangulating the evidence across the studies did not provide compelling support for causal effects of diet/physical activity on depression/anxiety or vice versa.

Type
Original Article
Creative Commons
Creative Common License - CCCreative Common License - BYCreative Common License - NCCreative Common License - ND
This is an Open Access article, distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivatives licence (http://creativecommons.org/licenses/by-nc-nd/4.0), which permits non-commercial re-use, distribution, and reproduction in any medium, provided that no alterations are made and the original article is properly cited. The written permission of Cambridge University Press must be obtained prior to any commercial use and/or adaptation of the article.
Copyright
© The Author(s), 2025. Published by Cambridge University Press

Introduction

Mental health disorders (particularly depression and anxiety), along with unhealthy lifestyle behaviours (such as poor diet and physical inactivity) are pressing challenges in today’s society (GBD 2019 Diseases and Injuries Collaborators, 2020; NCD Countdown 2030 collaborators, 2018; World Health Organization, 2022). Between 1990 and 2019, the estimated past-year global prevalence of depression and anxiety increased from 171 to 280 million and from 195 to 301 million, respectively (GBD Mental Disorders Collaborators, 2022). Additionally, the global prevalence of insufficient physical activityFootnote 1 has increased from 23% in 2000 to 26% in 2010 and 31% in 2022 (Strain et al., Reference Strain, Flaxman, Guthold, Semenova, Cowan, Riley and Stevens2024) and over several decades dietary quality has worsened globally (e.g., increased uptake of processed foods, away-from-home meals and sugar-sweetened beverages) (Popkin, Adair, & Ng, Reference Popkin, Adair and Ng2012). Accordingly, understanding factors contributing to the development and recovery of these conditions/behaviours is important.

Previous studies have suggested that mental health and lifestyle behaviours might be interconnected and possibly influence each other (although results are mixed). Several studies have found support for bidirectional or unidirectional causal pathways between diet or physical activity and depression or anxiety (Choi et al., Reference Choi, Chen, Stein, Klimentidis, Wang, Koenen and Smoller2019; Iob et al., Reference Iob, Pingault, Munafò, Stubbs, Gilthorpe, Maihofer and Danese2023; Liu, Yan, Li, & Zhang, Reference Liu, Yan, Li and Zhang2016; Mammen & Faulkner, Reference Mammen and Faulkner2013; McDowell, Dishman, Gordon, & Herring, Reference McDowell, Dishman, Gordon and Herring2019; Molendijk, Molero, Sánchez-Pedreño, Van der Does, & Martínez-González, Reference Molendijk, Molero, Sánchez-Pedreño, Van der Does and Martínez-González2018; Pasman et al., Reference Pasman, Bergstedt, Harder, Gong, Xiong, Hägg and Lu2024; Pearce et al., Reference Pearce, Garcia, Abbas, Strain, Schuch, Golubic and Woodcock2022; Rebar et al., Reference Rebar, Stanton, Geard, Short, Duncan and Vandelanotte2015; Roshanaei-Moghaddam, Katon, & Russo, Reference Roshanaei-Moghaddam, Katon and Russo2009; Saghafian et al., Reference Saghafian, Malmir, Saneei, Milajerdi, Larijani and Esmaillzadeh2018; Schuch et al., Reference Schuch, Stubbs, Meyer, Heissel, Zech, Vancampfort and Hiles2019; Schuch et al., Reference Schuch, Vancampfort, Firth, Rosenbaum, Ward, Silva and Stubbs2018; Tuck, Farrow, & Thomas, Reference Tuck, Farrow and Thomas2019; Wanjau et al., Reference Wanjau, Möller, Haigh, Milat, Hayek, Lucas and Veerman2023; Yan, Xu, Li, & Liu, Reference Yan, Xu, Li and Liu2023), while others have not (Appleton, Boxall, Adenuga-Ajayi, & Seyar, Reference Appleton, Boxall, Adenuga-Ajayi and Seyar2024; T. T. Chen, Chen, Fang, Cheng, & Lin, Reference Chen, Chen, Fang, Cheng and Lin2022; Choi et al., Reference Choi, Chen, Stein, Klimentidis, Wang, Koenen and Smoller2019; De Moor, Boomsma, Stubbe, Willemsen, & de Geus, Reference De Moor, Boomsma, Stubbe, Willemsen and de Geus2008; Iob et al., Reference Iob, Pingault, Munafò, Stubbs, Gilthorpe, Maihofer and Danese2023; Moreno-Peral et al., Reference Moreno-Peral, Pino-Postigo, Conejo-Cerón, Bellón, Rodríguez-Martín, Martínez-Vizcaíno and Bellón2022; Pasman et al., Reference Pasman, Bergstedt, Harder, Gong, Xiong, Hägg and Lu2024; Yan et al., Reference Yan, Xu, Li and Liu2023). It must be noted that these studies focussed on either diet or physical activity in relation to mental health. None of these studies took into account that diet might confound the relationship between physical activity and mental health, and vice versa that physical activity might confound the relationship between diet and mental health. Consequently, this limits the ability to provide more robust estimates on direct effects between diet/physical activity and depression/anxiety. Additionally, not all pathways have been thoroughly investigated. For example, studies involving diet have mostly focused on the intake of healthy foods (e.g., fruits and vegetables) but neglected studying the intake of unhealthy foods (e.g., sweet or savoury snacks).

Various research designs have been employed in previous studies (e.g., randomized controlled trials, prospective cohort studies, co-twin control designs (CTCDs)Footnote 2 and Mendelian randomization (MR)Footnote 3). Given that no single-methodology approach can provide definite evidence for causal pathways on its own (Hammerton & Munafò, Reference Hammerton and Munafò2021), the ultimate approach to gain more robust insights on this complex causal question is to apply “triangulation” (Munafò & Davey Smith, Reference Munafò and Davey Smith2018; Patton, Reference Patton1999). This refers to using various study designs, each with unique strengths and potential weaknesses, to address the same research question. In a triangulation framework, the results across the included methods are compared to help overcome the biases arising from the use of one single method, to focus on overarching patterns, and to help achieve empirical consensus (Munafò & Davey Smith, Reference Munafò and Davey Smith2018; Patton, Reference Patton1999).

Consequently, in this study, we aimed to unravel the potential causal effects of diet (intake of sweet snacks, savoury snacks, fruits and vegetables)/physical activity on depression/anxiety and vice versa. This was done in unadjusted models (including one lifestyle and one mental health measure) and adjusted models (including one diet, one physical activity and one mental health measure; to explore possible confounding by diet/physical activity and obtain more robust direct effect estimates of the exposure of interest on the outcome). We employed a triangulation framework including three distinct lines of evidence, to utilise the different strengths and acknowledge potential weaknesses and sources of bias of each method. We focused on: (1) exploring temporal associations using random intercept cross-lagged panel models (RI-CLPMs), (2) separating genetic and environmental confounding using CTCD mixed-effects models, and (3) utilising genetic variants as instrumental variables with MR. Study 1 used longitudinal data from the Healthy Student Life (HSL) project (n = 9,276 Dutch University students) (van Hooijdonk, Simons, van Noorden, Geurts, & Vink, Reference van Hooijdonk, Simons, van Noorden, Geurts and Vink2023). Study 2 used cross-sectional data from the Brisbane Longitudinal Twin Study (BLTS; n = 341 monozygotic and n = 415 dizygotic Australian adult twin pairs) (Mitchell et al., Reference Mitchell, Campos, Rentería, Parker, Sullivan, McAloney and Scott2019). Study 3 used summary statistics from genome-wide association studies (GWASs; n varied between 17,310 and 447,401).

Methods

Triangulation framework

In the current study, we employed a triangulation framework to examine causal pathways between diet/physical activity and depression/anxiety. This involves answering the same causal question by integrating results from different statistical methods which have different strengths, limitations and (preferably) unrelated sources of potential bias (Lawlor, Tilling, & Davey Smith, Reference Lawlor, Tilling and Davey Smith2016; Munafò, Higgins, & Davey Smith, Reference Munafò, Higgins and Davey Smith2021). In the current triangulation framework, we used three complementary methods (referred to as Studies 1, 2 and 3), that provide insight on potential causal effects form different perspectives (as illustrated in Figure 1):

  • RI-CLPMs (Study 1). We estimated individual-level temporal associations which show how changes in one variable might predict changes in another variable over time.

  • CTCD (Study 2). We separated genetic and environmental confounding by examining and comparing the association between an exposure and outcome at population-level and within twin pairs (who have overlap in genes and shared early environment).

  • MR (Study 3). We leveraged genetic variants as instrumental variables to estimate the potential population-level causal effect of an exposure on an outcome.

Figure 1. Concept of triangulation.

Note. Illustration of how our three complementary methods help to strengthen causal inference.

When the findings in a triangulation framework converge, this provides greater confidence in the conclusions and potential causal effect, since it is unlikely that the same bias effects all three methods. Divergent findings will help to identify sources of bias which require further investigation (Munafò et al., Reference Munafò, Higgins and Davey Smith2021). Table 1 provides a detailed comparison of how the three methods in the triangulation framework complement each other in addressing different sources of potential bias. To investigate converge/divergence in our triangulation framework, we evaluated the magnitude, direction and margin of error of all effect estimates and compared these across the three methods.

Table 1. Overview of the studies included in the triangulation framework

Note. All studies use observational studies to some extend which has known potential biases like reporting/recall bias, measurement errors, selection bias, and social-desirability bias. DZ = dizygotic. MZ = monozygotic.

All methods in the triangulation framework were preregistered in the Open Science Framework: https://osf.io/e4d5b/ (Supplementary Material A describes deviations). Analyses were performed in R version 4.3.1 (R Core Team, 2023). Reporting followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) and STROBE using mendelian randomisation (STROBE-MR) guidelines (Skrivankova et al., Reference Skrivankova, Richmond, Woolf, Davies, Swanson, VanderWeele and Langenberg2021; Von Elm et al., Reference Von Elm, Altman, Egger, Pocock, Gøtzsche and Vandenbroucke2007). Details of the methods per study have been provided below.

Study 1: random intercept cross-lagged panel models

The RI-CLPM is a structural equation modelling approach used to model within-person directional effects of one variable on another variable over time, and vice versa (Hamaker, Kuiper, & Grasman, Reference Hamaker, Kuiper and Grasman2015). In the RI-CLPM, observed scores of all constructs are decomposed into (1) grand means (time-varying or fixed means over all individuals per occasion; the paths from the triangles (constants, with value fixed at 1) to the squares (observed scores) in Figure 2); (2) stable between-person variance (random intercepts, an individual’s time-invariant deviation from the grand means, see “Between” in Figure 2) and (3) fluctuating within-person variance (differences between an individual’s observed measurements and their expected score based on the grand means and random intercept, see “Within” in Figure 2) (Mulder & Hamaker, Reference Mulder and Hamaker2021). The within-person cross-lagged effects (red arrows Figure 2) illustrate how one variable potentially influences another variable over time within a person and are of interest when studying possible causal effects.

Figure 2. Example of a random intercept cross-lagged panel model (RI-CLPM) assessing the bidirectional pathways between variable A and variable B, including between-person and within-person components at three survey waves.

Note. The squares indicate observed variables, while the circles represent latent (unobserved) variables. U and V represent the residual variance. The triangles represent constants for the mean structure. To improve readability, no covariates are presented. Based on Hamaker et al. (Reference Hamaker, Kuiper and Grasman2015) and Mulder and Hamaker (Reference Mulder and Hamaker2021).

Procedure and participants

Data from the HSL project were used (van Hooijdonk et al., Reference van Hooijdonk, Simons, van Noorden, Geurts and Vink2023). This questionnaire study follows Dutch Radboud University students and assesses their mental health and lifestyle. In this study, we utilized data from Wave 1 (October–November 2021, collected during the COVID-19 pandemic; Ninvited = 25,035), Wave 2 (May–July 2022; Ninvited = 23,994) and Wave 3 (May–July 2023; Ninvited = 23,425). The analytical sample included 9,276 students (6,004 participated in one questionnaire; 2,208 in two; 1,064 in all). The study was independently reviewed by Radboud University’s Social Sciences Ethics Committee, and there is no formal objection to this study (ECSW-2021-086). Supplementary Material B includes additional information.

Measures (self-reported)

For the main analyses, we used continuous data on sweet snack intake (van den Broek, Larsen, Verhagen, Burk, & Vink, Reference van den Broek, Larsen, Verhagen, Burk and Vink2020), savoury snack intake (van den Broek et al., Reference van den Broek, Larsen, Verhagen, Burk and Vink2020), fruit intake (van den Broek et al., Reference van den Broek, Larsen, Verhagen, Burk and Vink2020), vegetable intake (van den Broek et al., Reference van den Broek, Larsen, Verhagen, Burk and Vink2020), physical activity (IPAQ Research Committee, 2005), depressive symptoms (Van de Velde, Levecque, & Bracke, Reference Van de Velde, Levecque and Bracke2009), anxiety symptoms (Donker, van Straten, Marks, & Cuijpers, Reference Donker, van Straten, Marks and Cuijpers2011) and age. Additionally, information on gender (male/female/other) was used. For descriptive purposes, we also used data on living situation, relationship status, parental educational type, body mass index (BMI; kg/m2), and overall perceived physical and mental health. Supplementary Material C includes additional information.

Statistical analyses

Descriptive statistics were used to explore all measures. Additionally, intraclass correlation (ICCs) were calculated using the lme4 package (Bates et al., Reference Bates, Maechler, Bolker, Walker, Christensen, Singmann and Grothendieck2009b) to gain insight in the proportion of variance explained between persons (ICC) and within persons (1-ICC) across the waves (Aarts, Verhage, Veenvliet, Dolan, & Van Der Sluis, Reference Aarts, Verhage, Veenvliet, Dolan and Van Der Sluis2014). Next, RI-CLPMs were applied to the data using the lavaan package (Rosseel, Reference Rosseel2012). We ran ten RI-CLPMs including one diet/physical activity and one mental health measure (Supplementary Table ST1). In the results, we refer to these analyses as unadjusted models. Moreover, we ran eight RI-CLPMs where the unadjusted temporal associations were adjusted for either diet (when unadjusted model included physical activity) or physical activity (when unadjusted model included diet), taking into account potential confounding by diet/physical activity (Supplementary Table ST2). Each model included one diet measure, physical activity, and one mental health measure. In the results, we refer to these models as adjusted models.

In all models, five parameters were estimated: (1) between-person covariance random intercepts, (2) within-person stability (or autoregressive) effects, (3) within-person cross-lagged effects, (4) within-person concurrent covariance, and (5) time-invariant covariate associations of gender (male/female) and Wave 1 age (extension 1 in Mulder and Hamaker (Reference Mulder and Hamaker2021)). When participants joined after Wave 1, Wave 1 age was estimated and gender reported at first participation was used. Full information maximum likelihood (FIML) was used to handle missing data and the robust estimator maximum likelihood with robust standard errors (MLR) was used to handle non-normality (Enders, Reference Enders2001; Graham, Reference Graham2009). Per model, the fit was assessed and considered acceptable when: (1) the scaled chi-square test was non-significant (p > .05), (2) root-mean-square error of approximation < .06, (3) standardized root mean square residual < .08, and (4) comparative fit index > .90 (Hu & Bentler, Reference Hu and Bentler1999; Kline, Reference Kline2023). To investigate potential support for causal effects, the magnitude, direction and margin of error of the within-person cross-lagged effects were studied.

Last, sensitivity analyses to examine the robustness of the findings were performed using complete cases for all RI-CLPMs (n varied between 776 and 841). These were added as a large proportion of HSL participants only joined one wave and these might be different from participants who participated every time (e.g., healthier lifestyle/better mental health).

Study 2: co-twin control design

The CTCD is applied on the assumption that monozygotic (MZ) twins share 100% of their genetic material and 100% of their shared (early) environment (e.g., prenatal exposures, childhood environment and other family influences), and that dizygotic (DZ) twins share, on average, 50% of their genetic material and 100% of their shared (early) environment (Vitaro, Brendgen, & Arseneault, Reference Vitaro, Brendgen and Arseneault2009). In population-level observational studies, the association between exposure and outcome can be confounded by multiple factors. This limits the possibility of drawing valid conclusions on causal inference. In contrast, equivalent association estimates in discordantFootnote 4 MZ and same-sex DZ twin pairs control for age, sex and varying degrees of familial confounding, thus helping to differentiate between causation and confounding (Eaton et al., Reference Eaton, Krueger, South, Gruenewald, Seeman and Roberts2012; Mosteller & Boruch, Reference Mosteller and Boruch2004; Shadish, Cook, & Campbell, Reference Shadish, Cook and Campbell2002).

The CTCD consists of three sub-models, testing the association between an exposure and outcome: (1) at the population-level (using all data without considering zygosity/twin pair discordance), (2) within DZ same-sex twin pairs discordant for the exposure, and (3) within MZ twin pairs discordant for the exposure (Gonggrijp et al., Reference Gonggrijp, van de Weijer, Bijleveld, van Dongen and Boomsma2023). The mean difference in the magnitude of the association between exposed and unexposed individuals in these sub-models can be compared to infer the contributions of unmeasured genetic and shared (early) environmental confounders and, thus, indirectly infer possible causation. Figure 3 further illustrates this. Scenario 1 points to possible causation, as the mean difference in the magnitude of the association between exposed and unexposed individuals is equal in all three sub-model (regardless of overlap in genes/shared (early) environment within the DZ and MZ twin pairs). Scenario 2 points to genetic confounding, as the mean difference in the magnitude of association between exposed and unexposed individuals is intermediate within DZ twin pairs (50% overlap in genes) and zero within MZ twins (100% overlap in genes). Scenario 3 points to environmental confounding, as the mean difference in the magnitude of association between exposed and unexposed individuals is zero within DZ and MZ twin pairs (both 100% overlap in shared (early) environment). Scenario 4 points to a combination of genetic and environmental confounding, given that the mean difference between exposed and unexposed individuals is not zero in all three sub-models and reduces depending on the degree of genetic overlap.

Figure 3. Hypothetical outcome scenarios of a co-twin control design (CTCD).

Note. Population-level (PL) = the mean difference in the magnitude of association between exposed and unexposed individuals; DZ = the mean difference in the magnitude of association within dizygotic same-sex twin pairs discordant for the exposure; MZ = the mean difference in the magnitude of association within monozygotic twin pairs discordant for the exposure. Scenario 1: the magnitude of the association is similar across all groups (regardless of shared genes or (early) environment), inferring possible causality. Scenario 2: the association between exposure and outcome is entirely inferred by genetic confounding, as the association within same-sex DZ twins (who share 50% of their genetic material) is intermediate, and the association within MZ twins (who share 100% of their genetic material) is zero. Scenario 3: the association between exposure and outcome is entirely inferred by environmental confounding, as the association within same-sex DZ and MZ twins (who share 100% of their (early) environment) is zero. Scenario 4: the association between exposure and outcome is partly inferred by both genetic and environmental confounding. Based on Gonggrijp et al. (Reference Gonggrijp, van de Weijer, Bijleveld, van Dongen and Boomsma2023).

Procedure and participants

We used data from the BLTS, 25UP project (Mitchell et al., Reference Mitchell, Campos, Rentería, Parker, Sullivan, McAloney and Scott2019). The 25UP project utilized questionnaires to collect information on mental health conditions, general/physical health, psychosocial items, and general demographic information. Data were collected between 2016 and 2018, and participants included twins and their non-twin siblings from South-East Queensland (Australia). Recruitment details have been reported by Mitchell et al. (Reference Mitchell, Campos, Rentería, Parker, Sullivan, McAloney and Scott2019). For this study, data from complete twin pairs were used (ntotal individuals = 1,512, including 341 MZ, 217 DZ same-sex and 198 DZ opposite-sex twin pairs).

Measures (self-reported)

For the main analyses, we used data on sweet/savoury snack intake (dichotomized) (Mitchell et al., Reference Mitchell, Campos, Rentería, Parker, Sullivan, McAloney and Scott2019), fruit/vegetable intake (dichotomized) (Mitchell et al., Reference Mitchell, Campos, Rentería, Parker, Sullivan, McAloney and Scott2019), physical activity (continuous and dichotomized) (IPAQ Research Committee, 2005) and symptoms of depression/anxiety (continuous and dichotomized) (Andrews & Slade, Reference Andrews and Slade2001). Additionally, age and gender (male/female) were used. For descriptive purposes, we also used information on living situation, study/work status, highest level of education, current relationship status, BMI (kg/m2), and overall perceived physical and mental health. Supplementary Material D includes additional information.

Selection of twins per sub-model

For sub-model 1 (population-level), we selected all available data without considering zygosity or twin pair con/discordance on the exposure of interest (dichotomized sweet/savoury snack intake, fruit/vegetable intake, physical activity and symptoms of depression/anxiety). For sub-model 2, we selected data from DZ same-sex twin pairs who were discordant on the exposure of interest (within one pair: one twin exposed and one twin unexposed to the exposure of interest). For sub-model 3, we selected data from MZ twin pairs who were discordant on the exposure of interest.

Statistical analyses

Descriptive statistics were used to explore all measures. Next, the CTCD was utilized by running linear (lmer) and logistic (glmer) mixed-effects models using the lme4 package (Bates et al., Reference Bates, Maechler, Bolker, Walker, Christensen, Singmann and Grothendieck2009a). We fitted six mixed-effects models, each including the three sub-models (mentioned above), including one diet/physical activity measure and the combined measure for symptoms of depression/anxiety (Supplementary Table ST3). In the results, we refer to these models as unadjusted models. Moreover, we fitted eight mixed-effects models to adjust the unadjusted models for either diet (when unadjusted models included physical activity) or physical activity (when unadjusted models included diet), taking into account potential confounding by diet/physical activity (Supplementary Table ST4). In case this variable was continuous, the variable was standardized. Again, each model consisted of three sub-models and each sub-model included one diet measure, physical activity, and symptoms of depression/anxiety. In the results, we refer to these models as adjusted models. In all models, family clustering was taken into account (by adding family as a random effect). Sub-models 1 were corrected for age and gender, Supplementary Material E provides additional information on model settings, calculations of p-values and confidence intervals. To investigate potential causal effects, the magnitude, direction and error margin of the association between exposure and outcome across the three sub-models in each model were studied.

Sensitivity analyses were performed for all models to assess the reliability of the chosen discordance thresholds (Supplementary Material D). In these analyses, slightly different discordance criteria were used (given that no set cut-offs were available).

Study 3: mendelian randomization

In two-sample univariable Mendelian randomization (UVMR), genetic variants (single nucleotide polymorphisms; SNPs) robustly associated with an exposure are leveraged as instrumental variables to estimate the potential population-level causal effect of an exposure on an outcome (Davey Smith & Ebrahim, Reference Davey Smith and Ebrahim2003), see Figure 4A. The basis of MR is built on Mendel’s laws of random segregationFootnote 5 and independent assortmentFootnote 6 (Evans & Davey Smith, Reference Evans and Davey Smith2015). MR is less prone to unmeasured confounding or reverse causation than observational studies as genetic variants are randomly assigned during conception and typically unaffected by environmental/lifestyle factors later in life (Carnegie et al., Reference Carnegie, Zheng, Sallis, Jones, Wade, Evans and Martin2020; Gupta, Walia, & Sachdeva, Reference Gupta, Walia and Sachdeva2017). To gain valid estimates, several core assumptions need to be satisfied: (1) the genetic instrumental variable is robustly associated with the exposure (relevance assumption), (2) there is no confounding between the genetic instrumental variable and the outcome (independence assumption), and (3) the instrumental variable is not associated with the outcome other than via the exposure (exclusion restriction) (Lawlor, Harbord, Sterne, Timpson, & Davey Smith, Reference Lawlor, Harbord, Sterne, Timpson and Davey Smith2008). Multiple MR methods exist that evaluate these assumptions. The consistency of effect estimates across these methods strengthens the evidence.

Figure 4. Directed acyclic graphs for (A) Univariable mendelian randomization and (B) Multivariable mendelian randomization.

Multivariable Mendelian randomization (MVMR) is an extension of UVMR which facilitates using multiple exposures (e.g., diet and physical activity) to estimate the direct independent effects of multiple exposures on an outcome (Burgess & Thompson, Reference Burgess and Thompson2015), conditional on the effect of the other exposure on the outcome (Figure 4B).

Data sources

Publicly available summary statistics from GWASs (European ancestry) were used to select SNPs associated with sweet snack intake (Elsworth et al., Reference Elsworth, Lyon, Alexander, Liu, Matthews, Hallett and Hemani2020), savoury snack intake (Elsworth et al., Reference Elsworth, Lyon, Alexander, Liu, Matthews, Hallett and Hemani2020), fruit intake (Cole, Florez, & Hirschhorn, Reference Cole, Florez and Hirschhorn2020), vegetable intake (Elsworth et al., Reference Elsworth, Lyon, Alexander, Liu, Matthews, Hallett and Hemani2020), physical activity (Klimentidis et al., Reference Klimentidis, Raichlen, Bea, Garcia, Wineinger, Mandarino and Going2018), depressionFootnote 7 (Howard et al., Reference Howard, Adams, Clarke, Hafferty, Gibson, Shirali and McIntosh2019), and anxietyFootnote 8 (Otowa et al., Reference Otowa, Hek, Lee, Byrne, Mirza, Nivard and Hettema2016). See Supplementary Table ST5 and Supplementary Material F for information per phenotype.

Statistical analyses

Supplementary Material G explains the SNPs selection and Supplementary Table ST6 (UVMR) and ST7 (MVMR) contain all selected SNPs. First, twenty UVMR analyses were run using the TwoSampleMR package (Hemani, Zheng, et al., Reference Hemani, Zheng, Elsworth, Wade, Baird, Haberland and Haycock2018). Each model included one diet/physical activity and one mental health measure (Supplementary Table ST8). The inverse-variance weighted (IVW) method was used as the main method (Burgess, Butterworth, & Thompson, Reference Burgess, Butterworth and Thompson2013), providing valid estimates when horizontal pleiotropy is balanced or absent (Hemani, Bowden, & Davey Smith, Reference Hemani, Bowden and Davey Smith2018). Next, several sensitivity analyses (mostly available in the TwoSampleMR package) were performed (Supplementary Material H): MR-Egger (Bowden, Davey Smith, & Burgess, Reference Bowden, Davey Smith and Burgess2015), weighted median (Bowden, Davey Smith, Haycock, & Burgess, Reference Bowden, Davey Smith, Haycock and Burgess2016), simple mode (Hartwig, Davey Smith, & Bowden, Reference Hartwig, Davey Smith and Bowden2017), weighted mode (Hartwig et al., Reference Hartwig, Davey Smith and Bowden2017), MR Pleiotropy RESidual Sum and Outlier (MR-PRESSO) (Verbanck, Chen, Neale, & Do, Reference Verbanck, Chen, Neale and Do2018), MR using the robust adjusted profile score (MR-RAPS) (Zhao, Wang, Hemani, Bowden, & Small, Reference Zhao, Wang, Hemani, Bowden and Small2018)) and, MRlap (Mounier & Kutalik, Reference Mounier and Kutalik2023). The mean F-statistic was calculated to evaluate instrument strength (F < 10 may indicate a weak instrument (Burgess & Thompson, Reference Burgess and Thompson2011)), IVW and MR-Egger heterogeneity tests and MR-Egger pleiotropy tests were performed to assess horizontal and directional pleiotropy, respectively (Hemani, Zheng, et al., Reference Hemani, Zheng, Elsworth, Wade, Baird, Haberland and Haycock2018).

Moreover, twenty-four MVMR analyses were run using the TwoSampleMR package (Hemani, Zheng, et al., Reference Hemani, Zheng, Elsworth, Wade, Baird, Haberland and Haycock2018). This extension of the UVMR allows for adjustment of potential confounders, in this case adjusting the UVMR analyses including diet for physical activity and adjusting the UVMR analyses including physical activity for diet. Each model included one diet measure, physical activity and one mental health measure (Supplementary Table ST9). The MVMR-IVW method was used as the main method (Sanderson, Davey Smith, Windmeijer, & Bowden, Reference Sanderson, Davey Smith, Windmeijer and Bowden2018) and several sensitivity analyses were performed (Supplementary Material H): MVMR-EggerFootnote 9 (Rees, Wood, & Burgess, Reference Rees, Wood and Burgess2017) and MVMR-PRESSO (Verbanck et al., Reference Verbanck, Chen, Neale and Do2018). The F-statistic was calculated to evaluate instrument strength and pleiotropy tests were performed using the MVMR package to assess heterogeneity (Sanderson, Spiller, & Bowden, Reference Sanderson, Spiller and Bowden2021). For both UVMR and MVMR, the magnitude, direction and confidence intervals of the odds ratios (ORs) were studied to evaluate potential causal effects.

Results

Triangulated evidence

Table 2 provides a summary of all effect estimates and the magnitude, direction and margin of error of the effect estimates for Studies 1, 2 and 3 for the analyses performed with one lifestyle behaviour (diet/physical activity) and one mental health measure (depression/anxiety); unadjusted models. The findings of most assessed pathways across the three methods did not provide strong support for causal pathways between diet/physical activity and depression/anxiety (or vice versa). This is reflected by the magnitude of the effect estimates which were considerably small, pointing to weak or non-existing causal pathways. Converge for all three methods was observed for the models testing bidirectional causal effects between sweet/savoury snack intake and depression/anxiety, and the models testing causal effects of physical activity on depression/anxiety. The convergence across the three methods provides greater confidence in the conclusions and absence of causal pathways.

Table 2. Triangulated evidence of all unadjusted/univariable analyses

Note. Findings highlighted in bold present pathways were support was found for that potential causal path. b = unstandardized estimate, SE = standard error, b = regression coefficient for the exposure, 95% CI = 95% confidence interval, OR = odds ratio, SNPs = single nucleotide polymorphisms, W1 = Wave 1, W2 = Wave 2, W3 = Wave 3, PL = population-level, MZ = monozygotic twin pairs, DZ = same-sex dizygotic twin pairs.

a Only the cross-lagged within-person effects have been presented. Full results can be found in Supplementary Tables ST12–ST29.

b Only the regression coefficients of the exposure of interest have been presented. Full results can be found in Supplementary Tables ST32–ST45. In Study 2, combined measures of sweet and savoury snack intake, fruit and vegetable intake, and depression and anxiety were used.

c Only the results from the inverse-variance weighted (IVW) method have been presented. Full results can be found in Supplementary Tables ST47–48.

Some divergence was observed in the models assessing bidirectional causal effects between fruit/vegetable intake and depression/anxiety. Study 1 did not provide support (i.e., the (un)standardized betas are close to zero), Study 2 did provide support (i.e., the regression coefficients across the three sub-models are roughly equal), Study 3 did not provide support in most assessed pathways (i.e., the ORs are close to one), except for the pathway fruit intake on anxiety (OR = 1.99, 95% CI = 1.19–3.34). Additionally, some divergence was observed in the models assessing causal effects of depression/anxiety on physical activity. Studies 1 and 3 did not provide support (i.e., in Study 1 the (un)standardized betas are close to zero and in Study 3 the OR are close to one), while Study 2 did provide support (i.e., the regression coefficients of the MZ/DZ sub-models are larger than of the population-level submodel). A further reflection on how potential biases might have impacted these findings and contributed to the divergence is required, and has been provided in the discussion. Below, we discuss the findings from Studies 1, 2 and 3 in more detail.

Supplementary Table ST10 provides a summary of effect estimates and the magnitude, direction and margin of error of the effect estimates from Studies 1, 2 and 3 for the analyses were the causal pathways were adjusted for a potential confounder (the other lifestyle behaviour; diet or physical activity); adjusted models. Conclusions were in line with the unadjusted models.

Study 1: random intercept cross-lagged panel models

Sample description and intraclass correlations (ICC)

At Wave 1, the mean age of the sample was 23 years (SD = 4.2), 72% were female, 86% were not living alone, and 49% were single. The mean BMI score was 22.5 kg/m2 (SD = 3.5) and overall physical and mental health was perceived as very good or excellent by 32% and 25%, respectively. Descriptives for Wave 2 and 3 were similar (Supplementary Table ST11).

The ICC for depressive symptoms was 0.63, which shows that 63% of the variance in the three measurements of depressive symptoms can be explained by differences between persons (i.e., stable trait level) and the remaining 37% of the variance can be explained by fluctuations within persons (i.e., change over time). For anxiety symptoms the ICC was 0.53, for sweet snack intake the ICC was 0.64, the ICC for savoury snack intake was 0.64, the ICC for fruit intake was 0.73, the ICC for vegetable intake was 0.66, and the ICC for physical activity was 0.52.

Random intercept cross-lagged panel models

Table 3 shows all within-person cross-lagged effects. Supplementary Tables ST12-ST29 provide all estimates. The fit indices per model were acceptable (Supplementary Table ST30).

Table 3. Results within-person cross-lagged effects

Note. Other model estimates (between-person covariance random intercepts, within-person stability (or autoregressive) effects, within-person concurrent covariance, and time-invariant covariate associations of gender and age) per model are presented in Supplementary Tables S11-S28. W1 = Wave 1, W2 = Wave 2, W3 = Wave 3, b = unstandardized estimate, β = standardized estimate, SE = standard error.

a Adjusted models were corrected for physical activity.

b Adjusted models were corrected for all diet measures and these cross-lagged effects have been presented in Supplementary Tables S22–S29.

The within-person cross-lagged effects (unadjusted models) did not provide support for causal effects of diet or physical activity on depressive or anxiety symptoms, given that the (un)standardized estimates were close to zero with relatively large error margins. Similarly, the within-person cross-lagged effects (unadjusted models) did not provide support for causal effects of depressive or anxiety symptoms on diet or physical activity. This means that participants’ changes in depressive/anxiety symptoms, relative to their own expected scores, were not predicted by participants’ diet/physical activity at the previous wave (or vice versa). This pattern was consistent with the adjusted models and sensitivity analyses.

Study 2: co-twin control design

Sample description

The mean age of the total sample was 30 years (SD = 4.2), 62% were male, 53% lived with a partner (and/or children), 66% worked full-time, and 70% had a relationship (Supplementary Table ST31). The mean BMI score was 24.7 kg/m2 (SD = 4.8) and overall physical and mental health was perceived as (very) good by 76% and 72%, respectively. A total of 64% of the participants were “exposed” to sweet/savoury snack intake, 47% to insufficient fruit/vegetable intake, 19% to physical inactivity and 29% to symptoms of depression/anxiety.

Mixed-effects models

All results are reported in Supplementary Tables ST32-ST45 and patterns in regression coefficients are shown in Figure 5 and compared to the scenarios in Figure 3. Information on the DZ/MZ discordance rate per model has been provided in Supplementary Table ST46.

Figure 5. Results mixed-effects models - patterns in regression coefficients at population-level (PL), within dizygotic (DZ) same-sex twins and within monozygotic (MZ) twins. Asterisks represent associations between exposure and outcome per subgroup where p < .05.

Our results do not support causal pathways between sweet/savoury snack intake and symptoms of depression/anxiety (Figure 5A/5B). The association between exposure and outcome was present at the population-level (unadjusted models) but not within DZ/MZ twin pairs, and the regression coefficients within DZ/MZ twins were roughly equal with overlapping error bars. This suggests that Scenario 3 (confounding by shared (early) environment) most closely represents the observed patterns in regression coefficients (unadjusted/adjusted models). Although this pattern was observed in some sensitivity analyses, these findings were not fully consistent with the sensitivity analyses (Supplementary Tables ST32-ST33ST38-ST39).

Our results support causal pathways between fruit/vegetable intake and symptoms of depression/anxiety (Figure 5C/5D). The patterns in regression coefficients (unadjusted/adjusted models) most closely represent Scenario 1, as the regression coefficients across the three groups were roughly equal. In most sub-models, the association between exposure and outcome was present (except for one within MZ and all within DZ twins sub-models). These findings were consistent with the sensitivity analyses (Supplementary Tables ST34-ST35ST40-ST41).

Our results provided mixed support for causal pathway between symptoms of depression/anxiety and physical activity. In the models where symptoms of depression/anxiety were the exposure and physical activity was the outcome (Figure 5F), the pattern in regression coefficients (unadjusted/adjusted models) between the three groups most closely represents Scenario 1, supporting causal effects. In most sub-models, the association between exposure and outcome was present (except the population-level model adjusted for fruit/vegetable intake and all within DZ models). These findings were consistent with the sensitivity analyses (Supplementary Tables ST37ST44-ST45). In the models where physical activity was the exposure and symptoms of depression/anxiety were the outcome (Figure 5E), the pattern in regression coefficients (unadjusted/adjusted models) between the three groups most closely represents Scenario 3 (confounding by shared (early) environment). The pattern observed in the sensitivity analyses pointed more towards genetic confounding (Scenario 2) (Supplementary Tables ST36, ST42-ST43).

Study 3: mendelian randomization

Univariable and multivariable mendelian randomization

Supplementary Tables ST47-ST48 provide all UVMR and MVMR results. Figure 6 presents results from the IVW method. UVMR did not provide support for the hypothesis that the genetic liability for unhealthy diet/physical inactivity causally increases the risk of depression (Figure 6A) or anxiety (Figure 6C), given that most of the ORs were close to one with confidence intervals including one. Likewise, our results do not support the hypothesis that the genetic liability of depression (Figure 6B) or anxiety (Figure 6D) causally increases unhealthy diet/physical inactivity. One exception, UVMR does suggest fruit intake might causally increase the risk of anxiety (Model 11AUVMR; ORIVW = 1.99; 95% confidence interval (CI) = 1.19 to 3.34; p = .009). This effect was consistent across all sensitivity analyses (Supplementary Table ST47). The UVMR IVW (p = .023) and MR-Egger (p = .019) heterogeneity tests for Model 11A did provide evidence for heterogeneity (Supplementary Table ST49), while the MR-Egger pleiotropy test (p = .918) did not provide evidence for directional pleiotropy (Supplementary Table ST50). The MVMR findings were in line with the UVMR findings (Supplementary Tables ST48, ST51). Most instruments were sufficiently strong (F-statistics >10; Supplementary Tables ST52 (UVMR) and ST53 (MVMR)).

Figure 6. Results of the univariable and multivariable mendelian randomization.

Note. (A) Exposure = diet (sweet snack intake, savoury snack intake, fruit intake, vegetable intake) or physical activity. Outcome = depression. (B) Exposure = depression. Outcome = diet or physical activity. (C) Exposure = diet or physical activity. Outcome = anxiety. (D) Exposure = anxiety. Outcome = diet or physical activity. Method = Inverse variance weighted. UVMR = Univariable Mendelian randomization; MVMR = Multivariable Mendelian randomization; OR = odds ratio; 95% CI = 95% confidence interval; SNP = single nucleotide polymorphism.

Discussion

Principal findings

Triangulating evidence from three distinct methods did not provide compelling support for causal effects of diet/physical activity on depression/anxiety (or vice versa). These conclusions remained when adjusting the analyses including diet and depression/anxiety for physical activity and when adjusting the analyses including physical activity and depression/anxiety for diet.

Comparison with previous studies

Sweet/savoury snack intake and depression/anxiety

Convergent evidence from our triangulation framework did not support causal effects of sweet/savoury snack intake on depression/anxiety (both in unadjusted and adjusted models). No previous studies focused on these specific pathways, although our results are in line with an MR study that did not find a causal effect of “Never eating sugar or foods/drinks containing sugar” on depression (Du et al., Reference Du, Guo, Sun, Zhou, Jiang, Shen and Zhou2023). In contrast, meta-analyses of prospective/cohort studies did support causal effects of Western-style dietary patterns (including high consumption of sweets), sugar-sweetened beverages and ultra-processed foods on increased depression/anxiety (Lane et al., Reference Lane, Gamage, Du, Ashtree, McGuinness, Gauci and Marx2024; Li et al., Reference Li, Lv, Wei, Sun, Zhang, Zhang and Li2017; Y. Wang et al., Reference Wang, Zhao, Wang, Zhao, Zhu and Tian2022). These effects could be enhanced by unmeasured confounders or differences could be explained by using different measurements and including participants with different ancestries or from different cultures (with different eating habits).

Convergent evidence from our triangulation framework did not support causal effects of depression/anxiety on sweet/savoury snack intake (both in unadjusted and adjusted models). This is in line with some studies focusing on the relationship between sugar intake and negative mood (Cardi, Leppanen, & Treasure, Reference Cardi, Leppanen and Treasure2015; Knüppel, Shipley, Llewellyn, & Brunner, Reference Knüppel, Shipley, Llewellyn and Brunner2017), although it is in contrast with previous work indicating that some individuals cope with negative emotions by overeating energy-dense, nutrient-poor and palatable foods (Burnatowska, Surma, & Olszanecka-Glinianowicz, Reference Burnatowska, Surma and Olszanecka-Glinianowicz2022; Dakanalis et al., Reference Dakanalis, Mentzelou, Papadopoulou, Papandreou, Spanoudaki, Vasios and Giaginis2023). Assessment of foods (for example high in sugar) might not reflect an individual’s general diet, which requires more nuanced assessments to shed new light on potential causal effects.

Fruit/vegetable intake and depression/anxiety

Consistent in the unadjusted and adjusted models, Study 1 did not support causal effects of fruit/vegetable intake on depression/anxiety, while Study 2 did support causal effects between increased fruit intake and increased depression/anxiety and Study 3 did support a causal effect from increased fruit/vegetable intake to increased anxiety (Study 3). Contradicting insights were also observed in previous work (Appleton et al., Reference Appleton, Boxall, Adenuga-Ajayi and Seyar2024; T. T. Chen et al., Reference Chen, Chen, Fang, Cheng and Lin2022; Liu et al., Reference Liu, Yan, Li and Zhang2016; Molendijk et al., Reference Molendijk, Molero, Sánchez-Pedreño, Van der Does and Martínez-González2018; Saghafian et al., Reference Saghafian, Malmir, Saneei, Milajerdi, Larijani and Esmaillzadeh2018; Tuck et al., Reference Tuck, Farrow and Thomas2019; Q. Wang et al., Reference Wang, Ou, Chen, Li, Chen and Ye2024; Yan et al., Reference Yan, Xu, Li and Liu2023). Possibly, the time intervals in Study 1 (6–12 months) were too long to detect causation (Singh et al., Reference Singh, Verhulst, Vinh, Zhou, Castro-de-Araujo, Hottenga and Neale2024). Additionally, the variance in the study measures across all waves that could be explained by fluctuations within persons varied between 27% to 48%, which might limit the opportunity to detect possible temporal associations in Study 1.

Studies 1 and 3 did not provide support for causal effects of depression/anxiety on fruit/vegetable intake while Study 2 support causal effects between depression/anxiety and fruit/vegetable intake (both in unadjusted and adjusted models). When comparing these findings to previous literature, only two MR studies were identified which did not support a causal effect of depression on fruit intake and/or vegetable intake (T. T. Chen et al., Reference Chen, Chen, Fang, Cheng and Lin2022; Yan et al., Reference Yan, Xu, Li and Liu2023).

Possibly, limitations of the designs used in Studies 2 and 3 could explain the suggestive findings mentioned above. E.g., a small sample size, arbitrary discordance definitions or confounding by non-shared factors in Study 2 (Frisell, Öberg, Kuja-Halkola, & Sjölander, Reference Frisell, Öberg, Kuja-Halkola and Sjölander2012) and potential violations of the MR assumptions (e.g., horizontal pleiotropyFootnote 10) in Study 3 (Davies et al., Reference Davies, Howe, Brumpton, Havdahl, Evans and Davey Smith2019; Hemani, Bowden, & Davey Smith, Reference Hemani, Bowden and Davey Smith2018; Spiga et al., Reference Spiga, Gibson, Dawson, Tilling, Davey Smith, Munafò and Higgins2023). This emphasizes the need for triangulation and developing more complex methods where the strengths of different designs are combined (e.g., within-family MR).

Physical activity and depression/anxiety

Convergent evidence from our triangulation framework did not support causal effects of physical activity on depression/anxiety (both in unadjusted and adjusted models). Although this is in line with some studies (various designs, mostly using self-report) (Choi et al., Reference Choi, Chen, Stein, Klimentidis, Wang, Koenen and Smoller2019; De Moor et al., Reference De Moor, Boomsma, Stubbe, Willemsen and de Geus2008; Iob et al., Reference Iob, Pingault, Munafò, Stubbs, Gilthorpe, Maihofer and Danese2023; Moreno-Peral et al., Reference Moreno-Peral, Pino-Postigo, Conejo-Cerón, Bellón, Rodríguez-Martín, Martínez-Vizcaíno and Bellón2022; Pasman et al., Reference Pasman, Bergstedt, Harder, Gong, Xiong, Hägg and Lu2024), most existing studies do provide support for a protective causal effect of physical activity on depression and/or anxiety risk (Choi et al., Reference Choi, Chen, Stein, Klimentidis, Wang, Koenen and Smoller2019; Iob et al., Reference Iob, Pingault, Munafò, Stubbs, Gilthorpe, Maihofer and Danese2023; Mammen & Faulkner, Reference Mammen and Faulkner2013; McDowell et al., Reference McDowell, Dishman, Gordon and Herring2019; Pearce et al., Reference Pearce, Garcia, Abbas, Strain, Schuch, Golubic and Woodcock2022; Rebar et al., Reference Rebar, Stanton, Geard, Short, Duncan and Vandelanotte2015; Schuch et al., Reference Schuch, Stubbs, Meyer, Heissel, Zech, Vancampfort and Hiles2019; Schuch et al., Reference Schuch, Vancampfort, Firth, Rosenbaum, Ward, Silva and Stubbs2018). Some of these studies (using MR) found different results for self-reported or accelerometer-based physical activity (Choi et al., Reference Choi, Chen, Stein, Klimentidis, Wang, Koenen and Smoller2019; Iob et al., Reference Iob, Pingault, Munafò, Stubbs, Gilthorpe, Maihofer and Danese2023), which suggests future studies could explore this difference further using various other designs.

No support for causal effects of depression/anxiety on physical activity was provided by Studies 1 and 3 while Study 2 did support possible causal effects between increased depression/anxiety and decreased physical activity (both in unadjusted and adjusted models). Mixed results were also observed in previous work. For instance, a systematic review (longitudinal studies) showed that depression was associated with being less active over time (Roshanaei-Moghaddam et al., Reference Roshanaei-Moghaddam, Katon and Russo2009). The authors suggested this might be due to lower motivation and energy to exercise. A recent MR study also found a causal effect of depression on decreased accelerometer-based physical activity (Pasman et al., Reference Pasman, Bergstedt, Harder, Gong, Xiong, Hägg and Lu2024). In contrast, this was not found in other MR studies (including both self-reported and accelerometer-based physical activity) (Choi et al., Reference Choi, Chen, Stein, Klimentidis, Wang, Koenen and Smoller2019; Iob et al., Reference Iob, Pingault, Munafò, Stubbs, Gilthorpe, Maihofer and Danese2023). Pasman et al. (Reference Pasman, Bergstedt, Harder, Gong, Xiong, Hägg and Lu2024) used a more recent and larger GWAS for depression (Als et al., Reference Als, Kurki, Grove, Voloudakis, Therrien, Tasanko and Børglum2023) (>1.3 million individuals including 371,184 cases; identifying 243 risk loci) compared to Choi et al. (Reference Choi, Chen, Stein, Klimentidis, Wang, Koenen and Smoller2019) and Iob et al. (Reference Iob, Pingault, Munafò, Stubbs, Gilthorpe, Maihofer and Danese2023) who used a smaller GWAS (Wray et al., Reference Wray, Ripke, Mattheisen, Trzaskowski, Byrne, Abdellaoui and Sullivan2018) (including 344,901 controls and 135,458 cases; identifying 44 risk loci). Consequently, this may explain the differences in findings and suggest sufficient sample sizes are needed in future studies. Alternatively, other potential biases (described above) might also contribute to the divergent results.

Strengths, limitations & future research

Although triangulation is not new, triangulation in mental health research is limited (Hammerton & Munafò, Reference Hammerton and Munafò2021). To our knowledge, this study is the first to triangulate evidence to assess the causal pathways between diet, physical activity, depression, and anxiety. No single method can provide definite evidence for causal pathways on its own (Hammerton & Munafò, Reference Hammerton and Munafò2021). In our study, triangulation strengthened our conclusion of no causal effects for sweet and savoury snack intake, since this finding was consistent across methods that address different types of confounding. However, for fruit/vegetable intake and physical activity, divergent findings across methods revealed potential biases that single-method studies might have missed. This demonstrates triangulation’s value in both confirming null effects and identifying method-specific biases needing further investigation. We also extended previous work by considering mutual confounding between diet and physical activity, since both play substantial roles in health maintenance and disease prevention, rather than treating them as independent behaviours.

However, several limitations exist. First, retrospective triangulation was applied using existing data. Although the measures across the three studies were aligned as much as possible, not all measures were exactly the same. Consequently, this could influence the comparability of the studies. In line with this, the three studies included in the triangulation framework used data originating from different countries (i.e., The Netherlands and Australia). It could be that country-specific dietary habits or physical activity norms have impacted the compatibility of the three studies, which could explain some divergence in the findings. However, we do not expect that the difference in the origin of the data had a major impact, given that all individuals were from European Ancestry. Future studies could adapt prospective triangulation approaches (Munafò et al., Reference Munafò, Higgins and Davey Smith2021; Treur, Lukas, Sallis, & Wootton, Reference Treur, Lukas, Sallis and Wootton2024), where study measures, sample populations and timing of data collections are aligned before data collection. This will result in even stronger confidence in the conclusions and help to avoid divergence in the results. Second, all studies used observational self-reported data to some extent. Therefore, reporting/recall biasFootnote 11, measurement errors and selection biasFootnote 12 could not fully be ruled out (Hammerton & Munafò, Reference Hammerton and Munafò2021; Pandis, Reference Pandis2014). Future studies can consider additional/other approaches less prone to these biases (e.g., the MR and the Direction of Causation twin model (Castro-de-Araujo et al., Reference Castro-de-Araujo, Singh, Zhou, Vinh, Verhulst, Dolan and Neale2023; Minică, Boomsma, Dolan, de Geus, & Neale, Reference Minică, Boomsma, Dolan, de Geus and Neale2020) or within-family GWAS/MR). Third, Wave 1 of Study 1 was collected during the COVID-19 pandemic (Fall 2021). During this unique time, several measures aimed to reduce the spread of the coronavirus were implemented (e.g., working from home, limits on number of students in classrooms at schools/universities, 1.5-meter spacing rule, and wearing face masks; National Institute for Public Health and the Environment (n.d.)). These measures most likely impacted the lifestyle behaviours and mental health reported during Wave 1. E.g., other studies do report that physical activity levels reduced and depressive symptoms increased during the COVID-19 pandemic (Caroppo et al., Reference Caroppo, Mazza, Sannella, Marano, Avallone, Claro and Sani2021; P. J. Chen, Pusica, Sohaei, Prassas, & Diamandis, Reference Chen, Pusica, Sohaei, Prassas and Diamandis2021; Park, Zhong, Yang, Jeong, & Lee, Reference Park, Zhong, Yang, Jeong and Lee2022). However, this does not necessarily imply that the associations between variables in this study are different during the COVID-19 pandemic compared to other periods, as in previous work physical activity-depressive symptoms associations before and during the pandemic could be constrained to be equal over time (van den Broek et al., Reference van den Broek, Maran, Beckers, Burk, Verhagen, Vink and Larsen2024). Given the different time intervals between the waves in the RI-CLPMs in Study 1, we were unable to empirically test whether we could constrain during and after the pandemic. Despite this, we have little reason to suspect that the within-person effects would be different for the two time intervals, given that all effect sizes were considerably small/close to zero. Fourth, Studies 2 and 3 might have limited power. In Study 2, the sample sizes are relatively small (especially DZ twin pairs). Consequently, larger twin studies are needed. In Study 3, this is reflected by smaller GWASs of sweet snack intake, savoury snack intake, vegetable intake and, anxiety. Consequently, fewer robustly associated exposure SNPs were available at the genome-wide significance level and a less stringent p-value threshold was used to select exposure SNPs. This could have introduced weak instrument bias (Burgess & Thompson, Reference Burgess and Thompson2011). Consequently, larger GWASs are needed. Fifth, sex-specific analyses were not included. However, it is known that depression/anxiety are more common among females than males (GBD Mental Disorders Collaborators, 2022) and a recent study in adolescents suggested sex-specific effects from physical activity on depressive symptoms (van den Broek et al., Reference van den Broek, Maran, Beckers, Burk, Verhagen, Vink and Larsen2024). Future studies could investigate these possible sex-specific effects. Last, post-treatment bias might bias the results of the performed adjusted models, in case the exposure has an causal effect on the included confounder (Acharya, Blackwell, & Sen, Reference Acharya, Blackwell and Sen2016).

Implications

Integrating triangulation approaches (also beyond the scope of this study) in scientific work more systematically, instead of single-methodology approaches, will greatly impact the weight of scientific output (Hammerton & Munafò, Reference Hammerton and Munafò2021). Institutions, organisations and policymakers can also use triangulated evidence to make more confident decisions on policy/strategy development, innovations to foster and resource allocation. With regard to the implications of the current study, given the acknowledged limitations of our triangulation approach, it is too early to provide strong recommendations for practice. However, triangulating evidence adds an important piece of the bigger puzzle of finding the true answer to the question if causal pathways exist between diet/physical activity and depression/anxiety. Before we can provide clear practical recommendations, additional pieces of this bigger puzzle are needed. As mentioned, these can be obtained by e.g., using complementary methods with other unrelated sources of potential biases or conducting larger twin studies/GWASs. Although in the current study no strong support for causal pathways between diet/physical activity and depression/anxiety was found, offering prevention/intervention services to e.g., stimulate physical activity, healthy diet and mental health will remain important (regardless of potential causal pathways), given their role in the disease risk reduction of multiple chronic diseases (Afshin et al., Reference Afshin, Sur, Fay, Cornaby, Ferrara, Salama and Murray2019; Warburton, Nicol, & Bredin, Reference Warburton, Nicol and Bredin2006).

Conclusion

Triangulated evidence from three distinctive methods (with unique strengths, weaknesses and key sources of potential bias) did not provide strong support for causal effects of diet/physical activity on depression/anxiety or vice versa, neither in the unadjusted or adjusted models (where diet was adjusted for physical activity and vice versa). Future studies could apply prospective triangulation to gain even more robust insights and confidence in answering complex causal questions.

Supplementary material

The supplementary material for this article can be found at http://doi.org/10.1017/S0033291724003349.

Data availability

Data from the Healthy Student Life project, used in Study 1, will be available via the data repository of Radboud University. Data from the Brisbane Longitudinal Twin Study, used in Study 2, are available upon request via the co-author (). Summary statistics from genome-wide association studies, used in Study 3, are publicly available (for access see Supplementary Table ST5).

Acknowledgements

We would like to thank all participants who contributed to Study 1, Study 2 or Study 3. We would also like to express our gratitude to all researchers/contributors involved in the Healthy Student Life study (special word of thanks to prof. Sabine Geurts, dr. Tirza van Noorden and, dr. Sterre Simons), the Brisbane Longitudinal Twin study (special word of thanks to prof. Nicholas Martin) and the different genome-wide association studies consortia we have used GWAS results from in Study 3.

Author contribution

CRediT Taxonomy. Conceptualization – KvH, ZR, NvdB, MS, HS, NG, MM, JV. Data curation – KvH. Formal analysis – KvH, ZR, NvdB, MS, JV. Project administration – KvH, JV. Visualisation – KvH. Supervision – ZR, MM, JV. Writing original draft – KvH. Writing review & editing – KvH, ZR, NvdB, MS, HS, NG, MM, JV.

Code availability

Codes are available on the Open Science Framework (https://osf.io/e4d5b/).

Funding statement

Kirsten van Hooijdonk received travel grants from Radboud University, Simons Foundation Fund and Cultuurfonds to support her work visit to the University of Bristol. Zoe Reed and Marcus Munafò are supported by the UK Medical Research Council Integrative Epidemiology Unit at the University of Bristol (Grant ref: MC_UU_00032/7). Madhurbain Singh is funded by the U.S. National Institutes of Health (NIH) grants R01DA049867 and R01MH125938. The Healthy Student Life project is funded by Radboud University. The Brisbane Longitudinal Twin Study, 25UP project was funded by NHMRC Project Grant APP1069141. The funders had no role in the design, methods, results, interpretation, writing, or conclusions of this manuscript.

Competing interest

The author(s) declare none.

Ethical statement

This study was performed in line with the principles of the Declaration of Helsinki. The Healthy Student Life study (Study 1) has been reviewed by Radboud University’s Social Sciences Ethics Committee, and there was no formal objection to this research (ECSW-2021-086). The Brisbane Longitudinal Twin Study, 25UP project (Study 2) assessment protocols were approved by the QIMR Berghofer Medical Research Institute Human Research Ethics Committee. Additional ethics approval was not required for the current study.

Footnotes

1 Insufficient physical activity is defined as not adhering to 150 minutes of moderate-intensity activity, 75 minutes of vigorous-intensity activity, or an equivalent combination per week.

2 In co-twin control designs monozygotic and dizygotic twin pairs are leveraged to account for genetic and environmental confounding (Gonggrijp, van de Weijer, Bijleveld, van Dongen, & Boomsma, Reference Gonggrijp, van de Weijer, Bijleveld, van Dongen and Boomsma2023).

3 In Mendelian randomization studies genetic variants robustly associated with an exposure are used as instrumental variables to study its causal effects on an outcome (Davey Smith & Ebrahim, Reference Davey Smith and Ebrahim2003).

4 Discordance refers to the fact that one member of the twin pair is exposed while the other person is unexposed to an exposure of interest.

5 Mendel’s law of random segregation: alleles separate at meiosis, and a randomly selected allele is passed from the parents to the offspring (Evans & Davey Smith, Reference Evans and Davey Smith2015).

6 Mendel’s law of independent assortment: alleles for separate traits are transmitted independently of one another (Evans & Davey Smith, Reference Evans and Davey Smith2015).

7 Including three depression-related phenotypes: broad depression, probable major depressive disorder (MDD), and International Classification of Diseases (ICD, version 9 or 10)-coded MDD.

8 Including lifetime diagnosis for any of the five core anxiety disorders: generalized anxiety disorder, panic disorder, social phobia, agoraphobia, and specific phobias.

9 MVMR-Egger was performed twice, once oriented so that the exposure 1 SNPs were positive and once oriented so that exposure 2 SNPs were positive (Sanderson et al., Reference Sanderson, Davey Smith, Windmeijer and Bowden2018). For each exposure, only the results for the relevant (positive) orientation have been reported.

10 Horizontal pleiotropy: The SNPs employed might affect the outcome through pathways unrelated to the exposure.

11 Reporting/recall bias: bias introduced by incomplete or inaccurate reporting.

12 Selection bias: bias introduced by the selection of participants.

References

Aarts, E., Verhage, M., Veenvliet, J. V., Dolan, C. V., & Van Der Sluis, S. (2014). A solution to dependency: Using multilevel analysis to accommodate nested data. Nature Neuroscience, 17(4), 491496. doi:10.1038/nn.3648CrossRefGoogle ScholarPubMed
Acharya, A., Blackwell, M., & Sen, M. (2016). Explaining causal findings without bias: Detecting and assessing direct effects. American Political Science Review, 110(3), 512529. doi:10.1017/S0003055416000216CrossRefGoogle Scholar
Afshin, A., Sur, P. J., Fay, K. A., Cornaby, L., Ferrara, G., Salama, J. S., … Murray, C. J. L. (2019). Health effects of dietary risks in 195 countries, 1990-2017: A systematic analysis for the Global Burden of Disease Study 2017. The Lancet, 393(10184), 19581972. doi:10.1016/S0140-6736(19)30041-8CrossRefGoogle Scholar
Als, T. D., Kurki, M. I., Grove, J., Voloudakis, G., Therrien, K., Tasanko, E., … Børglum, A. D. (2023). Depression pathophysiology, risk prediction of recurrence and comorbid psychiatric disorders using genome-wide analyses. Nature Medicine, 29(7), 18321844. doi:10.1038/s41591-023-02352-1CrossRefGoogle ScholarPubMed
Andrews, G., & Slade, T. (2001). Interpreting scores on the Kessler Psychological Distress Scale (K10). Australian and New Zealand Journal of Public Health, 25(6), 494497. doi:10.1111/j.1467-842x.2001.tb00310.xCrossRefGoogle ScholarPubMed
Appleton, K. M., Boxall, L. R., Adenuga-Ajayi, O., & Seyar, D. F. (2024). Does fruit and vegetable consumption impact mental health? Systematic review and meta-analyses of published controlled intervention studies. British Journal of Nutrition, 131(1), 163173. doi:10.1017/s0007114523001423CrossRefGoogle ScholarPubMed
Bates, D., Maechler, M., Bolker, B., Walker, S., Christensen, R. H. B., Singmann, H., … Grothendieck, G. (2009a). Fitting linear mixed-effects models using lme4. Journal of Statistical Software, 67(1), 148. doi:10.18637/jss.v067.i01Google Scholar
Bates, D., Maechler, M., Bolker, B., Walker, S., Christensen, R. H. B., Singmann, H., … Grothendieck, G. (2009b). Package ‘lme4’. http://lme4.r-forge.r-project.org.Google Scholar
Bowden, J., Davey Smith, G., & Burgess, S. (2015). Mendelian randomization with invalid instruments: effect estimation and bias detection through Egger regression. International Journal of Epidemiology, 44(2), 512525. doi:10.1093/ije/dyv080CrossRefGoogle ScholarPubMed
Bowden, J., Davey Smith, G., Haycock, P. C., & Burgess, S. (2016). Consistent estimation in mendelian randomization with some invalid instruments using a weighted Median estimator. Genetic Epidemiology, 40(4), 304314. doi:10.1002/gepi.21965CrossRefGoogle ScholarPubMed
Burgess, S., Butterworth, A., & Thompson, S. G. (2013). Mendelian randomization analysis with multiple genetic variants using summarized data. Genetic Epidemiology, 37(7), 658665. doi:10.1002/gepi.21758CrossRefGoogle ScholarPubMed
Burgess, S., & Thompson, S. G. (2011). Avoiding bias from weak instruments in Mendelian randomization studies. International Journal of Epidemiology, 40(3), 755764. doi:10.1093/ije/dyr036CrossRefGoogle ScholarPubMed
Burgess, S., & Thompson, S. G. (2015). Multivariable Mendelian randomization: The use of pleiotropic genetic variants to estimate causal effects. American Journal of Epidemiology, 181(4), 251260. doi:10.1093/aje/kwu283CrossRefGoogle ScholarPubMed
Burnatowska, E., Surma, S., & Olszanecka-Glinianowicz, M. (2022). Relationship between mental health and emotional eating during the COVID-19 pandemic: A systematic review. Nutrients, 14(19). doi:10.3390/nu14193989CrossRefGoogle ScholarPubMed
Cardi, V., Leppanen, J., & Treasure, J. (2015). The effects of negative and positive mood induction on eating behaviour: A meta-analysis of laboratory studies in the healthy population and eating and weight disorders. Neuroscience & Biobehavioral Reviews, 57, 299309. doi:10.1016/j.neubiorev.2015.08.011CrossRefGoogle ScholarPubMed
Carnegie, R., Zheng, J., Sallis, H. M., Jones, H. J., Wade, K. H., Evans, J., … Martin, R. M. (2020). Mendelian randomisation for nutritional psychiatry. The Lancet Psychiatry, 7(2), 208216. doi:10.1016/S2215-0366(19)30293-7CrossRefGoogle ScholarPubMed
Caroppo, E., Mazza, M., Sannella, A., Marano, G., Avallone, C., Claro, A. E., … Sani, G. (2021). Will nothing be the same again?: Changes in lifestyle during COVID-19 pandemic and consequences on mental health. International Journal of Environmental Research and Public Health, 18(16), 8433.CrossRefGoogle ScholarPubMed
Castro-de-Araujo, L. F. S., Singh, M., Zhou, Y., Vinh, P., Verhulst, B., Dolan, C. V., & Neale, M. C. (2023). MR-DoC2: Bidirectional causal modeling with instrumental variables and data from relatives. Behavior Genetics, 53(1), 6373. doi:10.1007/s10519-022-10122-xCrossRefGoogle ScholarPubMed
Chen, P. J., Pusica, Y., Sohaei, D., Prassas, I., & Diamandis, E. P. (2021). An overview of mental health during the COVID-19 pandemic. Diagnosis (Berl), 8(4), 403412.CrossRefGoogle ScholarPubMed
Chen, T. T., Chen, C. Y., Fang, C. P., Cheng, Y. C., & Lin, Y. F. (2022). Causal influence of dietary habits on the risk of major depressive disorder: A diet-wide Mendelian randomization analysis. Journal of Affective Disorders, 319, 482489. doi:10.1016/j.jad.2022.09.109CrossRefGoogle ScholarPubMed
Choi, K. W., Chen, C. Y., Stein, M. B., Klimentidis, Y. C., Wang, M. J., Koenen, K. C., & Smoller, J. W. (2019). Assessment of bidirectional relationships between physical activity and depression among adults: A 2-sample mendelian randomization study. JAMA Psychiatry, 76(4), 399408. doi:10.1001/jamapsychiatry.2018.4175CrossRefGoogle ScholarPubMed
Cole, J. B., Florez, J. C., & Hirschhorn, J. N. (2020). Comprehensive genomic analysis of dietary habits in UK Biobank identifies hundreds of genetic associations. Nature Communications, 11(1), 1467. doi:10.1038/s41467-020-15193-0CrossRefGoogle ScholarPubMed
Dakanalis, A., Mentzelou, M., Papadopoulou, S. K., Papandreou, D., Spanoudaki, M., Vasios, G. K., … Giaginis, C. (2023). The association of emotional eating with overweight/obesity, depression, anxiety/stress, and dietary patterns: A review of the current clinical evidence. Nutrients, 15(5). doi:10.3390/nu15051173CrossRefGoogle ScholarPubMed
Davey Smith, G., & Ebrahim, S. (2003). ‘Mendelian randomization’: can genetic epidemiology contribute to understanding environmental determinants of disease? International Journal of Epidemiology, 32(1), 122. doi:10.1093/ije/dyg070CrossRefGoogle Scholar
Davies, N. M., Howe, L. J., Brumpton, B., Havdahl, A., Evans, D. M., & Davey Smith, G. (2019). Within family Mendelian randomization studies. Human Molecular Genetics, 28 (R2), R170R179. doi:10.1093/hmg/ddz204CrossRefGoogle ScholarPubMed
De Moor, M. H., Boomsma, D. I., Stubbe, J. H., Willemsen, G., & de Geus, E. J. (2008). Testing causality in the association between regular exercise and symptoms of anxiety and depression. Archives Of General Psychiatry, 65(8), 897905. doi:10.1001/archpsyc.65.8.897CrossRefGoogle ScholarPubMed
Donker, T., van Straten, A., Marks, I., & Cuijpers, P. (2011). Quick and easy self-rating of generalized anxiety disorder: Validity of the Dutch web-based GAD-7, GAD-2 and GAD-SI. Psychiatry Research, 188(1), 5864. doi:10.1016/j.psychres.2011.01.016CrossRefGoogle ScholarPubMed
Du, Z., Guo, S., Sun, Y., Zhou, Q., Jiang, Y., Shen, Y., … Zhou, H. (2023). Causal relationships between dietary habits and five major mental disorders: A two-sample Mendelian randomization study. Journal of Affective Disorde, 340, 607615. doi:10.1016/j.jad.2023.08.098CrossRefGoogle ScholarPubMed
Eaton, N. R., Krueger, R. F., South, S. C., Gruenewald, T. L., Seeman, T. E., & Roberts, B. W. (2012). Genes, environments, personality, and successful aging: Toward a comprehensive developmental model in later life. The Journals of Gerontology. Series A, Biological Sciences and Medical Sciences, 67(5), 480488. doi:10.1093/gerona/gls090CrossRefGoogle Scholar
Elsworth, B., Lyon, M., Alexander, T., Liu, Y., Matthews, P., Hallett, J., … Hemani, G. (2020). The MRC IEU OpenGWAS data infrastructure. bioRxiv, 2020.08.10.244293v1. doi:10.1101/2020.08.10.244293CrossRefGoogle Scholar
Enders, C. K. (2001). A primer on maximum likelihood algorithms available for use with missing data. Structural Equation Modeling, 8(1), 128141. doi:10.1207/S15328007SEM0801_7CrossRefGoogle Scholar
Evans, D. M., & Davey Smith, G. (2015). Mendelian randomization: New applications in the coming age of hypothesis-Free causality. Annual Review of Genomics and Human Genetics, 16, 327350. doi:10.1146/annurev-genom-090314-050016CrossRefGoogle ScholarPubMed
Frisell, T., Öberg, S., Kuja-Halkola, R., & Sjölander, A. (2012). Sibling comparison designs: Bias from non-shared confounders and measurement error. Epidemiology, 23(5), 713720. doi:10.1097/EDE.0b013e31825fa230CrossRefGoogle ScholarPubMed
GBD 2019 Diseases and Injuries Collaborators. 2020 Global burden of 369 diseases and injuries in 204 countries and territories, 1990–2019: A systematic analysis for the Global Burden of Disease Study 2019. Lancet, 396(10258), 12041222. doi:10.1016/s0140-6736(20)30925-9CrossRefGoogle Scholar
GBD Mental Disorders Collaborators. (2022). Global, regional, and national burden of 12 mental disorders in 204 countries and territories, 1990–2019: A systematic analysis for the Global Burden of Disease Study 2019. The Lancet Psychiatry, 9(2), 137150. doi:10.1016/S2215-0366(21)00395-3CrossRefGoogle Scholar
Gonggrijp, B. M., van de Weijer, S. G., Bijleveld, C. C., van Dongen, J., & Boomsma, D. I. (2023). The co-twin control design: Implementation and methodological considerations. Twin Research and Human Genetics, 18. doi:10.1017/thg.2023.35Google ScholarPubMed
Graham, J. W. (2009). Missing data analysis: Making it work in the real world. Annual Review of Psychology, 60, 549576. doi:10.1146/annurev.psych.58.110405.085530CrossRefGoogle ScholarPubMed
Gupta, V., Walia, G., & Sachdeva, M. (2017). ‘Mendelian randomization’: An approach for exploring causal relations in epidemiology. Public Health, 145, 113119. doi:10.1016/j.puhe.2016.12.033CrossRefGoogle ScholarPubMed
Hamaker, E. L., Kuiper, R. M., & Grasman, R. P. (2015). A critique of the cross-lagged panel model. Psychological Methods, 20(1), 102116. doi:10.1037/a0038889CrossRefGoogle ScholarPubMed
Hammerton, G., & Munafò, M. R. (2021). Causal inference with observational data: the need for triangulation of evidence. Psychological Medicine, 51(4), 563578. doi:10.1017/S0033291720005127CrossRefGoogle ScholarPubMed
Hartwig, F. P., Davey Smith, G., & Bowden, J. (2017). Robust inference in summary data Mendelian randomization via the zero modal pleiotropy assumption. International Journal of Epidemiology, 46(6), 19851998. doi:10.1093/ije/dyx102CrossRefGoogle ScholarPubMed
Hemani, G., Bowden, J., & Davey Smith, G. (2018). Evaluating the potential role of pleiotropy in Mendelian randomization studies. Human Molecular Genetics, 27 (R2), R195r208. doi:10.1093/hmg/ddy163CrossRefGoogle ScholarPubMed
Hemani, G., Zheng, J., Elsworth, B., Wade, K. H., Baird, D., Haberland, V., … Haycock, P. (2018). The MR-Base platform supports systematic causal inference across the human phenome. elife, 7, e34408. doi:10.7554/eLife.34408CrossRefGoogle ScholarPubMed
Howard, D. M., Adams, M. J., Clarke, T. K., Hafferty, J. D., Gibson, J., Shirali, M., … McIntosh, A. M. (2019). Genome-wide meta-analysis of depression identifies 102 independent variants and highlights the importance of the prefrontal brain regions. Nature Neuroscience, 22(3), 343352. doi:10.1038/s41593-018-0326-7CrossRefGoogle ScholarPubMed
Hu, L. t., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling: A Multidisciplinary Journal, 6(1), 155. doi:10.1080/10705519909540118CrossRefGoogle Scholar
Iob, E., Pingault, J. B., Munafò, M. R., Stubbs, B., Gilthorpe, M. S., Maihofer, A. X., & Danese, A. (2023). Testing the causal relationships of physical activity and sedentary behaviour with mental health and substance use disorders: a Mendelian randomisation study. Molecular Psychiatry, 28(8), 34293443. doi:10.1038/s41380-023-02133-9CrossRefGoogle ScholarPubMed
IPAQ Research Committee. (2005). Guidelines for data processing and analysis of the International Physical Activity Questionnaire (IPAQ)-short and long forms. https://sites.google.com/view/ipaq/scoreGoogle Scholar
Klimentidis, Y. C., Raichlen, D. A., Bea, J., Garcia, D. O., Wineinger, N. E., Mandarino, L. J., … Going, S. B. (2018). Genome-wide association study of habitual physical activity in over 377,000 UK Biobank participants identifies multiple variants including CADM2 and APOE. International Journal of Obesity (Lond), 42(6), 11611176. doi:10.1038/s41366-018-0120-3CrossRefGoogle Scholar
Kline, R. B. (2023). Principles and practice of structural equation modeling. Guilford publications.Google Scholar
Knüppel, A., Shipley, M. J., Llewellyn, C. H., & Brunner, E. J. (2017). Sugar intake from sweet food and beverages, common mental disorder and depression: Prospective findings from the Whitehall II study. Scientific Reports, 7(1), 6287. doi:10.1038/s41598-017-05649-7CrossRefGoogle ScholarPubMed
Lane, M. M., Gamage, E., Du, S., Ashtree, D. N., McGuinness, A. J., Gauci, S., … Marx, W. (2024). Ultra-processed food exposure and adverse health outcomes: umbrella review of epidemiological meta-analyses. BMJ, 384, e077310. doi:10.1136/bmj-2023-077310CrossRefGoogle ScholarPubMed
Lawlor, D. A., Harbord, R. M., Sterne, J. A., Timpson, N., & Davey Smith, G. (2008). Mendelian randomization: Using genes as instruments for making causal inferences in epidemiology. Statistics in Medicine, 27(8), 11331163. doi:10.1002/sim.3034CrossRefGoogle ScholarPubMed
Lawlor, D. A., Tilling, K., & Davey Smith, G. (2016). Triangulation in aetiological epidemiology. International Journal of Epidemiology, 45(6), 18661886. doi:10.1093/ije/dyw314Google ScholarPubMed
Li, Y., Lv, M. R., Wei, Y. J., Sun, L., Zhang, J. X., Zhang, H. G., & Li, B. (2017). Dietary patterns and depression risk: A meta-analysis. Psychiatry Research, 253, 373382. doi:10.1016/j.psychres.2017.04.020CrossRefGoogle ScholarPubMed
Liu, X., Yan, Y., Li, F., & Zhang, D. (2016). Fruit and vegetable consumption and the risk of depression: A meta-analysis. Nutrition, 32(3), 296302. doi:10.1016/j.nut.2015.09.009CrossRefGoogle ScholarPubMed
Mammen, G., & Faulkner, G. (2013). Physical activity and the prevention of depression: a systematic review of prospective studies. American Journal of Preventive Medicine, 45(5), 649657. doi:10.1016/j.amepre.2013.08.001CrossRefGoogle ScholarPubMed
McDowell, C. P., Dishman, R. K., Gordon, B. R., & Herring, M. P. (2019). Physical activity and anxiety: A systematic review and meta-analysis of prospective cohort studies. American Journal of Preventive Medicine, 57(4), 545556. doi:10.1016/j.amepre.2019.05.012CrossRefGoogle ScholarPubMed
Minică, C. C., Boomsma, D. I., Dolan, C. V., de Geus, E., & Neale, M. C. (2020). Empirical comparisons of multiple Mendelian randomization approaches in the presence of assortative mating. International Journal of Epidemiology, 49(4), 11851193. doi:10.1093/ije/dyaa013CrossRefGoogle ScholarPubMed
Mitchell, B. L., Campos, A. I., Rentería, M. E., Parker, R., Sullivan, L., McAloney, K., … Scott, J. (2019). Twenty-five and up (25Up) study: A new wave of the Brisbane Longitudinal Twin Study. Twin Research and Human Genetics, 22(3), 154163.CrossRefGoogle ScholarPubMed
Molendijk, M., Molero, P., Sánchez-Pedreño, F. O., Van der Does, W., & Martínez-González, M. A. (2018). Diet quality and depression risk: a systematic review and dose-response meta-analysis of prospective studies. Journal of Affective Disorders, 226, 346354. doi:10.1016/j.jad.2017.09.022CrossRefGoogle Scholar
Moreno-Peral, P., Pino-Postigo, A., Conejo-Cerón, S., Bellón, D., Rodríguez-Martín, B., Martínez-Vizcaíno, V., & Bellón, J. (2022). Effectiveness of physical activity in primary prevention of anxiety: Systematic review and meta-analysis of randomized controlled trials. The International Journal of Environmental Research and Public Health, 19(3). doi:10.3390/ijerph19031813CrossRefGoogle ScholarPubMed
Mosteller, F. F., & Boruch, R. F. (2004). Evidence matters: Randomized trials in education research. Brookings Institution Press.Google Scholar
Mounier, N., & Kutalik, Z. (2023). Bias correction for inverse variance weighting Mendelian randomization. Genetic Epidemiology, 47(4), 314331. doi:10.1002/gepi.22522CrossRefGoogle ScholarPubMed
Mulder, J. D., & Hamaker, E. L. (2021). Three extensions of the random intercept cross-lagged panel model. Structural Equation Modeling: A Multidisciplinary Journal, 28(4), 638648. doi:10.1080/10705511.2020.1784738CrossRefGoogle Scholar
Munafò, M. R., & Davey Smith, G. (2018). Robust research needs many lines of evidence. Nature, 553(7689), 399401. doi:10.1038/d41586-018-01023-3CrossRefGoogle ScholarPubMed
Munafò, M. R., Higgins, J. P. T., & Davey Smith, G. (2021). Triangulating evidence through the inclusion of genetically informed designs. Cold Spring Harbor Perspectives in Medicine, 11(8). doi:10.1101/cshperspect.a040659CrossRefGoogle ScholarPubMed
National Institute for Public Health and the Environment. (n.d.). Tijdlijn van coronamaatregelen 2021. https://www.rivm.nl/gedragsonderzoek/tijdlijn-van-coronamaatregelen-2021Google Scholar
NCD Countdown 2030 Collaborators. (2018). NCD Countdown 2030: Worldwide trends in non-communicable disease mortality and progress towards Sustainable Development Goal target 3.4. Lancet, 392(10152), 10721088. doi:10.1016/s0140-6736(18)31992-5CrossRefGoogle Scholar
Otowa, T., Hek, K., Lee, M., Byrne, E. M., Mirza, S. S., Nivard, M. G., … Hettema, J. M. (2016). Meta-analysis of genome-wide association studies of anxiety disorders. Molecular Psychiatry, 21(10), 13911399. doi:10.1038/mp.2015.197CrossRefGoogle ScholarPubMed
Pandis, N. (2014). Bias in observational studies. American Journal of Orthodontics and Dentofacial Orthopedics, 145(4), 542543. doi:10.1016/j.ajodo.2014.01.008CrossRefGoogle ScholarPubMed
Park, A. H., Zhong, S., Yang, H., Jeong, J., & Lee, C. (2022). Impact of COVID-19 on physical activity: A rapid review. Journal of Global Health, 12, 05003.CrossRefGoogle ScholarPubMed
Pasman, J. A., Bergstedt, J., Harder, A., Gong, T., Xiong, Y., Hägg, S., … Lu, Y. (2024). Causes and consequences of major depressive disorder: An encompassing Mendelian randomization study. medRxiv, 2024.2005.2021.24307678. doi:10.1101/2024.05.21.24307678CrossRefGoogle Scholar
Patton, M. Q. (1999). Enhancing the quality and credibility of qualitative analysis. Health Services Research, 34 (5 Pt 2), 11891208.Google ScholarPubMed
Pearce, M., Garcia, L., Abbas, A., Strain, T., Schuch, F. B., Golubic, R., … Woodcock, J. (2022). Association between physical activity and risk of depression: A systematic review and meta-analysis. JAMA Psychiatry, 79(6), 550559. doi:10.1001/jamapsychiatry.2022.0609CrossRefGoogle ScholarPubMed
Popkin, B. M., Adair, L. S., & Ng, S. W. (2012). Global nutrition transition and the pandemic of obesity in developing countries. Nutrition Reviews, 70(1), 321. doi:10.1111/j.1753-4887.2011.00456.xCrossRefGoogle ScholarPubMed
R Core Team. (2023). R: A language and environment for statistical computing. https://www.R-project.org/Google Scholar
Rebar, A. L., Stanton, R., Geard, D., Short, C., Duncan, M. J., & Vandelanotte, C. (2015). A meta-meta-analysis of the effect of physical activity on depression and anxiety in non-clinical adult populations. Health Psychology Review, 9(3), 366378. doi:10.1080/17437199.2015.1022901CrossRefGoogle ScholarPubMed
Rees, J. M. B., Wood, A. M., & Burgess, S. (2017). Extending the MR-Egger method for multivariable Mendelian randomization to correct for both measured and unmeasured pleiotropy. Statistics in Medicine, 36(29), 47054718. doi:10.1002/sim.7492CrossRefGoogle ScholarPubMed
Roshanaei-Moghaddam, B., Katon, W. J., & Russo, J. (2009). The longitudinal effects of depression on physical activity. General Hospital Psychiatry, 31(4), 306315. doi:10.1016/j.genhosppsych.2009.04.002CrossRefGoogle ScholarPubMed
Rosseel, Y. (2012). lavaan: An R package for structural equation modeling. Journal of Statistical Software, 48, 136. doi:10.18637/jss.v048.i02CrossRefGoogle Scholar
Saghafian, F., Malmir, H., Saneei, P., Milajerdi, A., Larijani, B., & Esmaillzadeh, A. (2018). Fruit and vegetable consumption and risk of depression: accumulative evidence from an updated systematic review and meta-analysis of epidemiological studies. British Journal of Nutrition, 119(10), 10871101. doi:10.1017/s0007114518000697CrossRefGoogle ScholarPubMed
Sanderson, E., Davey Smith, G., Windmeijer, F., & Bowden, J. (2018). An examination of multivariable Mendelian randomization in the single-sample and two-sample summary data settings. International Journal of Epidemiology, 48(3), 713727. doi:10.1093/ije/dyy262CrossRefGoogle Scholar
Sanderson, E., Spiller, W., & Bowden, J. (2021). Testing and correcting for weak and pleiotropic instruments in two-sample multivariable Mendelian randomization. Statistics in Medicine, 40(25), 54345452. doi:10.1002/sim.9133CrossRefGoogle ScholarPubMed
Schuch, F. B., Stubbs, B., Meyer, J., Heissel, A., Zech, P., Vancampfort, D., … Hiles, S. A. (2019). Physical activity protects from incident anxiety: A meta-analysis of prospective cohort studies. Depress Anxiety, 36(9), 846858. doi:10.1002/da.22915CrossRefGoogle ScholarPubMed
Schuch, F. B., Vancampfort, D., Firth, J., Rosenbaum, S., Ward, P. B., Silva, E. S., … Stubbs, B. (2018). Physical activity and incident depression: A meta-analysis of prospective cohort studies. American Journal of Psychiatry, 175(7), 631648. doi:10.1176/appi.ajp.2018.17111194CrossRefGoogle ScholarPubMed
Shadish, W. R., Cook, T. D., & Campbell, D. T. (2002). Experimental and quasi-experimental designs for generalized causal inference. Houghton Mifflin Boston, MA.Google Scholar
Singh, M., Verhulst, B., Vinh, P., Zhou, Y. D., Castro-de-Araujo, L. F. S., Hottenga, J. J., … Neale, M. C. (2024). Using instrumental variables to measure causation over time in cross-lagged panel models. Multivariate Behavioral Research, 59(2), 342370. doi:10.1080/00273171.2023.2283634CrossRefGoogle ScholarPubMed
Skrivankova, V. W., Richmond, R. C., Woolf, B. A., Davies, N. M., Swanson, S. A., VanderWeele, T. J., … Langenberg, C. (2021). Strengthening the reporting of observational studies in epidemiology using mendelian randomisation (STROBE-MR): Explanation and elaboration. BMJ, 375. doi:10.1136/bmj.n2233Google ScholarPubMed
Spiga, F., Gibson, M., Dawson, S., Tilling, K., Davey Smith, G., Munafò, M. R., & Higgins, J. P. T. (2023). Tools for assessing quality and risk of bias in Mendelian randomization studies: A systematic review. International Journal of Epidemiology, 52(1), 227249. doi:10.1093/ije/dyac149CrossRefGoogle ScholarPubMed
Strain, T., Flaxman, S., Guthold, R., Semenova, E., Cowan, M., Riley, L. M., … Stevens, G. A. (2024). National, regional, and global trends in insufficient physical activity among adults from 2000 to 2022: A pooled analysis of 507 population-based surveys with 5· 7 million participants. The Lancet Global Health, 12(8), e1232e1243. doi:10.1016/S2214-109X(24)00150-5CrossRefGoogle ScholarPubMed
Treur, J. L., Lukas, E., Sallis, H. M., & Wootton, R. E. (2024). A guide for planning triangulation studies to investigate complex causal questions in behavioural and psychiatric research. Epidemiology and Psychiatric Sciences, 33, e61. doi:10.1017/S2045796024000623CrossRefGoogle ScholarPubMed
Tuck, N. J., Farrow, C., & Thomas, J. M. (2019). Assessing the effects of vegetable consumption on the psychological health of healthy adults: A systematic review of prospective research. The American Journal of Clinical Nutrition, 110(1), 196211. doi:10.1093/ajcn/nqz080CrossRefGoogle ScholarPubMed
Van de Velde, S., Levecque, K., & Bracke, P. (2009). Measurement equivalence of the CES-D 8 in the general population in Belgium: A gender perspective. Archives of Public Health, 67, 115. doi:10.1186/0778-7367-67-1-15CrossRefGoogle Scholar
van den Broek, N., Larsen, J. K., Verhagen, M., Burk, W. J., & Vink, J. M. (2020). Is adolescents’ food intake associated with exposure to the food intake of their mothers and best friends? Nutrients, 12(3). doi:10.3390/nu12030786CrossRefGoogle Scholar
van den Broek, N., Maran, P. L., Beckers, D., Burk, W. J., Verhagen, M., Vink, J. M., & Larsen, J. K. (2024). Examining the bidirectional associations between adolescents’ physical activity and depressive symptoms before and during the COVID-19 pandemic. Mental Health and Physical Activity, 27, 100618. doi:10.1016/j.mhpa.2024.100618CrossRefGoogle Scholar
van Hooijdonk, K. J. M., Simons, S. S. H., van Noorden, T. H. J., Geurts, S. A. E., & Vink, J. M. (2023). Prevalence and clustering of health behaviours and the association with socio-demographics and mental well-being in Dutch university students. Preventive Medicine Reports, 35, 102307. doi:10.1016/j.pmedr.2023.102307CrossRefGoogle ScholarPubMed
Verbanck, M., Chen, C. Y., Neale, B., & Do, R. (2018). Detection of widespread horizontal pleiotropy in causal relationships inferred from Mendelian randomization between complex traits and diseases. Nature Genetics, 50(5), 693698. doi:10.1038/s41588-018-0099-7CrossRefGoogle ScholarPubMed
Vitaro, F., Brendgen, M., & Arseneault, L. (2009). The discordant MZ-twin method: One step closer to the holy grail of causality. International Journal of Behavioral Development, 33(4), 376382. doi:10.1177/0165025409340805CrossRefGoogle Scholar
Von Elm, E., Altman, D. G., Egger, M., Pocock, S. J., Gøtzsche, P. C., & Vandenbroucke, J. P. (2007). The strengthening the reporting of observational studies in epidemiology (STROBE) statement: Guidelines for reporting observational studies. The Lancet, 370(9596), 14531457. doi:10.1016/S0140-6736(07)61602-XCrossRefGoogle ScholarPubMed
Wang, Q., Ou, Z., Chen, J., Li, L., Chen, Y., & Ye, D. (2024). Association between vegetable intake and major depressive disorder: results from National Health and Nutrition Examination Survey 2005-2018 and bidirectional two-sample Mendelian randomisation. Public Health Nutrition, 27(1), e220. doi:10.1017/s1368980024001691CrossRefGoogle ScholarPubMed
Wang, Y., Zhao, R., Wang, B., Zhao, C., Zhu, B., & Tian, X. (2022). The dose-response associations of sugar-sweetened beverage intake with the Risk of stroke, depression, cancer, and cause-specific mortality: A systematic review and meta-analysis of prospective studies. Nutrients, 14(4). doi:10.3390/nu14040777Google ScholarPubMed
Wanjau, M. N., Möller, H., Haigh, F., Milat, A., Hayek, R., Lucas, P., & Veerman, J. L. (2023). Physical activity and depression and anxiety disorders: A systematic review of reviews and assessment of causality. AJPM Focus, 2(2), 100074. doi:10.1016/j.focus.2023.100074CrossRefGoogle ScholarPubMed
Warburton, D. E., Nicol, C. W., & Bredin, S. S. (2006). Health benefits of physical activity: The evidence. CMAJ, 174(6), 801809. doi:10.1503/cmaj.051351CrossRefGoogle ScholarPubMed
World Health Organization. (2022). World mental health report: Transforming mental health for all. https://iris.who.int/bitstream/handle/10665/356119/9789240049338-eng.pdf?sequence=1Google Scholar
Wray, N. R., Ripke, S., Mattheisen, M., Trzaskowski, M., Byrne, E. M., Abdellaoui, A., … Sullivan, P. F. (2018). Genome-wide association analyses identify 44 risk variants and refine the genetic architecture of major depression. Nature Genetics, 50(5), 668681. doi:10.1038/s41588-018-0090-3CrossRefGoogle ScholarPubMed
Yan, Z., Xu, Y., Li, K., & Liu, L. (2023). Increased fruit intake is associated with reduced risk of depression: Evidence from cross-sectional and Mendelian randomization analyses. Frontiers in Public Health, 11, 1276326. doi:10.3389/fpubh.2023.1276326CrossRefGoogle ScholarPubMed
Zhao, Q., Wang, J., Hemani, G., Bowden, J., & Small, D. S. (2018). Statistical inference in two-sample summary-data Mendelian randomization using robust adjusted profile score. Annals of Statistics. doi:10.1214/19-AOS1866Google Scholar
Figure 0

Figure 1. Concept of triangulation.Note. Illustration of how our three complementary methods help to strengthen causal inference.

Figure 1

Table 1. Overview of the studies included in the triangulation framework

Figure 2

Figure 2. Example of a random intercept cross-lagged panel model (RI-CLPM) assessing the bidirectional pathways between variable A and variable B, including between-person and within-person components at three survey waves.Note. The squares indicate observed variables, while the circles represent latent (unobserved) variables. U and V represent the residual variance. The triangles represent constants for the mean structure. To improve readability, no covariates are presented. Based on Hamaker et al. (2015) and Mulder and Hamaker (2021).

Figure 3

Figure 3. Hypothetical outcome scenarios of a co-twin control design (CTCD).Note. Population-level (PL) = the mean difference in the magnitude of association between exposed and unexposed individuals; DZ = the mean difference in the magnitude of association within dizygotic same-sex twin pairs discordant for the exposure; MZ = the mean difference in the magnitude of association within monozygotic twin pairs discordant for the exposure. Scenario 1: the magnitude of the association is similar across all groups (regardless of shared genes or (early) environment), inferring possible causality. Scenario 2: the association between exposure and outcome is entirely inferred by genetic confounding, as the association within same-sex DZ twins (who share 50% of their genetic material) is intermediate, and the association within MZ twins (who share 100% of their genetic material) is zero. Scenario 3: the association between exposure and outcome is entirely inferred by environmental confounding, as the association within same-sex DZ and MZ twins (who share 100% of their (early) environment) is zero. Scenario 4: the association between exposure and outcome is partly inferred by both genetic and environmental confounding. Based on Gonggrijp et al. (2023).

Figure 4

Figure 4. Directed acyclic graphs for (A) Univariable mendelian randomization and (B) Multivariable mendelian randomization.

Figure 5

Table 2. Triangulated evidence of all unadjusted/univariable analyses

Figure 6

Table 3. Results within-person cross-lagged effects

Figure 7

Figure 5. Results mixed-effects models - patterns in regression coefficients at population-level (PL), within dizygotic (DZ) same-sex twins and within monozygotic (MZ) twins. Asterisks represent associations between exposure and outcome per subgroup where p < .05.

Figure 8

Figure 6. Results of the univariable and multivariable mendelian randomization.Note. (A) Exposure = diet (sweet snack intake, savoury snack intake, fruit intake, vegetable intake) or physical activity. Outcome = depression. (B) Exposure = depression. Outcome = diet or physical activity. (C) Exposure = diet or physical activity. Outcome = anxiety. (D) Exposure = anxiety. Outcome = diet or physical activity. Method = Inverse variance weighted. UVMR = Univariable Mendelian randomization; MVMR = Multivariable Mendelian randomization; OR = odds ratio; 95% CI = 95% confidence interval; SNP = single nucleotide polymorphism.

Supplementary material: File

van Hooijdonk et al. supplementary material 1

van Hooijdonk et al. supplementary material
Download van Hooijdonk et al. supplementary material 1(File)
File 250.3 KB
Supplementary material: File

van Hooijdonk et al. supplementary material 2

van Hooijdonk et al. supplementary material
Download van Hooijdonk et al. supplementary material 2(File)
File 87.3 KB