Aim, Design and Setting
The GLAD Project is a collective, multi-study initiative led by Deakin University’s, Food & Mood Centre, which is situated within the IMPACT (Institute for Mental and Physical Health and Clinical Translation), School of Medicine, Barwon Health, Geelong, Australia. The GLAD Project will be completed by the GLAD Taskforce, comprising the Project Team, the Advisory Group, and the Working Group containing all Member Studies (Figure 1).
The GLAD Project and the following methods have been prospectively registered on Open Science Framework (https://doi.org/10.17605/OSF.IO/ZBG6X). Our methodology comprises three essential stages, which will be repeated for other lifestyle exposures of interest (Figure 2):
- Conducting a systematic search to identify relevant studies and recruit study leads to contribute data (Phase 1).
- Generating estimates of study-level associations of dietary risk factors with CMDs using harmonised data analysis protocols developed by the GLAD Taskforce and approved by the GBD (Phase 2).
- Pooling the outputs from each study to provide the GBD with robust estimates to calculate the diet-CMD risk-outcome pairs in the first instance (Phase 3).
The methods outlined herewith are based on the most up-to-date definitions used by the GBD, at the time of writing this protocol. Given that the GBD is subject to change, future iterations may vary in accordance with such amendments. The initial iteration of the GLAD Project will focus on dietary exposures, and the methods described within this protocol refer specifically to this first iteration. Subsequent updates to the protocol may be made when additional lifestyle risk factors are considered in future iterations of the GLAD Project.
Member Studies (identified in Phase 1) apply a harmonised study protocol and data analysis plan, developed by the GLAD Project Team, to their own datasets to concurrently calculate associations between dietary intakes (derived from the GBD study, see Table 2) and depression and/or anxiety outcomes (Phase 2). The GLAD Project Team will then apply meta-analytic techniques to pool the results and assess the strength of evidence for each GBD-defined dietary intake and CMD outcome (Phase 3).
Eligibility and Study Identification
Potentially eligible epidemiological studies with data on dietary risk factors and CMD outcomes were initially identified through a database search, using combinations of search terms and based on the search terms from the 2019 GBD systematic review of mental disorders (1). Additional studies were identified through previous collaborations of the GLAD Project Team and researchers in the field of Nutritional Psychiatry (e.g. via the International Society of Nutritional Psychiatric Research), as well as promoting the project on Food & Mood Centre social media sites. Data custodians of identified studies were contacted, and those with relevant data and who had capacity to participate in the GLAD Project completed an Expression of Interest via the Food & Mood Centre website. An overview of the included studies is provided in Table 1, and additional details are available on the Food & Mood Centre Website (https://foodandmoodcentre.com.au/projects/global-burden-of-disease-lifestyle-and-mental-disorders-glad-taskforce/).
To be eligible for inclusion, Member Studies must have dietary intake (Table 2) and depression and anxiety outcomes (Table 3) consistent with GBD definitions, or the ability to recode existing variables accordingly. Any study design, in any location, with any sample size, is eligible to be a Member Study and can complete the required data analyses.
Table 1: Overview of the studies participating* in the GLAD project.
Study Name
|
Study Location
|
Maximum Possible Sample Size†
|
Age Range
|
The African-PREDICT study (18)
|
South Africa
|
1,202
|
20-30
|
Child to Adult Transition Study (CATS) (19)
|
Australia
|
1,239
|
8+
|
Dunedin Study (20)
|
New Zealand
|
1,037
|
0-45
|
Environmental Risk (E-Risk) Longitudinal Twin Study (21,22)
|
United Kingdom
|
2,232
|
0-18
|
Fragility in the Elderly Lombardy Study (FELS) (23)
|
Italy
|
639
|
65+
|
Geelong Osteoporosis Study (GOS) (24)
|
Australia
|
1,518
|
30+
|
Health4Life (25,26)
|
Australia
|
6,639
|
11-13; 14-17
|
Healthy Life in an Urban Setting (HELIUS) (27)
|
The Netherlands
|
24,789
|
18-70
|
Longitudinal Aging Study Amsterdam (LASA) (28)
|
The Netherlands
|
3,805
|
55-85
|
Lothian Birth Cohort 1936 (29,30)
|
United Kingdom
|
1,091
|
60+
|
Melbourne Collaborative Cohort Study (31)
|
Australia
|
41,500
|
40+
|
Moli-sani Study (32)
|
Italy
|
24,325
|
35+
|
Netherlands Study of Depression and Anxiety (NESDA) (33)
|
The Netherlands
|
3,348
|
18-65
|
Northern Ireland Cohort of Longitudinal Ageing (NICOLA) (34)
|
United Kingdom
|
8,500
|
50+
|
NutriNet Brasil (35)
|
Brazil
|
109,245
|
18+
|
NutriNet-Santé (36)
|
France
|
171,000
|
18+
|
Pacific Obesity Prevention in Communities Project (OPIC) (37)
|
Fiji, Tonga, New Zealand and Australia
|
14,000
|
12-19
|
Piccolipiù/Piccoli+/Piccolipiù in Forma (38)
|
Italy
|
3,328
|
0-4
|
REgistre GIroní del COR (REGICOR; English: Girona Heart Registry) (39,40)
|
Spain
|
11,158
|
26+
|
*Additional studies may be participating or providing data for the GLAD Project and not be listed here. †Maximum possible sample size refers to the largest reported sample size from study publications, study websites or study protocol/profile papers. Actual sample size to be included in the GLAD Project may vary.
Table 2: Dietary exposure definitions and required coding for the GLAD Project*
Dietary Risk Factor
|
GBD Exposure Definition
|
Required coding†
|
Fruits
|
Average daily consumption of less than 310-340 grams of fruit including fresh, frozen, cooked, canned, or dried fruit, excluding fruit juices and salted or pickled fruits.
|
grams/day
|
Vegetables
|
Average daily consumption of less than 280-320 grams of vegetables, including fresh, frozen, cooked, canned, or dried vegetables and excluding legumes and salted or pickled vegetables, juices, nuts and seeds, and starchy vegetables such as potatoes or corn.
|
grams/day
|
Legumes
|
Average daily consumption of less than of 90-100 grams of legumes and pulses, including fresh, frozen, cooked, canned, or dried legumes.
|
grams/day
|
Wholegrains
|
Average daily consumption of less than 140-160 grams of whole grains from breakfast cereals, bread, rice, pasta, biscuits, muffins, tortillas, pancakes, and other sources.
|
grams/day
|
Nuts and seeds
|
Average daily consumption of less than 10-19 grams of nuts and seeds.
|
grams/day
|
Milk
|
Average daily consumption of less than 360-500 grams of milk including non-fat, low-fat, and full-fat milk, excluding plant derivatives.
|
grams/day
|
Red meat§
|
Any intake of red meat including beef, pork, lamb, and goat but excluding poultry, fish, eggs, and all processed meats.
|
grams/day
|
Processed meat
|
Any intake of meat preserved by smoking, curing, salting, or addition of chemical preservatives
|
grams/day
|
Sugar-sweetened beverages
|
Any intake of beverages with ≥50 kcal per 226.8 gram serving, including carbonated beverages, sodas, energy drinks, fruit drinks, but excluding 100% fruit and vegetable juices.
|
grams/day
|
Fiber
|
Average daily consumption of less than 21-22 grams of fiber from all sources including fruits, vegetables, grains, legumes, and pulses.
|
grams/day
|
Calcium
|
Average daily consumption of less than 1.06-1.1 grams of calcium from all sources, including milk, yogurt, and cheese.
|
grams/day
|
Seafood omega-3 fatty acids
|
Average daily consumption of less than 430-470 milligrams of eicosapentaenoic acid and docosahexaenoic acid.
|
milligrams/day
|
Polyunsaturated fatty acids
|
Average daily consumption of less than 7-9% total energy intake from polyunsaturated fatty acids.
|
% total daily energy intake
|
Trans fatty acids
|
Any intake of trans fat from all sources, mainly partially hydrogenated vegetable oils and ruminant products.
|
% total daily energy intake
|
Sodium
|
Average 24-hour urinary sodium excretion greater than 1-5 grams
|
grams/day
|
Ultra-processed foods (optional)
|
Any intake of ultra-processed foods, as defined by the Nova system, in grams and percentage of total daily energy intake (42).
|
grams/day
|
*Based on the GBD risk exposure definitions (41; Supplementary Material 1, page 217-218).†Analyses using continuous exposures are required by the GLAD Taskforce. The units listed here are recommended, however these may be re-scaled to more meaningful units in papers (e.g., per 10 grams) provided the original units are available to the Project Team. §The GBD 2021 results will not be imposing a strict zero consumption cut-off, with results to be released shortly. For now, the GLAD Taskforce will be using red meat in grams per day, and this change should not impact the analyses required by the GLAD Taskforce.
Table 3: Mental health definitions and recommended coding for the GLAD project*
Mental Disorder
|
Outcome Definition
|
Diagnostic Reference
|
Recommended coding†
|
Major depressive disorder (MDD)
|
Involves the presence of at least one major depressive episode, which is the experience of either depressed mood or loss of interest/pleasure, for most of every day, for at least two weeks.
|
DSM-IV-TR:
ICD-10:
|
0 = no MDD
1 = MDD
|
Anxiety disorders (any/all subtype)
|
Involves experiences of intense fear and distress, typically in combination with other physiological symptoms. Anxiety disorders will be modelled as a single cause for “any” anxiety disorder to avoid the double-counting of individuals meeting criteria for more than one anxiety disorder. Epidemiological estimates reporting an outcome for “any” or “total” anxiety disorders will be included if they reported on at least three anxiety disorders.
|
DSM-IV-TR:
- 300.0-300.3
- 208.3
- 309.21
- 309.81
ICD-10:
- F40-42
- F43.0
- F43.1
- F93.0-93.2
- F93.8
|
0 = no anxiety disorder
1 = anxiety disorder
|
*The definitions and diagnostic references presented in this table are the current definitions used by the GBD (1; Supplementary appendix, page 4). Since GLAD Members may use alternative methods to ascertain disorder status, definitions or diagnostic references may vary slightly. For example, somatic forms of depression may be captured more by symptom scales, or studies using prescriptions may use Anatomical Therapeutic Chemical (ATC) codes to determine incident depression and/or anxiety. †Outcomes must be binary (condition present vs condition absent). For studies which only use symptom scales, validated cut-offs can be used to determine condition present vs condition absent (see Additional File 1). MDD=Major Depressive Disorder; DSM-IV-TR=Diagnostic and Statistical Manual of Mental Disorders, 4th Edition, Text Revision; ICD-10=International Classification of Diseases 10th Revision.
Processes, Interventions and Comparisons
Exposure and Outcome Definitions
Definitions of dietary intakes and CMDs are based on definitions used by the GBD (1,17). In addition to the 15 dietary intakes used by the GBD, ultra-processed food intake will be included where practicable given the increasing literature in the field linking this dietary exposure to a wide range of health outcomes (43–47). In doing so, we will use the most widely adopted food classification system based on the purpose and extent of industrial food processing, Nova, to define ultra-processed foods (42). Furthermore, we will follow best practice guidelines when categorising foods according to the Nova food classification system (48).
CMDs can be measured differently across epidemiological studies. They are most commonly assessed using self-reported symptom scales with cut-offs for diagnosis, from medical records, or using diagnostic interviews. Supplementary Table 1 outlines some commonly used tools and their recommended cut-offs for identifying symptoms of depression and anxiety. The sensitivity, specificity, and diagnostic accuracy of these tools largely depend on the population. When undertaking meta analyses, we may conduct a sensitivity analysis to determine if the magnitudes of associations change based on exposure and outcome measurement or ascertainment type.
Covariates and Confounders
To address the anticipated heterogeneity of Member Studies’ available data, a set of core confounders are provided for the purpose of generating a consistent minimally adjusted analysis. The conduct of additional analyses (e.g. using additional variables, fitting sensitivity models or exploratory analyses) is at the discretion of the study author(s) based on available data and/or reviewer requests when studies are submitted for publication.
To be included in the GLAD project, Member studies must adhere to the following iteratively adjusted models:
- Model 1: Unadjusted model, with the dietary variable (as listed in Table 2) as the exposure and depression or anxiety as the outcome.
- Model 2: Minimally adjusted model, with the dietary variable of interest, depression or anxiety, and with baseline mental disorder status (or lifetime history of mental disorders for cross-sectional studies), age, sex, and a measure of socio-economic status (SES), such as household income, employment status, or education.
- Sensitivity Model: Same as model 2, but with adjustment for total energy intake.
Of particular relevance in nutritional epidemiology is methods for adjusting for energy intake. Intake of specific nutrients are correlated with total energy intake, and so appropriate adjustment is required to disentangle the effect of the nutrient from the effect of energy. For example, an individual may alter intake of a specific nutrient by changing dietary composition, and not by changing total energy intake. As such, controlling for energy intake can reduce confounding, and failure to appropriately control for total energy intake may nullify associations between nutrient intakes and disease outcomes (49). Given the correlation between nutrient intake and energy, the application of Willett’s residual method is recommended to account for total energy intake (49). This involves fitting a regression model with the dietary variable as the outcome and total energy intake as the exposure, predicting residuals from this model, then using these predicted residuals in the analysis models (Model 1 and Model 2) (49).
Statistical Analyses
Baseline and Demographic Characteristics
To describe participant characteristics, authors of Member Studies include a table containing descriptive statistics, given as n (%) for categorical variables, and mean (standard deviation) or median (25th-75th percentile) for continuous variables. A column containing p-values will not be included, as statistical tests do not meaningfully assess differences between samples or populations, nor should they be used as the basis for assessing confounders (50).
Risk Ratios, Odds Ratios and Hazards Ratios
Each study is required to use risk ratios (estimated by Poisson regression with robust standard errors) as the primary means of assessing the association of dietary intake with CMDs. Where the estimation of a risk ratio is not possible, for example, where the total number of people exposed at baseline is not available (such as for retrospective or case-control studies) (51), odds ratios are estimated using standard logistic regression techniques. Where only time-to-event data are available, hazards ratios are used, estimated by a Cox proportional hazards model. For rigour, all effect estimates will be accompanied by a 95% confidence interval and exact p-value (52).
Models can be fitted using any statistical analysis software, including R, SAS/STAT, Stata, and SPSS. Prior to fitting the relevant models, Member Studies must assess model assumptions. Should any model assumptions be violated, an alternate approach to analysis may be selected. For example, non-linearity of continuous variables should be investigated, and transformations or non-linear models fitted where necessary. Similarly, should any of the assumptions of Cox proportional hazards models be violated, alternative strategies will be deployed to account for these violations. In addition to the models listed above, the following subgroups shall be fitted separately for each exposure-outcome pairing, where data allows:
- Sex
- Age – studies should explore the relationship of the risk factor and mental disorder with age to see if a continuous measure is appropriate. Where a continuous measure is not appropriate, studies should generate appropriate age categories for their dataset by inspecting the frequency of outcomes by age groups. As a guide, the GBD currently uses the following age categories (53):
- 0-6 days (early neonatal)
- 7-27 days (late neonatal)
- 28-365 days
- 1-4 years
- 5-9 years
- 5-year intervals from 10–95 years
- 95 years and older
- Year (for longitudinal datasets with multiple waves of follow-up)
- Country (for multi-centre studies)
- All symptom scales and diagnostic measures (for studies with depression and anxiety assessed via multibple measures).
Sensitivity Models
Multiple Testing
Given the need for multiple models to be fitted within each study, the GLAD Taskforce recommends a p-value adjustment procedure, such as the Simes method, to minimise the chance of a type 1 error and determine whether multiple testing may be influencing results (54). This should be applied to all models in a study where multiple models are required.
Outliers and Influential Data
Initial models include all available data points, with sensitivity analyses included for any influential or outlying observations. Influential observations are determined by predicting residuals, Cook’s distance, or DFBETA after fitting the relevant model. Although several cut-off points have been proposed to identify influential observations, conventional cut-offs are applied: 3 for standardised residuals (55); 4/(n-p) for Cook’s Distance (56), where n=sample size and p=number of parameters in the model; and 2/(√n) for DFBETA (57).
Missing data
Each study initially uses a complete-case model, whereby individuals with missing observations are excluded from the model. Member Studies with missing exposure, outcome and/or covariate observations conduct a missing data analysis to determine the influence of missing data. While multiple imputation approach is preferred, other methods such as k-nearest neighbour imputation or inverse probability weighting may be used where an imputation approach is not possible or appropriate (58–63).
Meta-Analyses
Upon completion of the analysis phase of the GLAD Project (Phase 2), the GLAD Project Team and Advisory Group will lead a meta-analysis (Phase 3). A separate protocol will be created for this phase of the GLAD Project. Briefly, the GLAD Project Team will conduct a systematic literature review to identify any relevant studies not already participating in GLAD. After screening abstracts and full text, data will be extracted from all identified studies and studies participating in GLAD. A random effects meta-analysis will be performed to pool results from all eligible studies. We will perform a meta-regression to quantify how the risk can vary by demographic factors, and methodological biases (64). Where studies have used different assessment methods for mental health outcomes, we will perform cross-walking or sensitivity analyses to ensure comparability between measurement scales (65).
Data Availability
The GLAD Project is a multi-country collaborative effort for which individual collaborators apply a harmonised protocol developed by the Taskforce to their own data. Data are not pooled in a shared repository or location, nor publicly available. The data custodians of each dataset may or may not make their data available upon request.
Ethics and Dissemination
Each member study will have obtained ethical approval from institutional ethics review boards. Details for ethics for each study can be found in references (18–40). Results from the GLAD project will be disseminated 1) to the GBD study, which may be used by governments worldwide to inform policy and global researchers to guide future research; 2) by presenting at local, national and international conferences; and 3) via peer-review publications, including separate results from each member study and a large meta-analysis.