1 of 26

Dr Roopal Desai

Clinical Fellow

University College London (UCL)

roopal.desai.15@ucl.ac.uk

Dr Verena Zuber

Lecturer in Biostatistics

Imperial College London

v.zuber@imperial.ac.uk

Background

Literature review

Results and Discussion

MR study

Results and Discussion

Examining the Lancet Commission on risk factors for dementia using Mendelian Randomization

2 of 26

Combined PAF of 45%

Early-life risk factor

Education

Lancet Commission review

.

5

1

2

4

3

Risk factors

for

Dementia

Lancet Commission Risk Factors for Dementia

Mid-life risk factors

Hearing loss

Hypertension

Obesity

Alcohol*

TBI*

High LDL cholesterol**

Late-life risk factors

Smoking

Depression

Physical inactivity

Low Social Contact

Diabetes

Air Pollution*

Vision Loss**

3 of 26

Lancet Commission report 2024

Lancet Commission report 2020

4 of 26

PAF assumption

Causal associations

Are the associations causal?

Epidemiological studies

Reverse causality
Confounds

Critique of Lancet Commission

5 of 26

Mendelian Randomization

Randomized control trial

Mendelian Randomization

Randomization

Random allocation of alleles

Intervention

Control

Effect allele

Control

Confounders equal between groups

Outcomes compared between groups

6 of 26

What do MR studies tell us about modifiable risk factors and dementia?

Since 2015 37 studies using MR

Ten of the Lancet Commission risk factors

No studies on air pollution and traumatic brain injury

Six different dementia outcomes

160 separate MR analyses

136 unique analyses

7 of 26

What do MR studies tell us about modifiable risk factors and dementia?

Evidence coming from MR studies can be graded:

‘robust’
‘probable’
‘suggestive’
‘insufficient’

106 analyses evaluable
Almost 50% of these effect sizes are
non-significant

8 of 26

What do MR studies tell us about modifiable risk factors and dementia?

Educational attainment (17 analyses)

Excessive alcohol consumption (8 analyses)

Hypertension (6 analyses)

Obesity (19 analyses)

9 of 26

What do MR studies tell us about modifiable risk factors and dementia?

Hearing loss (3 analyses)

Low social contact (12 analyses)

Physical inactivity (19 analyses)

Depression (3 analyses)

10 of 26

What do MR studies tell us about modifiable risk factors and dementia?

Smoking (7 analyses)

Diabetes (15 analyses)

11 of 26

What does all of this tell us?

Overall evidence coming from MR studies is weak

Almost half of the estimates across the studies indicated non-significant effect sizes

Relatively strongest genetic evidence is for diabetes and less education being causal risk factors

Where effect sizes are significant for many risk factors the evidence is contradictory

What explains these results?

Methodological issues with MR studies?
Poor construct definition?

Associations observed in epidemiological studies are artefacts?
Reverse causality

Risk factors or disease markers?

12 of 26

13 of 26

AIMS

Use MR to study links using most up-to-date data with the largest datasets a standardized analysis pipeline
Use different dementia outcomes

AD, FTD & DLB

14 of 26

Study design

GWAS data for 10/12 LC risk factors
Two-sample MR

Data sources

Publicly available GWAS datasets

Mendelian randomization analysis

SNPs selected
F statistics >10
Clumped & pruned
Wald ratio calculated and combined using the IVW method

Sensitivity analyses

Methods

15 of 26

Results – Alzheimer’s Disease

16 of 26

Results – Dementia with Lewy Bodies

17 of 26

Results – Frontotemporal Dementia

18 of 26

Revisiting: Results – Alzheimer’s Disease

Protective effect of established risk factors

19 of 26

Open question: Why are some MR effect estimates of certain risk factors on dementia protective?

Answer 1 based on shared biology:

Colocalization in the ACE-gene region (Bone et al 2021)

Answer 2 from life course epidemiology: Survivor bias

Birth Early life Mid life Late life

Average age of first heart attack

Average age of onset of dementia

0 65 80

20 of 26

Open discussion: How can we make MR more robust to survivor bias?

G

X₁

X₂

Y

U₁

U₂

Multivariable MR:

Extends the standard univariable MR paradigm with one exposure to multiple exposures
Estimates the direct effect of one exposure conditional on other (correlated) exposures

Aims:

To select likely causal exposures

IV selection: Based on genetic variants associated with any of the exposures

To account for pleiotropic pathways

IV selection: Based on the primary exposure

21 of 26

Open discussion: How can we make MR more robust to survivor bias?

Multivariable MR:

Extends the standard univariable MR paradigm with one exposure to multiple exposures
Estimates the direct effect of one exposure conditional on other (correlated) exposures

Aims:

To select likely causal exposures

IV selection: Based on genetic variants associated with any of the exposures

To account for pleiotropic pathways

IV selection: Based on the primary exposure

G

X₁

X₂

Y

U₁

U₂

22 of 26

Open discussion: How can we make MR more robust to survivor bias?

Multivariable MR implementation:

Primary exposure of interest: Blood pressure
Potential pleiotropic pathway: Coronary artery disease

IV selection:

Based on the primary exposure of interest:

-> 219 independent genetic variants associated with blood pressure

G

Genetic variants associated with blood pressure

X₁

Blood pressure

X₂

Coronary artery disease

Y

U₁

U₂

23 of 26

Example: Are blood pressure and smoking protective?

24 of 26

Open discussion: How can we make MR more robust to survivor bias?

Limitations of multivariable MR to adjust for potential survivor bias:

Arbitrary choice which pleiotropic pathways to adjust for

Question: Which competing events are necessary to adjust for?

Less powerful than univariable MR (more parameters to estimate)
Odds ratios from binary (disease) phenotypes introduce issues of non-collapsibility

Other ways to tackle survivor bias in MR:

Individual-level MR crossed with a competing risks time-to-event model
Limitation: Data availability and long follow-up

25 of 26

Discussion of Lancet Commission Risk Factors for Dementia

Apparent protective effects of increased obesity, blood pressure, & smoking

Survivor bias?
Null-results from multivariable MR

Educational attainment inconsistent findings

Heterogeneity in GWAS due to use of proxy diagnosis?

Limitations of current MR methodology

Weak instrument bias/no genetic instruments available for some risk factors
Univariable MR and all sensitivity analysis are not immune to survivor bias
Dynastic effects can only be uncovered using within-family design

Limitations of genetic data

Heterogeneity in phenotypes
Conceptualization of risk factors

-> Call for better data and better methods

-> More efforts into triangulation using different methodologies and data sources

26 of 26

With thanks to:

The Adapt Lab @ UCL
Prof Joshua Stott
Prof Georgina Charlesworth
Dr Emma Anderson
Dr Amber John
Dr Natalie Marchant
Dr Verena Zuber