Loading [MathJax]/jax/output/CommonHTML/jax.js
+ - 0:00:00
Notes for current slide
Notes for next slide

Considering questions before methods in dementia research with competing events and causal goals


L. Paloma Rojas-Saunero MD, PhD
Postdoctoral scholar
Mayeda Research Group, Department of Epidemiology

1

Hill et al. Ethnicity and disease. 2015

2

www.express.co.uk, 2016

3

Medscape, 2019

4

Cause-specific vs. subdistribution HR

Austin et al. Circulation.2016

5

Systematic review

Searching criteria

  • Original research published between Jan/2018 to Dec/2019

  • Dementia/AD & longitudinal/cohort & hazard/risk

  • Alzheimer’s and Dementia, Annals of Neurology, BMJ, Neurology, JAMA, Jama Neurology, Lancet, Lancet Neurology

6

Systematic review

Searching criteria

  • Original research published between Jan/2018 to Dec/2019

  • Dementia/AD & longitudinal/cohort & hazard/risk

  • Alzheimer’s and Dementia, Annals of Neurology, BMJ, Neurology, JAMA, Jama Neurology, Lancet, Lancet Neurology

Eligibility criteria

  • Time-to-dementia/AD as primary or co-primary outcome

  • With a clear exposure/intervention, and uses methods to handle confounding

  • Not a descriptive or predictive aim

6

Results

Out of 57/209 papers included:

7

Results

Out of 57/209 papers included:

  • 56% report death numbers, 18% death by exposure level
7

Results

Out of 57/209 papers included:

  • 56% report death numbers, 18% death by exposure level

  • 47% do not include any description about death in the methods section, 26% consider it a sensitivity example, 14% only mention it was treated as a censoring event

7

Results

Out of 57/209 papers included:

  • 56% report death numbers, 18% death by exposure level

  • 47% do not include any description about death in the methods section, 26% consider it a sensitivity example, 14% only mention it was treated as a censoring event

  • 87% use Cox PH models, 93% present hazard ratios

7

Results

Out of 57/209 papers included:

  • 56% report death numbers, 18% death by exposure level

  • 47% do not include any description about death in the methods section, 26% consider it a sensitivity example, 14% only mention it was treated as a censoring event

  • 87% use Cox PH models, 93% present hazard ratios

  • 86% innacurate interpretations (e.g. "risks")

7

8

9

10

Total effect


What is the risk of dementia at 20 years of follow-up had all individuals stopped smoking, compared to had all individuals continued smoking?

Pr[Ya=120]Pr[Ya=020]

11

Total effect


What is the risk of dementia at 20 years of follow-up had all individuals stopped smoking, compared to had all individuals continued smoking?

Pr[Ya=120]Pr[Ya=020]

11

Controlled direct effect


What is the risk of dementia at 20 years of follow-up had all individuals stopped smoking and not died throughout the study period, compared to had all individuals continued smoking in adulthood and not died throughout the study period?

Pr[Ya=1,d19=020]Pr[Ya=0,d19=020]

12

Controlled direct effect


What is the risk of dementia at 20 years of follow-up had all individuals stopped smoking and not died throughout the study period, compared to had all individuals continued smoking in adulthood and not died throughout the study period?

Pr[Ya=1,d19=020]Pr[Ya=0,d19=020]

12

Controlled direct effect


What is the risk of dementia at 20 years of follow-up had all individuals stopped smoking and not died throughout the study period, compared to had all individuals continued smoking in adulthood and not died throughout the study period?

Pr[Ya=1,d19=020]Pr[Ya=0,d19=020]

13

Identifiability assumptions

Assumption Total Effect Controlled direct effect
Exchangeability assumption needed for death (competing events)? Not needed At each k + 1, conditional on the measured past, death is independent of future counterfactual outcomes had everyone followed A = a and death was eliminated.
14

Identifiability assumptions

Assumption Total Effect Controlled direct effect
Exchangeability assumption needed for death (competing events)? Not needed At each k + 1, conditional on the measured past, death is independent of future counterfactual outcomes had everyone followed A = a and death was eliminated.
Positivity assumption needed for death (competing events)? Not needed For any possibly observed level A = a and covariate history amongst those remaining uncensored (alive) and free of dementia diagnosis through k, some individuals continue to remain alive through _k + 1_.
15

Identifiability assumptions

Assumption Total Effect Controlled direct effect
Exchangeability assumption needed for death (competing events)? Not needed At each k + 1, conditional on the measured past, death is independent of future counterfactual outcomes had everyone followed A = a and death was eliminated.
Positivity assumption needed for death (competing events)? Not needed For any possibly observed level A = a and covariate history amongst those remaining uncensored (alive) and free of dementia diagnosis through k, some individuals continue to remain alive through _k + 1_.
Consistency assumption needed for death (competing events) Not needed An intervention that “eliminates death (competing events)” is well-defined.
16

Application

  • Participants from Rotterdam Study I, recruited between 1990-1993 and with follow-up data

    • Current and former smokers

    • No prior history of dementia diagnosis

    • Complete information at baseline

  • Final sample size of 4179 participants

  • Mean age at baseline of 62 years

  • 368 developed dementia and 1318 died

17

Analysis plan

  • For confounding: Inverse probability weighting for treatment (IPTW)
18

Analysis plan

  • For confounding: Inverse probability weighting for treatment (IPTW)

  • Total effect: Cause-specific cumulative incidence / Aalen-Johansen estimator + IPTW

18

Analysis plan

  • For confounding: Inverse probability weighting for treatment (IPTW)

  • Total effect: Cause-specific cumulative incidence / Aalen-Johansen estimator + IPTW

  • Controlled direct effect: Kaplan-Meier + Inverse probability censoring weights + IPTW

18

Analysis plan

  • For confounding: Inverse probability weighting for treatment (IPTW)

  • Total effect: Cause-specific cumulative incidence / Aalen-Johansen estimator + IPTW

  • Controlled direct effect: Kaplan-Meier + Inverse probability censoring weights + IPTW

  • Bootstrapping for confidence intervals

Semi-parametric or parametric alternatives are also possible

18

Results

Causal effect Risk Difference (95%CI) Risk Ratio (95%CI)
Total effect on dementia 2.1 (-0.1, 4.2) 1.21 (0.99, 1.50)
Controlled direct effect on dementia (with IPCW for death) -2.6 (-6.1, 0.8) 0.86 (0.72, 1.05)
Total effect on mortality -17.4 (-20.5, -14.2) 0.68, (0.63, 0.72)
19

Results

Assumption Risk Difference (95%CI) Risk Ratio (95%CI)
Evoking unconditional exchangeability assumption for censoring -0.7 (-3.3, 2.2) 0.96 (0.82, 1.16)
Evoking conditional exchangeability assumption on baseline covariates for censoring -1.5 (-4.6, 1.8) 0.92 (0.78, 1.12)
Evoking conditional exchangeability assumption on baseline and time-varying covariates for censoring -2.7 (-6.1, 0.8) 0.86 (0.7, 1.1)
20

21

Other possible estimands

  • Survivors average causal effect: The risk of dementia on a subgroup of individuals who would never experience the competing event.
22

Other possible estimands

  • Survivors average causal effect: The risk of dementia on a subgroup of individuals who would never experience the competing event.

  • Separable effects: Effects of modified treatments motivated by the physical decomposition of the exposure assumed to operate on dementia and death through separate pathways.

22

Other possible estimands

  • Survivors average causal effect: The risk of dementia on a subgroup of individuals who would never experience the competing event.

  • Separable effects: Effects of modified treatments motivated by the physical decomposition of the exposure assumed to operate on dementia and death through separate pathways.

  • Composite outcome of dementia and death

22

Conclusion

  • When competing events are present there is more than one way to consider them as part of the primary research question.
23

Conclusion

  • When competing events are present there is more than one way to consider them as part of the primary research question.

  • Let the question guide the most appropiate methods and estimators.

23

Conclusion

  • When competing events are present there is more than one way to consider them as part of the primary research question.

  • Let the question guide the most appropiate methods and estimators.

  • For various reasons, risks and survival curves should be preferred over hazards.

23

Conclusion

  • When competing events are present there is more than one way to consider them as part of the primary research question.

  • Let the question guide the most appropiate methods and estimators.

  • For various reasons, risks and survival curves should be preferred over hazards.

  • Collaborative work between clinical researchers, epidemiologists and statisticians should narrow the gap between methods development and applied research.

23

Thank you! Gracias!



 lp.rojassaunero@ucla.edu

  @palolili23

  @palolili23

24

Hill et al. Ethnicity and disease. 2015

2
Paused

Help

Keyboard shortcuts

, , Pg Up, k Go to previous slide
, , Pg Dn, Space, j Go to next slide
Home Go to first slide
End Go to last slide
Number + Return Go to specific slide
b / m / f Toggle blackout / mirrored / fullscreen mode
c Clone slideshow
p Toggle presenter mode
t Restart the presentation timer
?, h Toggle this help
Esc Back to slideshow