Hypothetical blood-pressure-lowering interventions and risk of stroke and dementia
L. Paloma-Rojas Saunero
Epidemiology Department
Causal Inference, Neuro-Epi Group
Hello My name is Paloma Rojas-Saunero, I am a PhD candidate at Erasmus MC and today I will present my work called: Hypothetical blood-pressure-lowering interventions and risk of stroke and dementia
Introduction
- Lowering blood pressure reduces the risk of stroke and is discussed as a potential intervention to reduce dementia.
Introduction
- Lowering blood pressure reduces the risk of stroke and is discussed as a potential intervention to reduce dementia.
- Clinical trials have focused on high-risk subjects, have limited follow-up and focus primarily on the intention to treat effect.
Introduction
- Lowering blood pressure reduces the risk of stroke and is discussed as a potential intervention to reduce dementia.
- Clinical trials have focused on high-risk subjects, have limited follow-up and focus primarily on the intention to treat effect.
- There is limited evidence on the sustained effect of interventions related to lowering blood pressure in a general population.
Aim
To estimate the sustained effect of hypothetical interventions lowering systolic blood pressure (SBP) on the 15-year risk of stroke and dementia.
To accomplish this goal we will emulate a target trial using data collected in the Rotterdam Study
Target trial specifications
| Section | Target trial | Emulation using observational data |
|---|---|---|
| Eligibility criteria | < 80 years old, no cognitive impairment, no history of stroke, dementia diagnosis and other related diseases. | Same + MMSE above 26 at baseline |
To emulate the target trial we first need to specify each of the study design sections:
So first:
We would include all individuals younger than 80 yo, with no history of stroke, dementia, cognitive impairment or related diseases.
In our emulation defined as eligible to all individuals who met this criteria, and excluded those with MMSE above 26
Target trial specifications
| Section | Target trial | Emulation using observational data |
|---|---|---|
| Eligibility criteria | < 80 years old, no cognitive impairment, no history of stroke, dementia diagnosis and other related diseases. | Same + MMSE above 26 at baseline |
| Treatment strategies | 0. Natural course (comparison arm) 1. Keep SBP < 120 mmHg 2. Keep SBP < 140 mmHg 3. Reduce SBP by 20% if >140 mmHg 4. Quit smoking 5. Joint 3 + 4 |
Same |
Next, we specified the following interventions:
- First the natural course, in which there is no intervention implemented, this is the comparison arm for the rest of strategies
- We defined 2 strategies keeping blood pressure under 120 and 140 over time, as defined in previous trials...
- Next we defined a strategy in which we would lower sbp by 20% if above 140, which we thought it would be more realistic
- We additionally added an intervention on quit smoking
- and a joint intervention that combines lowering bp and quit smoking... All these interventions are sustained over 15 years of fu and are dinamic which means that for example... if a person enters the trial with sbp below 140 no intervention is forced at that time point, but if in the future timepoints the person meets a criteria, then blood pressure would be lowered as defined in the intervention
Target trial specifications
| Section | Target trial | Emulation using observational data |
|---|---|---|
| Eligibility criteria | < 80 years old, no cognitive impairment, no history of stroke, dementia diagnosis and other related diseases. | Same + MMSE above 26 at baseline |
| Treatment strategies | 0. Natural course (comparison arm) 1. Keep SBP < 120 mmHg 2. Keep SBP < 140 mmHg 3. Reduce SBP by 20% if >140 mmHg 4. Quit smoking 5. Joint 3 + 4 |
Same |
| Follow-up | From year of first visit until 15 years of follow-up, or year of stroke/dementia or death, which ever happened first | Same + visit process simulation |
| Outcome | Stroke and Dementia (Death as a competing event) | Same |
For this reason our follow up is of...
Target trial specifications
| Section | Target trial | Emulation using observational data |
|---|---|---|
| Eligibility criteria | < 80 years old, no cognitive impairment, no history of stroke, dementia diagnosis and other related diseases. | Same + MMSE above 26 at baseline |
| Treatment strategies | 0. Natural course (comparison arm) 1. Keep SBP < 120 mmHg 2. Keep SBP < 140 mmHg 3. Reduce SBP by 20% if >140 mmHg 4. Quit smoking 5. Joint 3 + 4 |
Same |
| Follow-up | From year of first visit until 15 years of follow-up, or year of stroke/dementia or death, which ever happened first | Same + visit process simulation |
| Outcome | Stroke and Dementia (Death as a competing event) | Same |
| Causal contrast | What would have been observed if all individuals adhered to their assigned strategy over 15 years (Per protocol effect) | Same |
Last, we defined the causal contrast as the sustained effect, which resembles the PPE
Rotterdam Study
Population-based cohort, 4930 eligible participants
SBP, smoking and other covariates were collected at baseline (1990 - 1993) and follow-up visits (1993-1995, 1997-1999, 2002-2005)
Outcomes come from different sources: screening at each visit (plus further evaluation) + electronic clinical records + municipal registries
Rotterdam Study
Population-based cohort, 4930 eligible participants
SBP, smoking and other covariates were collected at baseline (1990 - 1993) and follow-up visits (1993-1995, 1997-1999, 2002-2005)
Outcomes come from different sources: screening at each visit (plus further evaluation) + electronic clinical records + municipal registries
57% women, 66 (SD: 7) years
24% current smokers
51% achieved a higher education
57% had history of hypertension
27% with hypertension medication
7% history of heart disease
13% history of diabetes
Statistical analysis
What would have happened had everyone was randomized and had adhered to each intervention (g)?
Parametric G-formula
Allows presence of time-varying confounding feedback
To answer the question:... we implemented the gformula... In brief, in our data we know that A (sbp) affects comorbidities and behavior over time, which subsequently affect BP at the next visits...and we cant use traditional methods to account for this confounding feedback While, given all collected information over time, the g-formula helps us to account for tv confounding and simulate treatment regimes...
Statistical analysis
What would have happened had everyone was randomized and had adhered to each intervention (g)?
Parametric G-formula
Allows presence of time-varying confounding feedback
To answer the question:... we implemented the gformula... In brief, in our data we know that A (sbp) affects comorbidities and behavior over time, which subsequently affect BP at the next visits...and we cant use traditional methods to account for this confounding feedback While, given all collected information over time, the g-formula helps us to account for tv confounding and simulate treatment regimes...
Statistical analysis
What would have happened had everyone was randomized and had adhered to each intervention (g)?
Parametric G-formula
Allows presence of time-varying confounding feedback
To answer the question:... we implemented the gformula... In brief, in our data we know that A (sbp) affects comorbidities and behavior over time, which subsequently affect BP at the next visits...and we cant use traditional methods to account for this confounding feedback While, given all collected information over time, the g-formula helps us to account for tv confounding and simulate treatment regimes...
Statistical analysis
What would have happened had everyone was randomized and had adhered to each intervention (g)?
Parametric G-formula
Allows presence of time-varying confounding feedback
A = SBP, L = Fixed covariates: age, sex, education, SBP, history of diabetes and heart disease. Time-varying covariates: visit process, smoking status, SBP, BMI, HT medication, total cholesterol and diagnosis of diabetes, heart disease, Parkinson disease, Parkinsonism, TIA, dementia or cancer, Y = Stroke / Dementia
To answer the question:... we implemented the gformula... In brief, in our data we know that A (sbp) affects comorbidities and behavior over time, which subsequently affect BP at the next visits...and we cant use traditional methods to account for this confounding feedback While, given all collected information over time, the g-formula helps us to account for tv confounding and simulate treatment regimes...
So we considered several fixed and timevarying covariates that relate to clinical comorbidities and
Results: Stroke
At 15 years we observe that all strategies would reduce the risk of stroke, and joing strategy that reduces blood pressure and smoking would lower the risk even more.
Since we considered death as competing event, part of the effect we observe is mediated through the effect of our interventions in the risk of death...as such...
in this case were compared the joint strategy with the natural course, we observe that the risk of death is also lower if everyone follows the treatment strategy
Subgroup analysis for stroke
This is relevant because, in the subgroup analysis, as we observe, the effect was similar for different subgroups but the joint strategies would have a higher impact among the group of people who is younger which may be because mortality is smaller in this group.
Results: Dementia
In contrast, we observe that non of the strategies had an impact on the risk of dementia, and as we observe, the point estimates are above one
However, we observe that in this case, the strategies would be preventing the risk of death
Subgroup analysis for dementia
We believe that this is the reason why, in the subgroup analysis, among women, who have a mortality distribution that is lower, the point estimate is below one, and in men who have a higher burden of mortality... the risk of dementia under the joint strategy is even higher compared to the natural course
Discussion
Lowering blood pressure and quitting smoking reduces the risk of stroke, as described previously in literature.
We did not observe the same effect in dementia, though the effect may be mediated on how interventions decrease the risk of death.
Discussion
Lowering blood pressure and quitting smoking reduces the risk of stroke, as described previously in literature.
We did not observe the same effect in dementia, though the effect may be mediated on how interventions decrease the risk of death.
Estimates rely on important assumptions:
Time-varying data was sufficient
No modeling misspecification (consider data generation process)
Discussion
Lowering blood pressure and quitting smoking reduces the risk of stroke, as described previously in literature.
We did not observe the same effect in dementia, though the effect may be mediated on how interventions decrease the risk of death.
Estimates rely on important assumptions:
Time-varying data was sufficient
No modeling misspecification (consider data generation process)
We didn't specified how SBP would be lowered, this represents a weighted average of strategies, determined by the frequency of these in the studied population.
in this case, the gformula is very sensitve so had to be very careful on specifying how data was simulated depending on how data was generated, so for data that was collected at the visits we modeled the visit process, and for data collected from ECR a different modelling strategy
while there are certainly limitations in terms of ambiguity to the interventions studied in the current paper, this is a step forward into the types of interventions we may consider in practice at a population level, and we need to do our best to match the research questions to the data
Discussion
Lowering blood pressure and quitting smoking reduces the risk of stroke, as described previously in literature.
We did not observe the same effect in dementia, though the effect may be mediated on how interventions decrease the risk of death.
Estimates rely on important assumptions:
Time-varying data was sufficient
No modeling misspecification (consider data generation process)
We didn't specified how SBP would be lowered, this represents a weighted average of strategies, determined by the frequency of these in the studied population.
Future studies are needed to disentangle treatment variation relevance and build upon this initial study.
in this case, the gformula is very sensitve so had to be very careful on specifying how data was simulated depending on how data was generated, so for data that was collected at the visits we modeled the visit process, and for data collected from ECR a different modelling strategy
while there are certainly limitations in terms of ambiguity to the interventions studied in the current paper, this is a step forward into the types of interventions we may consider in practice at a population level, and we need to do our best to match the research questions to the data
Acknowledgments
Saima Hilal
Eleanor J. Murray
Roger W. Logan
Arfan Ikram
Sonja A. Swanson