Download as pptx, pdf, or txt
Download as pptx, pdf, or txt
You are on page 1of 35

How to deal with missing information:

advanced statistical methods


Dealing with unmeasured confounding

Antoine Pariente
EUR DPH 2019-20
Plan

1. Propensity scores – some more info

2. Disease Risk Score

3. pre-Conclusion

4. Other techniques

2
1. Propensity scores

3
Definition

› Theoretical probability of receiving treatment

› With respect to several covariates

› Score between 0 and 1 which represents the relationship


between covariates and exposure (binary)

› In theory: for patients with a comparable score covariate-


related characteristics are similarly distributed between the
exposed and unexposed group

4
Definition

Exposition Event

Propensity Confounders
score

5
Conception

› In several stages
- Estimation of the propensity score
• Construction

• Quality check

- Using the propensity score

- Estimate of treatment effect

6
Conception

› Score construction: choice of covariates


- All types possible
• Demographic characteristics

• Medical characteristics

• Co-prescriptions ...

- Any significant or potential confounders that create an optimal


balance between groups

- Review of the literature +++

7
Conception

› Score construction
- Characteristics prior to exposure

Exposure +
follow-up start

Covariates Follow-up
assessment

8
Conception

› In practice
- Logistic regression
- Exposure: variable to explain / covariates: explanatory variables

› Validation of the score


- Classical logistic regression
• Suitability: Hosmer-Lemeshow
• Discrimination: AUC statistic or c

9
Conception

› Score validation
- Score distribution between 2 groups

- Covariates balance between 2 groups


+++ for main confounders
10
Check covariate distribution

EUR DPH 2019-20


11
Check for balance:
use of standardized difference; threshold: 10%

EUR DPH 2019-20


12
How to use the PS

› Several possibilities
- adjustment
• continuous variable in the regression model

- pairing
• exposed pair - unexposed

- stratification
• percentile, quintile, ...

13
How to match on PS

› optimal matching

› greedy nearest neighbor matching


w/wo
› greedy nearest neighbor matching replacement
within specified caliper widths

Caliper:
› Maximal acceptable distance

EUR DPH 2019-20


14
Advantages and limits

› Control of confusion bias (indication +++)


- Interesting for rare or multiple events

› limitations
- Do not control unmeasured or unknown confounders (at least under
these uses)

- Do not control time-dependent confounders

- In case of matching: loss of information / accuracy due to unmatched


subjects

15
High dimension Propensity Score
(hd-PS)
Hd PS

› But why
- PS performance depends on its ability to represent unmeasured
confusion
- The more variables considered for PS score, the better?
• IF these variables do not all represent the same thing -> several dimensions
• IF they are sufficiently associated with the event
• IF they are sufficiently frequent

› So how
- select variables that are sufficiently frequent in each dimension
- Select variables sufficiently associated with the event in each dimension
+/- sufficiently associated with the exposure

17
Hd PS

› In the medico-administrative databases

› Automatic selection of potential covariates of interest from


thousands of covariates in the database

› Set of several hundred covariates

› Benefit: control of unmeasured confounders

18
Master 2 de Pharmacologie – Université de Bordeaux
19
Hd PS example:
GI complications of t-NSAIDs vs. COX-2 inhibitors

20
Let’s go through the paper

› Garbe et al. 2013

EUR DPH 2019-20


21
Marginal structural models

› Use IPTW weighting


- IPTW = Inverse Probability of Treatment Weights

› For each time interval


- Likelihood of receiving treatment conditional on previous treatment and
covariates
- Weighting in the model

› Benefit: control of time-dependent confounders

› Disadvantage: difficult to implement

22
2. Disease Risk Score

23
Definition

› Theoretical probability of presenting an event

› With respect to several covariates (risk factors)

› Score between 0 and 1 which represents the relationship


between the covariates and the occurrence of the event
(binary)

› In theory: for patients with a comparable score, covariate-


related characteristics are distributed similarly between those
presenting the event and those not presenting the event

24
Definition

Exposition Event

Confounders Disease risk


score

25
Conception

› Similar to PS

› Construction of the disease risk score


- Unexposed Patients
- Whole cohort
- Historical cohort

Adapté de J. Gagne Harvard School of Public Health


26
Conception

› Construction of the disease risk score


- Non-exposed patients
- Estimation of score in non-exposed patients and application of score in
non-exposed and exposed patients

› Advantages
- Stronger and simpler

› Limitations
- In most cases no true non-exposed

› Pb if rare event among non-exposed (number of variables to


be included may be limited)
Adapté de J. Gagne Harvard School of Public Health
27
Conception

› Construction of the disease risk score

› Whole cohort
- Estimation of score in non-exposed and exposed

› Advantages
- More events,
- Score built and applied on the same population

› limitations
- Score "contaminated" by the effect of treatment (no bp if treatment does
not change the effect of covariates on the occurrence of the event)

Adapté de J. Gagne Harvard School of Public Health


28
Conception

› Construction of the disease risk score


- Historical Cohort
• Comparison for 2 drugs Application of
disease risk score

Estimation of
disease risk score Study
Population

Drug marketing

Adapté de J. Gagne Harvard School of Public Health


29
Conception

› Construction of the disease risk score

› Historical Cohort
- Advantages
• In theory: more events than in the entire cohort

• Can be used to assess the level of basic confusion bias

- limitations
• Data availability for the historical cohort

› Pb if the association between the risk factors and the occurrence of the
event changes over time
Adapté de J. Gagne Harvard School of Public Health
30
Advantages and limits

› Idem propensity score, except


- SP: interesting for rare or multiple events but binary exposure

- DRS: interesting for rare or multiple exposures but binary event

31
pre-Conclusion

32
pre-Conclusion - 1

› Interesting analysis techniques compared to classical


regressions

› BUT remain methodological tools

› Do not control all the biases

33
pre-Conclusion - 2

› And other techniques exist


- Statistics
• Prior event rate adjustment
• Instrumental variable
• Sensitivity analysis
• Propensity score calibration, Two-stage sampling, Multiple Imputation
- Epidemiological
• Case-only design
- Self-controlled case series
- Case Cross-over
- Case time-control

• Negative control
• Ecological studies ... (so-so for confounders)

FORMeDOC © 2014 - Tous droits réservés 34


So there might be use of PS that allows dealing with
UC

› Let’s go through Silenou et al.

EUR DPH 2019-20


35

You might also like