Professional Documents
Culture Documents
Bossuyt - E&E 2021 - 3. Conditioning - Handout
Bossuyt - E&E 2021 - 3. Conditioning - Handout
Conditioning &
Standardization
Patrick M Bossuyt
Outline
1. Single confounder
2. Population versus Conditional Effects
3. Multiple confounders
Hypothetical example: breastfeeding
SES
breastfeeding IQ at 7.5
Breastfed
112.5
98 99 100 100 101 102 110 113 115 115 117 120
Not Breastfed
100
93 94 94 96 96 97 103 107
98 99 100 100 101 102 110 113 115 115 117 120
? ? ? ? ? ? ? ?
Not Breastfed
? ? ? ? ? ? ? ? ? ? ? ?
93 94 94 96 96 97 103 107
100
98 99 100 100 101 102 110 113 115 115 117 120
5
95
93 94 94 96 96 97 103 107
98
93
93 94 94 96 96 97
93 94 94 96 96 97 103 107
107.5
98 99 100 100 101 102 110 113 115 115 117 120
7.5
100
100 115
98 99 100 100 101 102 110 113 115 115 117 120
5 10
95 105
93 94 94 96 96 97 103 107
100 115
98 99 100 100 101 102 110 113 115 115 117 120
5 10
95 105
93 94 94 96 96 97 103 107
100 115
98 99 100 100 101 102 110 113 115 115 117 120
5 10
95 105
93 94 94 96 96 97 103 107
12/20 8/20
IQ at 7.5
Population effect - standardization
100 115
98 99 100 100 101 102 110 113 115 115 117 120
5 10
95 105
93 94 94 96 96 97 103 107
.6 .6 × 5 + .4 × 10 = 7 .4
Population effect - standardization
• Estimate conditional effect in each stratum
• Equivalent alternative:
estimate outcome for each participant with exposure – average
estimate outcome for each participant without exposure – average
difference between these averages is population effect
Population effect - standardization
98 99 100 100 101 102 110 113 115 115 117 120
5 10
93 94 94 96 96 97 103 107
12×5+8×10
12 =7 8
20
IQ at 7.5
Single confounder
Population effect - Inverse Probability Weighting (IPW)
Population effect - IPW
.5 .75
98 99 100 100 101 102 110 113 115 115 117 120
.5 .25
93 94 94 96 96 97 103 107
IQ at 7.5
Population effect - IPW
1 1 1 1
× 98 + × 99 + … + × 110 + ⋯ + × 120
.5 .5 .75 .75 = 106
20
1/.5 1/.75
98 99 100 100 101 102 110 113 115 115 117 120
1 1 1 1
.5
×93+.5
×94+⋯+ .25
×103+.25
×107
= 99
20
1/.5 1/.25
93 94 94 96 96 97 103 107
106 – 99 = 7
IQ at 7.5
Population effect - IPW
1/.5 1/.75
98 99 100 100 101 102 110 113 115 115 117 120
1/.5 1/.25
93 94 94 96 96 97 103 107
IQ at 7.5
Population effect - IPW
1/.5 1/.75
98 99 100 100 101 102 110 113 115 115 117 120
1/.5 1/.25
93 94 94 96 96 97 103 107
IQ at 7.5
Population effect - IPW
1/.5 1/.75
98 99 100 100 101 102 110 113 115 115 117 120
1/.5 1/.25
93 94 94 96 96 97 103 107
93 94 94 96 96 97
IQ at 7.5
Population effect - IPW
1/.5 1/.75
98 99 100 100 101 102 110 113 115 115 117 120
1/.5 1/.25
93 94 94 96 96 97 103 107
IQ at 7.5
Population effect - IPW
1/.5 1/.75
98 99 100 100 101 102 110 113 115 115 117 120
1/.5 1/.25
93 94 94 96 96 97 103 107
IQ at 7.5
Population effect - IPW
• Assign weight to each participant:
inverse of probability of being exposed (for those exposed)
inverse of probability of being unexposed (for those unexposed)
• “Table 4 shows our estimate of the effect of statin treatment on the risk of
a stroke in the next 12 months for women aged 60-80 with a transient
ischemic attack, without history of cardiovascular disease. Statins lead to a
statistically significant reduction in the stroke risk.”
• Risk factor
• Risk indicator
• Risk marker
• Determinant
Multiple confounders
Multiple confounders
Conditional effect - restriction
Multiple confounders
Matching
Matching
• Can be challenging with many confounders
Examples:
- Logistic regression
- Cox proportional-hazards regression
- Linear regression
- Poisson regression
Logistic regression
𝑝𝑝
ln = 𝛼𝛼 + 𝛽𝛽1 𝑥𝑥1 + + 𝛽𝛽2 𝑥𝑥2 + ⋯ + 𝛽𝛽𝑛𝑛 𝑥𝑥𝑛𝑛
1 − 𝑝𝑝
Conditional effect: Modeling
• Beware:
• Adding covariates changes estimand (the effect that is estimated) –
different conditional odds ratio
• Beware
• Model misspecification
• E.g. “adjusted for age” ?
Modeling: prediction versus causal modeling
• Prediction:
• “Anything goes”
• Causal modeling:
• Only essential set of confounders
Multiple confounders
Propensity scores
Randomized Clinical Trial
Active Outcome
Study
Population Randomize
Group
Control Outcome
Randomisation
• Pr(E | Z1,Z2,….)
• X: exposure Y: outcome
Balance diagnostics
continuous proportions
• Advantages:
• Modeling blinded to outcome status – minimize bias
• With propensity as covariate: closer to population effect
• More robust against model specification
• But variance may be larger than fully specified model
• Both may suffer from unmeasured confounders
• Always positivity check needed
(evaluate overlap between groups)
Propensity Scores
• Must include confounders!
98 99 100 100 101 102 110 113 115 115 117 120
1/0.5 1/0.75
93 94 94 96 96 97 103 107
propensity : 0.5 propensity : 0.75
IPW
• creates two pseudo-populations
that represent the study groups we would have observed if:
(i) everyone had been exposed and
(ii) no-one had been exposed