antonella mazzei abba, university of neuchatel jixian
TRANSCRIPT
Antonella Mazzei Abba, University of Neuchatel
Jixian (Jason) Wang, Celgene
PSI conference, London, 2017
Big data and observational data
The problem of confounding and missing covariates
An introduction to propensity scores
Propensity score calibration for data with missing covariates
Our approach: calibration by Bayesian bootstrap
Simulation results.
Overview
What are big data?
Big sizes (in terms of the number of subjects and/or variables)
Complex structure
Messy (e.g., incomplete data of different patterns)
Often big data are observational, or a mixture of trial and observational data.
No randomization or randomization lost -> confounding.
Causal inference using big data is a big and hot topic .
Big data, observational data
One approach to eliminating confounding in big data is the propensity score Suppose there are 300 (150 male/150 females) patients in a population. They take a drug at either high or low dose (not randomized!). We want to compare the outcomes of both doses. Gender affect the outcome, and males are twice likely to get the high dose as
females. The propensity score is the probability of getting high dose given gender. Suppose we have observed 100/50 males/females on high dose. How to adjust for the confounding?
Stratification by gender <-> by PS. Weighting by the inverse PS (IPW) (1 female represents 2 in high dose group).
How about we have 30 (instead of 1) potential confounders? Estimate the PS given the 30 covariates. Use the PS to stratify or to weight.
The propensity score (PS)approach
Another way of using PS is to add the PS as a covariate (PS-as-cov ) in the model for treatment and outcome.
Similar to the direct adjustment with all covariates in the model.
It is the easiest PS approach, and can be used for all types of models.
The justification:
PS acts as a summary of confounding effect in the model.
Under some situations PS-as-cov is equivalent to stratification/matching.
Under some situations PS-as-cov is equivalent to IPW.
Propensity score as a covariate
Missing/incomplete covariates
1 2
3
Cov set 2Cov set 1
n2
n1
A,Y Missing data are very common in big data.
Missing patterns may be quite different.
Big data are often a combination of subsets from different sources.
Data pattern on the right: Cov set 1 may be expensive/inconvenient to
measure, hence only n1 patients have them
Cov set 2 are routinely collected, so all have them
Within block 3, there could be further missings
A=treatment allocation, Y=outcome
Assume Cov sets 1-2 contain all confounders
Fitting a model with sets 1-2 covariates to A for pop n1 gives the gold standard (right) PS(GS)
Fitting a model with set 2 covariates to A for pop n1+n2 gives an error prone (wrong) PS(EP).
Adjusting with PS(EP) may results in bias.
But using PS(GS) we throw away the n2 pop.!
Propensity score with missing covs
1 2
3
Cov set 2Cov set 1
n2
n1
A,Y
Can we calibrate the estimator adjusted by the “wrong” PS(EP)?
Assuming PS(GS)= c + d PS(EP) +Error (1) one can apply measurement error models for calibration (Sturmer et al.)
Is (1) a correct model, given PS(EP) and PS(GS) are all within (0,1)?
In general the relationship is nonlinear, depends on the distribution of Cov sets
Example on the right (C=set 1, X=set 2):
Calibration depends on the outcome (Y) models
Propensity score
calibration
1 2
3
Cov set 2Cov set 1
n2
n1
T,Y
Lin & Chen’s (2015) approach based on Chen & Chen (2000).1. Estimate treatment effect (B) in the outcome model
with PS(GS) adjustment from n1 pop.
2. Estimate treatment effect (B*) in the outcome model with PS(EP) adjustment from n1 pop.
3. Estimate E(B*) using PS(EP) for n2 pop.
4. The calibrated estimate is B + K(B* - E(B*)), assuming (B, B*) ~Normal
K=cov(B, B*)/var(B*), not easy to calculate.
Lin & Chen have SAS macros fitting common outcome models from scratch and with complex calculations
A more robust calibration
1 2
3
Cov set 2Cov set 1
n2
n1
A,Y
One easy approach to calculating cov(B, B*) is bootstrapping1. Take n1 samples with replacement from the n1 pop.
2. Fit outcome models with PS(EP) and PS(GS) adjustment to get B* and B.
3. Repeat 1 and 2 many times then calculate sample cov(B, B*)
4. If n2 is not very large, do 1. and 2. with PS(EP) in n2 pop for better E(B*).
We use Bayesian bootstrap for simplicity and smoothness.
It just needs a single line “weight= - log(ranuni(seed));” (SAS syntax) in the bootstrap loop and use weight when fitting the two outcome models
Ordinary bootstrap weights by integers and Bayesian bootstrap weights by real numbers, hence is smoother.
Bayesian bootstrap calibration
Simulation results• We performed extensive simulations to evaluate the proposed approach• Results below are for Poisson regression model for outcome• Total samples n=n1+n2 with 20% of n1, 2 covs in Cov set 1 and 2 in Cov set 2• 100 Bayesian bootstrap samples and 1000 simulations for each scenario
• Var(emp)=sample variance in simulation
• Var(est)=Estimated variance
• Coverage=Covarage of 95% CI
• PSC=PS calibration• BB=Bayesian bootstrap (BB)
We have assumed a constant treatment effect over the whole population
What if treatment effects are different between the pops n1 and n2 populations?
In this case, the approach estimates the effect in the n1 population
In general the effect of the whole population may not be estimable
We have also assumed the same distribution for Cov set 2 for n1 and n2 populations (can be checked).
Warnings: assumptions we made
Assume that n1=200 and n2=50,000 with 10 repeated outcome measures each subject.
Complex missing data approaches (likelihood, EM algorithms, multiple imputation) may not be feasible.
For our method, the model with PS(EP) needn’t to be correct. So we can
Use a good mixed effect model with PS(GS) to fit data to pop. n1
Use LS to fit a simple model with PS(EP), ignore subject effects
As n2 is very large, no bootstrap is needed to estimate E(B*)
The estimation is still valid, although less efficient than using an equally good model with PS(EP).
Another advantage when dealing with
big data
An advantage when dealing with big
data: simulation results• We use simulation to evaluate the approach • Assume that n1=200 and n2=800 with 2-6 repeated measures each subject,
generate data with subject and effects and measurement errors• Use GEE for n1 pop. and LS for n=n1+n2 pop.
Incomplete covariates are common in big data
Propensity score calibration is one simple way to mitigate this issue
Lin and Chen’s approach does not depends the model between PS(EP) and PS(GS), hence is more robust
The approach can be easily applied to any models with the help of Bayesian bootstrap
The framework can be extended to other adjustment approaches
Summary
Sturmer T, Schneeweiss S, Avorn J, Glynn RJ. Adjusting effect estimates for unmeasured confounding with validation data using propensity score calibration. Am J Epidemiol 2005; 162:279-289.
Sturmer T, Schneeweiss S, Rothman KJ, Avorn J, Glynn RJ. Performance of propensity score calibration: a simulation study. Am J Epidemiol 2007; 165:1110-1118
Lin HW, Chen YH. Adjustment for missing confounders in studies based on observational databases: 2-stage calibration combining propensity scores from primary and validation data. Am J Epidemiol 2014; 180:308-317.
Chen YH, Chen H. A unified approach to regression analysis under double-sampling designs. J R Stat Soc B 2000; 62:449-460.
Rubin DB. The Bayesian bootstrap. Annals of Statistics 1982; 9:130-134.
Wang J, Mazzei Abba A. A practical and robust propensity score calibration approach based on bootstrapping. Unpublished manuscript.
Reference