1 / 10 1 CVIVA, Research Center For Vitamins & Vaccines, Statens Serum Institut. 2 Department of Biostatistics, University of Copenhagen.
2 / 10 Outline How can a be used? What are the challenges in the current Ph.D project?
3 / 10 What do we want to model? We want to compare intensities - possibly time-dependent. not vac (0) α 02 (t) α 01 (t) dead (2) vac (1) α 12 (t)
4 / 10 A is a probability. Examples: 1. The probability of recieving a treatment. 2. The probablity of having recieved measles vaccine at 12 month of age. 3. The probablity of having recieved measles vaccine at age t (time-dependent -). Often the will depend on covariates e.g., sex, birth weight, homebirth, mothers age at birth, etc.
5 / 10 1/3 Observational studies and randomized studies typically, does not estimate the same parameter. Observational studies (regression analysis) estimate conditional effects: Given all others covariates equal what is the effect of being vaccinated (effects assumed equal for all subgroups). Randomized studies estimate causal effects: Compare outcome in two equal populations which only differ on vaccination status (effects possibly different in subgroups).
6 / 10 2/3 Using - techniques in analysis of observational data one can mimic a randomized study. Mimic means: We estimate the same parameter (the average causal effect) as in a randomized study. Note: Propensity analysis can not mimic balance of unmeasured confounders.
7 / 10 3/3 Other advantages of approach compared to ordinary regression analysis: 1. If outcome is rare (e.g. death) and treatment common (e.g. vaccination) a much richer model is allowed. 2. In general the model is more robust to misspecification. 3. Easier to detect non-comparable subjects or subgroups (they have a close to 0 or 1). Note: an ordinary regression potentially involves a lot of extrapolation.
8 / 10 In general 4 different methods: 1. Stratification on. 2. Match (1:1) exposed to unexposed based on. 3. Create pseudo population based on weighting by the inverse of the. 4. Include the as a covariate in a regression model. All 4 have been thoroughly studied in simple setups (time-invariant s) and are more or less interpretable.
9 / 10 Ongoing research Some themes in the current Ph.D project: 1. How to build an appropriate time-dependent - for vaccination status. 2. How to use this -. Matching? Create pseudo-population? 3. How to build models which can include and handle different vaccination-histories. (multistate models) 4. How to handle survival-bias (caused by missing vaccination cards) without throwing away information (landmark approach). death 1. visit 2. visit
10 / 10 References Rosenbaum P.R., Rubin D.B. The central role of the in observational studies for causal effects. Biometrika, 70(1):41 55, 1983. Williamson E., Morley R., Lucas A., and Carpenter J. Propensity s: From nave enthusiasm to intuitive understanding. Methods in Medical Research, 21(3):273 293, 2012. D Agostino R.B. Propensity methods for bias reduction in the comparison of a treatment to a non-randomized control group. Statistics in Medicine, 17(19):2265 2281, 1998.