You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
As of now, we have "vanilla" synthetic control working with cp.pymc_experiments.SyntheticControl as the experiment class, and that is fed the cp.pymc_models.WeightedSumFitter as the model.
It is the cp.pymc_models.WeightedSumFitter which does the vanilla synthetic control model - as in weightings which sum to 1, and that is done via a Dirichlet distribution.
We want to add the ability to do augmented synthetic control. This will still use the cp.pymc_experiments.SyntheticControl experiment class, but instead we will feed it a new model, something like cp.pymc_models.AugmentedSyntheticControlModel. (However, see below because we may no need a new model)
Implementation notes
As far as I understand the algorithm for augmented synthetic control is along the lines of:
Based on the pre-treatment data, fit vanilla synthetic control model where weights are constrained to sum to 1.
Calculate the residuals between the model pre-treatment predictions and the observations
Fit these residuals with a model
Use the predictions of that model to adjust the synthetic control predictions
That need not be done in separate steps. What you could do is to have a model where the weightings of the control groups are constrained to sum to 1, but then simply add in more components to the model, such as an intercept and trend. For example, the model formula in one of the examples is currently:
Though you would have to ensure that the weights of the control units are constrained to sum to 1, but the 1 and trend predictors are weighted by 'unconstrained' coefficients.
So practically we might want to keep the original model formula but add a new residuals ~ 1 + trend, or something similar. Though it could just be simpler to do a custom model with something like:
As of now, we have "vanilla" synthetic control working with
cp.pymc_experiments.SyntheticControl
as the experiment class, and that is fed thecp.pymc_models.WeightedSumFitter
as the model.It is the
cp.pymc_models.WeightedSumFitter
which does the vanilla synthetic control model - as in weightings which sum to 1, and that is done via a Dirichlet distribution.We want to add the ability to do augmented synthetic control. This will still use the
cp.pymc_experiments.SyntheticControl
experiment class, but instead we will feed it a new model, something likecp.pymc_models.AugmentedSyntheticControlModel
. (However, see below because we may no need a new model)Implementation notes
As far as I understand the algorithm for augmented synthetic control is along the lines of:
That need not be done in separate steps. What you could do is to have a model where the weightings of the control groups are constrained to sum to 1, but then simply add in more components to the model, such as an intercept and trend. For example, the model formula in one of the examples is currently:
but you could implement augmented synthetic control with something like
Though you would have to ensure that the weights of the control units are constrained to sum to 1, but the
1
andtrend
predictors are weighted by 'unconstrained' coefficients.So practically we might want to keep the original model formula but add a new
residuals ~ 1 + trend
, or something similar. Though it could just be simpler to do a custom model with something like:control_units = ["Austria", "Belgium", "Bulgaria", "Croatia", "Cyprus", "Czech_Republic"]
residuals ~ 1 + trend
The text was updated successfully, but these errors were encountered: