missing data

[hyp...@gmail.com]

How does CI deals with missing data in the dependent and  the independent variables?

[Kay Brodersen]

Missing values in the dependent variable (response time series) are not a problem. The more missing values, the wider your resulting posterior intervals.

Missing values in the independent variables (regressors) are not currently supported.

On 13 October 2014 15:14, <hyp...@gmail.com> wrote:

How does CI deals with missing data in the dependent and  the independent variables?

--
You received this message because you are subscribed to the Google Groups "CausalImpact" group.
To unsubscribe from this group and stop receiving emails from it, send an email to causalimpact...@googlegroups.com.
To post to this group, send email to causal...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/causalimpact/d275dc95-8e97-4523-9b8c-6f6949c3386d%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[JCStanton]

I am getting an error message when I have NA's in my dependent variable:

Error: NA values in y.post not currently supported

I have already interpolated out the NA's from all of the independent variables, the only NA's are in the first column. Error message disappears and function runs if I interpolate those NA's out of the first column. Am I missing something?

packageVersion("CausalImpact")
[1] ‘1.0.3’

[Kay Brodersen]

Hi Jessica,

Could you please send a minimal reproducible example?

Kay

--
You received this message because you are subscribed to the Google Groups "CausalImpact" group.
To unsubscribe from this group and stop receiving emails from it, send an email to causalimpact...@googlegroups.com.
To post to this group, send email to causal...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/causalimpact/c2e99e8e-e7b7-4ff6-85e1-cf1f24d40335%40googlegroups.com.

[hyp...@gmail.com]

Hi, Kay. What would you get if you get if you substitute the missing observations in the independent variables with the CI prediction for that values? Would the point forecast for the dependent variable be "right" by any mean(I'm guessing forecasting intervals would be wrong since they would ignore the uncertainty of the first stage prediction)? Thank you in advance
Israel

[Kay Brodersen]

Hi Israel,

Let me check whether I understand your question correctly - do you mean, in a first stage, to use CausalImpact or some other model to impute, or 'fill in', missing observations in your predictor variables, and then use these, in a second stage, as regressors in a CausalImpact analysis? I agree, this would lead to overconfident predictions since you'd be ignoring the uncertainty in your predictors. I wouldn't expect this to be much of a problem, though, especially if the fraction of missing values is small. If in doubt, a better approach would be multiple imputation, where you'd rerun this analysis many times, each time drawing the missing values from their posteriors rather than picking the expectation.

Kay

On 9 March 2015 at 22:14, <hyp...@gmail.com> wrote:

Hi, Kay. What would you get if you get if you substitute the missing observations in the independent variables with the CI prediction for that values? Would the point forecast for the dependent variable be "right" by any mean(I'm guessing forecasting  intervals would be wrong since they would ignore the uncertainty of the first stage prediction)? Thank you in advance
Israel

--
You received this message because you are subscribed to the Google Groups "CausalImpact" group.
To unsubscribe from this group and stop receiving emails from it, send an email to causalimpact...@googlegroups.com.
To post to this group, send email to causal...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/causalimpact/37d7109c-287d-414f-ad2e-0ec9bbd1fe77%40googlegroups.com.