setx with arima returns error, but not setx1

[Christopher Gandrud]

Looking into an issue reported in #298 by @mbsabath , it seems like plot for arima models--specifically when creating acf plots doesn't recognise setx simulations, but does find setx1:

library(zeligverse)
data(seatshare)
subset <- seatshare[seatshare$country == "UNITED KINGDOM",]
ts.out <- zarima$new()
ts.out$zelig(unemp ~ leftseat, order = c(1, 0, 1), data = subset)
ts.out$setx(leftseat = 0.25)
ts.out$sim()
plot(ts.out)

##  Error in obj$get_qi("acf", xvalue = "x1") : xvalue must be x. 

But

ts.out <- zarima$new()
ts.out$zelig(unemp ~ leftseat, order = c(1, 0, 1), data = subset)
ts.out$setx1(leftseat = 0.25)
ts.out$sim()
plot(ts.out)

returns the expected plot:

[Ben Sabath]

I'm working on this issue, and I fixed the issues in the plotting wrapper that lead to the initial error, but the issue remains that qi isn't generated when only setx is run. I have a theory on how to fix this, but I want to get feedback on my pseudocode before going forward.

If setx has been called, but setx1 has not:

• generate mm type object based on the algorithm used in set for an argument less setx (mostly average)
• generate qi using average as mm and the counterfactual as mm1

Do we anticipate any errors from this, or do we think this would be fine?

[Ben Sabath]

I have a fix that's working on my computer. see commit 8b8868d.

[Ben Sabath]

All checks are currently passing on the branch

[Christopher Gandrud]

Thanks for this @mbsabath. Could you bring this branch up-to-date with master? There seem to be some merge conflicts when brought up to speed with master.

[Ben Sabath]

Yup. I branched it off of the branch implementing the wrapper for time series. I'll merge things on my machine and get things into that pull request?

[Christopher Gandrud]

sounds good

[Ben Sabath]

Just updated the pull request

[Christopher Gandrud]

That seems to work great @mbsabath. Thanks so much for your work on this!

My only question now is what (if anything) is setx1 supposed to do with plot for arima? The following produces the same plot (e.g. there are no first differences for the second example) for the version in #289:

library(Zelig)
library(dplyr)

data(seatshare)
subset <- seatshare[seatshare$country == "UNITED KINGDOM",]
# Example 1
zelig(unemp ~ leftseat, data = subset, model = "arima",
      order = c(2,0,1)) %>%
    setx(leftseat = 0.75) %>%
    sim() %>%
    plot()

# Example 2
zelig(unemp ~ leftseat, data = subset, model = "arima",
      order = c(2,0,1)) %>%
    setx(leftseat = 0.75) %>%
    setx1(leftseat = 0.25) %>%
    sim() %>%
    plot()

Relatedly summary for an arima Zelig object with setx1 returns:

s.out <- zelig(unemp ~ leftseat, data = subset, model = "arima",
      order = c(2,0,1)) %>%
    setx(leftseat = 0.75) %>%
    setx1(leftseat = 0.25) %>%
    sim()
summary(s.out)

## Error in apply(qi, 2, quantile, c(0.5, 0.025, 0.975), na.rm = TRUE) : 
##  dim(X) must have a positive length

Question

Should setx1 be useable for arima models?

• If not we could throw an error when using setx1.

• If yes, then we should fix the summary error and have a meaningful change to the plot.

Any thoughts @tercer?

[Ben Sabath]

That summary error is persistent, I haven't quite figured out how to fix that yet. The issue is that qi for time series are not easily printed out. The lack of difference is definitely a bug, I can go into the code today to figure out where the issue is coming from (Probably just a mis-specified if condition).

My personal feelings on what setx1 should do are this: if only setx1 is called, and not setx:we throw an error saying setx hasn't been called. But if setx has been called, we use the values in it as the base values, with setx1 as the shock values.

[Christopher Gandrud]

Thanks for looking into this. Let's talk about this in the call today.

[Christopher Gandrud]

Resolved by #289

Excluding the summary error. Which is now addressed in its own issue: #305