Conditioning on partisanship

Glad to see somebody's worked out the numbers on this one.

It's bugged me how American papers have been quoting Trump's continued high approval ratings among Republicans without doing much to correct for that the people most disgusted by Trump will have stopped identifying as Republicans.

John Holbein points to this piece, which tries to put some bounds on the effects. Just conditioning on stated past voting record won't do it where recall bias matters and where folks might not want to remember having supported something that's now kinda ugly. So they try to adjust for 'missing' partisans of the President's party relative to a baseline measure. The key figure is below, followed by the authors' discussion.

In the lower panel of Figure 2 [pictured above], we plot observed partisan approval rates, the bounds on the compositionally-corrected partisan approval rate, and 95 percent confidence intervals for the upper and lower bound during the first 163 days of each presidential term.23 The marker is the observed partisan approval rate, the first capped line extending out from the marker are the bounds, and the second set of capped lines are the 95 percent confidence intervals on the lower and upper bounds.

Trump’s observed partisan approval rates are very low compared to the same period during Obama’s first term, but are roughly comparable to Obama’s second term. More relevant for our analysis is how the bounds evolve over time. The lower bound on the compositionally-corrected partisan approval rate is quite low during Trump’s presidency. In 14 of the 23 weeks, the lower bound is below 0.8. With only one exception, the lower bound on Trump’s compositionally-corrected partisan approval rate is lower than the lower bound from the analogous poll during Obama’s second term. The observed partisan approval rate is partially an artifact of missing respondents who would have previously reported Republican partisanship. While President Trump’s observed partisan approval rate has received much attention, the data are also consistent with the possibility that his partisan approval rate is quite low relative to recent presidential history.

I'm surprised the lower bound of the adjusted interval is as high as it is.