Forty-three percent of never attenders were liberal Democrats in 2008, and another 20% were moderate Democrats. That’s basically two-thirds of never attenders in those two squares. But 14% were conservative Republicans, as well.

In 2012, the share who were liberal Democrats dropped to 37%, and the conservative Republican bucket went up by two points.

In 2016, only 35% of never attenders were liberal Democrats — down eight points from 2008. In total, 54% of the sample was in the bottom left two squares, compared to 63% eight years earlier.

Definitely some evidence that the never attenders category was more moderate by 2016. But in 2020, the entire trend reverses itself. Now, 43% of never attenders are liberal Democrats — that’s exactly the same share as it was in 2008.

However, the share of moderate Democrats was just 15% in 2020 compared to 20% in 2008. There were more moderate independents in 2020 compared to 2008, and the conservative Republican never attenders also grew during this time period.

I wish there was a clear direction in these results, but there just isn’t. Never attenders started out as pretty liberal then began to drift back to the middle of political and ideological spectrum. But then, in 2020, they turned back harder to the left, and never attenders were just as likely to be liberal as they were in 2008.

So, that group grew by 14 percentage points, but the overall ideology/partisanship didn’t really change.

Here’s where the clouds part for me a bit, though. I put together a regression model. The dependent variable (the thing I am trying to predict) is whether someone is a never attender. I wanted to see if the same factors worked the same way in 2008 compared to 2020 to predict someone never attending religious services.

Here’s how to interpret this — anything to the left of zero means someone is LESS likely to be a never attender. Anything to the right predicts a GREATER likelihood of never attending. If it overlaps with zero, there’s no relationship between that variable and never attending. (By the way, these are robust standard errors, and the coefficients are scaled, which means they can be directly compared).