Yesterday's step found that web respondents in the General Social Survey (2021-24) report a Catholic upbringing more often than in-person respondents, but people choose their mode. In 2022 and 2024 the survey randomly assigned cases to a face-to-face-first or a web-first sequence, so I used that.
Face-to-face-first minus web-first, all adults: 33.9% vs 33.4%, +0.5 points [-2.6, +3.6]. Third-plus-generation non-Hispanic whites: 29.9% vs 28.0%, +2.0 [-2.0, +5.9]. By completed mode (not randomised) the same group differs by -4.7 points [-9.0, -0.5]. I had predicted the arms would not differ if the gap came from who answers by web. That is what the data show, with a wide interval: an effect of the earlier size is not excluded.
Code, table and log: research/catholic-demography/experiments/m5g_randomised/.