Another example is randomizing the person who answers the question, e.g. using self-reporting vs. using proxy respondents papers.ssrn.com/sol3/papers.... (again wording certainly changes since you have to switch the grammatical person of the question, but you also randomize who answers it).

in fact the first version of the graph (without texture) renders better for people with monochromacy. Still, I'd adapt the order of the legend, it might be a bit surprising otherwise (we'd expect "don't know" to be at the top of bars).

As a possible improvement, I'd match the order of colors in the graph to their order in the legend. The color of "don't know" and "don't trust very much" are a bit similar for people with monochromacy, so that might be a slightly tricky to immediately interpret the graph, though the texture helps.

Might be of interest to you www.tandfonline.com/doi/pdf/10.1..., with a few early examples of significance testing in observational studies. Another interesting ref. seems to be Raymond Hubbard's book "Corrupt Research" (2016), with some discussion and stats by discipline, page 22 and following.

Ma tentative moche faite complètement à l'arrache, mais on voit l'idée :