The issue, as I understand gjm, is whether there’s a confounding factor.
Treating the correlation between choosing nonviolence (call that N for short) and succeeding in a political campaign (call that S) as accurately representing causality works if one assumes causation runs exclusively as N → S. gjm observes that expected probability of success (P) could be a confounding factor: maybe we have the extra causal arrows N ← P → S, as well as N → S.
According to her blog comment, Chenoweth did try to tackle this by establishing that there’s no N ← P causal arrow. Reconstructing her apparent logic, she says she found that external structural factors (X), which activists presumably use to form their beliefs about the probability of success, are uncorrelated with the decision to use nonviolence: i.e., she supposes that X → P and observes that X is uncorrelated with N, so P → N is implausible (because otherwise we’d have X → P → N and X would correlate with N, assuming, uhhh, the causal Markov condition I think?). If we accept that argument, then P isn’t a confounder.
Chenoweth also implies that she compared the outcomes within campaigns which split into a nonviolent arm and a violent arm. That would arguably control for P, if one accepted that for any given campaign which splits into two, P is the same for the two new subcampaigns. I don’t think I do accept that, but I could be wrong, and if so this alternative analytical approach would address the issue.
It’s hard for me to come to a strong final conclusion because I haven’t read the book. Based on the blog comment, I wouldn’t trust that Chenoweth’s isolated the potential confounder, but I might change my mind if I read the book. Along similar lines, you could well be correct about what’s typical for campaigns, but I can’t conclusively say you are because I haven’t systematically studied a wide range of political campaigns from this perspective.
The issue, as I understand gjm, is whether there’s a confounding factor.
Treating the correlation between choosing nonviolence (call that N for short) and succeeding in a political campaign (call that S) as accurately representing causality works if one assumes causation runs exclusively as N → S. gjm observes that expected probability of success (P) could be a confounding factor: maybe we have the extra causal arrows N ← P → S, as well as N → S.
According to her blog comment, Chenoweth did try to tackle this by establishing that there’s no N ← P causal arrow. Reconstructing her apparent logic, she says she found that external structural factors (X), which activists presumably use to form their beliefs about the probability of success, are uncorrelated with the decision to use nonviolence: i.e., she supposes that X → P and observes that X is uncorrelated with N, so P → N is implausible (because otherwise we’d have X → P → N and X would correlate with N, assuming, uhhh, the causal Markov condition I think?). If we accept that argument, then P isn’t a confounder.
Chenoweth also implies that she compared the outcomes within campaigns which split into a nonviolent arm and a violent arm. That would arguably control for P, if one accepted that for any given campaign which splits into two, P is the same for the two new subcampaigns. I don’t think I do accept that, but I could be wrong, and if so this alternative analytical approach would address the issue.
It’s hard for me to come to a strong final conclusion because I haven’t read the book. Based on the blog comment, I wouldn’t trust that Chenoweth’s isolated the potential confounder, but I might change my mind if I read the book. Along similar lines, you could well be correct about what’s typical for campaigns, but I can’t conclusively say you are because I haven’t systematically studied a wide range of political campaigns from this perspective.