Hypothesis: those directly affected by the troll policy (trolls) are more likely to have strong disapproval than those unaffected by the troll policy are to have strong approval.

Hypothesis rejected when we operationalize 'trolls' as 'low karma':

R> lwtroll <- lw[!is.na(lw$KarmaScore),]
R> lwtroll <- lwtroll[lwtroll$TrollToll=="Agree with toll" | lwtroll$TrollToll=="Disagree with toll",]
R> # disagree=3, agree=2; so:
R> # if positive correlation, higher karma associates with disagreement
R> # if negative correlation, higher karma associates with agreement
R> # we are testing hypothesis higher karma = lower score/higher agreement
R> cor.test(as.integer(lwtroll$TrollToll), lwtroll$KarmaScore, alternative="less")
Pearson's product-moment correlation
data: as.integer(lwtroll$TrollToll) and lwtroll$KarmaScore t = 1.362, df = 315, p-value = 0.9129
alternative hypothesis: true correlation is less than 0 95 percent confidence interval:
-1.0000 0.1679 sample estimates:
cor 0.07653
R> # a log-transform of the karma scores doesn't help:
R> cor.test(as.integer(lwtroll$TrollToll), log1p(lwtroll$KarmaScore), alternative="less")
Pearson's product-moment correlation
data: as.integer(lwtroll$TrollToll) and log1p(lwtroll$KarmaScore) t = 2.559, df = 315, p-value = 0.9945
alternative hypothesis: true correlation is less than 0 95 percent confidence interval:
-1.0000 0.2322 sample estimates:
cor 0.1427

Plots of the scores, regular and log-transformed: