All of the statistics are misleading. The third one is particularly so, because they are weighting by s$population * s$homicides_per_100000, i.e. 100000 * homic

All of the statistics are misleading. The third one is particularly so, because they are weighting by s$population * s$homicides_per_100000, i.e. 100000 * homicides which makes it completely meaningless). Starting from the data defined in that article: > d$homicides = d$homicides_per_100000 * d$population / 100000 > d$firearms = d$firearms_per_100 * d$population / 100 > summary(lm(formula = d$homicides ~ d$firearms)) Call: lm(formula = d$homicides ~ d$firearms) Residuals: Min 1Q Median 3Q Max -12851 -2081 -1841 -978 50042 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 2.114e+03 5.592e+02 3.780 0.000231 * d$firearms 9.142e-05 2.253e-05 4.058 8.18e-05 * --- Signif. codes: 0 ‘’ 0.001 ‘’ 0.01 ‘’ 0.05 ‘.’ 0.1 ‘ ’ 1 Residual standard error: 6561 on 141 degrees of freedom Multiple R-squared: 0.1046, Adjusted R-squared: 0.0982 F-statistic: 16.46 on 1 and 141 DF, p-value: 8.179e-05 > s <- d[d$gdp_ppp_per_capita > median(d$gdp_ppp_per_capita),] > summary(lm(formula = s$homicides ~ s$firearms)) Call: lm(formula = s$homicides ~ s$firearms) Residuals: Min 1Q Median 3Q Max -4509 -2097 -1997 -1747 50543 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 2.096e+03 9.202e+02 2.278 0.0258 * s$firearms 6.158e-05 2.739e-05 2.248 0.0277 * --- Signif. codes: 0 ‘’ 0.001 ‘’ 0.01 ‘’ 0.05 ‘.’ 0.1 ‘ ’ 1 Residual standard error: 7591 on 69 degrees of freedom Multiple R-squared: 0.06827, Adjusted R-squared: 0.05477 F-statistic: 5.056 on 1 and 69 DF, p-value: 0.02774 In other words, the most obvious analysis to do shows strong evidence for a positive relationship between total homicides and total firearms - the link is in the same direction but much weaker when you only look at the upper half of countries by GDP. Correlation doesn't imply causation - it could be that in some countries, people obtain firearms because the homicide rate is high, or there could be some other confounding factor not even in the data that drives both. If this had been an experiment and each country had been assigned a firearm ownership rate, it would have been safe to conclude that each million firearm owners cause between 55 and 68 homicides (99% confidence interval).