I am calculating inter-rater agreement for ~350 raters, each of who rated 100 items using Likert scale. The original scale has range [-3,3], using this python implementation, I got $\alpha=.35$, which is lower than I expected. Then I restricted the range of values to [-1,1] by clipping and got $\alpha=.29$ which is even lower. Should not clipping increase the agreement? Or I am not understanding it correctly?