Table 2.

Discrete power-law exponents for the frequency distribution of hashtags, fitted by the max-likelihood method for each event, and for the two ways of defining the time periods.

Level	Dataset	Same time: exp. (error)			Same n∘ hasht. usages: exp. (error)
		Before	During	Sign.	Before	During	Sign.
Hasht	Noaltarifazo/ruidazonac..	1.957 (11)	1.943 (11)		1.955 (11)	1.943 (11)
	9n/9ngranmarchaporlaj..	1.822 (11)	1.760 (10)		1.818 (9)	1.760 (10)	$\ast$
	15m	1.960 (46)	1.825 (23)	$\ast \ast \ast$	1.793 (20)	1.825 (23)	$\ast \ast \ast$
	Charlie Hebdo	2.036 (16)	1.988 (14)	$\ast \ast \ast$	2.054 (13)	1.988 (14)	$\ast \ast \ast$
User	Noaltarifazo/ruidazonac..	1.994 (12)	1.989 (11)		1.989 (12)	1.989 (11)
	9n/9ngranmarchaporlaj..	1.788 (10)	1.758 (10)	$\ast \ast \ast$	1.822 (9)	1.758 (10)
	15m	2.125 (54)	1.870 (24)	$\ast \ast \ast$	1.840 (20)	1.870 (24)
	Charlie Hebdo	2.052 (16)	1.974 (14)	$\ast \ast \ast$	2.028 (13)	1.974 (14)

Standard errors are reported in parenthesis in units of the least significant digit. We show results by counting the hashtags every time a user posted it (“Hasht”, above), and only once (“User”, below), irrespective of how many times it was used by the same user. The $x_{\mathit{min}}$ parameters were also fitted by max-likelihood and lie in the interval [1, 4]. All power-law fits passed the Kolmogorov-Smirnov goodness-of-fit test at p=0.05. The significance columns (“Sign”) represent the significance level at which we can reject the hypothesis that the “before” and “during” samples come from the same distribution (Epps-Singleton test ²³), thus denoting that there is a change of regime during the event ($\ast :\le 0.05,\ast \ast :\le 0.01,\ast \ast \ast :\le 0.001$).