Table 3.

Correlations between valid tweets and sentinel-provided ILI^a rates.^b

	1. All tweets, r	2. Number of all tweets	3. P-value for all tweets	4. Valid tweets, r	5. Number of valid tweets	6. P-value for valid tweets	7. Fisher’s z transformation, P
Boston	−.05	17,370	.834	.10	3813	.67	<.001
Chicago	.33	21,655	.139	.64	5116	.002	<.001
Cleveland	.63	7152	.002	.60	1497	.003	.064
Columbus	.01	3288	.978	−.24	1034	.274	<.001
Denver	.76	5706	.003	.69	1942	.009	<.001
Detroit	.81	8417	.001	.76	2195	<.001	<.001
Fort Worth	.69	4755	.001	.85	1236	<.001	<.001
Nashville-Davidson	.77	5805	.001	.83	1630	<.001	<.001
New York	.44	64,340	.047	.55	12632	.01	<.001
San Diego	.78	8002	.001	.88	1808	<.001	<.001

^aILI: influenza-like illness

^bCorrelation coefficients between all tweets and valid tweets, as identified by the machine-learning classifier, with sentinel-provided ILI rates for each city. Comparisons between tweets and ILI began in Weeks 36-49 (weeks starting September 1, 2013 to starting November 24, 2013) as ILI data became available by city and ended in Week 9 (ending March 1, 2014).