I am doing a sentiment analysis task about the people's attitudes towards transportation services in Hong Kong. I collected Tweets near the railway stations and see if there is any difference between the sentiment level of people near the railway stations. I have successfully computed the sentiments on a monthly basis and sample outputs are given below:
Station1:0.35, 0.36, 0.25, 0.27,....,0.58,0.36
Station2:0.42, 0.52, 0.32, 0.33,....,0.54,0.26
Station3:0.58, 0.23, 0.54, 0.59,....,0.51,0.26
The percentages here actually mean the percentage of positive Tweets in a given month. Hence for each station, we have 12 percentages of positive Tweets. In total, I have 93 stations and I want to know whether any two stations are statistically significantly different on sentiment level based on the data like above. Hence, for any two stations A and B, we could construct the following hypothesis test:
H_0: The sentiment levels between station A and station B are not significantly different.
H_1: The sentiment levels between station A and station B are different.
I first apply the Augmented Dickey-Fuller test to test whether the sentiments in one station are time-dependent. Results from most stations suggest that the sentiment level is time-dependent. So what should I do next? The Mann–Whitney U test and chi-square test seem most similar to my issue but these tests suggest that the proportions are independent to each other. So what should I do to check the significance of the difference on sentiment level between any two stations?