Rynaa Grover, Joyojeet Pal
(Cite Grover, Rynaa, Pal, Joyojeet. 2021. Does Twitter Impact Electoral Outcomes of Candidates: Evidence from West Bengal 2021 Assembly Elections Online at https://joyojeet.people.si.umich.edu/twitterwb)
The answer to precisely that question is straightforward. No.
But does Twitter matter for campaigns? Very likely, and there are a number of things that suggest that serious candidates, particularly those that are likely to have an important public presence, plan to invest in social media. Moreover, there are regional factors related to social media use that highlight ways in which online political capital is distributed, we dig into some of these here.
Data distribution
First, we find, to high statistical significance, that the winning candidate in a constituency is more likely to have a Twitter account than a losing candidate. This is particularly true for the AITC. This does not imply that getting a Twitter account is more likely to make a candidate win, just that the stronger contenders also tend to have a social media presence. Ironically, we found, also to statistical significance, that losing candidates, on average tweeted more than winning candidates (see Appendix 3).
The table below shows the distribution of 280 candidates with known Twitter handles in West Bengal who contested the 2021 Legislative Elections and were active on Twitter. The total number of constituencies in West Bengal are 294. Thus, the percentage candidates who won and are active on Twitter is 50% (147 / 294), which is significantly higher (Appendix 1) than the percentage of all contesting candidates who are active on Twitter which is 31.7% (280 / 882). This statistic depends on our method as being fairly complete in terms of not missing any candidates’ handles, more details on this in our methodology section.
| First position | Runner-up | Lost | Total | |
| AITC | 121 | 40 | 0 | 161 |
| BJP | 26 | 67 | 7 | 100 |
| CPIM | 0 | 1 | 7 | 8 |
| INC | 0 | 0 | 11 | 11 |
| Total | 147*** | 108 | 25 | 280 |
Table 1 Data Distribution of Candidates Active on Twitter
Who is portrayed by candidates – Party, Leader or Self?
Who a candidate decides to focus on in their social media output is an important part of what branding they feel is most likely to make them win an election. To better understand this, we used a bag of words technique to find the percentage of tweets related to each category – party, leader and oneself.
We see in the table below that the majority of candidates on both the AITC and BJP sides were more likely to talk about themselves, their party, or their leader than they were to talk about the opposite side. In terms of negative campaigning, we notice that AITC candidates were more likely to mention the BJP much more than BJP candidates mentioning AITC. We see that winning candidates engage a bit less in negative campaigning than losing candidates. We find also that the percentage of tweets about Narendra Modi and Mamata Banerjee made by candidates on both sides are significant features (see Appendix 3).
The most talked about leader during the campaign was Mamata Banerjee. Overall AITC candidates mention Mamata Banerjee much more than the party itself, but there is no significant relationship between mentioning her and winning or losing an election. Nonetheless, it is clear that AITC politicians are much more likely to mention Mamata than BJP candidates were to mention either Modi or others from the party such as Amit Shah, Dilip Ghosh. In West Bengal, it would appear that the BJP also decided to underplay Modi compared to the party itself. This is in important distinction from past elections in which research showed that Modi was typically used more than the party or the candidate themselves for seats in various states by the BJP.[1] We also find in Table 2 that winning candidates on the BJP side focused a bit more on themselves, and on their own party than the losing candidates, and furthermore that losing candidates were more likely to focus on negative messaging, mainly through higher tweeting about Mamata Banerjee, than the winning candidate.
| AITC | BJP | |||||
| Winners | Losers | Total | Winners | Losers | Total | |
| Median percentage of tweets about one’s own party | 23.33% | 25.55% | 23.72%** | 48.7%** | 35.58%** | 40.61%** |
| Median percentage of tweets about opposition party | 16.66% | 18.02% | 17.14%** | 7.5% | 7.91% | 7.89%** |
| Median percentage of tweets related to Narendra Modi | 3.84% | 2.77% | 3.65% | 9.69% | 7.75% | 8.93%** |
| Median percentage of tweets related to Mamata Banerjee | 35.33% | 37.59% | 35.64%** | 4.55% | 7.15% | 6.53% |
| Median percentage of tweets related to candidates themselves | 51.4% | 61.01% | 53.5% | 49.22% | 43.21% | 44.02% |
| Number of candidates in the category | 121 | 40 | 161 | 26 | 74 | 100 |
Table 2 Party-wise Percentage of Tweets Related to Candidates, Party and Leader (see Appendix 3)
Close Contests
We looked at the social media profile of the close contests, those in which the margin of victory. Two or more candidates being on social media is correlated to contests being slightly closer. The constituencies with two active candidates on social media have a smaller margin of victory (median margin 18454.0) than constituencies with only one candidate active on social media (median margin 23512.0), although running a Welch’s t-test on these two distributions gives us a p-value = 0.053, thus the differences are significant, but not highly significant.
We counted 18 contests that were decided by a margin of 10,000 votes or less, and had two or more candidates standing for elections who were active on twitter. Of these 18, we found that in 14 cases the politicians with more Twitter followers lost. Of the four who won, one was a sitting member of parliament (Nisith Pramanik, who defected from AITC in 2019), and another was cricketer Ashoke Dinda. Thus high social media following not a predictor of election success, let us turn to celebrity candidates to understand this better.
Does Online Influencer status matter?
We also find that high following does not have a relationship with the likelihood of winning. In fact, eight of the ten highest followed politicians from the BJP (Srabanti Chatterjee, Swapan Dasgupta, Babul Supriyo, Paayel Sarkar, Parno Mitra, Tanushree Chakraborty, Yash Dasgupta and Locket Chaterjee) all lost their seats, with one exception, all with large margins of over 15,000 votes. Three of the losers were members of parliament – in fact two of the highest profile candidates – Babul Supriyo and Srabanti Chatterjee, lost by landslide margins of 50,000 votes or more.
3 of Trinamool’s 5 most Twitter followed politicians Mamata Banerjee herself, Sayantika Banerjee and Koushani Mukherjee all lost elections to other candidates less followed than themselves on Twitter. CPIM, which was decimated in the elections, had all its high-profile candidates on Twitter – Sujan Chaterjee, Mohammed Salim, Aishe Ghosh, and Dipsita Dhar all ended in massive defeats of over 35,000 votes behind the overall winner, in three of four cases, ending third.
The data suggests not only that being an online influencer does not matter in predicting election outcomes, but conversely, that online influencers may be propped up as candidates in constituencies in which a party is already weak.
Does location matter?
We find that the districts of southern West Bengal see a much greater concentration of politicians on social media, which is in line with the spread of population in the state. The state’s most populous district, North 24 Parganas, is also the district with the most politicians on social media, followed by South 24 Parganas, Hooghly, Murshidabad, and Howrah. Jhargram district, in the Chhota Nagpur plateau at the western edge of the state had only one candidate. The northern part of the state, in the Jalpaiguri division, had the fewest candidates on Twitter, only three of whom had 5000 followers or more, whereas the 20 most followed candidates on Twitter are all from the districts within a 100 km radius of Kolkata.
We plotted the total tweets from each district across West Bengal against the percentage of seats won from that district for each party (see Pearson correlation coefficient Appendix 2). In Figure 1 below, we see that AITC won a higher percentage of seats in districts where their Twitter activity was higher, especially in the southern districts. BJP, on the other hand, won a high percentage of seats in districts where their social media activity was higher than that of AITC, for instance Cooch Behar, Malda, Dakshin Dinajpur, and Birbhum.
| Aggregate sum of total tweets from each district | Percentage of seats won by each party
[number of seats won from the district / total number of seats won by party] |
 |
 |
 |
 |
Figure 2 District-wise social media activity and percentage of seats won by each party
Methodology
In the figure below, we see the timeline of data collection followed for this work. First, the list of all candidates contesting in the 2021 West Bengal Legislative Assembly Elections was collected from Wikipedia [2] along with information about constituencies and parties of affiliation. Next, Google was queried for Twitter handles of the election candidates using GoogleSearch API [3]. A string consisting of the candidate’s name, party, and state along with keyword ‘twitter’ was used to search, and the Twitter handle was extracted from the search result. Post this, the handles were manually verified, and a dataset of 280 candidates and their Twitter handles were curated. Finally, the tweets for these handles were pulled using the Twitter API [4].
For the larger set of nationwide politician, media and influencer intersections, we used NivaDuck. NivaDuck [5] is a scalable Machine Learning-based pipeline used to identify the Twitter handles of a country’s politicians and celebrities based on the account descriptions and the tweets made. On applying NivaDuck for India, we collected Twitter accounts of 36k Indian politicians and 21k Indian celebrities, including journalists, media houses, and influencers. This data was used along with the candidates’ tweets for various analysis carried out in this work.
Figure 2 Timeline of dataset collection
Appendices
- Chi-square test to prove that winning candidates are more likely to have a Twitter account
| Observed Value (O) | Expected Value (E) | (O – E)^2 | (O – E)^2 / E | |
| Winners | 147 | 93.198 | 2894.65 | 31.059 |
| Losers | 133 | 186.802 | 2894.65 | 15.49 |
| Chi square | 46.55 | |||
| Degree of freedom | 1 | |||
| P value | < 0.001 | |||
Appendix Table 1 Chi-square test for showing the higher likelihood of winning candidates to have a Twitter presence
Since the p-value is lesser than 0.001, we reject the null hypothesis and conclude that the percentage of winners active on Twitter is significantly higher than the percentage of contesting candidates active on Twitter.
- Pearson correlation coefficients of district-wise twitter activity with the number of winning candidates who have a Twitter presence
| Attribute | Number of seats won by candidates active on Twitter from the district |
| Ratio of rural to urban households | -0.444 |
| Aggregate sum of retweet count from the winning district | 0.585 |
| Aggregate sum of total tweets from the winning district | 0.845 |
| Aggregate sum of followers counts of candidates contesting from the winning district | 0.318 |
Appendix Table 2 Correlation of district-wise social media activity with the number of seats won
- Welch’s t-test for correlating election outcomes with social media activity of candidates
- Welch’s t-test for correlating election outcomes with social media activity of candidates
| Attribute | P value |
| Difference between followers count of winners and losers | 0.25240232910999905 |
| Difference between aggregate retweets of winners and losers | 0.2429998743703439 |
| Difference between aggregate number of tweets made by winners and losers | 0.03381527789474302** |
| Difference between median retweet count of winners and losers | 0.604136351136941 |
| Difference between percentage of Narendra Modi related tweets made by winners and losers | 0.029974342036102494**
|
| Difference between percentage of Mamata Banerjee related tweets made by winners and losers | 9.661666779854375e-10**
|
| Difference between percentage of BJP related tweets made by BJP winners and losers | 0.023452363978817253**
|
| Difference between percentage of Narendra Modi related tweets made by BJP winners and losers | 0.9572247637602638 |
| Difference between percentage of AITC related tweets made by AITC winners and losers | 0.20123011861723109 |
| Difference between percentage of Mamata related tweets made by ATIC winners and losers | 0.3367126740412517 |
| Difference between percentage of tweets about opposition party made by BJP and AITC candidates | 6.4170621019774965e-12** |
| Difference between percentage of tweets about one’s own party made by BJP and AITC | 1.8969288775960985e-07** |
| Difference between percentage of tweets made by AITC candidates about AITC and Mamata Banerjee | 5.563017511240308e-09** |
| Difference between percentage of tweets made by BJP candidates about BJP and Narendra Modi | 1.4743661153249999e-21** |
Appendix Table 3 Welch’s t-test values for finding social media factors that affect election outcomes significantly
Link to Candidates_data
References
[1] A. Rajadesingan, A. Panda, and J. Pal. Leader or Party? Personalization in Twitter Political Campaigns during the 2019 Indian Elections. International Conference on Social Media and Society, pp 174–183. 2020 https://dl.acm.org/doi/abs/10.1145/3400806.3400827
[2] “2021 West Bengal Legislative Assembly election,” [Online]. Available: https://en.wikipedia.org/wiki/2021_West_Bengal_Legislative_Assembly_election.
[3] “googlesearch-python 2020.0.2,” [Online]. Available: https://pypi.org/project/googlesearch-python/.
[4] “Twitter API Documentation,” [Online]. Available: https://developer.twitter.com/en/docs/twitter-api.
[5] A. Panda, A. N. Gonawela, S. Acharyya, D. Mishra, M. Mohapatra, R. Chandrasekaran and J. Pal, “NivaDuck-A Scalable Pipeline to Build a Database of Political Twitter Handles for India and the United States,” International Conference on Social Media and Society, pp. 200-209, 2020.