One of the greatest challenges in Indian psephology is the projection of seat counts for Vidhan Sabha and Lok Sabha elections. While survey data can reveal how much vote share each party will win with a high degree of accuracy, they do not show how that vote is distributed across a state. Projecting that a party is to win a certain percent of votes does not say how many seats it is expected to win in an assembly, and therefore whether that party will perform strongly enough to form a government. In cases where the vote share of one party or alliance is much larger than all others, predicting the winner is fairly simple. In cases where the vote share of different parties or alliances is quite close — as is anticipated by nearly all polls for the Bihar elections — the vote-to-seat conversion model is key to predicting the winner.
Because this is such a crucial and challenging step to the projection of election results, pollsters are especially reluctant to share how they do their vote-to-seat conversion. The exception is Professor Rajeeva Karandikar, who for a long time did the seat share projections for the Lokniti unit of the Centre for the Study of Developing Societies (CSDS). Professor Karandikar is very open about the principles undergirding his vote-to-seat conversion models. Last year, he wrote a piece for the Hindu Centre for Politics and Public Policy outlining CSDS’ sampling methodology and detailing the principles of the model he uses to convert votes into seats.
Rajeeva Karandikar is the Director of Chennai Mathematical Institute in Chennai. He obtained his PhD from the Indian Statistical Institute (ISI) in 1981, and was on the faculty at ISI till 2006. He had a brief foray into the commercial world at Cranes Software as Executive Vice President from 2006-2010. Karandikar has made significant contributions to various areas including stochastic calculus, filtering theory, limit theorems, Monte Carlo techniques, and the theory of option pricing. He is a fellow of the Indian National Science Academy and Indian Academy of Sciences, and was awarded the S S Bhatnagar prize by the Council of Industrial and Scientific Research in 1999. Karandikar has been involved with opinion polls and exit polls in India since 1998. He worked with India Today, Doordarshan, and TV Today (Aaj Tak) from 1998-2005, and with CNN-IBN from 2005-2014.
On Tuesday morning, I met with Professor Karandikar at the Indian National Science Academy in Delhi, where he was visiting for the day. Sitting in his hotel room, he explained to me how he found himself working in psephology, painstakingly went through each step of his vote-to-seat conversion model, and considered what might be required to ensure higher standards of methodology disclosure. The transcript of the conversation below has been lightly edited for length and clarity.
On his involvement in public opinion research
Sam Solomon: You are a mathematician by training. Could you please tell me about your background related to public opinion research and how you got into psephology?
Rajeeva Karandikar: I’m a mathematician/statistician and keen observer of politics, though I have nothing to do with it. Fifteen years after my PhD, I was on the faculty here in Delhi at Indian Statistical Institute, and I used to read whatever little was written in the media about the methodology of opinion polls.That may have not been whole truth but whatever little blurbs were coming in, I used to be fairly critical of that. Especially about sampling methodology.
We had colleagues who were economists. One of my colleagues, Bhaskar Dutta, had some contacts at CSDS. He had come to CSDS for some consultation. And that’s where Yogendra [Yadav] asked him, “We are now going to do a big national poll and we want a proper statistician to advise us on statistics.”
SS: This is when he was launching Lokniti?
RK: Yes, in ‘97. So Dutta came and asked me, “Would [you] be willing?” And I said, “Definitely.” That’s how he got me in touch with Yogendra.
At that point, polls were not [that common].
SS: They did them in the 60s and the 70s, right?
RK: Yeah. In the 60s and the 70s, they did them on a small scale because of money. They developed a robust methodology, questionnaire design, all these things they developed. But ‘97, they were going national. Or maybe they had done it once national, but now it was going to be sustained national polls coming in.
You know, questions like, do we go to the same people who we interviewed last time or do we pick fresh sample? What are the pros and cons? So Yogendra had such questions. And he is a fantastic applied statistician. Apart from being a political scientist, he has a great sense of applied statistics. But he wanted to know theoretically how sound it is and so on. So that’s how I got roped in.
And then within a month of our meeting the parliament got dissolved and the polls were announced for three months later. So we got frantically working for a national poll. That was ‘98. That is where it began.
And in that I got involved so intricately that from then on, it’s almost become my second profession. Almost.
SS: So when you described this in the Hindu Centre article, you had these extensive conversations with Yogendra Yadav but also with a British professor, Clive Payne, about how to do this research and how to convert vote shares into seat shares. You say in the article that you thought UK-style models would not work with the Indian political system. I’d like to know–
RK: A little more about that.
SS: Yes. The unique features of the Indian political system.
RK: In any such applied work, whenever it’s involving sampling, let’s say, the first and foremost important thing is what kind of data is available. By that, I mean the sampling frame, that is, the list of potential respondents. The population. How is it listed? That is number one.
Number two, what characteristics of these respondents are available to you? For example, do you know their age? Do you know their sex? Do you know their economic condition? Do you know their education? Do you know their other social background? In a list. Or, if not at the individual level, at what level are these characteristics available?
This is one of the big differences between the UK and India. In the UK, [for] precincts, which are roughly like polling booths or clusters of polling booths, the socioeconomic profile is available. Not for individual persons, but at least at the precinct level.
SS: By socioeconomic profile, what specifically do you mean?
RK: Education, economic condition. In the Indian context, it could be religious breakup. It could be caste breakup. It’s not all, but at least some of the socioeconomic variables — could be education level, could be income level — are available at the precinct level. That information could be used in modeling the final outcome, and when you can do that then that can also decide your sampling. So this whole thing has to be done together.
In India, that is not available. The census data is there, but census data is aggregated and published at the level of a district. So a lone booth or cluster of booths, even a constituency’s socioeconomic profile is not available because there is no unique identifier between districts and constituencies. A constituency can have parts of different districts and a district can be distributed across several constituencies. This is a reality one has to take into account for modeling purposes. So this is one thing which they use strongly in UK models that Clive Payne uses which we do not have here.
The other important difference was what I call volatility of opinion: what percentage of people change their opinion from one election to the next, say Lok Sabha to Lok Sabha, Vidhan Sabha to Vidhan Sabha. In the UK, talking to Clive Payne, at least at that point in time — it may have changed now [since] this conversation was seventeen years [ago] — the perception was that in the UK the volatility is rather low. Across large sections, all their lives they continue voting Labour or Conservative. Whereas in India, it’s highly volatile. Extremely volatile. As you can see from successive percentages, if you try to look at any kind of data historically, you can see wild fluctuations at the aggregate levels. So now you can imagine [the fluctuation] at the individual level.
At one point, that is ‘98, with that Lok Sabha poll, we had the luxury of doing the following: We did a pre-poll for the whole country just before the first phase, published the outcome. Then CSDS had something for a television program, a counting day program, where Yogendra’s idea was, and we did it rather successfully–we researched, we worked out the model, and the idea was his, but we worked it out: In those days, the counting used to take three days or two days. And the idea was that at any given time, looking at the counting data at that point, [to] make a prediction of the final parliament. So you take the opinion poll and you take the counting data. So for this purpose, Yogendra got money to do a post-poll. Though we were not going to go on air with our findings of the post-poll, we got money to do the post-poll, which was going to be used for this part. This part worked out beautifully. Anyway, that has lost meaning with the electronic voting counting and [after] four hours everything is over. So our ESP, as we called it, Early Seat Prediction or ESP, that has lost its sheen. But at [that] point we did have it.
Anyway, the point is we got the money to do the pre-poll and the post-poll and we went to the same respondents. What we found was that 30 percent of people had changed their vote. And the timeline was, one third of the country eight days, another one third sixteen days, another one third maybe twenty-six days. So on average a little over two weeks and 30 percent of people have changed their minds. What that shows is actually opinions get formed at the fag end just as voting day comes near.
So I have grave question marks on the predictive power of any opinion poll done by whatever, [even] the best of methodolog[ies]. In fact, for several years CSDS and me, that is, we were not doing any pre-election poll.
SS: This was the first year, with Bihar, that they got back into pre-polls.
RK: In the assembly. For the Lok Sabha, we were resisting, but we were told that if you don’t do it, we have to go to someone else. The channel cannot just stay out of that race. So we did that.
Predictive power is rather poor because the volatility is too huge. Some small thing, some statement by someone changes public opinion and everybody turns around and changes their mind. In fact, at a few points, in various times in CSDS polls, we started putting questions like, “Have you decided whom you are going to vote for?” Then, even if they said no, then our follow-up question is, “If the polls were tomorrow, whom are you going to vote for?” Again, 25-35% of people say they have not made up their mind in the pre-poll data.
SS: Did you find that the people who made up their minds later fit a certain demographic profile?
RK: We tried this, but not really. India is not really one homogeneous country. But as soon as you drill down to a state or one section then your sample size is too little to make any such inferential correlations. And it changes from one election to the next. Perhaps the more educated ones are less likely to change.
SS: Is that your own personal experience? Your friends, your family, do they switch their votes pretty regularly from one party to another?
RK: Not as much as we found in the surveys. My friends, my family–
SS: And yourself?
RK: Myself. We make up our minds fairly early. But this fraction is low.
With Bihar — now I have not looked at any post-poll data, I don’t know if you have — but without the post-poll data, I’m not putting too much weight on the pre-poll findings. At best, my statement is that while going into the polls they were neck-to-neck, both sides. And later on we are hearing all kinds of buzz by the media, most of it the English-language media, who are painting one kind of picture. Maybe that is the correct one. Maybe they have feedback from the post-polls or exit polls. I don’t know.
SS: I guess we’ll see.
RK: We’ll see.
On vote-to-seat conversion
SS: Based on your article, a key assumption of your model is that the change in percentage of votes for a given party from the previous election to the present is constant across a given state.
RK: Or some region in the state.
SS: Or some region in the state. This is actually different from what I’ve heard from some pollsters about elections in India and they have said that there isn’t really such a thing as uniform swing.
RK: It is true that any kind of statistical analysis post facto, or back testing as it is called, if we do this model it fails miserably. But this model is the basis which allows me to mathematically come up with a methodology, as opposed to just ad hoc. And this has proved to be reasonably good and no one has been able to show me a methodology which is better, other than their gut feeling.
SS: And a big part of it is that you’re predicting the aggregate number of seats, not particularly for any individual constituency.
RK: Absolutely not. Absolutely not for the constituency. In fact, when we do parliamentary polls I used to be extremely reluctant to even give [projections] at the state level. I would say that at best we will give the four regions: north, south, east, west. Not even at the state level.
SS: So you notice state swings but also regional swings within states. How do you group the regions of a state then?
RK: Grouping of regions is done politically by CSDS. So they’ll say, “Socioeconomically, this is [region] 1.” They used to do the splitting of states. It was generally geographic, but still how would they draw the boundaries? This is where their understanding of the socio-economic conditions across the country played a crucial role.
Let’s take Maharashtra. Four regions. Greater Bombay is one region. Now suppose we measure that the overall swing for BJP statewide is 5%, and in Bombay region it is 3%. We’ll assign some way, either half-half, or one thirds-two thirds, so I’ll take half of the state swing plus half of the region swing; that would be my swing for every seat in Bombay region.
Now these regions can also be extended to what we may call phases. Now that elections are going on for a month and what happened first phase and what happened in the fifth phase, lots of things have changed. So I’ll take one third of the state swing, one third of the phase swing, and one third of the region swing, let’s say.
SS: The phase swing, what would that come from?
RK: The phase swing would come from last election in this phase what was the percentage of votes, and in my opinion poll [for this election] what was the percentage of votes. So the difference will give you the phase swing.
SS: And you look at the phase across the entire state?
RK: Across the entire state.
SS: And so you take the last election, 2009, or 2014 and you’d say–
RK: No, no, no. So in Bihar, there are lots of interesting points which will emerge. Primarily, we take not 2014 but the 2010 Vidhan Sabha poll.
SS: For Bihar 2015, you take the 2010 Vidhan Sabha data.
RK: However, from 2010 to now, lots of things have happened. The alliance structure has changed dramatically. Somehow we have to take that into account and come up with what we call a simulated initial file. If nothing had changed, the 2010 [file] would have been the initial file on which we look at the swings. But now what we do is — this is ad hoc — but firstly, wherever BJP contested on behalf of the [former JD(U)-BJP alliance that governed Bihar from 2005 to 2013], I would split it, let’s say, 60/40. And if it is a JD(U) candidate, I will give it 60/40 [favoring] the JD(U).
SS: And how do you come up with those figures? You just have to make some sort of subjective judgment.
RK: Yeah. Some sort of subjective judgment. And it used to be easier when Yogendra was a partner because then he would have a very good sense of these things.
SS: Just from his own experience.
RK: Yeah. And over the years I think I have developed a keen [sense]. So now I don’t fret over it. I do it and proceed.
The other thing is it is very robust because it is only used to do the initial file. 60/40 or 55/45 or even 65/35 is not going to make a huge change in the final analysis for the total prediction at the state level. I’ll experiment with this. Of course if I do 10/90 or 90/10 it will not be [similar], but in a certain range it really doesn’t change dramatically.
SS: It doesn’t make a huge change at the aggregate level. Maybe at the individual constituency [level].
RK: Aggregate level.
So likewise, when two groups align, then we don’t simply add their votes, but there is going to be some friction. So we might take 85% or 90% or something like that.
Or if it were UP [Uttar Pradesh], let’s say, and in the previous election Mayawati and Congress fought separately, and let’s say in this election they are fighting together, Yogendra would say that, “Whatever were Mayawati’s votes, even if there was a Congress candidate, it will transfer 100%. But whatever Congress votes, if there was a Mayawati candidate, it will only transfer 50%.”
SS: And that’s just knowledge he has from being in Indian politics for a very long time.
RK: Yes. So we take into account whatever we can to bring in a subjective judgment about, “If this were the alliance picture [in the previous election], what would the vote have been the last time?”
Now, in Bihar, there is another challenge. We see that there was a Lok Sabha election which changed the realities tremendously. So for Bihar what I had been doing was, I took the file from 2010, applied the 60/40 [vote split between the JD(U) and BJP], and then I took the  Lok Sabha [file], and I took an average of these two.
SS: You took an average of the [re-weighted] 2010 Vidhan Sabha file and the 2014 Lok Sabha file as your base [file].
RK: Under my model, the level of support is determined by the opinion poll. I am not applying any subjectivity there. The whole thing is how is that vote distributed across the state? If you say the BJP is getting 41% in Bihar, how was it distributed across the constituencies? So my base file’s only role is to make an assessment about how it is distributed.
SS: How it’s distributed within each constituency.
RK: Within 243 constituencies [for Bihar]. Where is it that they are polling more than the average? Where are they lower than the average? Things like that.
SS: And then you’re going to take the uniform swing — or the uniform swing plus the regional swing plus the phase swing — to calculate what you think would happen in each constituency?
RK: Once again, the base file and the model together tell me how this 41% of votes is distributed across the constituencies. That model, I am not using it to say it is not 41%, it is 44%, no. The level is pegged by the observed data. The model and the base file altogether only help me in attributing that across constituencies. For the BJP as well as for others.
SS: At the aggregate level and not in individual constituencies.
RK: No, no. I do it at each constituency, I aggregate assigned votes. Then I would say who is winning where and I will add it up to get my final number.
SS: But the actual constituency level projections are not– the model doesn’t do well at the micro-level.
RK: No, no, absolutely not. Absolutely not.
SS: [So] in the aggregate the errors cancel each other out. Or at least, that’s the theory.
RK: That’s the theory. And at least I have observed [it] in practice reasonably well. If our estimates are good, this is good.
SS: I was hoping that we could — we kind of did it already — but we could walk through each stage of the process, slowly and clearly, just to explain the vote share to seat share conversion. So I have it all on [the] recording.
SS: So first, you do the survey to determine the uniform swing across the state. You also have a regional swing and a phase-based swing. And then you’re using that to calculate projected vote share in each constituency.
RK: For each party. For each of the major parties.
SS: So, for example, if the BJP is projected to win 5% more votes statewide in Bihar, and they won 36% of Constituency X in 2010. Then it becomes 41% in 2015. You just use Vidhan Sabha figures to build the model, but you just said that–
RK: In special cases like Bihar this time.
SS: And the reason for that is because the 2014 Lok Sabha was such a huge–
RK: Tectonic change. It’s a big change.
So that is why while there is one broad model, there is no relevant software where you just plug in data. Each time we may have to come up with some tweak, like this one in Bihar.
SS: So once you’ve got the projected vote shares in each constituency, you calculate the probability of victory for each of the parties for each constituency. Do you do this by doing a multiple proportion z-test for each constituency?
RK: Something like that, something like that. And there is again a subjective parameter, the sigma, or the standard error for the vote estimate.
SS: That’s subjective. You’re using the standard error for the entire sample but you’re using it for just a constituency.
RK: So therefore we put a factor of safety. So whatever is the sigma, I inflate it by 2 or 3 or something like that.
SS: You’ll inflate the standard error to allow for more uncertainty in the model. Okay.
But you only do this for the top three parties. Why do you stop at three? It seems like there might be some cases in India where more than three could be competitive in a constituency.
RK: The first time we did it, we did it with only two [parties]. And then we found that there were [many] places where the party [that was in third place] ended up winning. So we extended it to three. Now, rarely is someone who is fourth in my estimate actually winning. It’s very rare.
It can be done. It will make matters a lot more complex.
SS: And it doesn’t add a lot to the model.
RK: It doesn’t add a lot to the model. That’s the point.
To come up with these conversion tables and so on, all we had to was do it with the bivariate normal distribution. If I have to go to four, I will have to go to three-dimensional Gaussian, and that can be done. If it was necessary, it can be done. But it simply was not worth the effort.
SS: It sounds complicated.
And then you say it’s as simple as summing the probabilities of each constituency across all the constituencies. But wouldn’t you want to run the model many times and see what kinds of distributions you get?
RK: Of course. Of course. That I do.
So, for example, I can play with the different weights for state/region/phase [swings] and so on and so forth. I can play with different weights there. Then I can vary the sigma parameter, the inflating factor. I’ll run it several times and then get a sense and try to position myself somewhere in the center, normally.
Whatever are my subjective parameters, I vary those. I don’t vary the vote percentage. I vary the subjective parameters.
SS: Does that include the base file? For example, if there’s an alliance–
RK: To begin with, I might have made two or three base files. I don’t vary them tremendously, but I might have made two or three base files.
SS: If there is some unusual alliance, as there is in Bihar.
RK: For Bihar, what I have is one pure 2010 base file, one pure  Lok Sabha base file. That I will not use, but I will [weight them against each other by] one third-two third, half-half. So this is the variation I might try.
SS: With each of those, you do different projections based on changing state swing, or regional swing?
RK: No, weight for the regional swing.
SS: Ah, how much you balance the regional swing and the state swing.
RK: The main point is none of my things make a modification to the opinion poll data. It is just interpreting that vis-a-vis seats which is where I play with my subjective model and parameters.
SS: And again those would include how you’re weighting Vidhan Sabha versus Lok Sabha data for the base. It would include the different vote shares that the alliances are picking up. And then also you said the sigma, so how much uncertainty you add to the model for each constitutency. Any other subjective parameters that I’m missing there?
RK: Not really [editor’s note: we both forgot to mention here how the state, regional, and phase swing are weighted against each other].
SS: How does this work for Lok Sabha elections? You say in the article that there’s uniform swing within states, within regions of seats [editor’s note: this is incorrect. He uses uniform swing and regional swing as the basis of making projections but does not describe the seats as uniformly swinging one way or another], but not across the country. Do you treat each state as its own election?
RK: Absolutely. Absolutely. Because that’s how India is, at least electorally. [There are] districts across Tamil Nadu and Karnataka boundaries, or districts across UP and Bihar boundaries. Socioeconomically they may be very similar but political behavior is very different. That’s why there is no nationwide swing or nationwide effect. So I treat each state as separate.
SS: Those subjective factors that you listed in the model, [those] can account for a lot of [variance] in seat projections, right? For a lot of the polls for Bihar, a lot of the vote share projections have been pretty similar but the seat share projections can be pretty different.
RK: See, but nobody else even says how they are doing the seat projections.
SS: So you have no idea how the other–
RK: No, because they don’t say. I am the only one who says, and I go and give lectures, and my methodology is public [and] open.
SS: But no one else does this because it’s their “secret sauce.”
RK: I believe there is no “secret” and there is no “sauce.” They just do it ad hoc, is my guess.
SS: Do you think so?
RK: Yes. They will have some rule, “Oh, if it is two parties and only 5% swing then maybe…” Something like that. Some rules.
SS: But not transparent.
RK: No, certainly not transparent.
SS: And not methodical in the way that you’re describing here.
RK: At least, they have not disclosed it.
I would have no issue in debating with anyone my methodology versus theirs, but no one else comes forward.
On transparency of pollsters and disclosure of methodology
SS: You spoke at the beginning about transparency in disclosure of methodology, and how you were looking at this in the ‘90s. Has it changed a whole lot from the ‘90s to now?
RK: Not really. Not really. We zeroed onto a fairly good methodology and it’s these tweaks, little tweaks, taking into account the ground realities which we may be doing, but other than that the primary methodology has remained the same.
SS: In terms of the reporting, the disclosure of how these surveys are done. I’ve talked to people in media but also researchers themselves who say, “We need to get better about disclosing our methodologies.”
RK: Absolutely. The only thing is it cannot be mandated by law, but it should be voluntary by the media.
SS: Why can’t it be mandated by law?
RK: There is a very tricky issue. The regulation of media is a touchy issue. What can be regulated and what cannot be regulated is also debatable.
For television stations, the government can come out with a law because these are of recent origin and when they got licensed to uplink their signal from the land, they signed on [to] all kinds of norms, and opened themselves for regulation by the government. So they can be regulated. Otherwise, their license can be cancelled.
However, when it comes to print media — because print media predates independence by a century — print media cannot be regulated. In fact, Mr. N. Ram [the publisher] of The Hindu [has given] lectures in which he has categorically said that, “We obey your 48-hour [embargo on releasing polls before Election Day] of the Election Commission, because we think it’s a good idea. Not because we are bound to it.”
SS: So there has to be some kind of internal self-regulation.
RK: There is something called the Press Council in India. There is some other body for the television broadcasters association. These bodies can together come up — Yogendra had suggested this, and I strongly support that — at least, when was the survey done, what was the sample size, who led the survey, what [were] the names of the team members who did that. And a little more about the vote percentages / seat percentages methodology.
After Yogendra became a member of AAP, he in fact said that the entire survey data should be made public. That maybe is going a bit too far. But it’s like an audit. No company accounts are made public. However, it is subject to an audit by a professional. Likewise, we can be required to maintain records in a certain fashion that can be subject to audit by other professionals.
SS: Why do you think this hasn’t happened yet, if Yogendra has talked about it, if so many researchers have talked about it, [and] people in the media are talking about it?
RK: I’m not sure how many actually want it. Yogendra definitely wants it. I want it. But for both of us, I have to say, that for both of us it was easy to say and even practice because this was not our primary profession. We both have a different profession. We have our livelihood secure. Whether this opinion poll happens or doesn’t happen, whether we are engaged or not engaged, we don’t care.
With Rajdeep Sardesai and me and Yogendra, we had a very clear understanding that Rajdeep had no say in what we were going to say. He could say, “Hey, what you are saying doesn’t seem to be happening here. It’s not going to happen.” We would say, “Okay, we’ll examine our analysis once more.” But he could not tell us to reduce it or increase it, no.
SS: You had complete independence from media sponsors.
RK: Yes, absolutely.
On experience conducting surveys over time
SS: You got experience not just doing the vote-to-seat projection but also working on the surveys themselves. The specific focus of my research is systematic sampling error in the polls, whether there are certain populations, demographic groups — gender, religion, caste-based — that are more or less likely to be represented in these surveys. Did you find consistent patterns related to that?
RK: With CSDS data, I don’t think there was a consistent pattern [other than] underprivileged groups being slightly underrepresented.
SS: Which types of groups specifically?
RK: Rural areas, SC [Scheduled Castes], ST [Scheduled Tribes]. But this, Sanjay [Kumar, Director of CSDS] would have a much better sense than me.
But with other surveys, there are grave doubts about what happens, when they do convenience sampling. You know most others do convenience sampling, quota sampling.
I have [put] on record [several times] that whatever [are] my successes or limited successes is largely due to great quality data by CSDS. If the data is no good, my analysis cannot be good to begin with. And CSDS has done a phenomenal job.
SS: Have you found that certain states are harder to project than others?
RK: Not really, though our track record said that we did very well in east India and better in south India. Not so well in west India. (laughs) But it’s by chance. I am not sure why it happened.
In the east, [Lokniti and I] have a phenomenal record.
SS: The northeast?
RK: East, not so much northeast. [In the] northeast, only [in] Assam have we done surveys. Assam, we have done very well. Bengal and Bihar, we have got bang on when others did not expect it. Both victories of Nitish. the first time in 2005 that he gets a majority, and then in 2010 [when] they crossed 200 [Vidhan Sabha seats], nobody had dreamt. Nitish himself had not dreamt. And we got it.
SS: And you did very well for southern states as well?
RK: Reasonably well for southern [states]. Not so well for [the] west. I am from [the] west. (laughs)
SS: Which southern states and western states do you mean?
RK: Gujarat in particular, we had a checkered record. Not so good. Maharashtra, so-so. But [the] south: Tamil Nadu we have done well, Karnataka we have done well. AP [Andhra Pradesh] okay.
SS: Thank you very much.