Are marginals different? Evidence from British elections 1950–2015

If people vote only when expected benefit (utility from deciding the government times the probability that your vote will do so) exceeds the cost of voting, no rational person will vote. It is always open to political scientists to attribute bad statistical sense to voters, so that they exaggerate their chances of personally ejecting or saving their MP, or even the government itself. But ‘minimum rational choice’ and ‘general incentives’ have been the more popular options over the years. According to ‘minimum rational choice theory’ (Riker and Ordeshook 1973) voters also define their political efficacy ‘in terms of the influence of groups of people like themselves. Consequently, they are motivated to take action because they feel they can collectively make a difference’ (Clarke et al. 2004, p. 248). If these groups include your political allies within your own constituency, and collective efficacy is related to the likelihood that collective action will be pivotal, we have a rationale for higher voter turnout rates in marginal seats. The general incentives model (Whiteley et al. 1994; Whiteley and Seyd 2002) encompasses the minimum rational choice approach but adds other motives for voting, in particular the expressive motive first put forward by Buchanan (1954). And, in doing so, it provides not just a reason for higher turnout in marginals but a rationale for a different pattern of voting. If voters have a mixture of instrumental and expressive motives, the relative power of the latter will decline as the efficacy of the vote increases (Tullock 1971). Consequently, voters who would like to express a protest against the government, but still prefer it to the main opposition party, might give it a grudging vote in a marginal seat which they would withhold in a safe one.

This, then, is our first reason to expect marginal and safe seats to have different patterns of swing. Price and Sanders (1998) extend the analysis by attributing two alternative objectives to disenchanted government supporters

Either of these factors would reduce anti-government swings in marginal seats. But in that case they must reduce any pro-government swings in marginal seats, too. When the trend is towards the government (i.e., increases its majority), some voters in safe seats, previously protesting against the government, will return to it. So will some in the marginals, but not so many, because fewer deserted it in the first place. Whatever the direction of the national swing, then, its beneficiary will gain fewer votes in marginal seats than in safe ones.

Secondly, a generally positive personal vote for incumbent MPs will give parties particularly favorable swings in their own marginal seats taken as a whole. Unless a party captured no seats at all from its opponents at the previous election, its marginals will include a disproportionate number of candidates standing as the incumbent MP for the first time, and thus enjoying an ‘incumbency swing’ less prevalent in safe seats. Several of the election studies by Steed and/or Curtice (see below) notice an incumbent government holding rather successfully onto its marginals when it is losing ground in general. A particularly strong and persistent effect in their studies is how well a party generally does when its incumbent MP fights a seat he or she gained from an incumbent opponent at the previous election. This ‘double incumbency’ effect seems to have been particularly striking in the elections of 2005, 1992, 1987 and 1970. (Steed 1970, p. 205; Curtice and Steed 1988, pp. 333–334, 1992, p. 340; Curtice et al. 2005, p. 248). In the 1970 case, Steed finds that the double incumbency effect fully explains why the outgoing Labour government suffered a less hostile swing in its 100 most marginal seats than across the country as a whole. (In 1979, by contrast, when a Labour government was again defeated, but did less badly in its own marginals, Curtice and Steed 1980 found the double incumbency effect to explain only part of the phenomenon.)

The third reason why marginals are different is that parties campaign harder and spend more in marginal seats. Again the most obvious—and well-documented—effect is on turnout. Turnout in marginal constituencies has exceeded average national turnout in every British election since 1950, but the size of this margin has fluctuated substantially. Denver et al. (2003) trace the widening of the gap up to 1979, and its subsequent narrowing during the Thatcher/Major years, only for it to open up again in 1997 and 2001. Controlling for socioeconomic status (a significant determinant of turnout) alters the story insofar as the downturn after 1979 is now backdated to 1966. The question raised is whether higher turnout in marginals comes from the more intensive campaigning they undoubtedly attract or whether the expected closeness of the result would have called out more voters anyway. Denver, Hands and McAllister conclude that both of these factors matter, though their relative strength appears to vary significantly between one election and another.

In addition, Curtice and Steed (1980) have found that the turnout effect of being a marginal seat rises in elections that follow hard on their predecessors (1951, 1966, October 1974) and then drops back at the subsequent election (1955, 1970, 1979). Presumably this is because recent memories of the previous election include a memory of the constituency’s marginal status. If so, this is itself evidence that higher turnout in marginals is not explained solely by more spending and harder campaigning there. So, indirectly, is the finding by Clarke et al. (2004) that turnout is higher across the country when the national result is close, given the lack of any inverse correlation between national winners’ margins and combined campaign spending.

To assert that more people would vote in marginal constituencies even if no party spent a penny on them is not, of course, to say that campaigning is ineffective. Indeed, recent evidence suggests not only that it is more effective than was previously thought, but that the two main parties, even if they spend equal sums of money nationally, do not just neutralize one another’s efforts. Some of this is because parties tend to spend more in their own constituencies (Clarke et al. 2004). All of this means that we would expect to see each party performing better in the marginals that it holds than in those it is trying to win over, and Clarke et al. (2004) find exactly this in the 2001 election. In the 2015 election, the difference was particularly striking. Across the 50 most marginal seats, the Conservative-held ones swung by an average of 1.2% towards the Conservatives and the Labour ones by 2.5% towards Labour.

What the above facts imply is that good reasons exist to think that non-uniformity of constituency swings is likely to be systematic as well as random, and a simple relation whereby marginal seats feature smaller or larger swings than safe ones is unlikely. And certainly the successive analyses of election swings by Michael Steed and John Curtice do not add up to any such pattern. Aggregating their surveys, however, is not easy. The definition of a marginal changes between surveys; in 1987, for example, they looked only at government marginals; in elections that follow boundary changes they do not try to identify marginals at all. This is not a criticism. Different approaches are appropriate to different electoral situations. But it does mean that an aggregative study of very safe, moderately safe, marginal and bellwether seats reaching down from 1950 has yet to be conducted. That is what we are going to do.

In the model that follows, expressive voting (voting to demonstrate support for a party regardless of an election’s result) and instrumental voting (voting to try and affect the outcome of the election) both depend on (1) the parties’ respective valences and (2) the voter’s ideological distance from the parties.

However the ideological and valence motives for voting are related insofar as a voter’s ideology drives his or her perception of valence. It is not just that ideological location colors a voter’s view of how well a party is handling a particular issue (Sanders et al. 2011) and how competent a party’s leader is (Jacoby 2009). Left and Right also rank the importance of the issues differently. Thus, for instance, even if voters were unanimous that one party is better at handling unemployment and the other at keeping inflation down, their differing weighting of the issues themselves would ensure divergent valence scores across the electorate (Palmer et al. 2013).

Sanders et al. (2011) find that spatial (i.e., ideological) variables have little effect on voting until their indirect effect, operating through valence perception, is isolated and added to the story. In what follows, we lump the direct effects and indirect effects together and define them as ideological voting. Valence voting, therefore, means its ‘pure’ form, i.e., voting stemming from perceptions of valence that are independent of a voter’s ideology. For simplicity, we assume that this perception is common to all voters i.e., that different estimates of valence stem entirely from ideological differences.

So let voter i derive the following expressive utilities from voting for L and R (the two largest national parties) respectively:where v is ‘pure’ valence as defined above, θ is ideological location and \( \beta \) captures the effects of ideology both directly and indirectly through imputed valence.

Now, over and above any expressive utility, let voter i’s welfare gain from having L’s policies, not R’s, implemented be

Hence, i’s instrumental utility from voting for L is:where p(x) is the perceived probability that i’s vote will be decisive if i lives in constituency x.

Hence, putting expressive and instrumental utilities together, and writing \( v_{L} - v_{R} \) in future as v, i’s total excess utility from L rather than R is:or, for a voter between \( \theta_{L} \) and \( \theta_{R} \),

We now assume that voters will vote for a third party or abstain if \( |U_{i}^{{\prime }}| < y_{i} \), where \( y_{i} \) is uniformly distributed across the electorate from 0 to a maximum of \( \bar{y} \). (Thus a voter for whom \( y_{i} = 0 \) will always vote for L or R, while a voter for whom \( y_{i} \) is large enough will be loyal to a third party or abstain in all circumstances.)

Voter i, then will

We assume that voters in each constituency are normally distributed in terms of \( \theta \) around the constituency mean (and median) \( \mu_{x} \). We calibrate the \( \theta \)-axis so that \( \mu = 0 \) in the seat where the L and R votes currently are equal.

Now let \( \theta \left( y \right) \) be the ideological position of voters for whom U′ = y. It follows that \( \theta (\bar{y}) \) and \( \theta ( - \bar{y}) \) are the ideological points outside of which all voters support L or R, so that the swing in constituency x caused by a small change in v will depend on the number of voters in that constituency within the range \( [\theta ( - \bar{y}),\theta (\bar{y})] \).

Now let v increase by dv. Then from (1):

This is the size of the ideological interval within which voters previously below U′ will attain it. Now consider voters for whom U′ = y. These voters are at \( \theta \left( y \right) \); let their number in seat x be \( n_{x} \left( {\theta \left[ y \right]} \right). \)

Thus, the number of voters in seat x whose U′ reaches y as a result of the increase in v is: \( n_{x} \left( {\theta \left[ y \right]} \right).\frac{\alpha + p\left( x \right)\gamma }{{2\left( {\beta + p\left( x \right)\delta } \right)}} \) dv. Of these, given \( y_{i} \)’s uniform distribution between 0 and \( \bar{y} \), the proportion \( 1/\bar{y} \) will start voting L in consequence.

Thus L’s gain among voters who require U′ = y in order to vote for it isor, aggregating across all possible values of y, L’s gain in constituency x iswhere \( n[0,\bar{y}] \) is the number of voters between \( \theta = 0 \) and \( \theta = \bar{y} \).

By symmetry R’s loss (of voters whose \( U_{i}^{{\prime }} \) now exceeds, their −y _i ) is where the second term gives the number of ‘marginal’ voters in the constituency and the third term measures the effectiveness of a valence change in shifting them.

Since the median voter in seat x is at ideological position \( \mu_{x} \), while the median voter in the most marginal seat is at ideological position 0, it follows that that the greater is the absolute magnitude of \( \mu \) the less marginal is the seat. We now consider the sign of \( \frac{ds}{d\left| \mu \right|} \), i.e., whether swings will increase or decrease as seats become less marginal. Clearly the first term in (3) is independent of \( \left| \mu \right| \). We now show (1) that the second term is declining in \( \left| \mu \right| \) and (2) the third term might be either declining or increasing in \( \left| \mu \right|. \) which is negative (positive) when \( \mu_{x} > \left( < \right)0. \) Thus the further the constituency is from the knife-edge seat, whether to its left or to its right, the lower is this component of the swing. This is just saying that marginal seats have more marginal voters.

Thus, \( \frac{\alpha + p\gamma }{\beta + p\delta } \) is increasing in p and, hence (assuming \( \frac{dp}{d\left| \mu \right|} < 0 \)) decreasing in \( \left| \mu \right| \) iff \( \beta \gamma > \alpha \delta . \)

We now consider some implications. Throughout, we assume that aggregate voting has both an instrumental and an expressive component.

Here, both the second and third terms of (3) are decreasing in \( \left| \mu \right| \).

The logic goes as follows. The component terms of \( \beta \gamma > \alpha \delta \) are:

But if \( \beta \gamma > \alpha \delta \), then \( \beta /\delta > \delta /\gamma \). This means that, compared with their respective sensitivities to valence, expressive voting is relatively sensitive to, and instrumental voting relatively insensitive to, ideological differences between the voter and the party. But in a safe seat, the expressive motive is relatively strong and the instrumental one relatively weak. It follows that voting in these seats is ideology- rather than valence-driven, so that it requires a large change in valence to get constituents to switch their votes. In this case, then, a given change in valence will be less effective in safe seats than in marginal ones, and this reinforces their existing tendency towards smaller swings, given by the fact that they have fewer marginal voters in the first place.

In this case, the second term of (3) is decreasing, but the third term increasing in \( \left| \mu \right| \).

Why? This time the intuition is as follows. When \( \alpha \delta > \beta \gamma \) (i.e., \( \alpha /\gamma > \beta /\delta \)) it is instrumental voting that is relatively sensitive to, and expressive voting relatively insensitive to, ideological differences between the voter and the party. It is thus in the marginal seats (where instrumental voting is most important) where the vote is driven most strongly by ideology, and where it thus requires the largest change in valence to get voters to switch. In this case a given change in valence is more effective in safe seats than in marginal ones, and this effect counteracts the latter’s tendency towards smaller swings, given by the fact that there are fewer voters on the margin in the first place. With these two effects now working in opposite directions, a monotonic relation between existing majority and swing may no longer exist.

All of the foregoing considerations cast a new light on the proposition that a government falling out of favor may lose more support in safe seats (where voters feel free to express their feelings of protest) than in marginal ones (where some voters would like to protest but do not actually want the government to be ejected from office.) The implicit assumption behind this story is some voters are expressively anti-government but instrumentally pro-government. This could be the case—indeed it corresponds to the case where \( \alpha \delta > \beta \gamma \) in our model (as the government’s valence declines, expressive sentiments change faster than instrumental calculations). But it doesn’t have to be the case—it is equally possible that the declining valence has more impact on the instrumental motive than on the expressive one. In that case relatively few voters will be deserting the government as a protest, and many more doing so as a strategy. As the marginal seats are the ones where strategy is most preferred to protest, the government will do worst in the seats where it can least afford to lose ground.

But so far we have been assuming that national swings are caused by changes in the valences of the parties. What if voters are switching because one of the parties has shifted its ideological position?

Suppose that \( \theta_{R} \) rises by \( d\theta . \) Then from (1) the value of \( \theta \) needed for a voter to reach any given U′ rises by \( d\theta /2 \). Hence, \( d\theta /2 \) replaces \( \frac{\alpha + p\gamma }{{2\left( {\beta + p\delta } \right)}}dv \) in (2), so that (3) becomes \( s_{x} = \frac{1}{{\bar{y}}} \cdot n_{x} \left[ { - \bar{y},\bar{y}} \right]d\theta /4. \) Since, as we have seen, \( n_{x} \left[ { - \bar{y},\bar{y}} \right] \) is declining in \( \left| \mu \right| \) we have a monotonic inverse relation between existing majority and swing. Clearly, the same relation would also hold if both parties shifted ideology, or (which comes to the same thing) the electorate itself underwent a wholesale ideological move one way or the other.

To summarise then: (1) If swings are driven by valence changes, and valence is more dominant in instrumental voting than in expressive voting, the largest swings will be in the most marginal seats. (2) If swings are driven by valence changes, and valence is more dominant in expressive voting than in instrumental voting, then the ideological position of the swing voter shifts furthest in safe seats, compensating for the fact that they have fewer swing voters to start with, and possibly producing either an increasing or a hump-shaped relation between existing majority and swing. [It cannot be U-shaped because at \( \mu = 0 \), swing is increasing in \( \mu \) on account of the second term in Eq. (3) being independent of it—see Eq. (4)]. (3) If parties are gaining or losing ground because their own ideology, or that of the electorate, has shifted, again the swing will be larger the more marginal the seat. Therefore, should we find either an increasing or a hump-shaped relation between existing majority and swing, the evidence is consistent with (2) only.

Before leaving the model, note that the parameters \( \alpha ,\beta ,\gamma \) and \( \delta \) together ensure that the relation between p and swing can vary by several orders of magnitude. Inspection of Eq. (5) will show that a large (small) difference between the products \( \alpha \delta \) and \( \beta \gamma \) will mean that small (large) differences in p between different constituencies can produce large (small) differences in swing. This is important, given the finding in the British Election Study (2005) that the marginality of the seat in which a voter lives makes only a very small difference in the perceived probability that they can affect the result.¹ As will be seen in the results section of this article, the difference in average swing between marginal and safe seats, though not great, is greater than one might at first sight expect given the British Election Study’s results. This points to a large difference in the respective magnitudes of \( \alpha \delta \) and \( \beta \gamma \).

To analyze the impact of a seat’s marginality on swing we make use of all constituency results (Northern Ireland excepted) wherein either Labour or the Conservatives won between 1950 and 2015. However, the elections of February 1974, 1983, 1997 and 2010, all of which followed major boundary changes, are used only to calculate swings at the succeeding election (and to establish which seats being contested by incumbent MPs). Our basic regression model is run using OLS and is as follows:here j refers to the constituency and t to the general election. Swing is swing between Labour and Conservative,² with positive swing representing swing to the gainer (party favored by the national swing). M is a matrix of variables measuring the marginality of the seat, X is a matrix of conditioning variables, and D is a set of year dummies, i.e. fixed effects for each election. Fixed effects for each constituency would not be appropriate because the majority of constituencies consistently were safe or marginal between each set of boundary changes. Controlling for constituency characteristics would thus largely control for just the variable we need to isolate for analysis.

For the purposes of this article we use traditional or ‘Butler’ swing. Traditional or ‘Butler’ swing looks at parties’ shares of the entire vote and averages one party’s gain and another one’s loss. Two-party swing replaces shares of the poll by shares of the combined vote of the two parties concerned. In some contexts, two-party swing captures the dynamics of the political process better than traditional swing (For example, it does not find a spurious ‘swing’ between the two main parties as a result of voters switching between third parties and abstention). Nonetheless, we use the traditional version of swing in this article. That is because much of our analysis is to do with the possible behavioral effects of voters’ realization that they are, or are not, in a marginal constituency. Public discussion of the swing needed to topple an MP or a government is couched entirely in the language of traditional swing. ‘The Tories’ 96th most winnable seat’ is not the one where the Conservative-Labour vote ratio is the 96th nearest to 50% but the one where Labour’s percentage margin over the Conservatives, as a proportion of the entire vote, is the 96th smallest. Commentators on election night do not present two-party swings, nor do they appear to have them in mind when they say ‘that’s one Labour should have won if it’s seriously hoping to form a government’. Consequently, we use Butler swing as the one most likely to shape electors’ perceptions of the consequences of their vote.

A further important methodological point is how to model the kind of incumbency effects, postulated and confirmed by Steed and Curtice, whereby a sitting incumbent might specifically brake or accelerate the national electoral trend. What we do is create two dummy variables, INCUMBENT and LAG.INCUMBENT. INCUMBENT is entered as 1 (−1) if an incumbent MP standing at the current election belongs to the party that is gaining (losing) ground nationally. LAG.INCUMBENT is entered as 1 (−1) if an incumbent MP standing at the previous election belonged to the party that is gaining (losing) ground at the current election. Thus, if, for example, the same MP fought both elections we would expect the coefficients on INCUMBENT and LAG.INCUMBENT to sum to zero (no incumbency swing will be observed at this election because the MP already will have been enjoying an incumbency advantage). If incumbents of opposing parties respectively fought the two elections, we would expect a double incumbency effect, such that the product of INCUMBENT and its coefficient, and that of LAG.INCUMBENT and its coefficient, would have the same signs.³

We now consider some factors that might affect voters’ perceptions of their ability to affect the result. So far as these change p(x), they will alter the relative weights of expressive and instrumental voting and, hence, the relation between existing majority and swing. In the first place, voting may become more expressive and less instrumental when little doubt exists about who is going to win nationally. Swinging the result in a marginal seat may seem less important to the voter when the marginal seat has no apparent chance of swinging the national outcome. The difference in swing between marginal and safe seats would then be less pronounced. Thus, we create the interactive variables \( MAJ*NATLEAD \) and \( MAJ^{2} *NATLEAD \), where NATLEAD is the national percentage lead (over the runner-up) of the election winner. Should the coefficient on \( MAJ*NATLEAD \) be significant and opposite in sign to that on MAJ, that would be consistent with our conjecture—that the more decisive the national result, the less distinctive the voters’ behavior in marginal seats.

Column (1) of Table 2 shows some evidence for the conjecture. MAJ * NATLEAD and MAJ ²  * NATLEAD are both statistically significant and opposite in sign to MAJ and MAJ ² . It seems that marginal seats do behave more like safe ones when only the constituency result is at stake. The above does assume, however, that voters’ expectations are best proxied by the actual result. An alternative measure would be the average of opinion poll forecasts on Election Day. The variable POLLLEAD is the average percentage lead, on polling day and across all polls, of the party forecast to win. As column (2) shows, some support for the conjecture still is found, but the coefficients and significance levels of both interactive variables are down. It would seem that voters’ behavior is to an extent driven by their expectation of the national result, and that they are better at predicting this than the pollsters—which will surprise no one after the latter’s performance at the 2015 general election and 2016 EU referendum.

Because NATLEAD is more significant, we use it in the construction of Fig. 3. The lower position of the curve as NATLEAD increases shows how constituency swings become more uniform as the national result becomes more decisive, as conjectured above. The bunching of the curves as MAJ increases indicates the converse: that the national result matters less as seats become safer. Again, we might expect this: as the seat gets safer, voters will care less about the national picture as their hopes of influencing it fade.

The other notable thing about Fig. 3 is that at NATLEAD =15, the shape of the curve reverses and it becomes a U. Given that the largest lead in our series of elections was 11.4% (Conservative over Labour) in 1987—leads were larger (though still below 15%) in 1983 and 1997 but these elections followed constituency boundary changes and were thus used by us only to derive swings at the following elections—the curve is hypothetical and should be treated with caution. All the same, Eq. (3) yields at any rate a partial rationale for its shape.

In (3), the only factor preventing a monotonically downward sloping curve was the change in the third term brought about by changes in MAJ by way of changes in p. But, as argued already, p likely is less sensitive to MAJ when NATLEAD is large. Hence, a downward-sloping curve is more likely when the national result is very lop-sided. Why the curve for NATLEAD = 15 should turn up again when MAJ is very large remains a puzzle, but we are dealing here with both a small number of very safe seats and a notional overall result way beyond anything actually experienced since 1950.

Another possibility is that voters’ awareness that the seat is marginal and, hence, the way some of them vote, depends on how long ago the last election was. As we have already seen, Curtice and Steed have found that turnout in marginals is greater than in non-marginals and that turnout in marginals is greater still in elections that come quickly after their predecessors (1951, 1966, October 1974). But is there a parallel effect on the way the inhabitants of marginals vote, as opposed to the likelihood that they vote at all? When we run swing to gainer on majority and majority squared for the elections of 1951, 1966 and October 1974 only, the coefficients on both variables go up (column 3). Three elections comprise a rather meagre data set, but our finding, so far as it goes, implies that voters in marginal seats do behave more distinctively when a recent election reminds them that the seat is marginal.⁴

A fundamental objection to the proposition that ‘marginals should be different’ is that it is not the existing marginals that are going to swing the election, but rather those seats that would be marginal in the event of a close national result (the ‘bellwether’ seats). The foregoing argument assumes that voters will regard as crucial those seats currently having small majorities. What if they realize that the seats crucial to the outcome are the bellwether seats? If Labour needs to gain 100 seats to win the election, then, even with the certainty of non-uniform swing, the 100th most marginal Conservative seat is far more likely to be decisive than the most marginal one. Might it therefore be in the bellwether seats that we see the most distinctive voting behaviour?

To test this possibility, we re-run column 2 of Table 1, except that our main independent variable no longer is the existing majority in each seat, but the absolute difference between its percentage government majority (positive or negative) and the percentage government majority in the government’s nth most marginal seat, where 2n is its parliamentary majority over the main opposition party—i.e., the bellwether seat.

In column 1, the coefficients on BELL and BELL² (0.0757 and −0.0016) are little different from what they were on MAJ and MAJ² (0.0709 and −0.0013). But when a horse race is run between the two measures in column 2, bellwether wins decisively. Not only do the coefficients on MAJ and MAJ² shrink by an order of magnitude, they also lose most of their statistical significance (Table 3). If voters are better at predicting elections than are opinion polls and have the ability to distinguish between marginal and a bellwether seat, this article is giving them rather a good post-election report.

Finally, we compare Conservative-Labour marginals with the second most common form of marginals, those contested between Conservatives and Liberals, where Liberal means Liberal up to 1979, Liberal or SDP in 1983, the Liberal-SDP Alliance in 1987 and Liberal Democrat from 1992 onwards. (Over the period as a whole too few Labour-Liberal marginals exist for analysis to be possible, though they became much more common in and after 2005.) We can skip all the extensions of the last section: we can’t analyze nationally close versus decisive Conservative-Liberal results because there weren’t any close ones, nor were there enough close Conservative-Liberal contests in the three elections with a recent predecessor for us to ask if these elections were distinctive. And clearly the idea of bellwether seats makes no sense if no one thought the Liberals were liable to win any of the elections.

What we therefore do is take those seats where Conservative and Liberal were in first and second place, in whatever order, and ask if there is a relation between size of majority and swing to national gainer (as between Conservative and Liberal). As can be seen from columns (2) and (4) of Table 4, looking for quadratic functions produces no significant results. But when we look for a linear relation between majority and swing, we find a negative one, provided that we distinguish between Liberal- and Conservative-held seats (col. 3). When we regress swing simply on absolute majority, the significance disappears and the coefficient changes sign (column 1).