Wallace was always behind Humphrey. The smallest gap was in a poll taken at the end of September and beginning of October, when Humphrey had 26% and Wallace 19%, with Nixon at 41%. The iPOLL database is not completely comprehensive, so maybe there were surveys in which Wallace led. But I doubt it, given the substantial margin and general stability of the figures (which were from two different polling firms, Gallup and Harris). Maybe Alter was thinking of 1992, when I think Perot did actually lead in some polls?
I hadn't realized that Humphrey was generally leading in the polls until August. The Republican convention took place at the beginning of that month and the Democratic convention took place at the end. That connects to another issue that I was going to write about, so I might as well do it now. Back in the primaries, I was thinking about the question of whether nominating an "extreme" candidate would make a difference. People who say that an extreme candidate will lose votes always include George McGovern as a leading example. McGovern definitely had a liberal voting record in the senate, but so did some more successful Democratic nominees (John Kerry and Barack Obama).
That led me to wonder if the relevant issue might be a divided convention, rather than "extremism." The two things tend to go together, because the reason for a divided convention is often an ideological challenge where one or both sides feel strongly enough to fight on until the end. By divided, I mean one in which there was open conflict that was visible to the TV audience. I think that the following conventions are generally agreed to meet that standard: Republicans in 1964 and 1976 (and now 2016), and the Democrats in 1968, 1972, and 1980. You could argue that the whole idea of a divided vs. united convention doesn't apply before the TV era--the action took place behind the scenes, and multiple ballots were common. However, the Democratic convention of 1924, which took over 100 ballots and involved a giant rally by the KKK after a motion to condemn them had been narrowly defeated, might deserve to be included.
I added a variable for a divided convention to Fair's model and got the following results:
Model estimate std err
1. Fair plus divided (1968-80 only) -2.36 (1.45)
2. Fair plus divided (including 1924) -2.51 (1.37)
3. Modified plus divided -3.41 (1.59)
The third model modifies Fair's variables in two ways: Gerald Ford is counted as an incumbent president, and the Democrats are counted as getting 34.8% of the two-party vote in 1924, which is what they actually got. Fair does not count Ford as an incumbent on the grounds that he was not elected to either the presidency or the vice-presidency, which doesn't make sense to me. There was an important third-party candidate (Robert LaFollette) in 1924, and Fair estimates that most of his voters would have gone to the Democrats if he had not been in the race. That could be true, but it's not possible to be sure, and if you were adjusting totals it seems that a better case could be made for 1968, when surveys indicated that most of Wallace's voters would have gone to Nixon as a second choice.
The figures mean that a divided Democratic convention is estimated to cost the Democrats about 2.4% and a divided Republican convention to help them by 2.4% (in model 1). There's a good deal of uncertainty about the size of the effect, and it's not statistically significant at the 5% level in the first two models. Still, I'd say that there's evidence that a divided convention hurts a party.