The General Social Survey provides an alternative source for analysis of this issue, since it also has a lot of repeated questions. An advantage of the GSS is that it was conducted every year (until 1994) or two (starting in 1994), making it possible to track changes more closely. A disadvantage is that unlike the surveys in the Pew analysis, it doesn't repeat exactly the same questions. However, I did an analysis based on the following questions*:
1. should a woman be allowed to get a legal abortion if she is married and doesn't want more children
2. sex between two adults of the same sex is always wrong....not wrong at all
3. favor or oppose the death penalty for murder
4. favor or oppose requiring permit for gun ownership
5. favor or oppose school prayer
6. vote on (hypothetical) law requiring homeowner to sell to person of any race
6a. government spending on improving the conditions of blacks
7. should marijuana be legal
8. should government reduce income differences
Using a scale with items 1-5, 6, 7, and 8, the estimated polarization is:
Using a scale with 1-5, 6a, 7, and 8, it is:
In both, the estimates vary more from year to year before 1994, because the samples were smaller. Either way, there is a definite increase over the period. As far as the timing, the model of a trend over the whole period fits slightly better than the model of an increase only after 2004. However, the difference in fit is small; either one is compatible with the data.
The GSS questions go back to the 1970s, some as far back and 1972, so the analysis could be extended to cover a longer period, although they weren't all asked in the same year until 1988, making the analysis less straightforward. In fact, I did an analysis of that about 15 years ago. I didn't find evidence of a general trend; there were some changes in the association among opinions, but they were complicated, and I never figured out a satisfying interpretation, so eventually I gave up before publishing. But it seems like reality finally caught up to my hypothesis.
*The numbers on the vertical axis are the eigenvalue of the first principal component.