A Resource for State Preelection Polling

How we did this

This report provides state by state estimates for the educational profile of voters in the 2004, 2008, 2012 and 2016 U.S. presidential elections. Estimates are based on an analysis of the Current Population Survey (CPS) Voting and Registration Supplements conducted by the U.S. Census Bureau for each of those years. The CPS is administered using in-person and live telephone interviewing. Households are selected using a national sample of addresses produced through a stratified, multi-stage sample design.

Post-mortem analysis of the 2016 election found that a failure to adjust for overrepresentation of college graduates was among the reasons many state-level polls underestimated support for Donald Trump. Voters who graduated from a four-year college are more likely to answer surveys than other adults and, in recent years, they are also more likely to support a Democrat for president. If a battleground state poll does not adjust for having too many college graduates, it is at risk of overstating support for a Democratic presidential candidate (in this case, Joe Biden).

Since 2016, many pollsters heeded this lesson and added an education adjustment to their work. Additionally, most national pollsters as well as some state pollsters had been making the adjustment for many election cycles and continue to do so. But not all have fixed this issue. For example, a June poll appeared to show Biden with a massive 18-percentage-point lead in Michigan. But a look at the sample shows why: More than two-thirds (69%) of those interviewed were college graduates – nearly double the rate among Michigan voters in recent elections. Regardless, a high-profile polling aggregator fed this poll into its average for the state, demonstrating how readily problems from 2016 can repeat.

One challenge in adjusting for education is identifying the proper benchmark. Using the June poll example, a rate of 69% college graduates is clearly too high. But what is the “right” number? Technically, no one knows, because the goal is to align the survey with the education profile of those who will vote in an election that has not yet happened. While the precise number is unknown, historical data from a large, high-quality federal study ably fills this need. In the month or so following each presidential and midterm election, the U.S. Census Bureau conducts the Current Population Study (CPS) Voting and Registration Supplement. The study does not ask who people voted for, but it does ask whether they voted. With more than 90,000 interviews nationally, more than a third of which are done in-person, the CPS supplement is among the nation’s best measurements of the demographics of voters and nonvoters.

The state-by-state results are freely available to the public, but for many they are difficult to access as they require software and servers that can process large data files. This report provides the CPS data on the education profile of voters in all 50 states and the District of Columbia for the past four presidential elections. State pollsters can use this data to inform their weighting adjustments. Poll observers can use this data to determine whether the share of college graduates in a battleground state poll is reasonable.

There are several critical factors to keep in mind:

Polls should be judged based on their weighted sample. The issue is not whether raw poll samples have too many college graduates. It is almost a given that they do. The issue is whether the pollster has adjusted for the issue – weighting down college graduates proportional to their plausible share of voters in the upcoming election. If a poll’s methodology states that education was included as an adjustment variable, often that is enough to safely assume this issue was addressed. If a poll did not adjust for education, observers curious about quality can ask the pollster what share of the weighted sample were college graduates. Reputable pollsters will recognize why this information would be of interest and provide it. If a pollster is unwilling to provide this information, that is a strong sign that the poll may not be trustworthy.

The expectation should be plausibility, not perfection. The CPS data gives a reality check for the typical proportion of a state’s voters who are college graduates. But the proportion in an upcoming election could always be somewhat higher or lower than in the CPS data. One takeaway from the data compiled here is that large election-to-election changes (for example, more than 8 percentage points) in the college graduate rate are highly unlikely – in other words, implausible. Changes on the order of several percentage points, however, are to be expected. Observers should not expect that a poll exactly mimics prior elections’ education profile; they should only expect that it comes reasonably close. For example, the CPS shows that the share of presidential election voters in Florida who are college graduates has recently been in about the mid-30% range. A 2020 Florida preelection poll should, therefore, have a college graduate rate in its weighted sample of between about 30% and 45%. If the rate is well above 45%, the poll runs the risk of overestimating support for Biden and underestimating support for Trump.¹

A plausible education profile is important, but other factors matter too. A poll’s education profile is far from the only factor that observers should consider when evaluating quality. For example, ideally a poll draws its participants from a source that includes nearly everyone in the state (or in the country for national polls). Examples of such sources are registered voter files, telephone random-digit dialing and the U.S. Postal Service residential address database. Other factors that are important to a poll’s trustworthiness include the sponsor, sample size, question wording and adjustments on other variables such as age, sex, race and geography. In other words, a plausible education profile should be on the checklist for trustworthiness in battleground state polls – but there are other items on the list as well.

Ideally, an education adjustment accounts for multiple levels and variation between race groups. For clarity, this analysis focuses on whether college graduates are overrepresented in poll estimates. But for practitioners, additional layers of detail can be important. A college vs. non-college adjustment is good, but a more detailed adjustment aimed at achieving proper representation of more fine-grained levels can be even better. For example, a pollster can use the CPS data to adjust for the share with a high school education or less, the share with some college experience (which typically includes trade schools and two-year college degrees), the share with a four-year college degree, and the share with a graduate degree.

Similarly, in geographies with relatively large shares of Hispanic, Black or Asian American populations, a pollster may further improve accuracy by adjusting the education profile within the largest race and ethnicity groups. For example, Pew Research Center’s national polls are adjusted to ensure that education groups (high school or less, some college, college graduate) are represented properly among Hispanic, Black, White and Asian Americans.

The CPS trend lines generally are fairly stable and slowly increasing. The stability of the state-level CPS trends dispels the notion that a pollster cannot anticipate roughly what the college graduate rate among a state’s voters will be. While other voter demographics (for example, the share who live in rural areas) may shift noticeably, the share who graduated from a four-year college simply do not tend to fluctuate wildly, according to the CPS. Furthermore, to the extent that there is movement, it is somewhat predictable: the college graduate rate has tended to increase by about 2 to 3 percentage points in the last four elections in battleground states. State pollsters could reasonably factor in such a modest increase when adjusting polls this cycle.

While this report focuses on the CPS, there are other useful sources of information that can be used to improve or assess the representativeness of a poll. For example, pollsters sampling from registered voter files can use race, age, sex, political party and other variables on file to adjust their samples. While voter file data on those characteristics can be quite accurate, appended data about voters’ education level tends to be less so. A 2018 Pew Research Center study of five national voter files found that individuals’ education level was either missing or inaccurate 49% of the time, on average, across the files.

Some polls – particularly those releasing estimates for all U.S. adults – do not need weighting targets that are specific to likely or registered voters. An alternative source that works well for such polling is the American Community Survey (ACS). Unlike the CPS, the ACS does not provide data on those who voted in an election. It does, however, provide authoritative data on the shares of all adults with various levels of education at the state level and much lower.

Finally, it is worth reiterating that education is just one of several dimensions that tend to require adjustment is polls. A poll also needs to be representative with respect to geography, age, race, ethnicity, urbanicity, sex and potentially more. Adjustments for political partisanship and urbanicity are increasingly common in polling. As the polling field enters the heat of the 2020 election, it’s imperative that public polls are strong on all the fundamentals, since it may be difficult to predict what new challenge may arise.

Voter’s education distribution has remained relatively stable since in recent presidential elections Among voters in each state in each general election…

State	2004 HS or less	2004 Some college	2004 College grad	2008 HS or less	2008 Some college	2008 College grad	2012 HS or less	2012 Some college	2012 College grad	2016 HS or less	2016 Some college	2016 College grad
US	37	31	32	34	32	34	32	31	37	30	31	40
AL	42	31	28	44	31	24	39	32	29	35	34	32
AK	32	38	30	29	38	33	31	38	32	30	34	36
AZ	32	36	32	28	35	36	26	38	37	23	36	40
AR	44	30	26	44	29	26	41	27	32	37	29	34
CA	28	35	37	27	35	38	25	34	42	25	32	43
CO	26	30	44	25	31	43	23	34	44	22	30	48
CT	34	25	41	31	29	40	28	27	45	29	25	46
DE	39	29	31	41	28	31	35	28	36	36	26	38
DC	25	19	56	25	21	53	24	16	59	18	16	66
FL	37	31	32	34	33	33	34	31	35	31	32	36
GA	36	32	32	35	31	34	36	30	34	33	31	36
HI	28	35	37	32	30	38	30	31	39	26	31	43
ID	35	37	28	34	33	33	30	36	35	28	37	35
IL	35	31	34	34	31	36	30	29	41	28	30	42
IN	45	29	26	44	29	27	36	30	34	36	29	35
IA	38	37	26	33	33	34	34	35	32	30	35	35
KS	33	32	35	28	34	38	28	31	41	25	35	40
KY	45	30	25	38	36	26	40	34	26	36	33	31
LA	46	28	26	44	27	29	44	30	26	40	30	31
ME	44	29	27	40	30	30	36	29	36	34	31	35
MD	34	28	38	31	28	41	29	27	44	26	28	46
MA	34	24	42	29	24	46	28	24	48	25	25	50
MI	41	33	26	36	35	29	35	34	32	30	32	38
MN	30	36	35	28	36	36	27	35	38	24	34	41
MS	50	29	21	46	30	24	41	33	26	43	31	26
MO	40	30	30	39	34	27	35	34	31	36	33	31
MT	35	37	28	37	32	31	31	36	33	30	34	36
NE	35	33	32	30	35	36	30	33	37	27	36	38
NV	38	35	27	35	35	30	35	35	31	30	40	31
NH	36	27	37	31	31	38	29	30	40	30	28	42
NJ	39	25	36	35	25	41	29	27	43	29	24	47
NM	35	37	28	31	30	40	30	28	42	30	35	36
NY	38	27	35	34	29	37	32	28	41	28	27	45
NC	40	29	31	33	33	34	33	34	33	29	33	38
ND	31	41	28	32	37	31	28	35	37	29	34	36
OH	43	30	26	41	30	29	42	30	28	37	29	34
OK	42	29	29	37	32	30	32	32	35	33	28	39
OR	31	39	30	28	38	34	29	34	37	26	31	43
PA	42	25	34	41	27	33	37	28	35	35	28	36
RI	38	25	37	34	29	37	34	28	39	33	29	37
SC	37	36	26	43	30	27	34	33	33	34	30	35
SD	39	35	27	35	34	31	33	36	31	28	37	35
TN	41	31	28	42	28	31	36	29	35	33	29	38
TX	35	32	33	34	34	32	32	34	34	27	33	41
UT	30	37	33	26	44	30	24	41	35	20	37	43
VT	36	26	38	34	27	39	31	26	43	27	27	45
VA	33	27	40	34	25	41	30	28	42	28	30	42
WA	28	37	35	24	37	38	26	33	41	24	31	45
WV	50	26	23	49	30	21	46	26	29	43	27	30
WI	41	32	27	36	33	32	33	33	34	31	32	37
WY	37	39	23	36	38	27	37	36	27	28	40	32

Source: Current Population Survey Voting and Registration Supplement 2004-2016