Dewey Defeats Truman: What the Polls Missed in 1948

Three national polling organizations told the United States in the autumn of 1948 that Thomas Dewey would be the next president. Gallup's final poll, interviewed between 10/15/1948 and 10/25/1948 and published on 11/01/1948, put Dewey at 49.5 percent and Harry Truman at 44.5 percent. Truman won the popular vote with 49.5 percent. The industry had the right number attached to the wrong man.

The standard account of 1948 is that the polls were wrong. The evidence points somewhere narrower and more useful. The surveys were roughly right about the fortnight in which they were taken, and the failure sat in the procedure around them: when the industry stopped asking, and who its interviewers were allowed to pick. A forecast is only ever as good as its last day of fieldwork, and in 1948 that day was 10/25/1948 for the one organization still in the field at all.

What would prove this wrong is simple to state. If the surveys taken in mid-October had been badly off for mid-October as well as for election day, the failure would be one of measurement and the remedy would be different. The committee that reviewed the year concluded otherwise, and its reasoning is set out below.

What the 1948 polls said and what happened

Harry Truman won the 1948 election with 24,105,810 votes and 303 electoral votes against Thomas Dewey's 21,970,064 votes and 189 electoral votes, according to the American Presidency Project's compilation of the returns. Strom Thurmond, running on the States' Rights ticket, took 1,169,114 votes and 39 electoral votes, one more than the four states he carried, because a Tennessee elector voted for him although Truman carried Tennessee. Henry Wallace's Progressive ticket took about 1.16 million votes and carried nothing. Turnout was 52.2 percent of the voting-eligible population, by the United States Elections Project's series.

Gallup's final 1948 poll recorded a further 8.7 percent as undecided. Interviewed 10/15 to 10/25/1948 and published 11/01/1948, from a Gallup summary held in the Gerald R. Ford Presidential Library. Popular vote shares from the American Presidency Project, except Henry Wallace's share, from Wikipedia's compilation of the 1948 returns.

Candidate	Final poll	Result	Error
Harry Truman (Democratic)	44.5%	49.5%	+5.0
Thomas Dewey (Republican)	49.5%	45.1%	-4.4
Henry Wallace (Progressive)	4%	2.4%	-1.6
Strom Thurmond (States' Rights)	2%	2.4%	+0.4

Against that, the published figures. Elmo Roper's last number of the campaign had Dewey at 52.2 percent and Truman at 37.1 percent, as a Christian Science Monitor reporter who covered the campaign recalled it in 1980, and other accounts give the pair as 44 and 31, which is most likely the same survey reported with and without undecideds set aside. Archibald Crossley's final release had Dewey at 49.9 and Truman at 44.8 by the Monitor's account, while TIME reported in November 1948 that Crossley's final prediction was a Dewey win by 51 percent to 42. The accounts conflict and both are reported here. Nobody conflicts about the direction.

Gallup's final 1948 poll against the result

Gallup's last published figures of the campaign, interviewed 10/15 to 10/25/1948, beside each candidate's certified share of the popular vote.

The measurement error was small relative to the embarrassment. Gallup missed Truman's share by 5.0 points and Dewey's by 4.4, which is a bad poll rather than a broken instrument. The catastrophe came from what the industry was willing to say on top of a bad poll. Roper had stopped publishing entirely, and told reporters on election eve that he stood by his prediction and that Dewey was in. Crossley's release declared Dewey and Earl Warren assured of election. The Chicago Daily Tribune's earliest editions of 11/03/1948 carried the banner DEWEY DEFEATS TRUMAN, printed before most states had reported, because the paper's linotype operators were on strike over the Taft-Hartley Act and the substitute process forced an earlier press time. More than 100,000 copies went out, often given as 150,000. Truman held one up for photographers at St. Louis Union Station on 11/04/1948.

Gallup's 1948 error by candidate

Each candidate's certified share minus his figure in Gallup's final poll, in percentage points.

Why the pollsters stopped interviewing before election day

Every major 1948 polling organization stopped interviewing well before election day because the profession believed presidential preferences were settled by early autumn and would not move, and that single procedural choice cost more than any flaw in the sampling. Roper announced on 09/09/1948 that he would publish no further figures unless something significant changed, by TIME's contemporaneous account, 54 days before the vote. Gallup's interviewing ended 10/25/1948. Crossley's final figure came out 10/29/1948.

Days between the final published figure and election day, 1948 to 2010

A column assembled by hand from five sources, one per reading. For the Gallup rows the date is the last day of interviewing; for Roper and Crossley in 1948 it is the publication date, because their interviewing dates are not established in the available record.

The belief had an academic pedigree, which is what made it durable. Irving Crespi, a former Gallup executive, told the Deseret News in 1998 that Paul Lazarsfeld's panel research in Sandusky, Ohio had been read as showing that elections are decided by Labor Day, that Gallup had planned to field a poll in the final two weeks, and that the question "why waste the money?" settled it. Gallup's own post-mortem, in a summary held at the Gerald R. Ford Presidential Library, says the apathy visible in the campaign led the organization to assume most of the undecided would not vote at all.

Both halves of that assumption failed in the same direction, which is the general shape of a polling disaster. A single wrong assumption produces a small error. Two wrong assumptions that happen to agree produce a wrong winner.

How quota sampling worked and where it went wrong

Quota sampling in 1948 matched the sample to census proportions on characteristics such as age, sex, race, region and rent paid, and then let the interviewer personally choose which individuals to approach inside each quota cell. The totals came out looking like the country. The selection inside the cells did not. Interviewers were told to find so many young white men, so many elderly Black women, and so on, then picked the individuals themselves, sometimes on street corners. Across thousands of such choices they reached too many Republicans.

The flaw is instructive because it is invisible in the output. A quota sample's margins are correct by construction, so the usual sanity check passes. The bias lives in the step the published tables do not describe, which is how a particular human being ended up in the sample rather than a different human being who met the same description. The remedy adopted afterward was probability sampling at all levels of respondent selection and the elimination of quotas, which means the sample design, rather than the interviewer, decides who is asked.

That distinction is still the first question worth asking about any set of numbers. Every sampling method has a step where somebody or something decides who gets included, and the error always hides in that step. In 1948 it was an interviewer on a street corner. The modern versions differ in mechanism and not in kind.

What the undecided voters did in the last two weeks of the 1948 campaign

About 15 percent of 1948 voters made up their minds in the final two weeks of the campaign, by Dominic Lusinchi's account of the postmortem literature, after every major organization had stopped interviewing. Gallup's final poll had recorded 8.7 percent as undecided and the organization had assumed most of them would stay home. Gallup's own summary says that many of them voted and that they voted overwhelmingly for Truman.

The Social Science Research Council's Committee on Analysis of Pre-Election Polls and Forecasts, nine members assembled within a week of the election, finished in five weeks and issued its report on 12/27/1948, published the following year as the Council's Bulletin 60 with the analysis directed by Frederick Mosteller. The committee's position, as Lusinchi summarizes it, was that Gallup's and Crossley's last surveys were possibly not too far off the mark for the moment they were taken. That is the finding worth carrying forward. The surveys measured something real about mid-October. The profession then treated a mid-October measurement as a forecast of a November event, and the two weeks in between contained one voter in seven.

The committee's other criticisms were procedural in the same way: applying probability-sample error formulas to samples that were not probability samples, making arbitrary adjustments, mishandling the people who said they did not know, and claiming more than the data could support. Mosteller's chapter on measuring error proposed eight ways to score a poll against a result, and those measures are still the standard. His own table put the three final 1948 polls at an average 54.6 percent for Dewey on a two-party basis against his actual 47.7 percent.

How the profession rebuilt survey research after 1948

The industry's response to 1948 was to change sampling, change timing and change the language of publication, and the changes held. Pollsters moved to probability-based selection and began interviewing closer to election day. Gallup revised its procedures over the course of the 1950s and began measuring preferences continuously up to the eve of the vote. The Roper Center credits the committee's effect in large part to the cooperation of George Gallup, Archibald Crossley and Elmo Roper themselves, who opened their files to a panel convened to examine their failure.

Accuracy of final national presidential polls since the rebuild

Average absolute error on the popular vote margin in final national presidential polls, against the average across all elections since 1936.

The measurable result is in the chart above. AAPOR's evaluation of the 2016 polls puts the average absolute error of final national presidential polls on the popular vote margin at 4.4 points across all elections since 1936 and 2.7 points for elections since 1992, with 1.8 points in 2008 and 2.2 points in 2016. The same report notes that its figures for 1936 through 1960 rest on Gallup alone, so the early part of that series is one organization's record rather than an industry average.

Three habits came out of 1948 and all three are now ordinary. Interview late, because preferences move and the last week is not a formality. Take respondent selection out of the hands of whoever is doing the contacting. And report what the instrument measured, with its date attached, separately from any claim about what will happen.

What 1948 should change about how a campaign reads its own numbers

The most expensive error of 1948 came from describing a measurement as an outcome. Roper had a figure from September and announced that Dewey was in. The Tribune had a correspondent's judgment and set it in type as a result. In each case the underlying information was defensible and the statement built on top of it was not.

Campaigns make the same move every week, in a meeting rather than in print. A field program is reorganized in the last ten days on research collected a month earlier. The fix is cheap. Every number in circulation should carry the date it was collected and the question it answers, and anybody acting on it should know how much of the campaign has happened since.

Questions about the 1948 polling failure

Why did the polls get the 1948 election wrong?

The 1948 polls were wrong for three reasons that compounded. All three major organizations used quota sampling, which let interviewers choose respondents inside demographic cells and produced a sample with too many Republicans. All three stopped interviewing before the final week, so none of them measured the roughly 15 percent of voters who decided in the last two weeks. And each then described a mid-October measurement as a settled outcome.

What was the Dewey Defeats Truman headline?

DEWEY DEFEATS TRUMAN was the banner on the earliest editions of the Chicago Daily Tribune dated 11/03/1948, printed before most states had reported their returns. The paper's linotype operators were on strike over the Taft-Hartley Act, and the substitute printing process forced an earlier deadline. More than 100,000 copies carried the error, and Truman held one up for photographers at St. Louis Union Station on 11/04/1948.

What is the difference between quota sampling and probability sampling?

Quota sampling fills a target number of interviews in each demographic cell and lets the interviewer decide which individuals to approach. Probability sampling gives every member of the population a known chance of selection and takes that choice away from the interviewer. The Social Science Research Council committee that reviewed the 1948 polls recommended probability sampling at all levels of respondent selection and the elimination of quotas, and the industry adopted it.

How accurate are election polls now compared with 1948?

Final national presidential polls have missed the popular vote margin by an average of 2.7 points in elections since 1992, against 4.4 points across all elections since 1936, by AAPOR's accounting. Gallup's final 1948 poll missed Truman's share by 5.0 points and Dewey's by 4.4 in the same direction. The error narrowed because sampling and timing both changed after 1948.