Ninety years ago, on Oct. 31, 1936, Literary Digest, which had correctly forecast presidential election winners since 1916, predicted that Alf Landon would beat Franklin Delano Roosevelt in the upcoming election.

Roosevelt won handily. The Digest’s poll surveyed 10 million people via postcard, using telephone books and car registration lists for the information. Even though almost 2.4 million people responded — an impressive response rate — the misleading sample provided inaccurate results. Literary Digest folded shortly afterward.

Two years later, in 1938, Gallup started tracking presidential approval ratings. Even so, polls remained far from perfect. In 1948, the Chicago Daily Tribune’s embarrassing front-page declaration "Dewey Defeats Truman" stemmed in part from a misuse of polling. The paper relied on earlier polls that did not reflect the Election Day reality in making its prediction.

As the science of polling improved over time, campaign pros began to rely on pollsters. John F. Kennedy was the first president to be strongly associated with a polling firm, as he used Lou Harris and the Harris organization's polls in his 1960 campaign for president.

LEE CARTER: AMERICA IS ANGRY AT EVERYONE — BUT WHO WILL ACTUALLY VOTE?

In the 1970s, we witnessed the rise of the celebrity pollsters: Pat Caddell for Jimmy Carter, Richard Wirthlin for Ronald Reagan, and Stan Greeberg and Dick Morris for Bill Clinton. These pollsters were successful but not infallible.

Caddell, for instance, was largely responsible for Carter's 1979 "Crisis of Confidence" speech. He misjudged the public mood, and Carter’s "malaise speech" became one of the most disastrous presidential speeches of all time.

Morris, for his part, had a tempestuous relationship with Bill Clinton. They once reportedly came to blows while Morris was advising Clinton during his time as governor of Arkansas.

Morris was initially on the outs and Greenberg in charge of polls when Clinton won the 1992 election. But Morris secretly came back and started advising Clinton after the Democrats lost control of Congress in the 1994 "Republican Revolution." Morris even had a code name — "Charlie" — to keep his identity secret from the more liberal White House staff, which resented Morris’ pushing of Clinton to the center.

Morris would poll everything, even where Clinton should go on vacation. When Clinton left office after two terms in 2000, he had survived the Monica Lewinsky scandal — despite polls showing that Americans strongly disapproved of Clinton's behavior. The most important number for Clinton was the Gallup approval rating. He left office at 66%, one of the highest outgoing figures in history.

21st-century presidents would be thrilled to get approval ratings anywhere near Clinton's final number. George W. Bush hit 86% after the 9/11 terror attacks, but that did not last. For the most part, the "low and steady" phenomenon has meant that presidents in this century have struggled to get about 50% approval.

Still, the polling mishaps have continued. In 2004, exit polls predicted a win for John Kerry over George W. Bush. Campaign adviser Bob Shrum even said to Kerry on Election Day, "Can I be the first to call you, Mr. President?" The exit polls were wrong, and Bush won a second term.

In 2012, Mitt Romney’s campaign wrongly thought it was headed for victory when its pollster made overly optimistic weighting assumptions. The turnout model incorrectly assumed that Barack Obama would not generate the same level of youth and minority turnout that he had in 2008.

Polling took another big hit in 2016, when pollsters confidently predicted that Hillary Clinton would defeat Donald Trump. She didn't, and the pollsters’ justification that they had the popular vote right but got the election result wrong did not help matters.

In our current midterm elections, we have seen a new wrinkle. Partisan polls overstated progressive support in the primaries, skewing the results and causing some candidates to drop out in the face of the inaccurate data. This latest mishap reinforces the lesson that polling often fails, although rarely in the same way twice.

While technology and the science of polling are much better today than in 1936, the improvements complicate matters further. While nearly everyone has a phone, they also have caller ID and don't answer calls from unfamiliar numbers. Response rates shrink, increasing the size of the "silent majority." In this new environment, human modeling decisions ironically become more, not less, important. Polling errors come not from new techniques or technologies but from imperfect human decisions about how to deploy the new technologies. In making those decisions, pollsters can overreact to what went wrong beforehand. In this, pollsters are like generals, often fighting the last war.

In the end, polls are more accurate as snapshots of opinion rather than predictors of behavior, specifically, who will actually submit a ballot. This means that whatever the polls say, the public should be wary. To borrow from the psalms, put not your trust in polls, and get out and vote.