WASHINGTON - For weeks, opinion polls have consistently shown former US vice-president Joe Biden ahead of President Donald Trump for the Nov 3 election. And though the range has varied, he has remained in the lead.

But a recurring nightmare among Democrats is a repeat of 2016, when polls showed Mrs Hillary Clinton would triumph, only to see Mr Trump narrowly take a handful of key states and thus win the electoral college.

So, how reliable are these polls? How are they conducted? Which are the ones to follow? US Bureau Chief Nirmal Ghosh explains.

HOW RELIABLE ARE POLITICAL POLLS IN AMERICA?

There are good polls and bad, or sloppy, polls, and every poll has a margin of error.

A single polling question rarely captures a complicated human being's opinions on complicated issues, some of which the person may not even know much about. So it is always better to take an average of several polls that are considered at least methodologically sound.

In 2016, problems with polls in a few key Midwestern states led to many underestimating the chances of a Trump victory.

The immediate post-election assessment was that there had been a complete polling failure. But that is an overstatement. The 2016 polls were not wrong on a national basis. An average of final, publicly released national polls pointed to Mrs Clinton winning the overall popular vote by three percentage points - and she did, by two points.

"The polls were actually quite good," Mr John Zogby, senior partner at John Zogby Strategies - who did not poll in 2016 - told The Straits Times. Almost every poll came close nationally.

"I honestly believe that talking heads simply could not conceive of a Trump victory and the herd mentality just assumed a Clinton victory. They were blindsided by their own myopia - not by the polls."

WHAT WENT WRONG IN 2016?

One key mistake, according to an analysis published by Pew Research, was that many state pollsters did not adjust to reflect that college graduates are more likely to take surveys than adults with less formal education.

This mattered more than in previous years, when there weren't big partisan differences between the two groups. In 2016, the more educated went for Mrs Clinton, and the less educated, who had not finished high school, went for Mr Trump. State polls that did not adjust their data by education were left with a biased sample.

Also, undecided voters swung more towards Mr Trump than to Mrs Clinton in the last few days before the election - but most state polls did not detect that simply because they had stopped polling.

HOW ARE THEY CONDUCTED?

Polling is conducted largely by phone interviews over landlines or mobile phones and sometimes via both; automated "robopolls", technically called "interactive voice response"; and over the Internet.

SO SHOULD WE WATCH THEM - AND WHICH ARE THE BEST POLLS TO FOLLOW?

"I think the polls are doing well," Mr Zogby told ST.

"We pollsters have to assume a turnout model and that is where some of the differences arise. How many Democrats or Republicans should be in the sample? 18 to 29-year-olds? Non-whites? Some assumptions need to be made. And some pollsters are better than others. I believe in online polls - I was one of the pioneers. But I also think that live telephone polls can still be accurate."

But he added: "We need to recall that polls are still only a photograph of a moment in time - not a prediction, especially with so much left to the campaign."

The websites Five Thirty Eight, and Real Clear Politics, are among those that track a daily or weekly average of polls.