Truth and Krugmanism

For those such as the New York Times' Nobel Prize-winning Paul Krugman and Nate Silver's "Mini-Me" at the CBC, the notion that there is underlying bias in polling data against Donald J. Trump is an inconvenient truth.

Writing at the NYT, the resident "winner" offered the following bit of non-science:

Oh, and let's not make too much of any one poll. When many polls are taken, there are bound to be a few outliers, both because of random sampling error and the biases that can creep into survey design. If the average of recent polls shows a strong lead for one candidate -- as it does right now for Mrs. Clinton -- any individual poll that disagrees with that average should be taken with large helpings of salt.

This, of course, is wrong.

Indeed, the Real Clear Politics (RCP) polling averages do show a substantial head-to-head lead for "Mrs." B.J. Clinton over Trump, but as discussed here at AT in recent days, there is evidence of bias against Trump all over the place in these polls.

And when systematic bias is highly likely, averaging – as Krugman recommends – does absolutely nothing to remove the problem.  In fact, it reinforces it.

And when that systematic bias is present, individual polls that disagree with the average stand a very good chance of being far more accurate than the average – namely, an unbiased accurate poll will look like an outlier against a biased inaccurate average of polls.  That is Statistics 101.

If you want evidence of systematic bias against Trump in RCP polling averages, here you go.

The following table shows the "RCP Average spread" for each of the GOP primaries between Trump and John Kasich – where one was able to be calculated by RCP before the voting due to sufficient recent polling data being available – compared to the final voting results.

Well, that is inconvenient for those blindly following individual polling data and its averaging.

The Trump-Kasich spread is a good proxy for liberal polling bias, since more liberal members of the GOP are where most of Kasich's support resides.  And as previous articles have shown, many polls are oversampling from the more moderate/establishment side of the Republican base – which is then coupled to an overall oversampling of Democrats versus Republicans to provide some seriously skewed results for the general population.

If there were no systematic bias against Trump in this polling data, the average deviation between the final Trump-Kasich spread and the corresponding "RCP Average Spread" across these 23 states should be zero.  But it's not.  It is 3.6% against Trump.

Even worse, if there were no bias against Trump in the data, the number of states where the polling average spread overestimated Trump's support should equal – at least approximately – the number of states where the polling average spread underestimated Trump's support.

But it is not even close.  A full 16 of the state polling averages underestimated Trump's support relative to Kasich, while only 7 overestimated his support.  This is highly statistically unlikely without systematic bias in the data.

Mini-Mes and the "winners" who symbolize all that is wrong with modern academia can unite and uncritically cite flawed data all they want to achieve political goals, but the cat is out of the bag now as the American public realize that almost nothing they read in the mainstream media is correct.

For those such as the New York Times' Nobel Prize-winning Paul Krugman and Nate Silver's "Mini-Me" at the CBC, the notion that there is underlying bias in polling data against Donald J. Trump is an inconvenient truth.

Writing at the NYT, the resident "winner" offered the following bit of non-science:

Oh, and let's not make too much of any one poll. When many polls are taken, there are bound to be a few outliers, both because of random sampling error and the biases that can creep into survey design. If the average of recent polls shows a strong lead for one candidate -- as it does right now for Mrs. Clinton -- any individual poll that disagrees with that average should be taken with large helpings of salt.

This, of course, is wrong.

Indeed, the Real Clear Politics (RCP) polling averages do show a substantial head-to-head lead for "Mrs." B.J. Clinton over Trump, but as discussed here at AT in recent days, there is evidence of bias against Trump all over the place in these polls.

And when systematic bias is highly likely, averaging – as Krugman recommends – does absolutely nothing to remove the problem.  In fact, it reinforces it.

And when that systematic bias is present, individual polls that disagree with the average stand a very good chance of being far more accurate than the average – namely, an unbiased accurate poll will look like an outlier against a biased inaccurate average of polls.  That is Statistics 101.

If you want evidence of systematic bias against Trump in RCP polling averages, here you go.

The following table shows the "RCP Average spread" for each of the GOP primaries between Trump and John Kasich – where one was able to be calculated by RCP before the voting due to sufficient recent polling data being available – compared to the final voting results.

Well, that is inconvenient for those blindly following individual polling data and its averaging.

The Trump-Kasich spread is a good proxy for liberal polling bias, since more liberal members of the GOP are where most of Kasich's support resides.  And as previous articles have shown, many polls are oversampling from the more moderate/establishment side of the Republican base – which is then coupled to an overall oversampling of Democrats versus Republicans to provide some seriously skewed results for the general population.

If there were no systematic bias against Trump in this polling data, the average deviation between the final Trump-Kasich spread and the corresponding "RCP Average Spread" across these 23 states should be zero.  But it's not.  It is 3.6% against Trump.

Even worse, if there were no bias against Trump in the data, the number of states where the polling average spread overestimated Trump's support should equal – at least approximately – the number of states where the polling average spread underestimated Trump's support.

But it is not even close.  A full 16 of the state polling averages underestimated Trump's support relative to Kasich, while only 7 overestimated his support.  This is highly statistically unlikely without systematic bias in the data.

Mini-Mes and the "winners" who symbolize all that is wrong with modern academia can unite and uncritically cite flawed data all they want to achieve political goals, but the cat is out of the bag now as the American public realize that almost nothing they read in the mainstream media is correct.