I recently read a USA Today editorial saying Trump is in trouble in midwestern swing states. It notes 45-49% of voters will vote for any Democratic candidate rather than Trump. It also notes Trump had only a 36% approval rating just prior to the 2016 election. The editorial goes on to say Trump has serious work to do if he wants to win this year, and that the real issue is not the booming economy, it's Trump himself. Now we read that covid 19 has either doomed Trump or is making him unbeatable. Atlantic reports he knows he is losing. Gallup reports his approval at 49%, but a compendium of polls remains all over the place on his popularity. How could a man get elected president if he had an approval rating of only 36%? Why are polls so variable? These questions are what got me interested in political polling and analysis about 20 year ago. The answers lie in the history of the Civil War.
If you read most tertiary histories of the Civil War you will wonder how the South won so many battles. At almost every battle the North had a substantial numerical superiority, according to what these books publish. The average Southern soldier must have been an ubermensch to triumph against such terrible odds. To understand what really happened you have to know where the data comes from, and what it means.
Army returns can be found in multiple primary sources. The most comprehensive source is the Official Records of the War of the Rebellion. This titanic work is 127 volumes long, each about 800 pages. In it are the official reports of every officer from every battle of the war, along with congressional hearings and first-hand testimony about conditions on the field. It is a trove of information. I read the two volumes devoted to Gettysburg while designing a wargame on that battle. It was here I learned the real strength numbers.
Union and Confederate armies tallied strength differently. Confederate regiment and brigade commanders reported rifles, that is, the number of men present with a gun in their hand. That count was typically made right before going into action. Thus men detached for provost guard duty, or sent to artillery units to carry shells, were not counted. Officers also were not counted, and officers were 7-10% of total infantry strength.
The Union armies did a roll call every morning. The totals were reported in three columns- present, present for duty, and present for duty equipped. Present simply meant the man was on the active regimental roll. Regiments when formed had a strength of 1000 men, but by the time they got to an army their strength was rarely more than 650. The average regiment at Gettysburg was 350 men, most of the losses from disease rather than combat deaths. 'Present' was simply the raw strength of the regiment, including those in hospital or under arrest. 'Present for duty' meant the man was standing actually able to report. This included men detached for artillery work, guard duty or off foraging (a major activity during some campaigns). 'Present for duty equipped' was the total of men actually available to march and fight with the regiment that morning.
It is apparent that 'present for duty equipped' is the closest equivalent to the Confederate strength report, but it must be adjusted by the fact that the Union also counted officers. Union 'present for duty equipped' must therefore be reduced 7-10%, or Confederate rolls increased 7-10% to have a true apples-to-apples comparison of force.
Unfortunately, almost all tertiary reports follow the convention of Livermore (a secondary analyst and historian), who used 'present for duty' as the measure of Union strength. It is from this conceptual error that all the imagery of overwhelming Union hordes comes from. While the Union DID field more men, at the 'point of the spear' the difference was much smaller than popular histories make it seem.
Exactly the same problem exists for polling. Polls may be taken of all adults, registered voters, or likely voters. The composition of these populations is not the same, nor are their definitions consistent. Likely voters are often defined as those who say they definitely will vote in an election. Some polls are more restrictive, requiring the person to have voted in a previous election and say they are going to vote in the current one. Likely voters trend much more Republican than the general population. On average, likely voters are 1% more Republican than the final vote, but registered voters are 5% more Democratic than the final vote. In some years that difference is even larger, as much as 7%. Polls of all Americans are worthless for voting prediction, and are only valuable to get a coarse sense of a politician's popularity.
A further source of trouble can be the composition of the poll sample and its size. The largest polls use about 3-5,000 people to represent the entire US. There are statistical methods to make this small sample valid, but its composition is vital. If the electorate's composition is different from the poll's sample the poll will not be accurate. This plagued the 2016 polling, which presumed the same turnout among African-Americans for Clinton as there had been for Obama. Of course, it wasn't, as almost anyone could have predicted. The pollsters DID in fact predict that, but got heavy public criticism from Obama and the DNC for their sample composition. They changed it to reflect the 2012 electorate, which made the polls run 2-5% more Democratic than the actual election. This is a major reason the poling was so inaccurate for the 2016 cycle.
Another difficult issue is voter enthusiasm. At various times Republicans and Democrats have been much more motivated to vote. Since most self-identified party adherents vote for their party regardless of who is running, enthusiasm of the base makes a big difference in who turns out to vote and the result. Some polls modify their sample for this, but most do not, as it cannot be reliably measured.
It is thus essential to know who was polled before giving a result any credence. The USA Today editorial's real purpose becomes clear when you look at their sources. The 36% figure for Trump was measured among all Americans. Clinton's same measure was only 45%. The poll purporting to show Trump in trouble in Ohio was of registered voters. It shows him behind by 2-3%. Except registered voters run 5% or more Democratic compared to actual vote. In other words, Trump isn't in trouble at all, The paper just wants to make an impression.
When reading political news, it is essential to look at the source and the item being reported. Simply accepting the poll 'results' without looking at where it came from and its methodology is just like reading a headline and not the article. When teaching residents about how to read medical literature I tell them to read the method and data and then form their own conclusion, and only THEN see what the authors' concluded. This sharpens appreciation for methods and also the reader's ability to think. Polls have to be read the same way. Read the data, not the conclusion, then decide if the conclusion is rational based on the data. This same caveat applies to reading material on covid 19.
It is often possible to understand a headline just from who published it. No one should take seriously a CNN poll report saying Trump is the spawn of Satan, just as I don't take seriously FoxNews reports that say Trump is loved universally. The most valuable information is that which goes against the editorial bent of the source. I look for articles on CNN that praise Trump, those on Breitbart that laud Democrats.
There is a powerful first-hand report from one of Sherman's soldiers about how his apparent overwhelming numerical strength was used. Nearly 30,000 men were involved in guarding his supply trains! At Kenesaw (correct sp in 1864) and Atlanta Sherman's supposed giant numerical superiority was in fact parity. This was also true at Gettysburg. Just like these battles cannot be well understood without knowing the real data, neither can polls, the reports of which must be read with care and conclusions made with caution.
Right now polls are all of registered voters. They mean virtually nothing in predicting who is 'ahead' and 'behind'. These same samples showed Regan losing to Carter, Clinton crushing Trump, Kerry leading Bush. Every election turned out differently. I honestly don't know why anyone bothers publishing these things. They mean nothing, predict nothing, and like so much else in our 'news' do nothing except excite partisan tension, which we don't need. If you want political news, better to look at fundraising; polls of registered voters at this time of year are as helpful as a forecast of November's weather made in May.....
Ted
If you read most tertiary histories of the Civil War you will wonder how the South won so many battles. At almost every battle the North had a substantial numerical superiority, according to what these books publish. The average Southern soldier must have been an ubermensch to triumph against such terrible odds. To understand what really happened you have to know where the data comes from, and what it means.
Army returns can be found in multiple primary sources. The most comprehensive source is the Official Records of the War of the Rebellion. This titanic work is 127 volumes long, each about 800 pages. In it are the official reports of every officer from every battle of the war, along with congressional hearings and first-hand testimony about conditions on the field. It is a trove of information. I read the two volumes devoted to Gettysburg while designing a wargame on that battle. It was here I learned the real strength numbers.
Union and Confederate armies tallied strength differently. Confederate regiment and brigade commanders reported rifles, that is, the number of men present with a gun in their hand. That count was typically made right before going into action. Thus men detached for provost guard duty, or sent to artillery units to carry shells, were not counted. Officers also were not counted, and officers were 7-10% of total infantry strength.
The Union armies did a roll call every morning. The totals were reported in three columns- present, present for duty, and present for duty equipped. Present simply meant the man was on the active regimental roll. Regiments when formed had a strength of 1000 men, but by the time they got to an army their strength was rarely more than 650. The average regiment at Gettysburg was 350 men, most of the losses from disease rather than combat deaths. 'Present' was simply the raw strength of the regiment, including those in hospital or under arrest. 'Present for duty' meant the man was standing actually able to report. This included men detached for artillery work, guard duty or off foraging (a major activity during some campaigns). 'Present for duty equipped' was the total of men actually available to march and fight with the regiment that morning.
It is apparent that 'present for duty equipped' is the closest equivalent to the Confederate strength report, but it must be adjusted by the fact that the Union also counted officers. Union 'present for duty equipped' must therefore be reduced 7-10%, or Confederate rolls increased 7-10% to have a true apples-to-apples comparison of force.
Unfortunately, almost all tertiary reports follow the convention of Livermore (a secondary analyst and historian), who used 'present for duty' as the measure of Union strength. It is from this conceptual error that all the imagery of overwhelming Union hordes comes from. While the Union DID field more men, at the 'point of the spear' the difference was much smaller than popular histories make it seem.
Exactly the same problem exists for polling. Polls may be taken of all adults, registered voters, or likely voters. The composition of these populations is not the same, nor are their definitions consistent. Likely voters are often defined as those who say they definitely will vote in an election. Some polls are more restrictive, requiring the person to have voted in a previous election and say they are going to vote in the current one. Likely voters trend much more Republican than the general population. On average, likely voters are 1% more Republican than the final vote, but registered voters are 5% more Democratic than the final vote. In some years that difference is even larger, as much as 7%. Polls of all Americans are worthless for voting prediction, and are only valuable to get a coarse sense of a politician's popularity.
A further source of trouble can be the composition of the poll sample and its size. The largest polls use about 3-5,000 people to represent the entire US. There are statistical methods to make this small sample valid, but its composition is vital. If the electorate's composition is different from the poll's sample the poll will not be accurate. This plagued the 2016 polling, which presumed the same turnout among African-Americans for Clinton as there had been for Obama. Of course, it wasn't, as almost anyone could have predicted. The pollsters DID in fact predict that, but got heavy public criticism from Obama and the DNC for their sample composition. They changed it to reflect the 2012 electorate, which made the polls run 2-5% more Democratic than the actual election. This is a major reason the poling was so inaccurate for the 2016 cycle.
Another difficult issue is voter enthusiasm. At various times Republicans and Democrats have been much more motivated to vote. Since most self-identified party adherents vote for their party regardless of who is running, enthusiasm of the base makes a big difference in who turns out to vote and the result. Some polls modify their sample for this, but most do not, as it cannot be reliably measured.
It is thus essential to know who was polled before giving a result any credence. The USA Today editorial's real purpose becomes clear when you look at their sources. The 36% figure for Trump was measured among all Americans. Clinton's same measure was only 45%. The poll purporting to show Trump in trouble in Ohio was of registered voters. It shows him behind by 2-3%. Except registered voters run 5% or more Democratic compared to actual vote. In other words, Trump isn't in trouble at all, The paper just wants to make an impression.
When reading political news, it is essential to look at the source and the item being reported. Simply accepting the poll 'results' without looking at where it came from and its methodology is just like reading a headline and not the article. When teaching residents about how to read medical literature I tell them to read the method and data and then form their own conclusion, and only THEN see what the authors' concluded. This sharpens appreciation for methods and also the reader's ability to think. Polls have to be read the same way. Read the data, not the conclusion, then decide if the conclusion is rational based on the data. This same caveat applies to reading material on covid 19.
It is often possible to understand a headline just from who published it. No one should take seriously a CNN poll report saying Trump is the spawn of Satan, just as I don't take seriously FoxNews reports that say Trump is loved universally. The most valuable information is that which goes against the editorial bent of the source. I look for articles on CNN that praise Trump, those on Breitbart that laud Democrats.
There is a powerful first-hand report from one of Sherman's soldiers about how his apparent overwhelming numerical strength was used. Nearly 30,000 men were involved in guarding his supply trains! At Kenesaw (correct sp in 1864) and Atlanta Sherman's supposed giant numerical superiority was in fact parity. This was also true at Gettysburg. Just like these battles cannot be well understood without knowing the real data, neither can polls, the reports of which must be read with care and conclusions made with caution.
Right now polls are all of registered voters. They mean virtually nothing in predicting who is 'ahead' and 'behind'. These same samples showed Regan losing to Carter, Clinton crushing Trump, Kerry leading Bush. Every election turned out differently. I honestly don't know why anyone bothers publishing these things. They mean nothing, predict nothing, and like so much else in our 'news' do nothing except excite partisan tension, which we don't need. If you want political news, better to look at fundraising; polls of registered voters at this time of year are as helpful as a forecast of November's weather made in May.....
Ted
Comments
Post a Comment