Thread by @PJACaytonPhD

Ensemble approaches, also known as combination methods, exist in statistics and data science to try & make sense of predictions coming from different models by combining them through some scheme.

In some instances, combining models by averaging predictions can be accurate!

Ensemble predictions tend to perform well as a means to settle highly varying estimates.

When one model prediction is too high,

And another model prediction is too low,

Sometimes the average of the two may be just right.

Could we use them for election polls?

There have been means of combining poll results from different firms to get some insights.

Here is Wikipedia's approach, wherethey collated results & found a smoothed middle among the varied results thru LOESS.

en.m.wikipedia.org/wiki/Opinion_p…

Personally, different opinion research firms would require differing weights as a function of their quality, but again, sometimes average or middles may be sufficient for accuracy and insight.

Now, I will make some toy analysis trying to combine Pulse Asia's results and Google Trends, especially when trying to reconcile some representation issues.

So many heavy assumptions that can be taken as points of attack, and generally my reply to valid non-troll critique would be:

"I know, understand, and acknowledge your critique, and with that, let's continue playing with this toy."

Let's start with one assumption:

That Google Trends are strongly correlated with vote percentages.

Pretty heavy assumption. I know. I hurled inside my mouth trying to continue with this, but let's continue with playing with this radioactive toy.

Second assumption:

That Pulse Asia survey results are strongly correlated with vote percentages in May.

Still pretty heavy, but this has been supported from previous elections, so not that noxious.

Third assumption, but this one is malleable as we continue:

Both Pulse Asia and Google Trends results are captured from the same population but their methods lead to different results and can produce representation issues.

We will try to make a fix in the age issue later.

So here were Pulse Asia & Google Trends presidential results, shot as of May 3, 2022, 11:58pm.

By simple average,

Marcos = (56 + 24)/2 = 40%
Robredo = (23 + 55)/2 = 39%

In reconciling these two sources of data, it may seem that they are practically tied, but favors Marcos.

Again, note that these rely on VERY HEAVY ASSUMPTIONS and should be only seen as toys.

Now, let's continue.

Let's adjust the third assumption, but the fix that I will do only takes into account the age representation between Pulse Asia and Google.

Third Assumption, changed:

The Pulse Asia and Google Trends results are adjusted to reflect voting-age population projections by weighted average, increasing weights from underrepresented age groups and decreasing weights for overrepresented groups.

This necessitates two additional assumptions:

Fourth Assumption:

The Google Trends results are all the same for each age group.

There is no age breakdown in Google Trends, so I will have to make do with this.

Fifth Assumption:

Voting age population projections are perfectly reflected on COMELEC population of registered voters.

It may not be the case so I will accept that critique.

Though, please give me some leeway as COMELEC has bad age bins:
15-41, 42-55, and 56 and above. Nope.

Some of the population %ages have been discussed in the "Question of Age Representation" thread, shown below

https://twitter.com/PJACaytonPhD/status/1521044912460173313

And here is the toy monster:

Marcos is estimated to have 41.96% of the vote,
Robredo is estimated to have 38.35% of the vote

How were the reconciled results were computed?

The formula is:

For each age group,

For example:

For Marcos's age 18-24 reconciled result,

58.27 = (19.34/10.83 * 72 + 19.34/27.05 * 24) / (19.34/10.83 + 19.34/27.05)

The formula for the overall result?

Overall Result
= sumproduct(reconciled results, est. voting population %) / sum(est. voting population %)

If you would like to play around with the toy analysis:

docs.google.com/spreadsheets/d…

*How were the reconciled results computed?

Do these results hold some ground?

They depend on the assumptions, of which with some I would vomit for even considering.

Again, VERY HEAVY ASSUMPTIONS, and thus these are simply toy analyses.

Whatever helps me sleep, of which I need now.

PS:

I also made for the VP race between Sara, Kiko, and Tito.

It seems Sara may garner the majority vote.

Side comment: We made the youth (age 18-24) more ignorant of our history.