Question on Statistics and Polls

[Alex H]

Articles in this weeks news tell me that smoking in the UK is now down to 12% of the population

I realise that the people creating that statistic did not ask all of the population if they smoked or not.

In the dim memories of my mind , I seem to recall something like, you only have to ask 2000 people and then multiply up to get a reasonable, workable figure.

Is this true? and if so is it based on someones 'laws of statistics?

If so, can you point me in the right direction?

 

[Drago]

I believe you are correct in that they clearly extrapolate a figure, but a lot of polls don't get anywhere near 2000, and are often just a single exercise in one part of the country.

Clearly many aren't done in a statistically rigorous manner, as the bulk called the last two general elections wrong, predicted the result of certain referendum in 2016 wrong, and so forth (no political discussion intended, simply using those as very prime examples of polls turning out to be well off beam.)

Like you, now you've mentioned it I too would be interested in the sample size, but also how many samples from around -the country are needed to stand a chance of removing the uncertainty caused by regional variance.

 

[vickster]

Google confidence intervals and sample sizes.
if the data is from ONS, you should be able to find info on their sampling and analysis methodology
https://www.ons.gov.uk/peoplepopula...ased on APS data, the,20.2% of the population).

 

[MontyVeda]

I think at some point in the past, a statistician realised that surveying more than XXXX amount of people would largely deliver the same results, so there's no point sending out a million questionnaires when they'd get a very similar result from 10,000.

 

[annedonnelly]

Provided that the 2000 you choose are a representative sample. You don't want to bias the results by choosing people who are more or less likely to do the thing you're researching. For example, if you wanted to know what the population thought of cycling you shouldn't just ask on a cycling forum.

 

[Badger_Boom]

Provided that the 2000 you choose are a representative sample. You don't want to bias the results by choosing people who are more or less likely to do the thing you're researching. For example, if you wanted to know what the population thought of cycling you shouldn't just ask on a cycling forum.

Surely actively ‘choosing’ a sample is guaranteed to introduce bias?

 

[presta]

The margin of error is dependent on absolute sample size not relative sample size, so if you have a skip full of coins, you don't need to toss every last one to get an accurate answer to the probability of getting heads.
https://en.wikipedia.org/wiki/Margin_of_error.

 

[presta]

Surely actively ‘choosing’ a sample is guaranteed to introduce bias?

No, because the samples are chosen to be representative of the population as a whole, not to serve a vested interest. If you want a biased answer it's less trouble to just make figures up, but then polling companies would quickly end up with no clients.

 

[annedonnelly]

Surely actively ‘choosing’ a sample is guaranteed to introduce bias?

I think it's very hard not to.

 

[Alex H]

from @vickster 's link above

"The Annual Population Survey (APS) is a continuous household survey, covering the UK, with the aim of providing estimates between censuses of main social and labour market variables at a local area level. The APS is not a stand-alone survey, but uses data combined from two waves of the main Labour Force Survey (LFS) with data collected on a local sample boost. Apart from employment and unemployment, the topics covered in the survey include housing, ethnicity, religion, health and education.

The datasets comprise 12 months of survey data and are disseminated quarterly. The achieved sample size is approximately 320,000 respondents."

A bit bigger than I thought then!

Reading further, there is a great deal of effort put in to the survey to ensure a minimum of bias in selection.

However, it's one of those documents that will put you to sleep unless you are extremely interested in the contents. 😴

 

[albion]

It is probably easier to do surveys these days though they can still go wrong.
12% sounds high considering so many people managed to switch to vapes.

The age divide stat for smoking would be interesting.

 

[wafter]

Man, this brings me back to dim and distant memories of A-Level Sociology

 

[vickster]

The age divide stat for smoking would be interesting.

See ons data.
Also has data on vaping. Not sure ex smokers are switching to vapes, more those who have never smoked taking up vaping.
Why do you assume it’s easier to do surveys? Cheaper probably yes, not necessarily easier

 

[Ming the Merciless]

Clearly many aren't done in a statistically rigorous manner, as the bulk called the last two general elections wrong

If you ask a sample in a conservative majority constituency you’ll find they are going to get in. Ask a sample in a labour majority you’ll find they are going to get in. A statistical sample has to be representative and you’re not going to get that from a single constituency.

Elections aren’t won on the percentages of who votes for whom but on key constituencies that swing between the parties.

 

[Ming the Merciless]

Articles in this weeks news tell me that smoking in the UK is now down to 12% of the population

I realise that the people creating that statistic did not ask all of the population if they smoked or not.

In the dim memories of my mind , I seem to recall something like, you only have to ask 2000 people and then multiply up to get a reasonable, workable figure.

Is this true? and if so is it based on someones 'laws of statistics?

If so, can you point me in the right direction?

It’s called confidence intervals. Everything you want to know but were afraid to ask, enjoy.

https://www.scribbr.com/statistics/confidence-interval/