3 ms·

How does 365 data points give your conclusions any sort of statistical significance? Snapchat's numbers are allegedly over 30M, Facebook's are over 1.2B and Ins

How does 365 data points give your conclusions any sort of statistical significance? Snapchat's numbers are allegedly over 30M, Facebook's are over 1.2B and Instagram's are 150M. I suggest reading "How to Lie with Statistics"[1] [1] http://www.amazon.com/How-Lie-Statistics-Darrell-Huff/dp/0393310728 http://www.amazon.com/How-Lie-Statistics-Darrell-Huff/dp/039... Edit: I guess that sounded a little harsh. No doubt, Snapchat's growth has been explosive and if this was just another blog-spam post about that fact, I'd let it be. But you're selling a statistical software! I'd expect better from people who know what they're doing. Also, "How to Lie with Statistics" is a really good book, should be required reading for anyone dealing with numbers. Did not mean that as an attack on your product.

lclarkmichalek 13y ago
Gallup's daily tracking polls, which offer a +-4% margin of error, are based of 1000 data points (insert complaint about appeal to authority here). While 365 is significantly lower than that, there is nothing wrong with using the data to produce a value and a margin of error/confidence.
ckelly 13y ago
Hi Nathan, Thanks for your comment. And no offense taken. :) Most people are surprised how "few" respondents it takes to get to 5% margin of error at 95% confidence levels: https://en.wikipedia.org/wiki/Margin_of_error#Different_confidence_levels https://en.wikipedia.org/wiki/Margin_of_error#Different_conf... In this case, .98/sqrt(365) = 5.1% So, perhaps counterintuitively, 300-400 respondents gives good read on large populations (like Snapchat users...or the US population!).
- tomato_sausage 13y ago
  Assuming the respondents are uniformly randomly chosen, it takes surprisingly few people to get a pretty high confidence interval. I'm not convinced that the respondents are uniform, but I expect that there is a sufficient distribution that the results are still interesting. 3x vs. 1.5x seems like a small difference to me, but I am not an investor or user, so....
  - michaelmior 13y ago
    Yes, this exactly. The assumption of uniform distribution is not one to be taken lightly.

lclarkmichalek 13y ago

Gallup's daily tracking polls, which offer a +-4% margin of error, are based of 1000 data points (insert complaint about appeal to authority here). While 365 is significantly lower than that, there is nothing wrong with using the data to produce a value and a margin of error/confidence.

ckelly 13y ago

Hi Nathan, Thanks for your comment. And no offense taken. :) Most people are surprised how "few" respondents it takes to get to 5% margin of error at 95% confidence levels: https://en.wikipedia.org/wiki/Margin_of_error#Different_confidence_levels https://en.wikipedia.org/wiki/Margin_of_error#Different_conf... In this case, .98/sqrt(365) = 5.1% So, perhaps counterintuitively, 300-400 respondents gives good read on large populations (like Snapchat users...or the US population!).

tomato_sausage 13y ago
Assuming the respondents are uniformly randomly chosen, it takes surprisingly few people to get a pretty high confidence interval. I'm not convinced that the respondents are uniform, but I expect that there is a sufficient distribution that the results are still interesting. 3x vs. 1.5x seems like a small difference to me, but I am not an investor or user, so....
- michaelmior 13y ago
  Yes, this exactly. The assumption of uniform distribution is not one to be taken lightly.