When drawing an inference (from a sample statistic, about a population parameter), we cannot avoid errors. We inevitably must commit either a Type I or Type II error
Conceptually, it does not really matter that the distribution is normal. The significance (e.g., 5%) is the probability of a Type I error and, if you use a different distribution, the different shape is reflected in a different deviate. Instead of 1.645 associated with the normal, your deviate comes from your distribution
I like your tutorials but this one was a little confusing. May be this can help: Type I error - when we reject a “good” part as “bad” Type II error - when we accept a “bad part” as “good”. If the part is critical to safety we would prefer to make a type I error.
In a hypothesis testing the decision can go either way: your sample can align with the null or alternative. You can say "accept the null" if you fail to reject it but they recently dropped that because it might involve you with the type II error (accepting H0 when it is false) when it's unnecessary. so it safer for you to say that you have no enough evidence to reject the null. (leaves you the type 1 and that only if your test actually detected an effect enough to reject)
you're a life saver david, im using your videos to help study for the Level 1 exam and they really help crystallize a lot of the quant material specifically. Thanks so much!
@cleobabie not if one-tailed, i showed both, see (eg) @4:25. A CI (bound), like a test, can be 1 or 2-tailed (sided). A one-sided test, (eg., all VaRs), have the 5% significance in one tail
@DURound Yes, agreed ... I had thought the Census was a sample, but as other commentors pointed out, a Census is an attempted survey of a population, so I did use a bad example. Thanks!
@lbrown1956 However, it is practically impossible to include EVERY individual in a nation-state. For example, in Singapore, where I come from, its only 50km wide +/-, yet still, it is practically impossible to include every single individual. The first and utmost reason is the practicality of the situation, and second, it should be economically pragmatic; inexpensive.
you can post questions at our forum at bionicturtle(dot)com/forum ... but in my humble opinion, part of the problem is you are linking Type 1/2 to the type of distribution. If you want to start at fundamentals, i would abstract from any distribution, then you may realize that your concept applies to (is independent of) the particular distribution. thanks!
@bionicturtledotcom Hi..we say that it is not possible to minimise both the errors simultaneously..that is if we attempt to decrease type 1 error (alpha), the type 2 error increases. (beta)..but how can we prove this correlation between these two types of errors?? that is how can we show that a decrease in alpha increases beta..thanks..
What do you mean the census doesn't measure everybody, this is exactly wrong. They do measure the a set of specific characteristics of everybody. It is true that they don't measure certain characteristics of everybody but there is a base set that they do measure of everybody.
No I am not kidding. That is what a census is. It measures every individual and it is required by the constitution that we count every person individually. Now it is true that inherently we are going to miss some. So people with knowledge in statistics are pushing to try to include some manipulation of the data to deal with these issues. They hire 10's of thousands to go around to every address in the US. They try to count the homeless by picking a specific night to measure them.
@bionicturtledotcom medical statistics made easy. its the one my university recommends but I don't like it. It doesn't go into enough detail for me to actually apply it. When I used it statistics was an abstract unapplied thing that I couldn't actually use. If I was a bit cleverer maybe I could extrapolate from simple definitions and calculations but I need a helping hand to apply it to actual research papers and documents
Conceptually, it does not really matter that the distribution is normal. The significance (e.g., 5%) is the probability of a Type I error and, if you use a different distribution, the different shape is reflected in a different deviate. Instead of 1.645 associated with the normal, your deviate comes from your distribution
I like your tutorials but this one was a little confusing.
May be this can help: Type I error - when we reject a “good” part as “bad”
Type II error - when we accept a “bad part” as “good”.
If the part is critical to safety we would prefer to make a type I error.
In a hypothesis testing the decision can go either way: your sample can align with the null or alternative. You can say "accept the null" if you fail to reject it but they recently dropped that because it might involve you with the type II error (accepting H0 when it is false) when it's unnecessary. so it safer for you to say that you have no enough evidence to reject the null. (leaves you the type 1 and that only if your test actually detected an effect enough to reject)
you're a life saver david, im using your videos to help study for the Level 1 exam and they really help crystallize a lot of the quant material specifically. Thanks so much!
@cleobabie not if one-tailed, i showed both, see (eg) @4:25. A CI (bound), like a test, can be 1 or 2-tailed (sided). A one-sided test, (eg., all VaRs), have the 5% significance in one tail
@DURound Yes, agreed ... I had thought the Census was a sample, but as other commentors pointed out, a Census is an attempted survey of a population, so I did use a bad example. Thanks!
@lbrown1956 However, it is practically impossible to include EVERY individual in a nation-state. For example, in Singapore, where I come from, its only 50km wide +/-, yet still, it is practically impossible to include every single individual. The first and utmost reason is the practicality of the situation, and second, it should be economically pragmatic; inexpensive.
The 5% outside of the confidence interval is distributed between the 2 tails, which means that each tail bears 2.5%, not 5% David.
excellent demonstrations. I have leraned a lot when see your videos.
Incorrect about the Census --- at least, if the Census is done correctly. Indeed, the word census means individual registration in the original latin.
sweet! man u are awesome in explaining these concepts visually !! i am a visual thinker & this helps tremendously!!
you can post questions at our forum at bionicturtle(dot)com/forum
... but in my humble opinion, part of the problem is you are linking Type 1/2 to the type of distribution. If you want to start at fundamentals, i would abstract from any distribution, then you may realize that your concept applies to (is independent of) the particular distribution. thanks!
@bionicturtledotcom Hi..we say that it is not possible to minimise both the errors simultaneously..that is if we attempt to decrease type 1 error (alpha), the type 2 error increases. (beta)..but how can we prove this correlation between these two types of errors?? that is how can we show that a decrease in alpha increases beta..thanks..
your vedios are very helpfull.
What do you mean the census doesn't measure everybody, this is exactly wrong. They do measure the a set of specific characteristics of everybody. It is true that they don't measure certain characteristics of everybody but there is a base set that they do measure of everybody.
Doesn't the Census count "everyone" within the US, barring measurement difficulties?
@lewiscooney interesting, what are these "textbooks" to which your refer, they sound familiar?
No I am not kidding. That is what a census is. It measures every individual and it is required by the constitution that we count every person individually. Now it is true that inherently we are going to miss some. So people with knowledge in statistics are pushing to try to include some manipulation of the data to deal with these issues. They hire 10's of thousands to go around to every address in the US. They try to count the homeless by picking a specific night to measure them.
good video. One question. In real life aplications How high an acceptable %of conficence should be?
Is this what is referred to as "error in research"???
@bionicturtledotcom medical statistics made easy. its the one my university recommends but I don't like it. It doesn't go into enough detail for me to actually apply it. When I used it statistics was an abstract unapplied thing that I couldn't actually use. If I was a bit cleverer maybe I could extrapolate from simple definitions and calculations but I need a helping hand to apply it to actual research papers and documents
You can use Quantitative Methods board at bionicturtle(dot)com/forum, thanks
I am thinking of burning all my textbooks and just using youtube