11 ChoosingStatTest

[Music] now that we've talked about those p-values you might want to know how do we get a p-value how do the statisticians that work on all these studies calculate those numbers that they're going to put into the papers that you're going to read so we're going to talk today about hypothesis testing we're going to show you briefly how we choose which statistical test to use and give you some examples and we're really going to avoid all the math the purpose of this course is not to teach you how to be a biostatistician but so that when you see those words written in an article you'll know exactly what was going on now the choice of a statistical test reminds me of kind of using the right tool for the right job you have to each statistical test is a tool you got to kind of match it to the job it's doing i had a bit of an interesting experience with this last weekend when i was building a wooden leprechaun box with my daughter for a school project all right so the basic idea is that we're going to put these wires through the side here to represent sort of the bars of the leprechaun cage but in order to hold the wires we're going to put these little tacks in um so i'll do the first one and you can just you'll do the rest after that so basically like this like sorry now yeah dad yeah it's not gonna work yeah i can't find my hammer can you use your eyes and look around okay all right that's nice we didn't do so well on that in any case let's go back to that concept of difference you remember in the last lecture when i was talking to you about the medical internal medicine docs and the surgeons and how i was going to measure their weights and we could assign a p-value to the difference and see how weird it our results would be considering that they assuming that they just are the same weight so so just to remind you um we were asking are our surgeons heavier than internal medicine doctors and we measured a sample of them and we found that the average internal medicine dock weighs 175 pounds and the average surgeon weighs 180 pounds and then we ask this question right is is 175 different than 180 and of course you're all saying yes right because 175 does not equal 180 but in the world of statistics you'll see people using this language all the time they'll say oh are those two different they don't mean are they literally different of course they're different what they mean is are they statistically different are they distinguishable using mathematics and so oftentimes people say oh no those aren't different even though the numbers themselves are different they're talking about statistics now calculating those p values from coin flips is pretty easy right remember getting one head in a row assuming you had a regular old-fashioned coin would be a chance of 50 percent and you know more heads in a row the level kind of goes down by a stepwise fashion but but medicine isn't coin flips right most of our data is not someone flipping a coin you know the measurement of weight in our doctors for instance doesn't uh isn't isn't similar to a coin flip so the specific equations to calculate p-values are behind the scope of this course but you just need to know how a test gets chosen and and the good news here is that you really only need a couple pieces of pieces of information to choose which statistical test or to recognize if the right statistical test was being used in a given study you need to know what type of data is my exposure so remember back to our talk about types of data and you need to know what type of data is your outcome so we've talked before about linking exposures to outcome we just need to add that layer of what type of data is uh is my exposure and outcome so here is the master chart okay so what i've done here is i've listed the types of exposure types of data you can have for an exposure and the types of data you can have for an outcome and listed a statistical test i'm going to give you some specific examples of all of these but just to show you how to read this table if your exposure is a categorical variable like let's say gender and your outcome is a categorical variable like death you would compare those with a specific statistical test a specific tool in this case it would be called the chi-squared test or the fisher's exact test and those of you had some statistics in college might remember those tests the rest of you i just don't want you to be afraid if you read these in a medical paper and says oh we use the chi-square test to evaluate this all they are is saying this is the tool we use to generate that p-value we're reporting to you okay so it tells you something about the type of data and you can see if the exposure is continuous and normally distributed and the outcome is categorical you would use something called the student's t-test i'm not going to walk through each one of these we'll reproduce these for you and this is something you can always look back at i also want to note that this assumes generally that we're comparing one thing to another thing there are special tests that you can use when you have multiple groups so instead of like men versus women if you had multiple racial groups for example there's a kind of special set of tests there and this doesn't talk at all about multi-variable testing and so we'll get later in the course to the idea that instead of just comparing you know age to death rate or something like that you might want to adjust for other variables like smoking status or income or something and so you're putting more variables into these equations that's that's its own its own thing right now for this type of statistical testing we're just comparing one thing to another thing so let's go through some examples and i think you'll you'll get a feel for it so here's a really encouraging article i recently read um about a condition called hutchinson-gilford progeria this is a genetic syndrome of rapid aging and these kids they they uh they start aging very rapidly they lose their hair they have very low muscle tone the natural history of the disease unfortunately is that these children tend to die somewhere in their teen years um so there was a new drug that they were trying for this a drug called lonafarnib i won't go into the mechanism and they did a small study you can see the numbers of patients here and they gave some some of the kids got lunafarnib and some did not take the drug and then we look at the number that survived and the number that died okay and so what you can see here is that in the lone farney group 59 kids survived and four died compared to 46 survivors and 17 deaths in the nothing group so i just want you to feel in your heart how weird that data is assuming that lonafarnib does nothing does that data feel weird assuming the null hypothesis it does to me and it should to you too this is a little weird it intuitively feels weird and if we put this into a chi-square test because that chart tells us to put it into a chi-square test we could put these numbers in and we'd get a p-value of .002 so results this weird would only happen two point two percent of the time assuming that drug doesn't work so that is encouraging evidence that this drug might work remember this p-value just tells you how weird it is assuming the drug doesn't work but very encouraging study here all right let's talk about a continuous variable across groups so so this was an article looking at a drug called phlebancerin or it's being marketed as addy for a hypoactive sexual desire disorder in women so this is sort of a libido enhancing drug and the metric that they looked at in the two groups so one group of women got placebo and one group got this drug and they looked at this metric called the change in the number of sexually satisfying events per month and so they had women report how many sexually satisfying events they had over the course of a month and what you see here is that that number increased in the placebo group by 1.1 so there's some placebo effect here even the women on placebo had about one extra sexually satisfying event per month and uh but it went up by 1.9 almost twos sexually satisfying events per month in the phlebancerin group and you can see the distribution of the results here so the placebo group is in blue everyone is kind of shifted above zero because most people did better some people had fewer sexually satisfying events per month and then you can see people coming all the way up some people had even 10 extra sexually satisfying events per month very encouraging results of course we see that in the placebo too so you can tell with this continuous data when we're comparing it it's harder to intuit how weird this is this i don't have a good intuition just looking at the picture to say oh how much better is this i can't assign a p-value quite as easily using my head but we can use students t-test to go ahead and give us a p-value and in this case the p-value was reported as less than 0.01 so about 1 you'd get data this weird 1 of the time assuming this drug does nothing at all that's suggestive that the drug has some effect all right what about correlations so this was an interesting study looking at uh doctors by specialty and what it looked at was how empathetic they were on the y-axis here so higher is more empathetic more caring and how republican they are is on the x-axis here and so what you can see are these dots are the various specialties and there seems to be a correlation right the more empathetic a specialty is the less republican they are and sort of vice versa there's that kind of negative correlation and we can use a pearson's correlation test that's the proper tool for this job to generate a p-value associated with this correlation and this i do think you can kind of eyeball and say you know this seems like there's something here this data assuming there's no relationship between empathy and republicanism at all this data would be pretty weird right and indeed it is again the p-value was reported as less than one percent less than .01 it turns out psychiatrists are the most empathetic of the specialties and the most liberal which makes a lot of sense i know some psychiatrists and those surgeons and anesthesiologists are on on the other end in terms of empathy and liberalism so what i've done during this brief talk is just to show you that statistical tests are all about asking how different two samples are and trying to quantify that in some some way now we are never a hundred percent sure that they are different in that statistical sense but we can describe how weird the results are assuming they really aren't different again that's the concept of the p-value the correct statistical test is just dictated by the type of data present so all you have to decide is what kind of data you have and then you can just plug plug your data into into a statistical test calculator and get a p-value so now you know how those p-values come about i'll see you next time

Channel: YaleCourses

Views: 670

Rating: 5 out of 5

Keywords:

Id: Ts54ZF0yayU

Length: 11min 20sec (680 seconds)

Published: Fri Jul 31 2020