15 CaseControlStudy

[Music] i told you in the last lecture that cohort studies didn't work very well when there was a rare outcome and there's a solution to that they're called case control studies let's talk about them now i'm going to explain what a case control study is and how it's a particularly good design when outcomes are very rare i'm going to describe the achilles heel of those studies and how you can find it and show you that interpretation is pretty straightforward even though it might not seem that way to give you a real sense of how these studies work though let's consider a new rare disease that is emerging let's say i'm a researcher who wants to study a really unusual disease affecting maybe one in a hundred thousand people the disease presents with day night reversal the patient tends to sleep all day and are much more active at night there's substantial power along with unusual taste changes the patients additionally present with lengthening of the canine teeth and a particular revulsion to certain foods of the allium genus such as garlic i am of course discussing vampirism which occurs quite sporadically and for obvious reasons rarely reaches medical attention assuming that one in a hundred thousand individuals contracts vampirism a cohort design would have to be huge to generate a sample of affected individuals large enough to make some guesses about the cause i'd need to enroll 10 million people to find 100 cases not feasible but what if i find the hundred cases first i'll advertise at various journals put posters up at s m clubs that sort of thing i've just completed half of my study now i just need to find about 100 controls people without vampirism and start looking for systematic differences [Music] did you catch the efficiency there so instead of enrolling 10 million people to find my 100 i find my 100 first and then i just have to compare them to average everyday people so that is the genius of the case control study case control studies flip the script we've been talking about exposures and outcomes and in a cohort study you take an exposure like smoking and you link it to various outcomes like death or cancer or dialysis or dementia a case control study starts with the outcome find the outcome like people with alzheimer's disease and then look backwards in time and see what were they exposed to in the past that can link to this outcome now in a cohort study we know the exposure and are interested in what outcomes may occur in a case control study we know the outcome and we're interested in what exposures may occur this is particularly useful when many potential exposures are possible so when we don't know what the cause of a disease is right we see a new disease we have no idea what's going on we have no hypothesis what could be causing it really useful so let's look at a medical study that was done with this case control design this in this case we're talking about oral contraceptives that's the pill and the risk of venous thromboembolism that means blood clots in the legs or lungs so this is a really rare outcome only 14 out of 10 000 women who take the pill might develop a blood clot and so you would have to enroll a huge number of people to do this in a cohort type design you know maybe a million women taking the pill to find enough blood clots to make any meaningful inferences so that's an inefficient study design what's the case control design here well what the authors did is they found 10 000 women of child bearing age who had blood clots and then they matched them to 45 000 women of child bearing age who didn't have blood clots and then they simply asked the question how many of them had been taking the pill okay so flipping the script a different design and what they found is that the women who had blood clots were more likely to have been taking the pill yeah case control studies make your tenses very strange but the interpretation is exactly the same as a cohort study in fact you just you can you can look at the numbers exactly the same and what they found the graph shows you across a variety of different uh pill types a variety of different oral contraceptives but broadly speaking the risk was increased anywhere between two and four fold now that's a relative risk increase remember the absolute risk is extraordinarily low you know 14 and 10 000 so even a doubling in risk only goes up to 28 and 10 000 but they never would have been able to detect this with a cohort design case controls the appropriate choice when you have a rare outcome now the achilles heel of case control studies is that word control it is really easy to identify a case right you just look in your medical records and say like who had blood clots in their legs that's your case who are the controls well you know that they can't have blood clots in their legs that's that's easy otherwise they'd be a case but in a case control study what you are actually trying to do is select your controls from this theoretical cohort study that you can't afford to run so what you do is you imagine if i was doing this as a cohort study and i was enrolling 10 million women who would i enroll who would i allow to get to the start of the of the marathon okay who would i enroll in that huge cohort study and then i'd follow and some would get you know some would get blood clots and some wouldn't we want to pick our controls from that huge theoretical population that didn't end up getting blood clots and that's where people often go wrong in these study designs you have to choose the right control in theory the controls should come from the same theoretical gigantic cohort that you would have enrolled if you could have done a cohort study which means they would have the same exact inclusion and exclusion criteria as the cases right because if you were doing the cohort study you wouldn't know who's a case right you would just enroll people and watch and see what happens they'd all have to get enrolled based on the same rules in the blood clot study in the study i just showed you cases were excluded if they had been on blood thinners in the past 40 days controls if they had ever been on blood thinners that is a subtle difference but it is important it means that they are not being drawn from the same giant theoretical cohort because being on blood thinners in the last 40 days is not the same as never being on blood thinners it's just not okay so that's the achilles heel that's where people often go wrong let me give you kind of a classic example of how you might reach the wrong inference if you don't use the right control group so let's say we're interested in gastric cancer and the risk factors for gastric cancer so our cases are easy we're going to start with the outcome right our outcome is gastric cancer cases are easy individuals with gastric cancer and let's say i find them in the hospital they're hospitalized with gastric cancer okay who are my controls well if i pick controls who are hospitalized on the gastroenterology service without gastric cancer well they don't have gastric cancer that makes them a good control and hey they're right next door to the people admitted with gastric cancer that's pretty good alright so maybe that's inadequate control and what i find is that the cases were much more likely to be coffee drinkers so my conclusion is that coffee causes gastric cancer coffee causes stomach cancer so what went wrong here i'm going to tell you that's not true coffee doesn't cause stomach cancer what went wrong here is that i chose the wrong controls the people who are admitted on the gastroenterology service the gi service the stomach service who don't have gastric cancer are admitted for some other reason right maybe they have peptic ulcer disease maybe they have horrible indigestion who knows but they might be less likely to drink coffee on the basis of their other stomach disease okay so we're not actually seeing a causative effect here where i went wrong is that the controls i picked were not from the same theoretical huge cohort that gave rise to the gastric cancer cases what's the huge cohort that gave rise to the gastric cancer cases well presumably you know uh people from the community who um were at risk of gastric cancer for whatever reason right and they they got hospitalized after the gastric cancer diagnosis so my controls probably should have been people from outside the hospital potentially so by picking the wrong controls here i got the wrong conclusions you always have to ask yourself is this coming from the same theoretical cohort that gave rise to my cases if not it's not a good case control study what are the case control pitfalls well the exposures aren't randomized i mean only randomized trials have randomized exposures you hear me hitting this over and over again so confounding those third factors are always present the other thing to remember about case control studies is they are by definition retrospective remember we talked about cohort studies being prospective where you get to like measure whatever you want as the race is going or retrospective which is you're watching a videotape of the race well case control studies are always retrospective because the outcome has to have happened because that's how you define your cases so you really never get to go back in time and measure interesting novel things a lot of times you actually have to ask the patients what they were exposed to right so you take a patient with gastric cancer and you ask you know how much coffee did you drink that is subject to particular bias called recall bias which we'll talk about more in a later in a later lecture but you might imagine that people with gastric cancer might be thinking a lot harder about what they ate and drank and might remember drinking more coffee than people without gastric cancer right and so you you may actually find false associations based on that recall bias in the context of a case control study so what are our take-home points case control studies are useful designs when outcomes are rare or the relevant exposure is unknown they should be thought of as an efficient analysis of a larger theoretical cohort study that you weren't able to do they are necessarily retrospective and they depend entirely on identification of the proper controls that's where people go wrong i'll see you next time

Channel: YaleCourses

Views: 2,755

Keywords:

Id: vcY7ov6I8Ic

Length: 9min 59sec (599 seconds)

Published: Fri Jul 31 2020