my name is laura lee johnson, i'm the associate director for the biometrics 3, u.s. food and drug administration. the disclaimer, first. even though we work for them. i categorize i had a death in my family last week. i just got back today.
i know my slides are late. i know you don't like that, it's totalry my fault. hopefully you can forgive me for the the reason that that happened. if you're pulling the slides from online, it didn't say summary module one.
as i looked at the slides, i realized the title for the slide deck that i had used the -- all the backgrounds for. so this is in fact module one and this is what you should see on your third slide, which is an overview slide. another quick comment that i
wanted to make which is i wasn't able to get om over the weekend to pull up the questions that had not yet been answered in the online forums. so we only get about 2 weeks from the live lecture in order to answer questions so when they get posted later in that
timeframe, we always don't get a chance to log in and actually post an answer to your questions. any of the questions from module one that didn't get answered or you're still a little bit confused by something, post them on to the board for this summary
lecture. and if it's not a question for me but for another lecturer, just note that in your comment. we'll try to tag down the person who gave that lecture and make sure that we get you an answer. so any questions thank didn't get answered, i know there were
a couple in my lectures, at least that that happened. make sure to repost them to this board. so if you'll remember back to the first lecture that i gave, i talked about the fact we have a lot of different people taking this course.
kind of real time, we have over 7,000 people signed up for this course. those of you in the room don't see a lot of folks. for some folks this is a very early introduction, for others, this is a very advanced lecture. so we have to try to hit a lot
of different places. this really is more of a trick tips concepts. if you want to understand, me an epidemiologist, be a statistician, you probably need to take actual course work at a university or someplace like that in order to develop those
skills. what we try to cover is a lot of the types of stuff that didn't get taught in those classes. so general objectives are that in this first part of the course you all become better consumers of the medical and scientific literature.
as you go on and you start to hear the lectures about the institutional review board, the review off human subjects and you start to hear how to apply for grants, you can also bert review that information and get feedback and better write that information as you go on in
yourrish careers. but even if i don't have arish career, all you need to do is read the newspaper, we want you to be more educated reader there also. we're hoping to enhance a conversation inside the research teams, and with the first part
of this first model, also -- module, also with statisticians and epidemiologists. as you move forward in i don't course, this idea of enhancing conversations, it's really important. we hope that people will get this.
realistically, at the end of the day we all want better science. so when dr. gallin started the course, he told you that we had all these different modules. if you look at this, you saw a lot of information. that you all have learned. congratulations.
you've gone pretty far. you have a ways to go, but this is a really tough stuff. even the folks that teach later lectures talk about that. this part is hard but we think it's a fundamental aspect that will help you better understand a lot of the questions and
problems that come up in the other lectures. so woe talked a little bit about these clinical designs, how you choose a research question. one of the most important elements that comes up. what is your question. is something that i give a lot
of talks about. i sometimes want to shake my investigators on. and i ask myself that. you have this whole conversation what at the end of the day is the question you're trying to answer. we did overview of clinical
study designs, then went into these observational or epidemiologic study designs. we talked and you had a patient come in to talk to you about the patient perspective. this is a bigger push, to have that patient voice in the development, not only the
research questions and. trials, but also thinking about your endpoints which is something talked about the last few weeks. so kate stoney came in and talked about study participant selection. why do you decide to put certain people in your trial or not in
your trial? and that's true whether it's a randomized study or observational study. paul wakim talked more about issues and randomization. and gave you a brief over view of hypothesis testing. again, there are many years of
lects that could be given on how to do hypothesis testing. this was a general overview. we came in and talked about somesize and power, both for anything from a phase one trial to really large studies. you're going to hear more about that toward the end of the
course when we start talking about more community participatory research and some of those larger studies that happen in real world settings. we went on and talked about survival analysis. again, something you can spend years of your life on.
but we wanted you to be able to read the medical literature better and sometimes when you see those capon meyer curves, realize many that's not the analysis that should have been done or maybe it's perfect. i can do it on the napkin on top of that.
then we spent a lot of time talking about measures. so dave talked a little bit about the validity and reexhibit issues, how -- reliability issues, so important trying to choose what to measure and how it impacts other questions, your ability to do the hypothesis
testing, how big, small your sample size may need to be. then kevin came, talked about quality of life and patient reported outcomes. and other topics like that. barbara came in to talk about designing and testing questionnaires and there is
proposally a lot of -- probably a lot of information that seemed similar there, and it should be. whether you're talking about a survey, a patient reported outcome, clinician reported, anything like that is good to -- good development is pretty much the same.
then last week you heard from layton chan about using these large data sets for population health research. he does work with what's called secondary data. in the united states people may use medicare claims and information like that.
this is also work done in various parts of the world using different but similar datasets. and then the last lecture you had was chuck natanson talking about meta-analysis, a different type of secondary data analysis, this time using clinical trials. so what i did for the summary
lecture, usually i pull together a lot of information on my own. given that richness, daniel and i e-mailed all the other lecturers and said what should the students know. give me 2 or of slide summary, what you taught. if you had a chance to go back
and tell them something again what would you tell them? so i made a couple of edits on here but in general, these are actually their slides that they sent me. although i reordered them a little bit to tell the story. so one of the places we started
is that it's really easy to write that you're going to use a randomized double blind parallel arm design in intent to treat analysis. hopefully compared to october, you know what each of those words mean. it's always easy to say that
subjects and participants will be concepted. as you go through the next few months and modules of this course you'll realize it's really hard to adequately corn sent to people. consent to people. so it's really not easy to
implement and maintain the integrity of your randomization. dr. wakim talked about that quite a bit. to maintain the integrity of your blinding or masking. i have a study where because of the substance that they're taking in one of the study arms,
one of their basic blood values always change, their creatinine changes. every time you get blood work they're checking your creatinine. how do you blind that study? when people have very different side effect profiles, depending
on the therapies they're talking, how do you blind the study? so we talked about various different tricks that can be used. and your textbook talks about some of those, too. maybe you're going to have
assessors that are blinded or other folks that are blinded. it's hard to maintain your multiple study arms, especially let's say you're giving a psychological based therapy and control group, and there are some studies they look back and look at videos as a control arm
therapy session, and the new intervention arm therapy sessions, realize those are starting to look an awful lot alike. how do you maintain the integrity of your [indiscernible] over time. integrity of data collection,
when your power goes out and you don't have an internet connection? how are you going to do? there are a lot of issues that come up. trying to figure out how you have all these different fail safes for problems that can
occur. then trying to transfer data from -- to the regulatory and other groups that will need it. did you remember in the consent form to tell people that those regulatory agencies may see their data? that is more common than you may
think. the regulatory agency requests and requires in many cases that primary data. but you didn't tell people in the consent form you were going to send it or allow these regulatory agencies to look at a so what happens?
in many cases you have to reconsent everybody. but the real fundamental element here that i think almost all of us have talked about is how good is your primary research question? at the end of the day, you do all this research, data is
analyzed, well, the answer regardless of what it is, to the primary research question actually advanced scientific knowledge or change clinical practice. now, i'm sure that chuck talked a lot about the fact that don't let one study change everything
that you do in most cases, because that usually just leads to bad things. but sometimes we choose not a big enough or important enough question. we're so focused on what we want our own research to go along, we don't realize maybe it's time to
change course. so as a reviewer, one of the most important things you can do is tell people no. you don't want to just say no, fo, no, no, no. sometimes you say, you know, it's not that this is an incremental step that must
happen in order to take the next big step. sometimes that's the case. but nobody cares about the results of your study. you don't want to say that because your best friend is doing something in competition, but sometimes you look at these
studies, you're like you should just stop. and you have too much the guts to tell people -- to have the guts to tell people that that's the case. so good primary research question is answering a question that people care about.
not just you in your study team. so also, don't forget, we have pico, different ways that they look at it, but it's basically any time you try to describe a study, you want to list the population, or the disease, the intervention or variable of interest, the comparison group
that you're going to be describing, what is the outcome, and then you should get some measure of time. so an example of this, like in this population, how does this intervention compared to the control influence the outcome during some time period?
now, maybe it's not an intervention. maybe you say in people with secural cell disease, how -- sickle cell disease, how does male compare to being female influence this outcome during this team period? so this works, whether it's
randomized, non randomized, observational, et cetera study. don't need an intervention here. an example from still well is in patients 65 and older, that population. how does the use of an influenza vaccine, compared to not receiving the vaccine, that's
your control. influence their risk of developing pneumonia, how are we going to measure that outcome, an important thing to understand in your study during the flu season. this is something you need to define better, northern helm
sphere, southern helm sphere, who is defining the flu season. then you can drill down the little bits. so this gets back to this question, people get confused about what the controls and interventions. the specific question being
addressed in the study leads to the choice of your control group or groups. sometimes you need more than one, for the study. you have to make sure, when you're comparing people in the intervention group, in the control group, or in your case
and in your controls, that there is something that you can define that is different about them. maybe it's the fact that these folks went to yoga crass 6 days -- class 6 days a week for an hour. other folks just read a book 6 days a week for an hour.
there is something that has to define them as being a case or being in the intervention group verses being a control. many times it's a lot easier to decide what your interventions are supposed to be. people like oh, i spent all this time writing these manuals,
developing this drug, doing all this stuff. it's a lot easier to do that compared to figuring out what a good control is. you look back at those recollecter notes, you'll see -- at those lecture notes, you'll see are you controlling for
time, attention, the number of pills people have to swallow, how frequently they swallow? are you controlling for ingestion of something? if you're -- if you have an iv verses a pill, is it just really that method that you're administering the same drug?
what is it that you're trying to do? figuring out your control arms, non drug, non biological studies is really hard. but it's really important to make sure you can do your analysis at the end of the day. when in doubt, write up your
table. write up a mock article about the study. write your manuscript. make sure you can say and analyze everything you want to do before you enroll your first patient. or rather check it before unroll
your first patient. now you have a nice template that you can change when you're done with your study. so i talked about effect modification. this is a nice way of saying interaction or synergy. it could be larger or smaller,
so this is when the association between your outcome and some other variable, maybe intervention, maybe something else, is modified by different levels of a third variable. the exwe talked about was smoking asbestos in lung cancer. smoking increases your risk of
lung cancer by a certain amount, asbestos increases your risk of lung cancer by a certain amount. if you're both a smoker and have asbestos exposure, it's foot an additive amount. not an additive amount. so don't get effect modification confused with confounding.
so i think probably dr. stoney talked about confounding to say one reason you may have certain exclusion criteria is to reduce the confounding in the study. so confounding, you again have these 2 or more variables. you're lucky when they're known. unlucky when they're unknown in
and you have confounding when their effects on the common response variable or outcome are mixed together. so we talked about the coffee and the smoking, pancreatic cancer example. i think the simplist example is the match example, which we'll
go to on the next side. the idea in confounding, the outcome and exporer is misestimated. in effect modification, i may have different rates if it was going to be i was a smoker but didn't have asbestos exposure, i'd -- verses i wasn't a smoker,
verses i had both, asbestos and smoking. it's not additive, there is something extra that happens. in confounding i have a misestimate. very simple. so the problem is, and this is kind of the easy one to
remember, i've seen association between carrying matches in your pocket and lung cancer. but really, the caring the matches in your pocket that causes the lung cancer is because it's confounded by some other unmeasured variable, which is basically if i'm carrying
matches in my pocket, more likely i'm smoking cigarettes or cigars or something else, so i'm using the matches to light it. it's that substance that is actually leading to the lung. lung cancer. so remember, good primary outcome measures, any measure,
needs to be clinically meaningful and simple. you don't want 20 melan clauses to understand your measure. sometimes just saying matches in your pocket might be too simple of a measure. another one that used to come up, do they have yellowing on
their fingers? because that was from holding the cigarette so they had the tobacco left on it. but again, that's not what was actually the problem. so you have at this different types of research studies. you've got the observational
studies, where i'm going to observe, collect data on all these different characteristics of interest, but i'm not supposed to be influencing the environment, the disease course, the participant. i'm merely supposed to observe. then i have experimental
studies. here i'm deliberating trying to influence the course of events. i want to change something here to change the course of disease, to change what have you. carefully select my population of subjects. do that in both types of
but when we do these experimental trials ophumans, we call them clinical trials or clinical studies. some people will say clinical studies cover everything. clinical trials have to be randomized. kind of loosey goesy.
nih has very specific definitions, so do other bodies. observational studies. we talked about the case report and case series, remember, the hiv case series? that is a case series, we're writing up basically a set of patient reports, but you're
going to comment on the same characteristics, mesh irrelevanted in the same way -- measured in the same way, et cetera. we talked about cross sectional or prevalence surveys. i'm going to do a snapshot and ask people a series of question,
hopefully gone through by [indiscernible] so she made sure that everybody understands them. in a case control study, so case control study, you might collect a group of cases, let's say those women that had that very rare vaginal cancer, and set of controls.
people that don't have that vaginal cancer. so when you're choosing cases and controls it tends to be based on some disease. then you look through to see what characteristics are the same or different. between these groups of people.
in their case it was a drug their mothers took during gestation. a cohort study is kind of like those prevalent surveys, but now i'm doing it longitudenily. the cohort studies are movies. i'm going to keep asking the same question and doing the same
measures over and over and over. the cardiovascular health study. every 6 months people went through a battery of measurements. they had imaging, blood drawn, asked a bunch of questions. so they had some things done every 12 months, other things
done every 6 months. they did that for a decade. that's a longitudinal study. natural history studies. that's something that the nih clinical center does quite a bit of. you'll see this a lot in rare diseases.
some ways, kind of ties up a little bit crosser to thosecacy reports and case series. here what we're doing, it is a lot of times like these longitudinal cohort studies. we're writing up the natural history of disease. we're not quite sure what will
come yet. i may not have a fixed set of measures quite yet. i may me adding to those overtime. -- may be adding to those over time. then we have the ecological studies, when i'm going to take nation level data.
data on the population. those summaries instead of individual level data. some of the maternal health information you see are ecological studies, not individual level studies. if you're going to look at, say, death rates.
women during labor. or due to labor. so what is epidemiology? it assumes that disease has some causal preventive factors that can be identified through a systematic investigation. that's the basic definition. then you move into these quasi
experimental single arm not randomized experimental studies. so first i observe everybody. then start to intervene. but maybe i don't really have a control. might be early in the investigation. may be a rare disease.
can be lots of different reasons that i do this type of study. but it's pretty low level evidence here, right? sometimes folks say they're going to have a type of concurrent control group. so they're just going to choose one, say, side of a hospital
corridor, and treat all those patients in one way. people in rooms on the other side of the corridor, they're going to treat them in a different way. maybe that's whether they're giving them baths on a daily basis, something like that.
they're not randomizing it, they're just assigning it. at least you have some type of control but it's not a very good so you have what are called historically controlled studies. where you may have a lot of missing data, poor data, non comparability again.
historically controlled studies might be patients from the past, they could be the control arm from a previously randomized study. but somehow you have patients, you have data that you're recycling from one study or from hospital records to another.
you tend to see this a lot in your rare disease or paediatric trial. but any time you have to start cutting corners, urealize you could be getting the wrong information. so in this intervention based spectrum, you have these quasi
experimental studies and preclinical studies. that's like your low level it's useful for something, but you want to move forward. phase 1 tends to be that early dose toxicity type of phase 2 is when you start moving more toward efficacy.
although you're keeping an eye on safety. now you're exposing more people to this experimental therapy. and current stage -- i'm sorry, phase 2 steins, we tend to have control groups. in the past they were less frequent.
now we have control groups in a lot of phase one studies. but in phase 2, you're not just looking for such a bad toxicity that you stop. some of those studies have these maximum tolerated dose. phase 2 we've chose many 2 to 3 doses or a few more that we're
testing in larger groups of people. phase 3, supposed to be your definitive study. you better figure out everything you wanted to know in phase 2. because phase 3 is basically supposed to be confirming all that information you figured out
in phase 2. do i have the right dose, seeing the same impacts? sometimes in phase 2 woe may stop or do our study with an outcome more of a bio marker. so it's kind of -- our surrogate endpoint. it's part way there.
phase 3. might go for vile or long term end -- survival or long term phase 4, is more post marketing studies or something is releaseed to the public. see how it's doing. and a much broader group of so how does it behalf in the
real world? also your long term safety so what will happen in all those phases earlier, is people become very pristine. my opinion, many times, they're way too strict. so you are choosing it for safety, choosing it for a lot of
other reasons, but phase 3 even they made a -- not have anybody with any co-morbid conditions in you can only have the disease of interest. you're trying to study something like hypertension, that's a pretty rare group of people to have.
so part of us tend to push, phase, 3 you need a wider spectrum of people. even in phase 2. to figure out how are these drugs behaving. but in phase 4 is where you really figure it out. who is being prescribed
medication and how do they behalf be-- behave? toward the end of the course we'll talk about dissemination and implementation and comparative studies. those tend to happen in phase 4 but if you're smart you're thinking about these studies
early on and how to implement them. you don't want to develop a therapy that can never be used. you have to think about how phones going to be trained to perform this surgical procedure? can i produce more than 2 devices a year?
does it use such rare minerals that you can't do that? so you have to think about the production aspects for whatever your intervention is, even if it's training people on a certain type of yoga or meditation, how are you going to train a lot of people on that if
it does great things? but when you're trying to decide between a non randomized or randomized study, you need to remember, non randomized studies can only show association and you're never going to know all those possible confounders we talked about.
kind of easy to figure out you left out smoking, that's a rather but easy example. in a randomized study, you can show association and causality, if you have a well done non adaptive randomization. because the unknown confounders should not create problems.
remember, however, it doesn't mean that they won't create problems, it means they should not create problems. the reason i say non adaptive randomization here, is that many of those adaptive randomizations and adaptive designs rely upon your actually knowing ha those
confounders are, and how they're going to be behaving. otherwise, it screws up your actual randomization or allocation to the study arms. so whenever people are talking about studies, they have this gold standard, you have this treatment controlled trial with
parallel arms. there is a superiority hypothesis, it's prospectively designed so you're collecting your data over time, looking toward the future, it's double blind or masked. and it's randomized. so we went through a whole set
of different randomized studies. we talked about each of them with some examples. i want to talk a little bit more about the adaptive designs because they're so popular. but first i'm going to talk about intent to treat. no matter what study design
you're doing, you have to figure out who your anonymousing. intent to treat is mainly randomized studies but it applies to -- at what point in time is somebody included in the analysis? you have to define this in any plan you work on.
if you have a randomized study, though, the idea is that once randomized, always analyzed. you assume that they adhered to the study regimen you assigned them to. you assume they completed the even if you know the pharmacy mailed them the wrong study
drug. and by the way, they dropped out after 6 days. why? because that's real life. so if you're trying to think about how things will happen in real life, in real life, people do not do what you tell them to.
in real life, mistakes happen. in clinical trial you're probably in the most press teen environment ever. so if you can't find a difference there, it might be that you'll see one in real life, but the chances are probably pretty rare.
where it did happen, actually, was in some acupuncture studies in germany. all their really carefully run phase 3 trials were not showing many differences, when they actually did a large scale effectiveness study across the country, they did see
differences. so again you never quite know. but in general, i.t.t. is what you're supposed to do. this modified intent to treat analysis, m.i.t.t., sometimes it's brought up and they say we're only going to include people that started the
intervention they're assigned to. typically your statisticians will get jumpy when you start saying this. some folks say you can only do this when it's double blind. other folks say you should never do this at all.
because folks may go just because you think it's double blind doesn't mean someone is not gone and figure out what drug they're talking. they know they're in oga, but they wanted to take something else. you have to figure out if
they're dropping out for other reasons. completers or adherers analysis. these are proposed, especially for mechanistic studies. you have to define whether they compete or adhere, 70% of your classes? taking 50% or more of your
medication? that felling out your daily diary, 6 or 7 days every week, what is it? but then you're only studying the well behaved. and anyone who is a clinician or if you're honest about you behalf, probably knows that most
people are not so well behaved. so what's with the masking and blinding? well, here, it's less common in non randomized studies, you still sometimes non randomized studies want to mask the outcome assessors to the hypothesis. that way they're not asking the
questions of the cases, a lot more forcefully than they're asking the controls in that case control study you're running. but in any study you have to specify who is going to be masked? why are they masked? how are you accomplishing this?
and what are they blind to? what information are they not supposed to know? somebody has to know something. it's just you may not let anybody know all the you need to assess the effectiveness of masking. no, you consort took that out of
their diagram a few years ago. but realistically, it's early in your study, toward the end of the study, to assess the effectiveness of your masking. if you find out that everybody is unblind, then you need to own up to that and it might be impacting the results of your
you need to specify the criteria for unmasking. the people that will be unmasked. so let's say i'm in a drug and i need to have emergency surgery. it's a double blind trial. well, the surgeon, the
apthesologist, they may need to know what drug i'm on. i don't need to know. probably a lot of people involved in the trial don't need to know either, although the people that are doing the safety assessments and dsmb, the data safety monering board that will
be talked about in a couple months, they will probably want to know. so who knows what? how do we determine that they need to know it. >> is going to tell them? that all needs to be laid out in your protocol.
you want to mask the determination of the outcome typically for your studies so that reviewers are unaware of the treatment assignment. again, everything here is like playing secret agent. it's all need to know basis in a but in the end, somebody needs
so what is being adapted in those adaptive trials? sometimes you have adaptive randomization, sometimes it's adaptive dose finding. sometimes you drop the loser or pick the winner out of those different doses. sometimes you do an adaptive
seamless phase 2, phase 3 trial. sometimes doing bio marker adaptive. depending how you show up with this panel of biomarkers, i decide what treatment or what study arms you're allowed to go into. so ispy, ispy two and other
studies have done things like. this groups sequential methods, adaptive design. remember when i talked about taking peeks at the data, how youd to change your type one error for the end of the trial? it's part of what goes into your sample size calculations?
that's a randomized study design, a really good one. commonly used in clinical research. sometimes we're doing sample size recalculation. so we go in, we reestimate the variants. across the entire study
population for that outcome. and then we update our sample size. so depending on what's being adapted, depends on what you want to do and the problems you're going to have. but you need to think about, for all these adaptations, the
reproducibility. throughout everything in your study design you need a well detained cohort. would someone else make the same decision to include or exclude that same person? if you're hit by a bus, the next person running that study, going
to run it the same way. suddenly your yoga teacher's mother is sick and she -- you need someone to replace her to keep running those classes. is that person going to run the class the same way? you're drug manufacturer has to shut down suddenly.
somebody else going to be able to produce the drug the same way? or the control substance? the outcomes. you have to be really, really precise. i know that deaf talked about how you're going to measure the
outcome. not only is it a good measure, you've got to make sure if you have 6 different people that are running and taking those measurements, that they would take them the same way. if you're going to do blood pressure, for example, they're
different cuff sizes, different ways you can be laying or standing. all of this information impacts that outcome. are you taking it once, 3 times, averaging it together? are you dropping the lowest one? the highest one?
what are you going to do? the study data, data analysis. is everything written down in detail. what vari can't believes? -- variables, and the exact code they need to type in to make the false run? there is a -- analyst run?
there is a lot of places where this falls apart. sometimes you may give fancy laboratory name and folks say we know what that is, and one person will let it incubate for 30 minutes, another person lets it incubate for an hour. they use slightly different
temperatures, the same procedure. you now have very different study results. you need to be precise about everything. it is painful. but i'm serious. i shouldn't be able to reone
your statistical analysis or create a database for you. i should be able to run your entire study. i'm not a phlebotomist. i do not know how to use a machine, but seriously, you want to know like a high schooler coming in off the street could
run your study, pretty much, if they had the correct cendals. -- credentials. you want anybody to be able to do it. that's reproducibility of also a lot of things that happen-- biases that happen in clinical research. we have supposed remedies for
these. there is the selection and assignment bias that several of us talk about. we try to do randomization if we possibly can in to order to help compensate for that. these treatment and assessment biases.
we try to mask the research team to what that individual subject and study is taking. we have a response bias. we try to mask the participant. so they don't say, well, i know i was getting this, so i'll talk-- my pain feels much better. problem is, that sometimes
people feel guilty because those nice people have been trying to so hard, so they start saying they're feeling better anyway. sometimes it's just placebo response. so their biases that happen during that data cleaning. sometimes we still need to mask
that assigned treatment and all that prespecified information. these are the rules, when i'm going to check certain data points. these are the rules how i'm going to decide that that is an uninterpretable squiggle or interpretable squiggle.
because sometimes people say is that 180 or 130? they were in the new study remember a. pause they're on the new treatment, i bet that's a 130, not a 180. you don't want that tore your systolic blood pressure, for
example. during the analysis, the reason you want intent to treat analysis, and that prespecification of all those analyses, you don't want people to pick up biases and go hunting for the answer that they want to be true.
you kept looking. once you found what you were expecting to find you stopped looking, right? no, that's not what you want to do. wedite all the time but it's wrong. publication bias.
reason we have all these registries for trials around the world, w.h.o., there is one here, national library of medicine, clinical trials.gov. there are tons of them around. the reason that we worry is because we know, and i'm sure dr. natanson talked about that
last night. a lot of these studies never get published. then you find and abstract that someone gave. or maybe the animal data fills in the missing information. a rot of these studies are not in these trial registries.
then we have a reporting bias. you measured 40 different hormones, various other things in the blood. your paper talks about ten. you have to have prespecification and consist closure. people get worried because that
you're afraid you'll only report what looks significant, and not the whole story. so remember, we have all those folks with the disease condition or disorder, then you've got the peoplenisted in participating in your trial. then the people who meet your
inclusion, exclusion criteria, then those who consent, then the people you randomize. this is another part of bias, too. we may hold up randomized trial as that cold standard, you're still not -- gold standard, you're not always representative
of the full population. a little bit more about that randomization. what did dr. wakim cover? talked about what does the word random actually mean? ha are we talking about. what is randomization? why you randomize.
who and what to randomize, and how to randomize. he made a lot of recommendations. use the computer program or online tools. i actually like drawing numbers out of a hat. it's not a very good form of
randomize station. use an online stool. online tool. use small random block sizes, blocks of 8 and 10. or 3 and 9. let's say 6 and 9. do something like that. for mutch -- multi site clinical
trial, use site as a stratification variable. when you stratify, needs to be in your statistical model. do not use too many stratification variables. you don't want to stratify your study and randomization in particular on the biomarker, and
the site, you won't end up with enough people in each of those little holes. unless necessary, unless you have the resources to do it adequately, avoid adaptive randomization methods. are they more popular, more common?
yes. are they a lot easier to screw up, the answer is also yes. there are some institutions where people will do them very well, but some of the better field in the field will say if you don't have the support and don't know enough about the
population you're working in, don't go here. so let's say, though, you are going to do a complex randomization method. get ready to hire bio statisticianition who remembers how to do them. biostatisticians are like
doctors. they have a general knowledge. that doesn't mean that they specialize in certain things. so they may have some information, they may be able to read the information faster. but you don't want to necessarily ask a
gastroenterologist to deliver your baby. they may have done it a couple of times. but especially if they were in medical school and did their residency 20 or 30 years ago, it's probably not the right person unless you don't have a
choice. so you've got to think about -- there are some folks that are -- we call them jack of all strides here. they do a lot of everything. they might be able to know it or know who to contact. some of the stuff is really
scary, very quickly. so it might be easy, and uncomplicate, and no problem. but that's not always the case. so you have to be able to understand when you're getting in trouble, that's the hardest thing and true for everything we do in all our professions.
so what are the implement recommendations for randomization? we talk about this in the textbook. you've got to make it possible to reproduce the strain of treatment assignments. need to make sure you're not
cooking the books, is what number one is saying. when i do an audit off your number two. part of, again, making sure books weren't cooked. you have to be documenting the randomization method that you i don't want to talk to you
after and you tell me you continue know if i did a stratified randomization or not. i did that once. it was not pretty. put in place features that prevent the treatment assignment until conditions for entry in the trial or fully satisfied.
so i had a study ones, very illustrious group. they were not confirming inclusion exclusion until a month people were randomized. they labeled intent to treat you just under mined the entire you can't do this comparison. someone is making a face in the
rome. my face is not very pretty. they fought and fought and fought about this. we always do it this way. well, you're wrong. make sure that someone cannot be assigned, cannot beramized, until you're sure they're even
eligible for your study in the first place. you want to blind assignments to every one concerned. randomization and blinding are two different elements. they go hand in hand in helping to maintain the study integrity. you want to make it difficult,
if not impossible, to predict future assignments from past assignments. that's part of the point of it's not going abababab. the reason that you have multiple block sizes is that people can't figure out what the next assignment is going to be.
they don't figure out where they want to put their aunt flo into and you want to put in place procedures for monitoring the departures from established protocols. if someone got assigned at the wrong time, to the wrong arm, all of that, you need a
procedure to recognize that and to document that. so we had 3 different lectures on measurement and a few others that touched on the topic. they sent me a handful of slides. his big focus was reliability and validity, right?
so what are the problems in reliability? you have a lack for reliability introduces a lot of error into the measurement. a lot less sensitive statistics, need a much larger sample size. if the worse thing happens you have completely uninterpretable
results. so your minimum sample size you need, regardless of everything else, based on the reliability coefficient. so it showed that during his main lecture. how do you improve reliability? of your measures?
you provide the standardized procedures. i don't say it because i'm mean. i say it because if you really want to get good reliable data, you have really strict, well documented, everybody is trained on them standardized procedures. i mention the training and
retraining on a regular basis. because sitting on a book shelf or sitting in a file on your computer does not help produce standardization. you have to have standardized procedures but then people have to follow them in a standardized manner.
and if it's not going to work, you want to know that soon. kate stoney gave a lecture. what she was famous for was making her entire study staff run her through the study. she was patient zero of every study when she was a researcher? one, i should know what we're
doing the best. then she made other study [indiscernible] made in questions, not to upset the consent people. everybody is going to run through it. one, you need to understand what's going to be ton.
any problems or inconsistencies, we're going to figure this out now. that's part of the standardization, the training and fixing when there is a problem. people can't follow your procedures?
that's a problem. you train the raters, you monitor your raters. anybody taking a measurement, anybody who is running people through those study visits, what is happening. using multiple raters for each rating.
it might be as simple as taking their blood pressure 3 times. so i have this one assessment, they're trying to decide how smoothly somebody walks over, i think it's a ten meater space. meter space. do you have one gait specialist, 2 or 3 gait specialists? do you
video tape the walking, have that spent for adjudication? or every when out of ten videos have somebody else look at it to see if they agree with the assessor? you immediate to think of the ways to try to improve the reliability of your measurement.
you also want to take repeated observations. so this is a mixture of all of this that leads to improving your rye leability. -- reliability. people say that takes a lot of time, money, staff, or your study could be a bust.
decide which one wastes more money of your patient team. sensitivity to change, super the ability to detect improvement or worsening. you can assess sensitivity to change with effect size. remember, i personally hate effect size.
owen, haase this in his book, sis he doesn't like effect size ether. if you don't have the variance, you're not sure the difference, people think about this. but large effect size can be very different things depending on your variants.
usually you want to understand what are the two means and what is that standard deviation, understand that separately. but the general idea that if there is not very big of a difference or swamped out by the variance, you're not very sensitive to change.
you also need to think about the clinical relevance. i think both dave and paul talked about this a bit. definitely discussed he be your because it's such an important topic, i can promise you this type of information will be on your exam.
we talk about sensitivity and specificity. when people develop a new test, sensitivity, you have the listens, how often is the test positive. specificity, if you have no illness. how often is the test negative?
these are used because they don't depend on disease prevalence. problem is if you're doctor or a patient, all you know is you tested positive or negative. while sensitivity and specificity are great, that's mote the actual question of
interest for most folks. what most of us care about are positive or negative congestive values. if you test positive, how often do you have the illness? if you test nih, how of do you have no illness. if i test positive for hiv,
what's the probability i actually have hiv? that is the positive predictive value. however, the formula for that will say well, if you're in a high risk population, then you will have a higher positive predictive value.
let's say that your women in the united states that donate blood where the probability of being hiv, the prevalence of being positive is .01%, even the best hiv test we have has really bad positive predictive values. so because of that, before they tell you you're positive, they
split samples. and they do -- if you show up positive on the cheap fast test, then they do the more expensive longer test to make sure it's a true positive or was it a false same thing a negative predictive there are different risks for each of these.
but it's really important if you're trying to roll a test out into the population, you understand positive and negative predictive values. you think you're pregnant, you want to know, when you bed on that stick, are you -- peed on that tech, are you positive or
negative? that's what you really want to know. so instead of laboratory tests let's talk about self-report measures. several of the speakers talked about these. they may be used in
questionnaires. not easily observable. now, some folks say it's self-report measures are sometimes of true scores. because they say well, there is susceptible respondents mood, their motivation, memory, understanding, context,
collecting the data in. am i interviewing you? or are you writing it down on your own? social desirability is a big issue. so that gets back to that pain question that i had. or do i want to tell you that
i'm pregnant and i had three drinks? of alcohol? probably not. i know i'm not supposed to do that. so do i lie? do i tell the truth? problem is people say
self-report measures are just the answers are only figured out if you ask people. if you use a rigorous method, you can mitigate a lot of these pit falls. mote everything can be measured in the blood of urine. those tests many times have just
as many problems. so kevin talked about patient reported outcomes or pros. sometimes certain parts of the world they're called prompts instead of pros. proms. according to the fda guidance from december, 2009 these are
defined as a measurement based on a report that comes directly from the patient about the status of the patient's health condition without being amended or interpreted, by clinician or anyone else. so a p ro is not asking mom about their baby's pain.
pro is not having a conversation with your doctor and your doctor writing down about your pain. pro is you directly responding about your pain. he also went through this talk about this balance that about the delusion of effects of the biological intervention and the
correlation that can happen. if you have some characteristics of individuals, some characteristics of the environment, then some biological and physician logical variablens. so you end up with the symptom status.
you end up with a functional how well are they moving, things like that? they're general health perceptions. sometimes we talk about this overall quality of life. which also is these non medical factors.
there may be financial factors, a lot of other factors going in but problem is, also, like how distal are you measuring from what you actually care about? from the lecture after that we talked about putting together draft questionnaires, the focus trying to use existing, i would
say good instruments when possible. existing relevant instruments when possible. so barbara mentioned health measures.net which has tools developed by the national institutes of healths. med cap has a data management
system online and electronic, but also a library of a lot of common case report forms and measurement tools. sometimes people actually copy questionnaires from national surveys. in the uk you have the national health survey.
in the united states, nhis. so some of these studies will copy the exes and questions and responses and use those. that can at times be useful because then you can compare your study subjects to the general population. but it's good and there is
better and better international agreement to try to have these shared libraries of data collection instrument. because them it will make chuck's job doing these meta-analysis a lot easier. so what are the criteria for good survey questions?
i'll tell you, these apply to whether it's a patient, clippings or any other -- clinician, or any other type of reported outcome. when you're trying to figure out any way that you're collecting data. in the united states we tend to
say you want literacy belo9th grade, some say below the 6th grade reading level. stepping on the countries -- depending on the countries you work in, you may need to lower this limit. a lot of times we'll say let's do paediatrics first, because a
60 year old will understand what a 6-year old understands. i call it my sesame street principle. specific questions are better than broad questions. people are going to be less likely to misinterpret it. you want to be culturally
sensitive. a huge issue. you need to test in different curb -- cultures and different languages to figure out people will interpret your question and the answers the same way and to make sure you're not being insulting.
this happened with one questionnaire i was given and asked to deploy and inner city, philadelphia, pennslyvania. it was asking about out house usage. going to a separate facilitate to use the bathroom. out houses are not used anywhere
that i've seen in inner city, philadelphia. people were unhappy about me asking that question and i couldn't blame them. it was a bad questionnaire. are your scales consistent? are the terms well defined? do i understand what an outhouse
even is? instructions. are they clear? the reference period is clear. you say in the last month, in some countries, people think back just to the first of the month. they don't think back the last
four weeks. the response options, do they match the question? is your question asking about frequency and your response options are all about severity? not good. also the response options that overlap, also not g multiple
concepts separated? we sometimes call these double barreled questions. all on top of each other. make sure you're separating out all the concepts. you need to make sure that your questions will be interpreted accurately by people with a
range of different demographic characteristics. it goes for literacy, it goes for culture, it goes for age. goes for different regions inside the same country. male or female, et cetera. you also need make sure that it's capturing what the
researcher intended. one time i had -- i saw a bunch of qualitative results. every one was uniform in what they thought the questions and responses meant. it had nothing to do with what the researcher thought the question and responses meant.
so you passed on one part but failed to answer the research question so you have to start over again. as you're trying to develop these questions you want to try to avoid the socializerability effects, negative wording, you also don't want to flip flop
back and forth, negative and doesn't help, just confuses old school said you had to do that, new school says why are you confusing? that double barreled type of jargon. i may say something here. some folks that are listening
that have perfectly good english in egypt have no idea what i mean. used to joke, there is one of those online i will interpret your language. i was back in texas this weekend, and it didn't understand my cousin who haase a
very strong north texas accent. so you have to pick up a lot of different elements. you don't want ambiguous be precise. you don't want leading questions so if you're going to ask me about how much i drank, make sure someone is not going to be
judging me. and you also don't want to say understanding is that you shouldn't be drinking during pregnancy, how much alcohol did you have this weekend? you have to be careful about these things. what about the developmentvation
of pro measures? again, kevin was talking about what is the concept you want to measure and why? it's a different variation opwhat is your question. you have to collect that qualitative data to understand the meaning of the pro concept.
just like a survey, right? you have to write items that you think will pressure the concept. measure the concept. you have the test the items for understanding, cognitive interviews. administrator the items to a large sample of people.
then use something called psycho metric analyses, a form of statistical analyses to see how well the items are working. you've got to develop a scoring method. trick. do not just assign them all, 1, 2, 3, 4, like 0 through 4 and
assume the spacing between 1 and 1, 1 and 2, et cetera, are the same. a lot of times you have to have waiting in the scores. you've got to veil the reliability and validity as a measure which leads you back to dave's talk.
barbara wanted me 250 remind you all for translations, so promis has a paper on this but international society has a paper on this, they all pretty much come along the same way. the idea being you need to harmonize, use different words and languages, you have to still
have that same meaning. even though there are a lot of different idioms, i speak spanish and english. you can come up with a lot of different phrases. there is what we call the television english or television spanish that the news casts use.
everybody understands that what they mean. they don't confuse a baby with a bus. baby carriage with a bus. so there are certain words thank every one will know what you even though they may not be the most commonly used word inside
that language. you have to develop this harmonization, because you don't want to have -- if you can avoid it at all, different questions in england, the united states, australia action and new zealand. nor do you want them in spanish
versions used in the u.s., argentina and spain. you want to have as much harmonization in this universal approach as possible. when you have people from various countries, different dialects, when i was in south africa, there were 13 or 14
official languages. what is it that every one will understand? talking about item banks,s instead of giving everybody the exact same questions you're trying to measure their physical function, you may have a lawfirm collection of items that --
large collection of items that measure that single concept or domain. any and all items can be used to provide score for that domain. and what's nice about an item bank, it's dynamic, not fixed. so if you have patients that during the course of your study
are going to be able to not get themselves out of bed, that later will be walking around, doing a lot of things, you may want to use an item bank instead of a fixed questionnaire. because you use a fixed questionnaire to cover that broad range of physical
function, you're asking a lot of if you want to have any sensitivity. so he showed you this example that was put together of physical functioning item bank. people in folks to folks do you go the tour defrance, the running parathons, et cetera.
so sometimes in the military, it will matter whether you can run 26 miles or run 5 miles. for most of us it's just a question, can you run or only walk? so it depends where you're doing your research, what your population is.
how much of this scale you care about. problem is, some of the tools, some of these pros that you might use, only measure on very small portions of the scale. it could be very useful to use something very broad in order to also be able to compare all the
studies to each other over time. so traditional classical test theory based pro measure, you need all the items in order to compute a score. use an item bank, any or all subsets of items can create a score. this is nice if you think about
patient burden, they can answer 1-3 items, you have a score. every one has to take the same number of items in that traditional measurement scheme, different people can get different items. traditional, off the shelf, item banks, you need to create a
measure for specific use. so you can still create a fixed number, if you want. traditional. scores are not easily comparable. and itemback, it's easier to do crosswalks. so maybe there are 6 different
ways to ask about that running. you can do that in crosswalk and still compare scores. i think also talked bow differential item functioning. this comes up a lot in the scores of the patient reported outcomes. an item behaves differently for
2 or more groups. might be language, sex, age, a lot of different issues. so let's say that map between depression and crying item is different. this is the classic way to talk about what is called in the past 7 days did you cry?
how you actually score and having a depressed mood, depending on how you answer that question, for males verses females is very different. women have a tendency to cry more. regardless of depressed mood. the probability you cried in the
past 7 days that it's yes, to get the same level of depression, you have to be at 90% here. whereas the men, 30%. basically, as soon as the man says that he cried in the past 7 days you start trigger that this person might have depression.
not true for all men or women, but we're talking about these population issues, you develop these curves using a lot of so typically, we would throw that question out. or we would score it differently for men and women. so in summary, teal with
measurement, you have this questionnaire development requires a lot of careful planning. you have existing validated instruments when possible, use let somebody else do all that work. unless that is all the work you
want to do. rigorous methods reduce response error. pays to be rigorous, here. things that can really screw up your study result measurement, huge issue. there is a research continueem to a clinical trial.
how bigth study? who is ininvolved. phase 1 is not always healthy. sometimes in cancer trials, they're very unhealthy. but phase one in a lot of studies, you use healthy humans and see if you cause any problems.
then we give it to the less healthy people. but how much control you have varies on this continuum. and there is this constant balance of trying to maximize my generalizability. more worried about false positives?
am i trying to maximize my control? and i'm trying to worry about the false negatives. different times that you're going to think about different things. some one like me, always wants to maximize generalizability.
i see what happens when i maximize my control. usually means i had the running results and i crewed up. but other people think it's super important early in research to maximize that but as long as you're trying to choose who and what study, your
number one reason you have exclusion criteria, regardless of the type of study, is safety. if i cannot randomize somebody -- i have computer randomizing. chuck used to blame me as the statistician. i said no, you have told me who
could be in the study, and i assumed they can be in any arm of your study. it is not safe for them to be in the study you should be excluding them. i'm not talking about something like metaphysical, i'm worried, blah, blah, blah.
they really shouldn't be receiving these study arms or one of these study arms, you have to make sure they don't get that study arm. my computer don't -- it's not a dr. doesn't know that. but i also try to reduce potential confounding.
this is where that maximizing control also comes in. but you don't want to overdo it. so what are some of the secondary data? think about where data comes per. might be primary data, so that's what you get, generated directly
for those research purposes. in many cases that includes the disease industries and national you also get the secondary data. so i'm not talking about the secondary data analysis of this is stuff like administrative, billing, electronic health records type
of data. so if you have a national healthcare system, the united states, even if you think about medicare and medicaid, we collect data. because we are paying for that care. different countries, higher or
lesser quality. united states, i'm not saying has very high quality of some of but it's often generated, the secondary data, they're thinking about north korea utilization -- healthcare utilization. you have to make sure, do you have enough data and high enough
quality to make a meaningful population based conclusion? why do they study these medicare patients? because in the united states and really the world, med care is the largest purchaser of healthcare. part of that is the cost of
healthcare in the u.s. but we have 54 million eproleys. and spend more than half a trillion dollars in 2014 in this program alone. so this is a huge amount of money out of the u.s. system. so if you're trying to study the united states, this is a good
place to start. if you're trying to study your own country, there is additional information and a lot of other places, and a lot of groups around the world that trito compare these different countries. in the united states, for the
most part. we have 20 million different data systems. medicare as one data system. so latin went through all the benefits of using medicare data. this is healthcare in action. what are people prescribing? there are a lot of limitations.
we don't have that much data on the severity of the illness. these tend to not be the working population in the united states. these very different. you can't extrapolate to the u.s. as a whole or other there are a lot of errors and bias that happened, so people
code in order to make sure that something gets paid for. or they just write down the wrong number. you have limited outcome measures of interest. but it's still useful data or a so you may have all these observational studies, you might
not be avoiding your selection bias. you have a lot of other issues. but it can give you some hypothesis to start testing. so you want medicare administrative billing or encounter data, he gives you the website.
if you are saying, well, i want to know this about egypt, i want to know this about some other place, you want to look in your own country for similar type of information, and you may want to share it on the online chat board, so that other people know where to find that information.
but there are also countries that are trying to put this type of information together, collectives of countries in many continents that trying to put this information together. because a lot of times we need to try to understand how healthcare is working, and
collaborative environment, not just individual work. so briefly, for hypothesis testing, bottom line, statistical inference uses the results from a sample for the population of interest to draw cubs about the population. conclusions about the
i don't care just about the little individuals, i want a bigger picture. the null hypothesis is set up with the hope that you'll reject it. if you can't figure out how to write your null hypothesis, remember that. that alpha or
type one error, that's a chance of making that type one error. probability of concluding that there is a difference when there is not a difference. beta is the chance of making a type two error. so it's the chance of concluding that there is not a deference
when there is. and power is one minus beta, unless you're in parts of europe, in which case they flip you do not want high error rates, do you want high power. that power the chance is concluding there is a difference that's your whole goal.
investigator controls, chance of making a type one error and chance of making that type two error with your sample size. so i'm going to say what are my parameters? what am i willing to make, what errors. and i'll derive a sample size. the p value is the probability
of obtaining results as extreme or more than one obtained if there were no difference. under that null hypothesis. these are all assuming superiority. you have statistical significance. doesn't mean clinical
importance. confidence intervals are useful to better understand your multiplicity adjustments needed with more than one primary hypothesis for those phase 3 trials. phase 2, earlier, we sometimes let you get away with it without
worrying about it. that approach, gaining popularly, more intuitive but you have to have people trained that know what they're doing to pull that off. so what about the sample size? there are lots of different sample size formulas that i
mention on the next vied. but just remember, changes and the difference you want to detect, in the type one error, you're willing to accept or the two type error, changes in the variance, the number of samples you're collecting, you'll do a one or two-sided test, that all
impacts your sample size calculation. so this is the basic formula for a two armed study, but depending on your study design, you probably have to do simulations. so just remember, though, the basics, that powers effected by all these different elements.
we talked about that extensively, gave a lot of examples about that in the sample size and power lecture. so when in down, first, what is your question. what is your hypothesis test? that you want to run. now i can get an idea of what
that formula should be or the simulations to relationship. i'm going to make a table, whether i have a fixed sample size and i want to figure out the power i'm going to have with that or whether i'm going to see the different sample sizes i should have, depending how i
vary the standard deviation, the type one area, all these things. you might want a graph, a lot of graphs. this is not a single number. sample size in the end will choose a single no. but you have to look at and evaluate a lot of numbers to figure out if you
screwed up one of my assumptions, how much does it hurt me? when in doubt, go high on sample making inference about the event rate is a little different thank the other hypothesis tests that we're talking about, like logistic regression, et cetera.
survival analysis, that rate at some time t, the rate among those at risk at time t. that was that idea, independence was key. you have to have independence sensoring, and of competing risks. big trick, you will always look
at median sacrificial, next to never, you look at mean survival. mean survival means every one had the event. cox regression, the most robust kaplan meier curves do not have sensible interpretations for competing risks even though
they're lovely creatures. independence of measurement and outcomes is key. truncation is about entering the study, and censoring, about leaving the study. someone asked and additional question on survival analysis, one of the first questions.
about one of the examples i gave. go back and read the board about that question. i wrote an answer about that in order to try to clarify this topic. so john powers, one of his last slides i thought was a pretty
good conclusion to this first module of the course. you're trying to develop an efficient trial. that all starts with planning and a good research question. everything starts there. your question always comes first.
then you're going to deal with sample size and everything else. but if you can't come answer the question thank matters, you shouldn't do the study. i'm going to be honest with you. various methods are going to help increase your effect size, decrease variability of your
you can apply it in a different setting. you want to get valid reliable answers to important public health questions. all these things we talk about is to get you to that end. and for some diseases w it is going to be really important to
develop the tools to develop the better outcome measures to get better data and natural history, to better understand what you're dealing with before you do all the other trials. a good start to better trials, making sure you can measure an outcome that matters.
you can do it reliability, with high validity. it's understanding your patient population, how they're going to change over time. doing natural history studies in order to better plan your randomized studies. always, always, always remember
that your analyst follows design. analysis follows design. the question always comes first. questions on this module will go beyond what i'm talking about tonight. but this is a brief summary. hopefully that was useful to
you. if you think it's just all about medicine, i was on the plane i was reading the latest economist from 12, december, 2015. some might listen to this a few years from now. the latest from my plane trip. talking about randomized control
they have two different articles i've given urls too. one was talking about more editorial style doctors use evidence when prescribing treatments, policymakers should, and they were referring to an article where some of these government organizations are
giving out livestock and teaching people how to use the livestock. that improving and moving the very, very poor up in their and they were talking about the fact that you can use randomized control trials. this group out of m.i.t. and
other groups are using randomized trials to test and so they've gone in, done randomized trials to test, if i give people a cow, if i give them goats, and i teach them how to try to raise them, does that mean that they're in less poverty five or ten years later?
you can do a trial on almost anything. what we've talked about here, in a clinical context, can be used in lots of situations. the basic tenants are the same. now, another friend of mine died last week. and that was norman breslow.
i'm going to talk about him. he was one of the founding members of the williams tumor study group. this is a paediatric cancer. 90% of children died -- diagnosed with it died. beginning of the 20th century, 90% of the children diagnosed
with it -- 21 century, led relatively normal lives. they refined treatments and looked at the long term effects of radiation and chemotherapy ob patients. everybody across all disciplines, not just oncologists, were full members
of this group. it was one of the first cooperative groups to do that. and norm worked with a lot of epidemiologists and a rot of other folks. they placed a very high priority on data collection, electronic data, capture, careful follow-up
and registering all the the pathologists were able to do work with that data. they actually developed that case control matched sample size discussed in the textbook that helps inform your later so they have done a lot of different trials over the year,
if you want to read through but i think one of the important ideas that came out of this was was it ethical. they've now increased survival, that patients, women that have had wilm's tumor, they were not able to have healthy pregnancies.
there were a lot of additional tumors happening relatively early in life. whether it was breast cancers, or other types of cancers. they did randomized carps, can i imagine being that parent. there are a lot of ethics that go into these studies.
a lot of people that put their lives into this. that's what you're talking about next. good luck. if you have questions put them on the board for the lecture. good luck in the rest of your course and all your work.
take care.
0 komentar: