Computer generated transcript

Warning!
The following transcript was generated automatically from the content and has not been checked or corrected manually.

OK, doctor, we are live now. All right. Can you see my screen? Yes, we are seeing it perfect. All right. So um I hope that the previous session was clear enough. I hope that you don't have any questions. Now, we will move to the uh model number six for you. And we will talk a little bit about the questionnaires, the validity of measures and the biases. So it's a very, very important, extremely important part of the methods that are used. So in addition to the study design, in addition to the sampling methods and the sample size, you will have to look at your questionnaire, the tools that you use and you have to use good tools in order to obtain good results. It's very simple. The relationship is straightforward, you put garbage, you get garbage. So if you use the bad tools, you obtain bad results, you have to really use good tools in order to to obtain good results. So we're gonna look at the data collection tool structure. Uh We're gonna distinguish distinguish bias from random error, identify the most common types of biases and learn how to deal with bias in epidemiological research. So let us start first all measures and statistics. And of course, in any epidemiological study, they carry errors. This is something that we always repeat. And there are many types of errors. So you have to pay attention because if you work without really putting all your time and all your reflection, uh you will obtain erroneous results. Can we just uh avoid all types of errors? No, it's not possible you cannot avoid, but at least you can decrease, you can uh uh partially, let's say, avoid errors. Uh So there are two types of errors. First, what we call the random errors. And these are the ones that are really unavoidable. They are there, you can decrease them but you cannot avoid them. These are, these are precision errors. They are mainly due to sampling fluctuations and they are related to all the alpha and beta and P value that we just saw. But there is also something that we call systematic errors. So systematic error or non random errors. Well, they are inherently linked to the methods that you use and they will give you a validity issue in your work, validity issue. I mean by that you will have a bias and your result will not be correct, they will be false. So the random errors, well, this is the biostatistics field that's mainly solved by improving the sample size, looking at the P values, adjusting P values. Well, interpreting results. All this is a random thing. But the nonrandom, it's the work of a methodologist and it is here where you have to pay attention even before you start your work. So you have to think how to do a study that minimizes the risk of errors uh because it's not easy to correct biases, some biases can be corrected and some can not be corrected. Unfortunately, this is how it is. So the first thing to think about is your data collection tool, whether we are talking about a clinical research form that you will fill out from medical files from patients. Uh uh let's say measurements uh in the hospital or in a clinic or you are in an epidemiological setting in the general population, you would use questionnaire to collect data from participants. You can add to the questionnaire measures, lab results, other results. As I told you, for example, uh we took a portable spirometer and we did measurements. Once we took what we call a a carboxy meter to measure the exhaled air in the uh to measure the carbon monoxide in the exhale, exhaled air, you can go and just do some BP measurements. So all these can be done. In addition to the questionnaire, you can add the result to the questionnaire or to the clinical research form and this will only improve the validity of your results. Now, what is the structure of a data collection tool? It should always include an introduction particularly if you're talking about the general population, they need to read to understand what you're doing. And let me tell you uh that it's better not to put too many details because if the patient or the participant know your exact objective, they might give you the answer that you want to hear. Just to please you, this is what we call the social desirability. So please, whenever you are doing a data collection tool, uh uh don't put too many details, just say that we are doing a study that is related to uh this type of problem and the study will help us in finding the reasons of this without telling. Well, we want to uh uh uh uh assess the association between, let's say and uh uh uh survive CO PD. No, don't say what is your exact objective? Be a little bit vague because you want to have objective results and you don't want people to just tell you what you want to hear because they want to please you, then you add the informed consent and here please pay attention. Uh Any study that you wanna do, any study will need an IRB approval or at least a waiver, an IRB approval is whenever we are doing an experimental study, definitely or a an observational study that includes, for example, blood drawing or genetic material, things like that. Well, you need a full fledged IRB approval. Now, if you are doing an uh an, an observational study with just a small questionnaire. Well, here you don't need a full approval. We say you need an IRB waiver. So the IRB will look at your study and will tell you, well, there's no need to go for a full fledged uh uh approval. We just waive the need for an approval and they give you a number and you're fine with it. Um Then you have the sociodemographic characteristics and here also putting the identification characters such as the name, the phone number of the participant, it should be optional. It should not be uh uh mandatory except if you have a follow up to do. If you have a follow up ma ma you have to know the name and the phone number of the individual. But if there is no follow up, if it's just one time, you can even skip the identification. This make it easier for everyone without the name, without the phone number, just the sociodemographic characteristics, age, gender education, socioeconomic status and things like that. Then we go to the core of your questionnaire. You should define your dependent and major independent variables. And here listen to me, don't reinvent the wheel. Search for the most valid tool. Search for a validated scale in the literature. Search for the best definition of your dependent and independent variable and appliance. For example, if we're talking about diabetes, well, the diabetes definition is known you have to have some glycemia results in order to define diabetes, you cannot just say, well, I will just ask the patient if he has diabetes be because you know that the majority of people have diabetes without knowing it. Mm If you want to ask about AAA problem, that is more of mental or social uh such as uh mm, anxiety, depression, uh wellbeing quality of life. Any subjective thing that should be reported by the patient? Well, here you cannot just rely on a question that you invent or that you come up with from your head, you have to go and search for depression. There are specific scales that are used in the literature scales that are known to screen or to diagnose depression. Once you ask them for the patient or for the participant, you cannot use just one question and say well and say, oh this is a person with depression. This is not the way it goes. You have to go and search for a validated tool. You start your search by a tool validated 11. So if you find a tool that is validated in Lebanon, this is a good starting point. Sorry, I will show you some tools. I don't know if you know that I have a research team that is called INSPECT LB. Uh We are very active in research. We have been uh in first position for the second year, uh consecutive year. We are even uh ahead of universities head of AU BMC according to S ao institutional ranking. So we are very proud of that. 1st and 2nd. I would like to tell you that in on our website, we have some Lebanese validated scales. So these are scales that we validated and you go there are some related to physical health, some related to mental health, some related to lifestyle and behaviors and some related to health profession. I will take an example, mental health. If you go, you will find a list of scales that were validated in Lebanon related to alcohol, anxiety, fear of commitment, anxiety, geriatric, depression. Uh I don't know there are really lots of scales that we did ourselves and we published, we validated them. Uh For physical health. For example, you have the cognitive function, you have medication adherence, you have urticaria diagnosis, urticaria, uh activity, um also uh medication adherence, osteoporosis knowledge, San George respiratory questionnaire, asthma, uh risk factors and so on so forth. There are also many scales that can be useful. You can also search on pub very simple. You put Lebanon validated scale. Uh II don't know, it depends on your uh uh topic. And if you find a scale validated in Lebanon, it will be your priority. Now, if you don't find a scale validated in Lebanon, second will be to find a scale validated in the Arab region. Let the Arab region I know it will be in Arabic and you want to have a scale that is validated in Arabic. Uh you check it, you, you make sure that the Arabic is acceptable is clear and you can apply it, use it in your study. Now, what if I don't find neither in Lebanon nor in the region, then I will go for an internationally validated scale, an internationally validated scale. So it will be in English most probably maybe in other languages, but I don't need it in other languages. So I will take the scale in English, I will translate it to Arabic. Mm and and back translated to English by somebody else. And then you will compare the both the English versions and you will solve any problems. Uh And finally, you will try it on in a in a pilot sample. Let's say this is how you do it. So English, OK. Arabic translation, English back translation comparison of both version and improvement of the uh uh scale in Arabic if if necessary. Now most of the times, if you do a good translation, it will be OK, you will have a really good back translation and you will not have problems. No, if you're working with students, particularly university students, most of the time, you don't need to use the scale in Arabic, you will use it in English because as you know, all university students, they speak English and they have no problem answering a questionnaire in English. This is nowadays the the trend and there are references that can prove that using an English questionnaire among university students is OK. Uh What else should I want? Uh Should I tell you? Yes, I forgot to tell you. If you are working with Children, you need parents' approval. The IRB will not let you do the study unless you take parents' approval. So if you were working in schools, for example, you will prepare a letter that you will send to parents and you will tell them if you don't want your son to participate. Say so if you don't answer, this means that you agree. So most of the time people, they don't want to really uh uh uh bother and uh uh refuse that there's some Ries and, and and things like that. So they don't answer and you can, you have the right to include their son in or their daughter in your uh work. So all this is to tell you that you have to be very careful, careful when choosing your variables, careful on how to perform your work. Now, in addition to the major dependent and independent variables, you will also mm uh a assess other independent variables and here it's more or less important to have validated scales. But if you don't, it's OK because you cannot use validated scales for everything. Of course, this is something very well known. Uh I would like to uh uh also give you just an idea about uh the socioeconomic status and how to measure it. Sometimes it's possible to um, ask people about their income and it's easy to ask people about their income, particularly if we're talking about students. For example, yes, they would tell you that this is my income. But if you were, you were talking about people from the general population, about people from several socioeconomic status. Well, they will not always tell you uh what is their income? So you will not be able always to just use the income in order to come up with a socioeconomic status classification. So what to do? In this case, there are several ways of working. You might first use a validated scale. There are validated scales that are specifically designed for that. There is one scale that we designed. It's called the socioeconomic state uh uh uh status composites scale. And it includes question related to many things, income education uh have receiving help from others from NGO S and ho they enter in this tool. So this you would, you, you would use it freely. Your work is uh very directed towards the socioeconomic status particularly we're talking about the current crisis that we live in Lebanon. Now, if you want something faster, the best way to assess a, a baseline socioeconomic status before the crisis is to ask about what we call the crowding index. What do I mean by that? You ask about how many rooms there are in the house aside from the kitchen and the bathrooms. So how many rooms are available to live in and how many people live in the house. And you will divide the number of people by the number of rooms and you will obtain what we call the crowding index. The higher the crowding index, the more crowded the house is. So you have, for example, five people living in one room, this is too crowded or you have one person living in a big house. So you imagine that people who are living in bigger houses. Uh well, they are better off from the socioeconomic level. So this is a way to assess socioeconomic status, basic baseline. Now, what if I want to assess the current now after the crisis, what's ongoing, what are people suffering from what, how, what happens to their income? So here you can ask about the household income, I mean, the full family income and you put for them some brackets less than $300 300 to 505 100 to 1000, 1000 to 2000, let's say more than 2000. Hm. And you ask the question about the full household and again, you ask about the number of people who live based on that income and you do the ratio and you will see how much money every person is getting from that household income, regardless of who is working. So if a father has two Children and earns $1000 it's totally different from the fathers who earns the same amount and has seven or eight Children. So you can understand that this is totally different and this is why you need to divide the the household income by the number of people who live based on that income. So these two methods, there was a crowding index and the household income, they will help you to assess the socioeconomic status. And of course, for every additional variable you can search for the best way to assess it. Rosko don't invent the wheel go and see what other people are doing. Very simple. Uh Now one more thing, what should be the length of the questionnaire? Well, it's better to keep a questionnaire below 100 question above 100 question people don't really uh pay attention. So below 100 question in general, it might take around 15 minutes to uh to fill out and 15 minutes is considered acceptable. If you want to do a very big study, you can go up to 30 minutes. But it's better at that at this time than to do an interview not to give the people so that they just fill out the, the, uh the questionnaire alone by themselves. Anyway, in all cases, a good questionnaire will allow you to have valid results. And what do I mean by valid results? I mean that it correctly reflects reality. For example, a test that reveals HIV status of a positive person is a valid instrument. However, it will be invalid if it doesn't reveal his disease, and you should know that there are two types of validity, there is what we call internal validity and external validity. So what is internal validity of a tool or an instrument or it can apply to a method or a study, a full study, it's whenever it measures what it is supposed to measure. And of course, the lengths that follow from these measurements and parameter estimates and the internal validity will depend on systematic errors specific to the method. The bias, as we said, we should try to find tools that are not biased. It will also depend on the incorrect calibration of instruments. If you have a physical instrument, I mean, and the correct way of using study instruments while external validity, this applies mainly to to studies. Uh it's whenever there is a possibility of extrapolating the results that are observed at the level of the target population. Now, of course, the more representative the sample, the more the or the higher the external validity. For example, if you have a sample where you worked on people aged 15 to 25 you cannot extrapolate it to an age group of 55 to 70 or others. No, in contrast to validity precision is something that is purely random as we said. And it depends mainly on the intrinsic precision of the instrument and the influence of factors that cause fluctuation in answers or in measurements such as temperature pressure noise, et cetera. And as we said, precision, it's the field of random error depends on the sample size. It's related to biostatistics while validity. It's the field of epidemiology. It depends on biases on methods and you can avoid it and sometimes you can correct it. So this is a valid but non precise measure valid because the average will be just in the middle. So it's OK, this is a precise but nonvalid. So even if you do 1000 measurement, they will all be away from the truth that is in the middle. And this is the idea a precise and valid, precise, appropriate sample size, very low random error, violent no selection bias, no information bias and no confounding bias. So this is really ideal sometimes it's possible, sometimes it's not possible to do it. So based on this uh uh uh large introduction, let's say let us discuss a little bit biases, biases are internal validity problems. They are systematic errors. They are due to the method that you use the measure that you use the parameter, the parameters that you are measuring and they are in general divided into three part. What do we have selection information or measurement and confounding base? So selection wise as its name indicate, you have a sample that is not representative. This is what we were talking in the session before you have. This is the entire population and this is the small sample. The entire population includes blue, green, other colors. But your sample includes only red and yellow and one pink. So it's obvious that your sample does not represent the population. So this type of bias, it's a sampling bias. It's present in all types of studies. It might occur in cross sectional studies. So it's a nonrepresentative, but it might also occur in prospective studies due to lost of follow up in different groups. So, so if you have people who quit a prospective study, and these people uh uh uh uh they change by quitting the nature of the sample that stays within the study. Well, this is also a selection bias. What to do. The best way is to avoid it from the start if it is partial. Well, I can correct it through what I call a waiting procedure. And here I would like you to remember the stratification. So if you have a sample where the percentage, the distribution of the percentage is not really similar to the population percentages, there is a possibility of trying to do this waiting to change the percentages uh based on some statistical calculations, but it will only be useful for prevalence and incidence correction. But if you want to work on association, there's no need and even you should not do waiting, you should not apply waiting for checking associations. So here this type of bias is due to an inadequate choice of subject, as we said, because having a representative sample is really important, particularly if we're talking about cross sectional study or a cohort study. Now, there are many types of selection bias. We have sampling, admission, migration and healthy worker sampling by its whatever the sample does not represent, as we said. So the choice of subject is twisted. It happens commonly in cross sectional in case control studies also. And here you have to think very well about the choice of subjects and how it is related to the studied factor or disease survival bias. Well, it happens mainly in cohorts in cohorts. You should not only analyze survivors because sometimes mortality may be related to the cause. And whenever the mortality of people is related to the exposure in question, well, you might have a problem during your calculation and you might think that your exposure is not really causing a disease while it could have been the cause of the death of these people. So this is why whenever you are doing a cohort and you have people who are lost to follow up, people who are just disappearing, you have to go and check and see why, where are they? And if there is a death, if they don't want to come back anymore, you should ask why, what is the cause of death or if they are still alive, why did they leave? It's very important. A third selection wise is called the admission bias or the Berkson bias. And I guess that you all know, for example, there are some hospitals that have some level of fame in certain specialty. For example, if I told you, everybody will say ah it's for burns. So if you are doing a case control study, if you take cases among burned people, it's important to take controls from all over Lebanon. Why? Because your burn people will be coming from all over Lebanon. So in order to have an equilibrium in the cultural and the ge geographical representation, it's important to take into account uh the, the nature of the service, the nature of the uh uh uh um I don't know the, the specialty that you are taking your patients from. No, the last one, the migration bias. This is similar a little bit to the uh mortality uh issue. If you have people who are lost to follow up, you have to know why and sometimes you might want to incl in I include them in your study, the healthy worker effect. Also, this applies mainly if you are doing your study uh uh in a professional setting, an occupational setting in January. Every people in an occupational setting are more or less healthy people who got sick. Well, they may have quit the job. So don't just rely on the people who are already on the job, but also you can go back in time and try to see what type of individuals were there and quit and why did they quit? Because the healthy worker effect might fool you and you might think that, well, this exposure is not toxic, for example, while the truth is, it is toxic and it caused so many people to quit and to go and to, to leave the job. So this bias resembles a little bit survival V OK. One more word about selection wise, what to do. You have to try to avoid it from the beginning. You have to find a way to avoid election bias. You only apply waiting as we said for prevalence and incidents. If you know the percentage of distribution of a factor within a population and you want also to apply it to your sample. But most of the time it's better not to do waiting, particularly if you're trying to link an exposure to a disease. The second type of bias is the information bias. The formation bias is whenever you classify the people, whether from the exposure point of view or from the disease point of view, you classify them erroneously. So you have some people that are classified with the exposure and disease. Hhh. If the rate of errors is low, this misclassification will not affect a lot your result, but sometimes it is uh it will affect your result and it will lead to a change in your result. Some examples recall bias, Recoba is very well known. It's whenever you rely on individual, just individual's memory and you ask about previous exposure in the past. For example, now this is particularly true. It happens a lot in case control studies because we ask previous questions about their previous exposure. Uh We have also what we call the haen effect. It happens whenever we take care of somebody. So just because you are taking care of them, they might feel better. So pay attention and you should apply the same methods for everyone. For all groups of comparison, you should apply the same method in order to decrease this type of subjectivity that might occur due to the hawthorn effect subjectivity bias. One, sometimes people may be subjective, sometimes physician may be s subjective and this is why you, as I told you, it's better not to uh uh reveal all your objective and don't really tell the patient or the participant, what are your exact objective so that they remain objective whenever they are answering and they are not affected by their own subjectivity. Um If you are using a measure with low validity, so a measure that is not sensitive enough, not specific enough. Again, this will lead to an information bias or a detection bias if you have some tool that is not detecting well a a disease or something. So all these situation will lead to a misclassification. If the misclassification occurs in everybody, it's OK. It will bias the results toward the null. I mean it will increase the risk of nonsignificant effect. It will dilute the results, but it's ok, because you know that the result will be diluted. The problem is that whenever one of these biases is what we call differential bias. So a differential bias is unpredictable. It might happen in a group and not happen in another. And well, here it's difficult to correct and it's difficult to avoid. So this is why it's better to really think very well before you start your study. Uh, because you want to avoid at any uh uh uh price and information wise. No. Ok. Let me add something. Mm. So here it's a little bit what I was telling you. These are the nondifferential information bias that could be due to the low quality of measurement, non-qualified personnel and things like that. And the differential one, these are the ones that are most important and most dangerous because they might affect one group in a differential way versus another group. It's very well known for the recall bias in case control cases, remember much better than controls their previous exposure because they are sick. So they think all the time, what did I do to get this disease? What did I do? So that happened? So they have this kind of recall bias while people who are healthy, they tend to forget the past. They don't think a lot about what is happening and how it is happening. This is why you have this recall bias. Um So this is why it's better to use the same means the same performing means among both groups and try to keep the subjectivity of people in order to have good result. Now let me just, oh let us continue. And then I will tell you the other idea I was thinking about this is what happens to the ratio or the relative risk. Whenever you have a certain percentage of misclassification, this is the true relative risk. These are the possibilities if you have 1% misclassification. So you can see that there is a certain degrees, but it's not really problematic if you are at 5% starting to become alarming 10% 20%. Well, here you're losing your high odds ratio and relative risk and if you are at 30% so a 30% misclassification, we let you think that an odds ratio equal two is 1.3 an odds ratio equal five is 1.8 and so on so forth. And measurement errors can even mask a dose effect relationship. So if you have the higher the exposure, the higher the odds ratio and if you have an error, well, you will see that you cannot see this increase in arts ratio or relative risk, you cannot see it, it will become really diluted. So this is just to say that it i it's important, very important to have really well measured measurements, whether we're talking about dependent variable or whether we're talking about the independent variable. Now, before I move to the confounding bi I would like to add one thing. It is possible sometimes to correct for this information bias if you know, by how much you are making there. What does it mean? I will give you an example. Uh, an example of a study that we did. We wanted to assess obesity among, mm, I guess it was among young people, maybe among university students. So we if to assess obesity, the easiest way is to calculate the body mass index, you know that if you rely on declared weight and declared the height of, of young people, well, it's very well known. Uh people in general, particularly girls have a tendency to give you a lower weight than the truth. Uh and boy will give you uh a higher height than the truth. So what can we do if we don't want to bother and measure weight and height in everybody? So we can ask everybody about their height and uh about their weight and height as we ask the question and we take their declared weight and height and for a small sample, we will also measure their weight and height. So by that, I will know for every people for every boy or girl, for whom I did the measurement. But they who we, oh, but they who we m let's see, I will be able to calculate an equation for height for weight of girls and boys. And based on that, I will be able to correct the declared height and weight for everybody. Even those for whom I did not measure height and weight Pfizer, you can use corrective formulas if you know, and this will improve, of course, the information precision and will decrease the information error and information wise. So before II moved to the last uh type of bias, that is very important. I would like to know. Do you have any question about what we did until now? I don't think doctor uh that anyone has any question. If anyone wants to say anything, you can chat. OK. And all right. So anyone has any question? Uh he can he or she can uh write it on the uh feedback form and I can send them to you if you want. OK, perfect. So let us move to the third type of bias. That is very, very important. It is the confounding bias, whatever you are trying to compare two groups, you know that we say we need to compare apples to apples and oranges to oranges. So if the baseline factors between two groups are not alike, it's possible that the difference that you observe might not be to the exposure, it might be to due to these baseline differences, due to age difference, due to gender difference, due to socioeconomic education, whatever exposure might affect our result other than the exposure that we are studying. And this is what we call a potential confounder and a potential confounder can either overestimate or underestimate the risk. So it's hiding within the exposure factor. So here this is the exposure and the outcome and a third variable might come and just confound before I say some confusion. So this third variable, it should be associated with exposure without being a consequence of the exposure. It should be associated with outcome independently of exposure. So it should not be just an intermediate factor, it should not be a mediator. Let's see, it can be corrected through stratified and multivariable analysis. And this is what is uh let's say, uh not beautiful, but let's say what is good about it is that you can correct. So if you have a problem related to a potential confounder such as a difference in age, a difference in gender at baseline of the two groups that you are comparing. So you can go and do some types of statistical analysis and get adjusted results. No, e even if you do excellent multivariable analysis, there will still be a risk of residual confounding. This is something that you should put in mind. However, you try to decrease residual confounding as much as possible. Now, before I move to these corrective strategies, I want to show you no, I run the hell SS BSS how to perform this kind of analysis. OK. So file open data uh again. OK. So I'll go back to my database and I will show you first a simple analysis. Let's say, I want to see if I don't know. Uh So having chronic cough is associated with CO PT. So I will ask for a Chi square. Ok. So if you see here, those who don't go have 5.5% CO PD prevalence and those who cough, who have a chronic cough, they have 24.2%. And actually, this difference is significant. And you might think, well, maybe this chronic cough is a risk factor for CO PD. Is it true? Is cough the risk factor for CO PD? Uh in your opinion, it's a symptom. It's not a risk factor. So those who cough and those who don't cough, they might have other reason. Y there is more COPD among those who cough. Let's see, we will go and we will try to do a multivariable analysis. So I would take chronic cough but I will also adjust over tender current cigarette smoking, previous cigarette smoking, previous water pipe smoking. So, no, II don't need that. So I am trying to include in my model, all the factors that might have caused this maybe spurious association between chronic cough and throop. So let us see if regardless of gender and cigarette smoking and previous cigarette smoking and previous water pipe smoking and all the others that chronic cough still is associated with COPD. So this is a multivariable analysis. I will not go into the detail because it's too long. But what I want you to see is that here you have all these factors that we included in the regression that are positively associated with CO PD and still chronic cough is associated with CO PD. So it's not because of gender, it's not because of cigarette smoking. It's not because of previous water pipe. But there is still something that I'm not finding. Maybe age. Let us see. OK, so the more you include in your model, the more you can see that the results are getting lower and lower. So will three our five, but then we went to four, then we went to three. So all these factors are potential factors that predict CO PD in addition to the chronic cough, that is also a symptom of CO PD and it seems to be an independent symptom. So whether you are a man or a woman, whether you smoke or you don't smoke or you are a previous smoker and whatever is your age and all these still, if you have a chronic cough, you are still at higher risk or how higher odds let's say of having CO PD. So this is how we do for correcting over potential confounding. We include several factor in one model in order to come up with what we call the adjusted odds ratio here, this exponential beta is called the adjusted odds ratio between chronic cough and COPD. So going back to here, finally, there are lots of corrective strategies that I can do for biases. Well, I should pay attention during study planning if I'm talking about uh an observational study. So it's mainly these ones selection measurement and lead time bias. If I'm talking about a uh clinical trial, I should pay attention about the random sampling and allocation, the appropriate control group well defined exposure, ensure correct exposure and outcome classification and correct for lead time. Then you have the information bias, recall, observer bias, procedural bias, nonresponse all these. Also, they should be taken into account during the study implementation. I will try as much as possible. I will do some blinding for placebo. I will confirm self report with medical report, I will minimize loss to follow up. So I'll try to do my best here. And finally, for confounding there is also for missing data and some other types of problems I will do matching, I will do a stratified or Multivariate regression analysis. And this is what I showed you. And there are also other statistical uh uh procedures that could be used for other less common problems. So you can see that there is always a way to avoid or correct. It's never 100%. It's never perfect, but it's better than nothing. And in general, once you do your study, you will know that your study is not perfect. It's good to write in the limitations, write everything. Don't be afraid to say what are the limitation of your work? Because the reader and the reviewer, they need to see that, you know what you're doing that, although your study is not perfect, but you know that you are interpreting your results with caution, you know how to suggest further studies that take into account these problems and these pitfalls and limitations. And this is why if you find a statistically significant association, it does not not mean that there is a, there is a causal relationship. And actually, I think some of you may know this list of causality criteria. What do we mean by causality criteria? In fact, this is a list of items, the more items are fulfilled by your study, the better you will be close. If you, let's say to causality, you cannot prove causality with observational studies. Only a, as we said yesterday, only an experimental study, placebo controlled double blind randomized and and randomization is major. It's key only whenever you have a an excellent clinic, I'll try it. You will be able to prove causality. And for all the rest, it's not possible. So if you have an observational study, as we said, a study where let's say we are um how to say a study uh where we are observing a toxic exposure and see its association with a disease. So here you cannot do a clinical trial, you have to go for a an observational study. So how to interpret a positive result, a clinically significant result. First, you look at the association strikes, it increases whenever the odds ratio and the relative risk are far from one and significantly different from one. So the higher or the, the more far they are from one, the better it is. This means that a strong association, whether it is positive or negative, a strong association means that you are closer to causality. While an alteration or relative risk that is close to one, it's not good. It means that you are away from causality. Second specificity of the association, if one factor between others is associated with one disease between others, this reinforces causality. Exactly exposure to. Let's take an example. Clostridium difficile is associated to what pseudomembranous colitis. Very specific exposure to mycobacterium tuberculosis. It is associated with tuberculosis. Very specific exposes to asbestos is associated with uh uh mesothelioma. So very specific, you know, when you have such a specific association that occurs only among people who are exposed and and not among people who are not. But this means that this is a specific association. So it's closer to causality measurement adequacy. Whenever you don't have a bias or you have a minimal bias, you have a valid and precise measure uh for both exposure and disease. Very good. This is also in favor of causality. D if you are able to demonstrate that exposure comes before the disease, this is also uh uh towards causality. And this is why if you remember we talked yesterday about cross sectional studies and we said that it's major drawback is that doesn't uh prove temporality, you cannot know that exposure comes before the disease. Number five, dose response or dose effect relationship. If you are able that with the increase in the disease, uh uh with the increase in the exposure, you have an increase in the disease. Well, this is also in favor of causality. For example, the longer the duration of smoking, the higher the risk of co PD simple biological plausibility. So, causality is reinforced if there is a biological relationship between exposure and disease, a biological relationship that has been demonstrated on animals, on cell cultures, it's not necessary in humans. This could have been demonstrated in other species, seven stability of the association. So if the study results are similar to those of other researcher in a different context in place and time, different population, again, this is good sign of causality. And this is why whenever you uh uh uh you are writing your discussion, you always compare to what other people did in other places, other time and other population. And finally, the coherence with anterior knowledge, it is the least important criterion because it does not take into account in your discoveries. However, if your work is coherent with what we know, for example, you're working on water pibe and you discover uh um positive effect similar to cigarette. So there's a kind of coherence with what we know about tobacco and its effect OK. So this is also in favor of causality. No, are all factors found together in one study. No, not necessarily. But the more factors you find in your study, the better it is, the more let's say solid, your demonstration of causality criteria is and finally, uh judging the evidence to judge the evidence, whether we're talking evidence based medicine or evidence based public health. Well, they give a lot of importance to first the absence of bias, random bias, systematic bias, internal validity. A lot of importance to do effect relationship. This gradient effect that we talked about. This is also part from internal validity. Also to biological plausibility, coherence and external validity. This is part of external validity. And finally, consistency, I mean reproducibility with time and place and other researcher. And this is also part of external validity. So you can see that these are the four most important ones among all these. But still, it doesn't mean that you will find everything in one study, but you will do your best. You should do your best to have really a good study that fulfills as much as possible from these Bradford Hill criteria. And one thing whenever you want to write down your uh discussion, think about these criteria and use them in order to write down a good discussion, whatever the study that you have done. Uh before I thank you, I would like to finally just uh uh uh tell you about something that is very important in my opinion. And that will help you in writing, whether you want to write a proposal or whether you want to write uh uh your article and it will help you to come up with uh all the method, methodological aspects, write them down and think about them very well. I will go back to the website. I told you about the inspect lb.org and here you have uh what is it? This is a methodological checklist. So for every type of document, there are people who worked. It's not us, there are international document that guide you on how to write. For example, if you want, this is the one that we we we use a lot, the Strobe statement checklist. This is for observational studies. If you go there, you will find checklist for cohorts case controls, uh cross sectional study. And if you open this list, you will see that they will tell you step by step what to write in every part of your work title. Abstract introduction, background, rational objective Methods, particularly methods. How detailed they should be what you should write. Your method have to allow somebody else to repeat your study in the same way you did it. There is a new, it has to include all the details that are necessary. And this is why study design has to be presented the setting, the participant, the variables and here remember validated scales references. You should put references in the methods. So that anybody who's reading will remember and will know where to go and search for your uh tools, data sources and measurement biases. If you did something to decrease the bias at the level of the methods study size, explain the sample size. It's important to say, how did you decide that you want to take 400 people? Let's see. Then the quantitative variables, how they were grouped. If they were grouped the statistical methods, all the details, the missing data, the multivariable analysis, sensitivity analysis subgroup and everything should be there. Then the results, participant, descriptive outcome data may results and so on so forth. So you can see that for every type of study you have a checklist. Strobe is for observational study. Concert is for experimental clinical trials mainly. Uh then Prisma for systematic review and meta analysis. Nice for the qualitative study, a clinical practice. It's called the agree case report. It's called the care uh economic evaluation, the Cheers uh Clinical trial protocol, the spirit uh diagnostic accuracy study start. This is for validation and this is the for the patient reported outcome measure mainly for validating a tool. If you want to do the validation, you're not using a validated one, you want to develop and validate a tool. So this is the design checklist and so on so forth. So you can see that in the field of research, you're not alone, always go back and see what other people did what other people uh are suggesting how they can help you because these checklist, for example, they would help you a lot in doing whatever you want to do. And I guess that that's all what I wanted to tell you. Thank you a lot doctor. Really. It was really, I love, it was uh interesting uh important. Uh Thank you uh uh for listening and of course, I think I was quick in, in many parts, uh you can read about whatever uh point you want. And if you have any question, I'm ready to answer. And uh one final thing about the multivariable analysis. Well, this is a full course that we give. If you're interested, I might even send you uh powerpoints and videos that are related to by statistics. These are my courses, the ones that I give to master's student uh at the Faculty of Pharmacy and the Faculty of Public Health. So if you're interested, I can just send it to you. It's, it's fine. I think that a lot of students have them and uh you can have them and there's no problem. Yes, doctor, we will uh we really appreciate if you uh send us uh the powerpoint and we'll share them on the group. OK. And uh OK, really thank you doctor for the session and for the three sessions actually, it was really uh they were really insightful uh the participants really in the feedback. I just read the feedback of the last module, everyone just really love you so much. So, thank you, we really are really grateful for your presence here. Thank you. I II appreciate this initiative a lot. I already told you yesterday and I'm telling you again, I really appreciate this initiative Bravo. And no, uh I guess that you need it. Uh, every physician needs to, to know how to do at least some simple research, uh uh studies in order to improve your CV and to get really AAA position that, that you deserve for your future. Thank you. Really? Thank you doctor. I will send you actually, uh the feedback uh uh with the F on May. So we can see how really they are really motivated after your sessions and uh we'll see everyone. Uh Tu on Tuesday was in an arm with the last session. Uh Thank you again, doctor and everyone, please. Ok. Have a nice, have a nice weekend and have AAA, nice uh evening. All right. Thank you, doctor. Ok, bye bye bye bye.