Skip to main content
Loading Explore guides…
Causality and Bias in Epidemiology · Acedra
Home
Explore
Medicine
Causality and Bias in Epidemiology
Sign in to save
Chat
If YouTube asks you to sign in, open the video on YouTube.
Watch on YouTube
Reload player
Transcript
206 segments
Copy
Hide video
0:03
Hello and welcome to this video lecture on causality, association, correlation, random error, bias and confounding. In this session we will look to understand and explain clearly
0:22
causality, association and correlation,
0:26
random error, bias and confounding. We will look at the causes of bias and confounding and how we can deal with them when we're designing an epidemiological study
0:42
and we'll do some practical exercises in the apply session. So let's start with causation, association and correlation. Please watch this video. It will be in the playlist and you can find it also on Youtube, The danger of mixing up causality and correlation.
1:04
Causation is a complicated business. It is not so easy as we might think to say one thing causes another. We need good evidence from many research studies but even then we have to recognise that there may be multiple causes and multiple effects from those causes, so that the chain of causation from exposure to health outcomes can be very complicated.
1:35
So what is causation and what is association? In lay terminology, in everyday use we talk about causes leading to effects but in epidemiology we don't do that. We talk about exposures leading to health outcomes. And that relationship, that link is called an association. We don't say it's causal at the beginning but when we have enough evidence to believe that the association is causal then we can say that an exposure causes a health outcome.
2:13
For individual studies we will talk about identifying, recognising and finding associations, a link or relationship between an exposure and a health outcome and this is because we have to recognise that there are other factors, other causes that can confuse and confound the association, the relationship between one exposure and one health outcome.
2:45
And the reason for that is because we have proximal and distal causes and we need to think about in epidemiology all the different various causes and the causes of causes generate health and well-being and disease. So we start off with risk factors; we have proximal causes, we have medial causes and we have distal mitres
3:12
distant far away causes and the far away causes are industrialisation of food production for example or the global economy or changes in social and cultural practices at national level and international level, focus on economic growth in most political systems and societies and the role of the media especially social media like Facebook and Twitter and Instagram and these all interact with each other and these interactions then lead to medial causes
3:46
such as stress and anxiety, the relationship between family and friends and ourselves, the technological changes that are happening in our communities and societies, changes in the job market, the opportunities we have for recreation, the free time we have,
4:04
the transportation networks that we can use and the social benefits and social welfare system that we are a part of that support us when things are going wrong such as when we lose our job.
4:17
And these lead to proximal causes such as smoking more, drinking more alcohol, having a healthy diet or not having a healthy diet, lack of physical activity or doing physical activity, unsafe sex, not getting enough sleep, becoming overweight and obese, increasing air pollution in our environment causing a range of risk factors and health problems as we'll just see in a moment and also the type of housing we live in are all examples of proximal causes.
4:51
And these lead to risk factors such as high blood pressure, high cholesterol, high glucose in the blood and high uric acid in the blood
5:00
and these then lead to heart disease, diabetes, mood and anxiety disorders, chronic muscle and joint pain, cancers, sexually transmitted infections. So we need to, in epidemiology, look at all these various causes and not just think about the direct causes of disease such as proximal causes or risk factors and we looked at this in an earlier session.
5:30
So what is causation? It's an event, condition or characteristic. So it could be an event a specific thing happening, it could be a condition that is more long-term or it could be some characteristic of an exposure or a combination of these factors which plays an important role in producing a health outcome whether that's a positive health outcome or a negative health outcome.
5:57
Logically a cause must precede or come before a health outcome.
6:09
Now what's an association? How does it relate back to a cause? Association refers to the statistical dependence between two variables, an exposure and a health outcome. Those are the two variables we're interested in. So it's a statistical relationship between an exposure and an outcome.
6:32
However the presence of an association does not imply that the relationship that we see in our study is a cause and effect relationship. It just means that we have found that there is a link, a connection between an exposure and an outcome and it may be a cause and effect relationship but we would have to do more studies and more thinking and more analysis and more gathering of evidence from biology and medicine and other areas
7:06
to be able to say for certain that one exposure causes one health outcome.
7:24
Now what is a correlation? It's a weaker form of association and so it needs to be used with care. Correlation is the degree to which two variables change together and how closely two or more of these variables are related. So it's the degree to which the closeness of two or more variables
7:51
and often when one variable changes the other will change and it could be in the same direction so if the exposure becomes higher then the health outcome becomes higher or if the exposure becomes lower the health outcome becomes lower but it could also be a negative correlation which is when the exposure is high the health outcome is low and when the exposure is low the health outcome is high that is often called the negative correlation
8:23
while a high exposure leads to a high health outcome and a low exposure leads to a low health outcome that is often called a positive correlation.
8:33
It is sometimes used as a synonym for association but it is not and it needs to be used with care. It's a very weak form of relationship between an exposure and an outcome.
8:49
This is also another nice video that will be in the playlist. Please watch it and think about it.
9:00
So what's the problem with correlations and why are they not as good as associations.
9:07
This is an example from a website called Spurious correlations and it shows the divorce rate in the US state of Maine and per person consumption of margarine in the USA and as you can see the blue line is the divorce rate in Maine and the orange line is the per person consumption of margarine and it follows exactly and so the correlation coefficient is 99% so there is a 99% closeness between these two factors, divorce rate
9:41
and consuming margarine. From this if we believe the correlation we would say that eating a lot of margarine seems to be related to having a higher divorce rate so you're more likely to be divorced if you eat more margarine. Now clearly this is such a silly relationship between an exposure and an outcome. We know that can't be true, that eating margarine cannot lead to more divorces. So this is what we call a spurious, a false correlation and we have to be really careful
10:12
and aware that these kind of relationships can occur by chance, by random error.
10:21
Check out the website below by Tyler Vegan and you'll find other examples of spurious correlations. Now one way we can test for causality and check and be more certain that an association is causal is to use Bradford-Hill's nine criteria
10:44
and that is strength or the effect size, the strength of relationship between the exposure and the health outcome and this is the larger the size of the relationship between an exposure and a health outcome, the association between them, the more likely it is a causal relationship.
11:02
Consistency or reproducibility, the same findings found in different studies by different researchers in different places with different samples of participants about an exposure and health outcome, the more likely it is a causal relationship. So the more often we have studies that show the same thing in different places, done by different people and at different times, the more likely the findings, the relationship between an exposure and health outcome is true.
11:38
Specificity, if one exposure only causes one health outcome the more specific the relationship is the more likely it is a causal relationship and there are a few things that are like that, for example asbestosis and inhaling asbestos fibres leads very strongly to lung cancer, mesothelioma and generally asbestos does not cause other major health problems. Temporality if the exposure always happens before the health
12:14
outcome then again it is likely to be a causal relationship. Remember exposure should come before the health outcome otherwise we can't say it's a cause but if exposures always happen and can never happen after someone has a health problem then it is more likely that there is a relationship, a true relationship between the exposure and health outcome.
12:42
Biological gradient, if high levels of exposure lead to higher levels of health outcome and low exposure leads to low levels of health outcome then the more likely it is that we have a causal relationship between the exposure and health outcome.
13:02
Plausibility, if there is a believable and logical biological mechanism for how the exposure
13:09
leads or causes the health outcome then the more likely it is a causal relationship. Coherence, if the relationship between an exposure and health outcome aligns and is coherent with what we already know about how the body works and how things affect the body, the more likely it is that the exposure and health outcome has a causal relationship. Eight, experiment, if there is experimental evidence from good
13:40
quality randomised controlled trials or similar experimental study designs then the more likely it is that the exposure and health outcome has a causal relationship.
13:53
Nine, analogy, if the relationship between an exposure and health outcome is similar to other exposures and health outcomes then the more likely it is a causal relationship. So these nine criteria are not used as a rigid checklist but they are used as a way of thinking generally about the strength of evidence in relation to whether an exposure leads to a health outcome.
14:26
Hence, in epidemiology we think like this: any relationship we find in an epidemiological study, whether that's an ecological study, a cross-sectional study, a cohort study, a case control study or a randomised trial is first of all likely to be because of random error, by chance, we got that result just by chance, it's not a real result, it just happened by chance, is because we had bias in our study, we didn't do the study properly,
14:57
there was some other factor, some confounding factor that created the health outcome and our exposure it was just there at the same time and actually there is no real relationship between the exposure and the health outcome, it's actually the confounding third factor that is causing the relationship and causing the health outcome. And lastly if we can cross out random error, we can cross out bias, we can cross out confounding then we are likely to have
15:27
a true association, a true relationship between an exposure and a health outcome.
15:37
So now let's move on to random error.
15:44
So what is random error? Random error happens because we use a sample, only a part of the wider target population, the population that we are interested in and because of that the sample of people in our research study will not be fully representative of the wider target population.
16:05
It will be different in some ways from that wider population even if you have chosen them randomly and even if you have a large number of participants, so there will always be random error the question is, is there a large amount of random error or a small amount of random error?
16:33
Random error affects two key things that we are interested in about the wider target population.
16:40
The level of disease that is occurring in a population: the incidence, the prevalence, mortality rate, how much illness/death is actually occurring in a population and what we will find is that the participants in our research study are likely to have a slightly higher or slightly lower level of disease occurrence.
17:03
Similarly, the strength of association, the relationship between the exposure and health outcome, what level of health outcome does the exposure produce is also likely to be different, potentially slightly higher or slightly lower than the true level of effect in terms of how much health outcome is produced, the level of health outcome that is produced in the wider population that we are interested in.
17:36
So let's look at a visual demonstration of random error and this happens in every study.
17:44
So for example, if we're looking at lung cancer and the incidence of lung cancer per 1000 smokers and looked at exposure to the number of tobacco cigarettes smoked per day,
17:57
then what we might find is a range of different studies would find different findings
18:09
and so these are eight epidemiological studies and because the participants aren't fully representative of the target population we've got a range of different values but we can see that we could draw a line through those data points and that line would be the true value of lung cancer incidence if people smoke in the adult population of England.
18:40
What you will find because of random error is that the data points, the findings will never be on that green line, it will always be above or below, near or far from that line. So we have to estimate and use statistical techniques to generate that line to show what is likely to be the true value in the population. If we had no random error then all those data points would appear on that line and we would have a nice straight line to draw through all the x's.
19:21
And we use statistical techniques to assess the influence of random error and to take account of it in every study. So let's look at an example of what happens if you have high random error in a study. One of the things that can happen is that you don't find the relationship that is there when there is one. So again using the same example of incidence of lung cancer per thousand smokers and number of tobacco cigarettes smoked per day and we have similar data points as we
20:00
had previously but they're not exactly the same so these are our eight studies and these are our findings but because we've had random error and a large amount of random error in how we measured the exposure, how we measured the outcome and also in how we selected the participants, there's error. This is not biased it's just random errors, chance events that would happen in any study no matter how good it is.
20:25
Then what we may find is that there is actually no relationship because we've got so much random error that the results don't produce the line that we saw on the previous page and so this is a false study result across these eight studies, because of random error we find that cigarette smoking does not cause, is not related, it's not linked to lung cancer and we know that's not right.
20:54
So this would be the true relationship. We should have had a positive relationship but because of random error we didn't get that relationship occurring.
21:06
And the other problem with high levels of random error is we find a relationship when there is not one.
21:14
So here is incidence of lung cancer per thousand Spider-man movie watchers and here's number of hours of watching Spider-man movies per week and again because we've got random error in our eight studies we find that there is a relationship that watching Spider-man movies and watching a lot of them seems to lead to lung cancer. Now we know from biology that there is no way that watching movies will lead to lung cancer. It might lead to heart disease and heart attacks
21:47
because you're not active but it is not likely and is very unlikely to lead to lung cancer.
21:56
So this is a false study result because of random error that watching Spider-man movies causes lung cancer, it doesn't. What it should have been is a line, a straight line across showing there's no relationship between lung cancer and watching Spider-man movies.
22:20
Random error is the difference between the real life value, death rate, relative risk or odds ratio and our study findings,
22:33
and it's because that difference is being caused by chance events. This difference is called random error or sampling error because we've sampled, we've taken a sample of participants from the wider population and that has generated this random error in our study. Random error goes down to near zero as the study sample increases. So when we have lots of participants, tens of thousands, hundreds of thousands, the random error effect gets smaller
23:08
and our study is likely to be more accurate and we are likely to find a true finding from our study.
23:19
So let's look at a little question and think about what the answer might be. So imagine we are doing a study on the population of the UK. The UK has roughly 50 million adults living in it.
23:33
We want to study eating sugar every day and heart attacks. Does eating sugar and large amounts of sugar lead to a higher risk of heart attacks?
23:44
How many participants should we recruit to get the most accurate findings that are close to the real life incidence of heart attacks from eating sugar every day? Is it 10, 100, 1000, 10000 or 100 ,000?
24:06
Well if you worked it out, hundred thousand, well done. Yes it's hundred thousand because the more participants we have, the less random error there is going to be because the larger number of participants are likely to be more representative of the wider target population we are interested in.
24:30
Random error can be measured statistically by calculating the confidence interval
24:37
and the p-value for the study findings.
24:48
Let's look at what we mean by the 95% confidence interval. Imagine we did 200,000 studies each with a thousand different participants from England on eating sugar and the risk of getting heart attacks.
25:04
What would we find?
25:08
We would find that the findings of these 200,000 studies would cluster around the true value, the true risk of heart attacks in the wider target population.
25:23
We don't have the money or time to do 200,000 studies so we use some clever statistics to estimate what would happen if we did do 200, 000 studies and this is what 95% confidence intervals and p-values are about.
25:40
95% confidence intervals mean that we have confidence that 95% of all research studies that we undertake on that topic would result in findings that would be within the range of the confidence intervals within the low number and the high number that the confidence interval gives us and we will look at this in more detail in a future session.
26:12
Another way of saying that is that we are 95% confident that the true value within the wider target population is within this range. Now this is a less accurate way of saying it but you will see this being said and it's helpful in a learning context to think about it this way but the first version the one in purple that is the more accurate version. That we are confident that 95% of all research studies that we would undertake or carry out about sugar and heart attacks would
26:44
have findings that would fall within, that would be within, the confidence interval range of values. And confidence intervals can be used with all the key measures of disease: proportions, means, rates, risk ratios, rate ratios, odds ratios, risks rates odds and risk and rate differences.
27:11
So let's look at an example, a visual diagram of what a 95% confidence interval looks like and this is for an average or a mean value in a wider target population.
27:25
So here we have body weight in pounds and the percentage of the population who have that body weight and the highest point of this normal distribution or bell curve as this is called and this is how in real life height and weight form in the community in terms of averages. You have lots of people in the middle and a few people at each end, low body weight and very high body weight. We want to find the average weight of the UK population and we don't know this value. So what
28:00
we do is we do a number of studies where we take a thousand different participants living in the UK and we measure their weight in pounds and we find that the findings of each of those studies is different. What we find is when we do lots and lots of studies something interesting happens. The studies all start clustering around the average weight value
28:34
and the confidence interval
28:41
from one study tells us the range, the low end and the high end, within which the true value for the average weight of the UK population is likely to be. So we don't need to do lots and lots of studies we just need to do the statistic called 95% confidence interval and work that out and that will give us a range between a low value and a high value within which the true average weight of the UK population will be and we can do the same with an example of risk. So here we have
29:16
risk of heart attack and amount of sugar in grams per day and this is again not a real
29:25
study or set of statistics. It's just an example for us to look at. So again we do a number of different studies with a thousand different participants
29:38
and this time the findings are a little bit closer together and so you will find that the confidence interval is narrower and that's what we're looking for. We're looking for narrow ranges for the confidence interval.
29:57
Again there's a clustering around the average risk in the population and then we do a 95% confidence interval that tells us that within these ranges the true average risk of heart attacks in the UK population, if people eat a lot of sugar, will be found.
30:25
So how do we interpret 95% confidence intervals? Well we start, in epidemiology, with something called the null hypothesis. Because we know that there is random error we assume that there is no difference between the two groups. So we don't say oh we know there's a difference or we think there's a difference, we say well there is no difference between the two groups and it's up to the study to prove us wrong.
30:51
And so what we do is we can look at a difference in the risk or odds, whether someone is exposed to tobacco smoke or sugar and not exposed to tobacco smoke or sugar and we assume there is no difference. If you smoke or if you don't smoke your odds or risks of lung cancer are exactly the same and the same with sugar and heart attacks. We assume that there is no difference, people are not going to die
31:15
more if they eat more sugar, people are not going to die of lung cancer more if they smoke tobacco.
31:23
And we can also do that with heights and weights and other types of values that we might be interested in between two population groups.
31:33
And so the confidence interval tells us, and we will look at this in a later session, whether the null hypothesis is true or false, whether there is no difference or whether there is a statistically significant difference and that means that the null hypothesis is false for this exposure and this health outcome. So the 95% confidence interval tells us whether the difference between two or more groups in terms of exposure and outcome is a true difference or one that has been
32:10
caused by random error or some other bias. P-values do the same thing. P-values or probability values, they are less useful than confidence intervals. They only tell us whether a result is statistically significant or not. They don't give an interval range. We don't know what the low value for the risk might be or the high value for the risk might be or the low value for the body weight might be on average and the high value for body weight might be on average. It just
32:40
tells you whether the result that you found in your study is significant or not, that is a true difference or not. That's why it's less useful and less important than confidence intervals.
32:52
Now the value of p that we use is developed through a history of consensus among epidemiologists and other researchers and the value they've chosen is 0.05 or 5% or less. So when a p-value for a finding is 0.05 or less than 0.05 then we say that that result is statistically significant and that this is likely to be a true finding, that there is a real relationship between an exposure and a health outcome. And the way to explain this is if we did
33:32
100 studies, the p-value tells us that only 5 out of those 100 studies would get a different result than what we have in our study, so the chances of getting another result different from our study is very small. Only five out of 100, 95 studies would find a similar result to what we have found in our study.
34:04
It's important to remember that just because we've used the p-value there's still a five percent chance that we could be wrong. So while our findings are likely to be true and apply to the wider target population, there is a small chance that our study is wrong. We can use lower p values 0.01 but that generally requires a bigger number of participants or something that causes really big health problems. So 0.05 is the best compromise we have
34:36
currently so that we can do cost-effective studies that don't cost too much and are useful.
34:46
So now let's move on to systematic error or bias and confounding.
34:56
So systematic error is a systematic change or difference caused by some flaw in the design or implementation of a study and the key types of bias are selection bias, information bias and confounding, which is a special type of bias.
35:19
If we look at an example of information bias and systematic error in collecting information we might for example be measuring how much sugar our participants are eating. Our participants are John, Olaboya, Tahmina, Susan, Gifty and Femi and we find in our study that John, Olaboye, Tahmina, Susan, Gifty and Femi
35:45
are eating a certain level of sugar but unfortunately these values are biased. We have made an error in how we have measured the sugar. Now that for example could be that we've just asked them, we haven't watched them, we've just asked them how much sugar do you eat and it may be they're underestimating or overestimating how much sugar they really eat. So the true value in this case is higher, we're saying that they're underestimating it because
36:16
they feel embarrassed, ashamed, they want to show that they're healthy and they have a healthy diet so they're not telling us the true value and so we have a systematic error, systematic bias where all our measurements are lower than what we should have measured if we were watching how much sugar they ate every day for six months. And so this bias, this error cannot be changed, we can't change the
36:42
values, we'd have to do the study again. So this is a big deal that having errors like this is a big deal because it affects the results of our study, we're not going to find any true findings.
36:54
Now remember we do have random error and what's the difference between systematic error and random error and the difference is this, that if we were going to measure sugar that people eat
37:07
and we had the same set of participants. If we had random error which we will likely have, the results that we would get,
37:16
the true values that should have been measured
37:23
and what we measure in blue
37:28
would vary randomly. So what we might find is that there's some values that we've measured that are below the actual values in terms of how much sugar people were eating and some values would be above the true values and so I hope you can see that the values average out. The high values average out the low values. We get an average value that is similar to the true value. This is the difference between random error and systematic error. Systematic error
38:03
doesn't average out. Random error, there's a potential if the error is not large, it's a small error, that the average variations that happen will average out and we will get a pretty close value to the true value or in the wider population. So this is a less important error than systematic error.
38:27
So bias is a systematic change of or difference in measurements or results from the true value in the wider population.
38:40
It's an error/flaw in the design and implementation of a study or in the collection, analysis and interpretation of the data that we have collected.
38:53
This error leads to a systematic deviation from the true value.
39:04
So now let's look at the two key types of bias that we might come across. Selection bias
39:12
which is a flaw in the relationship between an exposure to an outcome generated by the way participants are recruited into or drop out of a study.
39:29
And let's look at an example. If a UK study only recruited rich people or people who are from a single ethnic background then that study would not be representative of the wider UK population. It would be biased and this bias would be a selection bias because we've selected, we've invited, we've recruited only rich people or people from one ethnic background. Similarly, if we had a study that had rich people and poor people but over time we lost all or
40:04
most of the poor people from the study and the remaining participants were all rich or well off then that population again, even though we might have started with a representative population, what's happened over time in the study is that the participants are no longer representative of the wider target population and so we have to ask the question why did poor people drop out and what's the implication of people dropping out. What does it do to the findings of our study?
40:33
So in both cases the findings of our study are changed from what we would have found if the participants were representative of the wider target population, in this case the UK population.
40:50
How does selection bias happen? Well it happens because of the way we've chosen participants to participate in our study and so the question we have to ask is, of the people we've chosen how representative are they of the wider target population, the population of interest?
41:10
We also have to look at who has agreed to participate and who has not agreed to participate. Who responds, who does not and why is there this difference in response. And again this kind of selection bias has a different name, it's called a response bias. And this is often an issue in cross-sectional studies.
41:35
So it's really important to know if those who are selected or who respond to our invitation to join our study, if they're not similar to the target population, what is the impact. And the impact is likely to be that the findings of our study will not apply to the wider target population but only apply to the sub-population, the group of people, the types of people that are in our study.
42:07
So what happens in cohort and randomised control trials? In these two studies it's less of a problem because we select the participants before the health outcome is known. This is unlike case control studies where selection bias is a big problem but in cohort and randomised control trials the key selection bias that's important is drop out or loss to follow up.
42:33
As people leave the study because they've moved, they get bored with the research, they find it a hassle or some other reason and they've dropped out of the study, we need to understand why they've dropped out and what kind of impact that has had on the participants who remain. Has it changed the representativeness/the diversity of the participants who are continuing on in the study?
43:00
Now let's look at selection bias in case control studies.
43:05
In case control studies we always need to ask a number of questions. In terms of the controls we need to ask if the person developed the disease, would they be included, automatically become part of the cases? That's a good way of checking whether this control, the person and the criteria you are using to identify controls is appropriate. If a person got the disease and they were a control and that meant that you would put them into the group of cases,
43:34
in your case control study, then you have chosen an appropriate control and are less likely to have selection bias. And if it was a case, we would ask if the person had not developed the disease would they have been included in the control group? And by asking these two questions we're making it less likely that we have selection bias in our study. If the answer to both these questions is no, that
43:59
we would not put them in these two groups then it is likely that we have some kind of selection bias and the criteria and method we're using to select controls of cases is not appropriate, not right.
44:14
And this selection bias is more often likely to happen where we get cases from a hospital or a health clinic or a factory or office than it is if we get the cases from the wider population.
44:30
And that makes sense because obviously
44:34
hospitals, health clinics, offices and factories have only certain types of people in them. They don't represent all the wide range of diverse people that make up a community.
44:49
So let's look at an example of what the problem is with selection especially in case control studies. So imagine we were looking at cancer of the oesophagus and alcohol drinking and looking at whether alcohol causes cancer of the oesophagus and we chose cases that drank alcohol and also had cancer of the oesophagus. Now what would happen if the controls we found were men employed in a brewery? Would that be a good set of controls to have?
45:22
Well the problem is no, this would not be a good group. First of all, they're men only and secondly and more importantly they're more likely to drink alcohol compared to the wider population, so therefore they are not a normal group of people.
45:44
Now let's move on to the different types of selection bias.
45:50
Selection bias is about the people who are included in our study, who are participants in our study not being representative of the wider target population
46:04
and in cohort studies the big problem is loss to follow-up, people dropping out of the study and also the healthy worker effect that the people who do join our study tend to be healthier, that people who are ill have long-term health problems less likely to say yes I'll take part in the study because it will be more of a hassle for them. In cross-sectional studies the big issue is non-response. People not responding, not taking up the invitation to take part.
46:31
If we send out questionnaires there will be many people who don't bother to fill it out
46:37
and in case control studies it's about bias in selecting controls especially if it's hospital-based or clinic-based or factory-based and less likely if it's population-based i.e you're picking cases and controls from the whole wide population that you're interested in.
46:58
And of course bias in the selection of cases is also a problem and that is because the way we're identifying and choosing people is flawed in some way and hence we're biasing the selection of cases.
47:14
Now let's move on to information bias or measurement bias.
47:23
Information bias are flaws or errors in the measurement of exposures and/or health outcomes.
47:36
It's a misclassification of the exposure and a misclassification of the outcome, either one or the other or both and there can be random misclassification also called non-differential misclassification and systematic misclassification or differential misclassification and recall bias is an example of systematic misclassification of information. As people forget things and so the information they give is often likely to be systematically
48:09
incorrect. It's important to recognise that the exposure misclassification has to relate somehow to the health outcome and the health outcome misclassification has to relate back to the exposure otherwise it won't be a systematic misclassification. There must be an impact. By misclassifying the exposure there must be some change in the health outcome.
48:35
And often this systematic misclassification leads to an under or over estimate of the relationship between the exposure and the outcome. We don't get a true relationship. We either overestimate and say it's higher than it is or we say it's lower than it is and underestimate the real relationship.
49:04
So what kinds of random misclassification are we talking about? Random misclassification does not depend on the exposure and outcome. It's just happening randomly. There's no
49:18
thing that you could change that would prevent random errors.
49:24
And the good thing about this is that random errors tend to find no relationship or no difference between two groups. There's a zero finding and the null hypothesis comes out as true that there is no difference and this is because the random variations that these errors bring will dampen down the real relationship. It won't overestimate it or underestimate it. Systematic or differential misclassification is worse because it does, it depends on the exposure and outcome, it's
49:58
related to them, there's some systematic errors that we're making in measuring exposures and measuring outcomes and that is generating a bias that could be in any direction. It could be high or it could be low but it's less likely to be a zero finding like it is with random misclassification and that's not good because we don't want to have false findings that a relationship is very strong,
50:21
it's high, there is a big effect on health or a low finding showing that there is little effect on health when actually the true relationship is different, it's either much worse or much better.
50:36
And this impacts because we might make decisions in policy and practice about what to do if based on this kind of flawed information.
50:48
Three core types of systematic bias that we talk about is recall bias, people not remembering things about what they ate or what they did or what happened to them, recall bias and the other part of that is social desirability bias where people want to impress the researcher or they're too ashamed or embarrassed. So for example if they have lots of sexual partners and have lots of sex every day they may because of social norms say that they have fewer sexual partners and have
51:16
much less sex every day than they actually do because they're embarrassed about it and that similarly can happen with drug use where people feel that they don't want to tell the researchers about their true drug use or it can also happen in the opposite direction where people are feeling, are in pain, are not feeling great but because the researcher is taking the time and effort and they're enjoying the attention they may say that they are
51:43
in much better health than they really are, that they aren't in as much pain as they really are, social desirability bias because they don't want to make a fuss about the pain that they're in. The other kind of bias is researcher bias where somehow the researchers are purposely or unintentionally making errors in the collection of exposure and outcome data.
52:14
So let's look at an example. This is what happens if you have differential misclassification, systematic misclassification and we're using the same example we had before. So similar to what we discussed previously in an earlier slide, we have a range of measurements we've taken about how much sugar people are eating in our study but these biased values
52:42
are systematically different from the true values that we should have measured and this generates
52:51
a bias in our measurements which leads to an under or over estimate of the true value or true risk for example of heart attacks in our population because we're underestimating how much sugar people are taking in.
53:07
Now if it was a random misclassification then like random error in a previous slide, the errors are randomly distributed so some could be higher, some could be lower
53:32
and so the errors average out across a large number of participants. In this case what tends to happen is that we don't find anything. We find zero findings. The null hypothesis is true that there is no difference between the participants that we are studying.
53:54
Random misclassification tends to lead to us not finding a relationship when there is one. So if there was a relationship between lung cancer and smoking, large amounts of random misclassification or measurement errors or biases in the information that we're collecting is likely to lead to not finding a relationship even though there is a relationship in the population.
54:26
So random misclassification pushes findings towards zero, no result or no difference found,
54:37
the null hypothesis.
54:44
With systematic miscapture we will find a relationship even when there is no relationship similar to high levels of random error but this is much more likely because we have systematic errors, systematic misclassification. We find a relationship between watching spider-man movies and having lung cancer which we would find quite unbelievable given that there is no biological relationship between the two and so rather than finding no relationship
55:21
the systematic error for some reason has created a false positive relationship.
55:31
So systematic misclassification gives a false finding, a difference that is not really there. This can be higher or lower, an overestimate or an underestimate of the true value in the wider target population.
55:49
And the two key aspects that we look at when we're looking at information bias and measurement devices and instruments like questionnaires is validity. Are we measuring what we're intending/ want to measure, and reliability, are we consistently measuring the same thing with the instrument so that if we repeat the questionnaire with the same participant and there is no difference in their status, they are exactly the same as the last time we talked to them
56:21
then the questionnaire or measuring device should produce the same results. And validity and reliability are really important in relation to measuring exposures and outcomes accurately
56:36
and reliability is not just in terms of the device or the questionnaire that's wrong it can also happen because the researchers, different researchers, are doing different things. They're not following the protocol that they should, the training that they've been given and hence there are differences between researchers in what they are collecting and it could also be that the same researcher is using different techniques at different times and therefore there
57:03
is differences in the same researcher over time in how they are collecting the exposure or outcome information. And one visualisation of this is that if we are looking at a target, an archery target, and the bullseye is what we are intending to measure then we can have a situation where the questionnaire we have is measuring something but not what we want so it's not valid but it is reliable i.e it is capturing information in a consistent and reliable way.
57:41
The second example would be where the questionnaire is valid but because the way it is being used by researchers or the way it is interpreted by participants means that the results are not reliable. So even though the participant is in the same situation as they were two weeks ago the person is filling the questionnaire out in a completely different way. So we don't get anywhere near a single clear result. It's a set of random findings because even
58:19
though the questionnaire is a good one because we're not measuring and collecting information reliably from that questionnaire, from that measuring device, weighing scale or blood pressure monitor or whatever it might be, then this is not giving us good results. It's giving us inaccurate results that are likely to be systematically misclassifying exposure or outcome. And the last one is what we want which is a valid measure that is measuring
58:48
what we want, the bullseye and it's reliable because when it's used on different participants and the same participants, it's producing the same results and the same findings.
59:03
Now this is an example of the different types of specific errors that can occur. I'm not going to go through it, this is something for you to pause the video at and have a closer look but I'm going to emphasise the five main categories. Errors in the design of the instrument/ the questionnaire or the measuring device that we're using. Errors in the protocol or the training of how to use the instrument by researchers or by participants if it's sent by post.
59:33
Errors in how the study is implemented and how the study is carried out, so it's a wider problem than just the instrument, the questionnaire. Limitations of the participants and errors when we're entering the information on a computer and when we're analysing that information.
59:57
So now let's move on to confounding. So what is confounding? It's a distortion in measuring the effect of an exposure on an outcome, a health outcome, because the exposure is linked to other factors that also leads to the health outcome.
1:00:20
The relationship between an exposure and an outcome is therefore generated, explained and accounted for by a third factor that affects the outcome but is not affected by the exposure but has some relationship to it.
1:00:41
So let's look at a visualisation of this. So we have an exposure
1:00:45
that leads to an outcome. We may have a mediating factor something that acts between the exposure and the outcome and then we have this third factor, a confounding factor,
1:01:02
that also causes the outcome and has some relationship to the exposure.
1:01:12
Now you might wonder what is a mediating factor. So a mediating factor could be for example the exposure is high levels of sugar and the mediating factor would be high levels of blood glucose or high levels of triglycerides or cholesterol in the blood that then leads to the outcome. So the mediating factor is on the pathway between the exposure and the outcome. So the exposure of sugar is leading to the high blood glucose which then leads to the outcome.
1:01:43
I's not a different factor, it's part of the pathway. A confounder is a completely different factor and the confounding factor must be linked in some way to the exposure.
1:01:55
The confounding factor must cause the outcome and the confounding factor is not on the causal pathway between the exposure and outcome ie it's not a mediating factor.
1:02:11
Now let's look at an example. So here is a very simple example, physical inactivity and heart disease. So there's a relationship between not doing a lot of exercise or walking and cycling and other activities and increased risk of heart disease and heart attacks but there is a factor. The confounding factor in this relationship is age. So if we don't take account of age what we would find is we would get a relationship that is not about physical activity but is about age
1:02:41
because age as we get older we have an increased risk of heart disease regardless of whether we are physically active or not so we do need to take account of and that's why in our questionnaire or in our interview with participants we would want to make sure that we look at age. Can you think of some other key factors that are confounding factors that might affect a relationship between physical inactivity and heart disease or between some kind of other exposure and heart disease?
1:03:15
If you said or guessed ethnicity, sex as in male or female or socio-economic group,
1:03:24
whether you have lots of money or whether you have little money or education, whether you have high levels of education or you don't have high levels of education, these are all confounding factors that we need to take account of and we need to measure and we need to collect information on to make sure that we are measuring the exposure and outcome relationship accurately.
1:03:49
But what kinds of things might not be confounding factors?
1:03:53
Well one example would be physical activity and coronary artery disease or heart disease as we talked about previously and arthritis causing painful joints. So here there's a relationship between the exposure and arthritis, physical inactivity and arthritis because if you've got painful joints you're less likely to be physically active, you're going to be inactive and that of course leads to coronary artery disease but that arthritis causing painful joints
1:04:23
is not linked to coronary artery disease directly. That is arthritis does not cause heart disease, so therefore it's not a confounding factor. It's just another factor that explains the exposure, physical inactivity, and why a person might be physically inactive.
1:04:44
Another example would be if we only had 80 year olds or participants who were between 80 and 89 years of age then age would not be a confounding factor because we are picking participants who have similar ages so there isn't the same relationship. Yes, age affects heart disease and the risk of heart disease. The confounding effects of age have been removed because 80 to 89 year olds will have a similar risk profile for heart disease. So the only active factor affecting
1:05:20
80 to 89 year olds will be the level of physical activity or level of physical inactivity.
1:05:31
Now let's look at another example. There was a study done looking at Down Syndrome babies per thousand live births and birth order. So what we can see here is when women had a lot of babies, four and five, then we see a very big increase in the risk of having Down Syndrome babies. So from this we might say ah women who have lots of babies are more likely to have babies with Down Syndrome. But hang on, is that, how does that work, what
1:06:03
is it about birth order that is causing that, it can't just be birth order can it?
1:06:10
Then we do a study with Down Syndrome and look at mothers' ages and we find women who are older 40 plus or 35 to 39 have much higher risk of having Down Syndrome babies than those who are much younger 34 years or less.
1:06:29
So which one is the correct causal relationship?
1:06:36
Well if you chose mother's age as the key causal factor then well done. Mother's age relates to how many babies a mother will have because the more babies you have the more likely you are to be an older mother because you've had babies when you were younger and after you get to four or five babies it's unlikely that you're going to be in your 20s. W.hen you have your fourth or fifth child it's more likely you're going to be in your 40s.
1:07:09
So how do we deal with confounding?There's a few broad ways we do it. We randomise. That's why randomised controlled trials are so good because it reduces confounding. We can match participants. We can match by age or by sex or by ethnicity to reduce confounding and we can restrict it. We showed in the previous example we could restrict the age group of people who are invited and recruited into our study or we could recruit women and not men or men and not women
1:07:37
to restrict, to reduce the confounding effect and we can deal with confounding afterwards in the analysis stage by stratifying by age group hence you can see that we collect age and then when we produce the results we normally create five to ten year age groups and we stratify the findings by age groups so that we account for the impact of age. Similarly we might look at men and women and the findings for them separately as well as overall to see if there's any differences
1:08:12
and the same for different ethnic groups and we can also do some standardisation in terms of what we are looking at or measuring against standard population and we can use some statistical modelling to take account of confounding by putting in all the factors and then doing things like regression modelling to find out what each factor, each factor's effect is on the health outcome so we can work out exactly which has a bigger effect and which has a smaller effect across
1:08:45
all the various factors confounding factors as well as the exposures that we are interested in.
1:08:55
Let's have a quick recap of random and systematic errors. So random errors are by chance. They're part of the variation in measurement that is not connected to any other measurement and occurs by chance. It's not connected to the exposure or the outcome.
1:09:14
Systematic errors on the other hand are systematic and consistent inaccuracies in the measurement that occur in one particular direction and is often linked to a specific source like a faulty or uncalibrated equipment. If you read some of your papers where they're using measuring devices, they will explain in great detail how they've calibrated the device and what kind of device it was and the name and the details of the device and it could
1:09:40
also be because of the way you are collecting the data and the way you are doing the measurements because the researchers are not following the protocol of how to take blood pressure or how to take body weight measurements or it could be because of systematic errors in people remembering facts about the past such as how much they ate, how much they drank, how much they smoked.
1:10:10
So the sources of error are two types, systematic errors which fall into three broad groups: selection bias, information bias and confounding and then random errors which are the general errors that affect all study populations and all study designs and systematic errors can be managed and controlled at the design stage that is the best place to do it and during the analysis. This is for confounding if we've collected information on confounding
1:10:42
this is much more difficult when we've got bias and information and generally we need to control those two things, selection bias and information bias, at the design stage.
1:10:56
And when we look at the different types of studies we can do a little table like this showing that ecological studies have high levels of confounding. Cross-sectional studies have high levels of information bias and medium levels of other types of bias.
1:11:15
Case control studies have high levels of selection bias and information bias because they are looking back into the past and they are selecting cases and controls which are difficult to identify and choose. Core studies have high levels or risks of selection bias.
1:11:35
And randomised controlled trials have
1:11:39
medium levels of selection bias, dropout rates and information bias.
1:11:47
I hope you enjoyed this session. Think about chance bias and confounding in the articles you're reading for your summative assignment and I look forward to seeing you in the apply session. Bye Bye.
Here to help you learn
Sign in to keep a private conversation with your course tutor.
Ask me anything about the material…
Sign in to chat