Clarification requests from KEYS

Clarification requests from KEYS

by Laura Koth -

KEY #2:

pg.7: second paragraph: I had a question about the first sentence of the second paragraph: For measurement error to occur, how we measure AC-use would have to be systematically linked to how we measure prostate cancer-specific mortality (see below). Question: the wording sounds like there is a requirement for exposure errors to be linked to outcome errors in order to have measurement bias. Is that what you are saying? I was under the impression that you could have bias if there were errors in the measurement of exposure OR outcome and still get bias. is that not true?

pg. 7: 3rd paragraph: can you elaborate on what you mean by "extra person-time in the exposed group" and how that causes immortal time bias?

pg. 7: can you expand on how prevalent users causes selection bias. I see the DAGS in the slides but conceptually not getting it. maybe it's just something you memorize?

pg.8 can you provide the calculated risk ratio so that I can work through the calculation?

pg. 10, first sentence: what do you mean by this phrase: It is concerning that they included prevalent users of AC because any baseline covariate included in the analysis should be carefully excluded as a potential mediator.

KEY #3:

pg 2: confused by the terminology. in the paragraph with blue text after question 2, there is a discussion about primary measures of association. Should we really be calling it "cumulative incidence" if there is no person time in the measurement?

last pg: part c. I don't follow why the statement "Association is not causation and patient X and W are not exchangeable." is the explanation to this question;


REGARDING OFFICE HOURS: for those of us who have other classes that conflict with current times of office hours, can you try to have office hours about 30 min before class?


Re: Clarification requests from KEYS

by June Chan -

Hi Laura,

I can be in the Epid kitchen tmrw around 12:30-1pm, before the Epi Tools workshop, for an ad hoc office "half hour". If you come with questions, we can try to use one of the rooms off the kitchen. So, please look for me either in the kitchen, or in one of the huddle rooms right off the kitchen on the 2nd floor. If we can't use one of those rooms, please look here and we'll post what room we land in.

Also, please note, François Rerolle changed his office hours on 2/20 and 2/27 to be at 12-1pm, to try and minimize conflicts others had mentioned. 

Monica, François, and I are working on responses to your questions above about the Keys and will get back to you soon. 

Best wishes,

JMC


Re: Clarification requests from KEYS

by Monica Ospina Romero -

Hi Laura,

Thank you for your questions. Our response to your question is below. Please notice that I am only answering your questions for HW2 Key. Francois will be replying to your questions related to HW3 soon.

pg.7: second paragraph: I had a question about the first sentence of the second paragraph: For measurement error to occur, how we measure AC-use would have to be systematically linked to how we measure prostate cancer-specific mortality (see below). Question: the wording sounds like there is a requirement for exposure errors to be linked to outcome errors in order to have measurement bias. Is that what you are saying? I was under the impression that you could have bias if there were errors in the measurement of exposure OR outcome and still get bias. is that not true?

Thank you for your question. I think it would be better if it read as:  “For measurement error bias to occur, how we measure AC-use would have to be systematically linked to how we measure prostate cancer-specific mortality….”

There are two types of measurement error: random error and systematic error. When the error is related to true values of other variables in the analysis, this measurement error is a systematic error. We also call it differential misclassification or measurement bias.

Random error is unrelated to the actual true value of what we are trying to measure or true values of other variables. This type of error is the one that we call non-differential misclassification. 

In question 7 of HW2, we were looking for an answer that explains whether you think measurement error occurred, and if so, if it is leading to any bias. If there is bias, then we were looking for the type of measurement bias (differential or nondifferential misclassification), and the scenario that created this error. If you think it is a systematic error, you should explain how the true value of other variables (the outcome or other variables that cause the outcome) influence a measurement error in the exposure.

I think the slides on measurement bias from Epi Methods 1 (EPI 203) is a good resource to study measurement error if you would like to review this topic.

pg. 7: 3rd paragraph: can you elaborate on what you mean by "extra person-time in the exposed group" and how that causes immortal time bias?

Remember the formula for Incidence rate (IR) = number of events / person-time at risk

We know that people in the cohort could start using anticoagulants (AC) at baseline or at any time during follow-up. When the authors classified the exposed group as ever use, all the follow-up time of each participant who started AC will go to the denominator of the incidence rate in the exposed group. This follow-up time includes the time when they were not taking AC and the time after they started ACs. When we talk about the extra person-time in the exposed group, we are referring to this time before they were exposed (started ACs). This time should not be included in the denominator of the incidence rate in the exposed group because they were not taking this medication. Since they include this extra person-time in the denominator, they are underestimating the incidence rate in the exposed group.

This causes bias when estimating risk ratios because the unexposed group does not have the same extra person-time.

pg. 7: can you expand on how prevalent users causes selection bias. I see the DAGS in the slides but conceptually not getting it. maybe it's just something you memorize?

No, you don’t have to memorize that.

I am going to give an example using diagrams from epi methods 1. Let’s say all prevalent users (people who were taking ACs before the baseline of the study) started taking ACs 1 year before baseline, this is the left side of the figure (with the arrow). The baseline for the study is in red. As you can see with the figure, some of the people that were taking ACs who potentially could be included in the study had the event before the baseline, therefore they cannot be included in the study. It is possible that the people who were actually included in the study have a different risk for the outcome than the people who were not included because they experienced the outcome earlier. Ideally, we would like a sample of all the cohort of people taking ACs. To prevent this selection bias, we should exclude prevalent users from the analysis, as in a randomized trial.

pg.8 can you provide the calculated risk ratio so that I can work through the calculation?

Remember that the risk ratio = ratio of 2 cumulative incidence estimates.

The 7-year cumulative incidence in the ACs group was 1%

The 7-year cumulative incidence in the reference group was 3%

Then Risk ratio = 1% / 3% = 0.33

pg. 10, first sentence: what do you mean by this phrase: It is concerning that they included prevalent users of AC because any baseline covariate included in the analysis should be carefully excluded as a potential mediator.

Variables that change over time (e.i. BMI) can be confounders or mediators depending on when they were measured. For BMI, at a one-time point, this variable can be a cause and, in another time point, a mediator. One way to avoid this problem is to measure these type of variables before the exposure. In that way, one can be sure that is not the exposure that caused a change in BMI.

When we use prevalent users, the exposure happened before we measure BMI, then we are not sure this variable is a confounder or a mediator.

 

 

 


Re: Clarification requests from KEYS

by June Chan -

HI Laura, please see responses below for your questions regarding HW3 key, mainly drafted by François. I also will be in the Epid kitchen today around 12:30 if you have follow up questions.

KEY #3:

pg 2: confused by the terminology. in the paragraph with blue text after question 2, there is a discussion about primary measures of association. Should we really be calling it "cumulative incidence" if there is no person time in the measurement?


Comment:  Please refer to Epi 203. In cohort study, there are two ways to measure incidence: 
  • Cumulative incidence. When comparing 2 groups, the ratio of 2 cumulative incidences gives the cumulative incidence ratio, often better known as risk ratio or relative risk
  • Incidence rate. When comparing 2 groups, the ratio of 2 incidence rates gives the incidence rate ratio, also called relative rate
If there are no person time measurement, you can still measure the cumulative incidence.

last pg: part c. I don't follow why the statement "Association is not causation and patient X and W are not exchangeable." is the explanation to this question;


Comment:  
The thought process should be as follows:
  • First, you calculated the true average causal effect (1.05) in b) because you were lucky enough to have access to the counterfactuals
  • Second, in c) we ask you to imagine that you have access to only 1 scenario, which is what would happen in real life. You would therefore compute the observe association (0.7 in scenario 1 and 1.4 in scenario 2).
  • Third, comparing your observed association to the true average causal effect, you realize that they are different and that it would have therefore been wrong to jump from association to causation in b), when you only had access to 1 scenario, I.e you performed a real study
  • Last, you realize that this makes sense since association is not causation and that association measures should always go under high scrutiny before making causal claims. In particular, bias can arise from lack of exchangeability between patients X and W (since the outcome of X under treatment (or no treatment) cannot be used to represent the outcome of W under treatment (or no treatment) and vice versa)
We hope this helps!