Problem of selecting controls (a study design question)

[Zihan Dong]

Hi!

I encountered a problem of exclusion of my study population. I used a nest case-control design.

The study population of my study is patients with disease A, the outcome is patients with disease B diagnosed at age >=18 years old. The baseline is diagnosis of disease A, the index date for case and corresponding controls is diagnosis date of disease B.

I was wondering, for the study population diagnosed with disease B under 18 years old, should I delete them or consider them as potential controls? I felt if I delete them, I kinda use information from the future, because in reality we don't know whether one person will develop disease B or not at baseline. If I keep them as potential controls, the study population diagnosed with disease B under 18 years old cannot develop disease B in their adulthood (as we only look at the earliest date of diagnosis of disease B), which also sounds incorrect.

Does anyone have any thoughts about it?

Thank you!

[Clyde Schechter]

If I understand your research goal correctly, people who are diagnosed with disease B before age 18 are not at risk for your outcome: incident disease B after age 18. Consequently, regardless of their exposure status in your study (whatever the exposure you are testing is), their outcome is predetermined to be negative, and is completely uninformative about the association of the exposure with the outcome. Consequently, those who were diagnosed with disease B before age 18 must be excluded from the study completely: they cannot serve as either cases or controls in your study.

This is not "using information from the future." By definition, they cannot be cases of incident disease B after age 18 based on their present information of already having developed disease B. To take a silly, but clear, example, suppose disease B were amputation of the left leg after age 18. If a person has already had his or her left leg amputated before age 18, you can infallibly deduce from that fact that they cannot undergo left leg amputation after age 18--consequently they cannot provide any information about risk of left leg amputation after age 18 in association with anything. They would have to be excluded entirely from any such study.

[Zihan Dong]

If I understand your research goal correctly, people who are diagnosed with disease B before age 18 are not at risk for your outcome: incident disease B after age 18. Consequently, regardless of their exposure status in your study (whatever the exposure you are testing is), their outcome is predetermined to be negative, and is completely uninformative about the association of the exposure with the outcome. Consequently, those who were diagnosed with disease B before age 18 must be excluded from the study completely: they cannot serve as either cases or controls in your study.

This is not "using information from the future." By definition, they cannot be cases of incident disease B after age 18 based on their present information of already having developed disease B. To take a silly, but clear, example, suppose disease B were amputation of the left leg after age 18. If a person has already had his or her left leg amputated before age 18, you can infallibly deduce from that fact that they cannot undergo left leg amputation after age 18--consequently they cannot provide any information about risk of left leg amputation after age 18 in association with anything. They would have to be excluded entirely from any such study.

Thanks for your reply! However, we don't know whether ppl will develop disease B (either before or after age 18) or not at baseline, so that we cannot remove the study population based on future information. We decided to include all the cases with disease B (without age restriction) to solve this problem.

[Clyde Schechter]

I'm confused. In #1 you said

If I keep them as potential controls, the study population diagnosed with disease B under 18 years old cannot develop disease B in their adulthood (as we only look at the earliest date of diagnosis of disease B)

If that is correct, then by your definition of the outcome, developing disease B after age 18, looking only at the earliest date of diagnosis of disease B, then those who developed it before age 18 cannot have the disease B after age 18 outcome.

In addition, it seems to contradict what you say in #3:

We decided to include all the cases with disease B (without age restriction) to solve this problem.

Please clarify.

[Zihan Dong]

I'm confused. In #1 you said


If that is correct, then by your definition of the outcome, developing disease B after age 18, looking only at the earliest date of diagnosis of disease B, then those who developed it before age 18 cannot have the disease B after age 18 outcome.

In addition, it seems to contradict what you say in #3:



Please clarify.

those who developed it before age 18 cannot have the disease B after age 18 outcome.

This is actually the main problem for my study design when I asked that question. My study population is people with disease A at any age, setting outcome as disease B over 18 will make the study population during follow-up under 18 (no matter with or without disease childhood disease B) have no chance of developing the outcome (i.e., adulthood disease B) (immortal time bias). This is incorrect in nested case-control study, because in this design the underling cohort(potential controls) should be at risk of the outcome at any time-point. In other words, the person-time under 18 years old cannot contribute to the risk set.

To solve this issue, we should either restrict our study population to 18+ years old, or retain the age range of our study population but include all cases with disease B without any age restriction as outcome. Both methods make all the study population at risk of the outcome at any time-point.

[Clyde Schechter]

To solve this issue, we should either restrict our study population to 18+ years old, or retain the age range of our study population but include all cases with disease B without any age restriction as outcome. Both methods make all the study population at risk of the outcome at any time-point.

Yes, I completely agree that you should do one of these two things.