0
$\begingroup$

I'm conducting a study examining the association between exposure to particular prescriptions and the odds of subsequently getting a disease (binary outcome). In my design, I have patients who get the disease and do not get the disease. Because I'm examining the relationship between the drug exposure and the disease, I am counting all individual prescriptions of the drug up to 1 year before the diagnosis of the disease. So the temporal cutoff is 1 year before the diagnosis of the disease, and any prescriptions after this time are not counted.

My question is: what temporal cutoff should I use for the patients who do not get the disease? The reason I ask is that if there is no temporal cutoff for the non-disease cohort, they naturally end up having more prescriptions in their record, and I believe this biases the analysis.

$\endgroup$
5
  • $\begingroup$ When you say "In my design, I have patients who get the disease and do not get the disease." do you mean you sampled patients with disease and an appropriate matched set of non-diseased patient "controls"? Did you match on other factors (age, sex, ...). What is the natural history of this "disease"? Is there a preclinical window where patients are "detectable" but asymptomatic like with cancer or herpes? $\endgroup$
    – AdamO
    Jan 31, 2020 at 18:25
  • $\begingroup$ I did not do any matching, I included age, sex, race and other variables as covariates. The disease is primarily seen in patients after age 65, with early onset being seen in patients as young as 30s-40s. $\endgroup$ Jan 31, 2020 at 18:30
  • $\begingroup$ I'm afraid there's a subtly to "preclinical window" that needs addressing. Age is not important, but the lag from infection, possible screening/detection, and symptoms is what matters. $\endgroup$
    – AdamO
    Jan 31, 2020 at 18:39
  • $\begingroup$ I see, so you would discard age as a variable? The disease is neuropsychiatric and probably wouldn't be screened for unless symptoms were reported. Would you represent time lag as the time from symptoms reported to diagnosis, or something else? $\endgroup$ Jan 31, 2020 at 20:38
  • $\begingroup$ In all respect, it's an analysis that will require some statistical consultancy to do correctly. If your institution has a statistics center, I would bring a full data codebook and discuss "marginal structural models". $\endgroup$
    – AdamO
    Jan 31, 2020 at 20:48

1 Answer 1

0
$\begingroup$

There are two possible explanations for a possible association between initiating (or ongoing use of) medication and disease status in a cross sectional sample. One optimistic link is a direct causal association. The other, a source of bias, is confounding by indication. That is, the research participant was predisposed to disease and initiated a treatment either as prophylaxis or prevention. As an example, a 40 year old woman with a family history of breast cancer may initiate preventative tamoxifen, hence the crude, or unadjusted, association would suggest that tamoxifen increases breast cancer risk. That would be false.

Unfortunately, we can in no way be assured that a longer lag between initiating therapy and disease status diminishes this bias. Alas, this is exacerbated by the limitations of the design, a cross-sectional study.

  • Beware if medications are ascertained by self-report and are retrospective or other limitations of EHR, even then compliance is of concern
  • Disease may feature lead-time bias, so the "non-diseased" sample may simply be in the clinic/hospital less, and hence have fewer medications, similar to a type of Berkson bias

The best design to inspect the clinical development of disease and preliminary use of medication is a prospective longitudinal design where the indications of therapy are known. The Cox model with time-varying covariates based on initation of treatment and indicators of that treatment. Alternately, marginal structural models. In both cases, a risk model for treatment should be known. If indications of therapy cannot be characterized then there is no unbiased analysis.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.