The future of the (M)SCCS method

[schuemie (Martijn Schuemie)]

We’ve started working on the SCCS R package, but there are many questions still open.

In the original MSCCS design we added all drugs to the model, we fitted the model once, and the beta for each drug was the relative risk estimate. We could however decide to fit the model once for every drug, and treat the drug-of-interest differently from the other drugs, which can be considered covariates. For example, we could define the risk window for the target drug to be only the length of exposure, but add x days to the end of exposures of other drugs. That would create a memory of drugs taken in the past to be used in assessing the risk for a particular day. We could add different covariates for the same drug, with different persistence windows, and we could add conditions, procedures, visits, observations, etc. to the covariate list. We could specifiy different persistence windows for the different types of covariates (drugs, procedurese, etc.), or even per concept. We can go nuts here, but should we?

Another issue is picking up causal intermediaries. One example that comes to mind is when I was trying to fit a model for UGIB, and the drug with the highest coefficient was the drug you take before your esophagogastroduodenoscopy to confirm the UGIB diagnose. Clearly that drug didn’t cause the UGIB, but it is masking whatever is. How do we detect these, and remove them from the model?

Yet another issue is contraindication. If a drug is contraindicated for a particular outcome, it will appear to have an increased relative risk because the outcome will be less prevalent before iniation of the treatment (else the doctor would not be following guidelines). Some people remove a period of observation time just prior to treatment initiation to get around this (e.g. Tata et al), but this doesn’t seem very optimal.

Lastly there’s the issue of outcome occurrences effecting the probability of future occurrences and/or end of observation. Some options here are censoring on date of first occurrence, eliminating all but the first occurrence (but keeping all observation time), or modeling the dependency.

I would hereby like to kick-off the discussion around these topics. Feel free to add! (especially @msuchard, @tshaddox, @David_Madigan, @Patrick_Ryan)

[nicolepratt]

In response to your contraindication question, I feel that it is important to partition the pre-exposure risk period out of the unexposed risk period. This becomes really clear in the opposite case to the one you mention. I have published on cases where the outcome of interest is a hospitalisation event and the event itself predisposes patients to be prescribed the medicine of interest (see risk of antipsychotics and hip fracture Pratt et al Drug Safety). I have suggested previously that an appropriate pre-exposure window here would be the length of time of an average hospitalisation stay. If the medicine of interest is likely to be initiated in hospital for a particular outcome, it will appear to have an “decreased” relative risk because the outcome will be “more” prevalent before initiation of the treatment.

Also, in response to your question around outcome occurrences effecting the probability of future occurrences - it is sometimes possible to do a sensitivity analysis in which the SCCS analysis is repeated in those patients who survived to the end of the study period.

Also, I’m not sure I understand the problem with defining risk windows for the targeted drug. Would you not always want to define these windows based on the length of exposure of the medicine of interest?

[schuemie (Martijn Schuemie)]

Interesting! I had not considered the case where the outcome might increase the probability of receiving the treatment.

I really like the approach in your paper, where you keep the data prior to treatment initiation, but basically give it its own beta (in contrast to the earlier mentioned paper by Tata et al. where this time was removed from the analysis altogether). I also like the fact that you used multiple periods instead of proclaiming that one must be the right one. We could even use an approach similar to earlier work where we fitted splines to the relative risk as a function of accumulated time of exposure, and extend the spline to time prior exposure.

An alternative approach would be the one proposed by Farrington, using the so-called ‘pseudo-likelihood’ I have no idea which approach would be better. Sounds like something we could evaluate in an experiment.

Concerning the risk windows: in our MSCCS approach we include all drugs in the analysis at once, so thousands of different drugs and corresponding betas to estimate.

For the drug that you really want to estimate the risk for, you typically use the length-of-exposure as risk window. However, if you had strong evidence the effect must be acute, you can redefine your risk window to just the first x days after start of exposure. Similarly, if you believe the time to detection of the outcome may be long you can redefine the risk window to a time longer than the length of exposure.

For the drugs that are included as ‘covariates’ (which might be proxies for underlying time-dependent factors affecting the risk) all bets are off. How you define the ‘risk window’ all depends on how the covariate is associated with time-dependent risk.