Precision and decision making

Yes thanks for pointing us back to the original question. When you mentioned sensitivity and specificity those were unnecessary to the discussion.

While having good precision in estimating ORs (and adjusted regression effects in general), the discussion really needs to be on the absolute risk scale with full adjustment for all relevant clinical variables. Then one can simply plot post-test probabilities against pre-test probs. to see relevance (where “test” = new information such as genetics) for clinical decision making, as is shown here.

I made is explicit in my original post that this is not the question. The question is, given a defined clinical threshold based on a risk, how precisely do we need to estimate this risk. I can only suggest you go back to my original question and reread it carefully.

I see your point now, and I believe we’re in the same page. This is my overall take: A precise OR would be best for assessing whether the new genetic test adds to the “standard” clinical data in predicting who will get the disease. This is a necessary step, before determining how make the best use, from the clinical perspective, of an improved risk score scale. Some form of cost-effectiveness or value of information analysis should be used in the latter step.

I’ll repeat, there is no cut point. This is about risk estimtion in women with a particular genotype.

I went over your question again. This is what through me off course

And then you added:

The emphasis on when to offer prophylactic surgery make me jump into “what information does one need to figure out if prophylactic surgery should be offer to patients with a positive genetic test?” And my guess is that one needs to know the predictive properties of the test, what proportion of the population has a positive test, what is the baseline incidence of the disease, and what are the costs and benefits of an oophorectomy. Basically, one needs to be in the position to make a cost-effectiveness analysis.

I apologize for misunderstanding your question. Still, the exchange has been useful for me and, hopefuly, for others.

I see precise estimation of an effect (e.g., an OR) as a first step, but impact for decision making cannot be judged using relative measures. Instead alterations in absolute risk distributions due to knowing the new information is more central to the original question. The assessment of added information needs to come from things like those discussed here.

Indeed, it is the absolute risk that is relevant to the decision - as I think I made clear in my original question - but the question remains. The distribution of the likely absolute risks and how certain we would like to be that it is above a given threshold is conceptually similar. In germline genetics, absolute risk estimates often based on application of odds ratios as approximations of relative risks to population rates.

What is the “lay” translation of the meaning of “Protein truncating variants were associated with an increased risk with an odds ratio of 3.0 (95% CI 1.6 – 5.7)”?

Does this statement mean that case-control studies showed that the “odds” that women with breast cancer were found to be carrying the allele in question were approximately 3-fold higher than the “odds” that women without breast cancer were found to be carrying the allele?

You seem to be highlighting uncertainty around how to interpret the confidence intervals around results from gene association case-control studies. Specifically, such studies might identify many alleles that appear to be “associated” with ovarian cancer. I think what I’m hearing you say is that point estimate magnitude from these types of studies is typically used as the primary guide to decide which alleles might warrant a closer look. And I think you’re asking whether a point estimate with a wide confidence interval should be considered as compelling (with regard to deciding about “next steps”) as a point estimate of similar magnitude but with a narrower confidence interval?

Since I’m not trained in this area, I only have a vague understanding of what the “next steps” might be after deciding which alleles warrant a closer look. Clearly, there are many other considerations that might factor into estimating an individual woman’s absolute risk of developing ovarian cancer, quite apart from consideration of whether she carries any particular allele…But I don’t think you’re asking about next steps- rather, I think you’re asking about the best way to “triage” all the ORs you might get from gene association case-control studies.

This is a bit of a late reply, Frank, but I believe that in specific circumstances sens and spec for prognostic questions make sense.

Assume the following. You have in mind women at risk of developing breast cancer, and have a prognostic model/marker. You have developed a more intensive screening program, to detect cancer early. You are willing to invite marker positive women to your program.

In this case, (prognostic) sensitivity expresses the proportion of women who will develop breast cancer that are in your program, and specificity is the proportion of women who will not develop breast cancer and are not invited to your more intensive program.

In this case, I am considering the cumulative 5-year sensitivity and the dynamic specificity.

For guiding decision making, these will probably be helpful statistics. Of course, a fully informed decision about more intensive screening will never rely on just sens/spec, but on the full range of consequences.

Patrick

Since sensitivity and specificity condition on outcome status, they are only applicable to retrospective studies such as case-control studies. And they require test results to be purely binary and outcomes to be purely binary.