Statisticians: does the way we determine genetic health risks work for very small or large families?

In My Humble Opinion
1

For example:

[QUOTE=Angelina Jolie's Mastectomy: BRCA Testing in the Spotlight - Page 2]
Although there are no standardized criteria for selecting candidates for BRCA counseling, the National Cancer Institute, the National Comprehensive Cancer Network, and the US Preventive Services Task Force outline family history red flags, which generally point to first- and second-degree relatives with breast and/or ovarian cancers, especially at young ages:

2 first-degree relatives (mother, daughter, or sister) diagnosed with breast cancer, one of whom was younger than 50 years

3 or more first- or second-degree relatives (aunt or grandmother) diagnosed regardless of age

combination of first- and second-degree relatives diagnosed with breast and ovarian cancer regardless of age

first-degree relative with bilateral breast cancer

breast cancer in a male relative

combination of 2 or more first- or second-degree relatives with ovarian cancer

For women of Ashkenazi Jewish descent: any first-degree or 2 second-degree relatives on same side of family diagnosed with breast or ovarian cancer

[/QUOTE]

One thing I wonder about the rubric described for determining breast cancer risk and the need for further screening: how to we, or do we, amend that screening for women who come from very small or very large families? The “3 or more second degree relatives,” for example… I only have three female second degree relatives, and two of them (my paternal grandmother and her daughter, my aunt) have breast cancer. That’s only 2, and we’re not Ashkenazi, so by that rubric, I’m not at increased risk. I’m an only child, so I only have 1 adult first degree female relative (my mother) and she doesn’t have it (yet), so by that rubric, I’m not at increased risk. But that’s hardly much of a sample size.

I can’t help but think that this might statistically need to be looked at differently than a woman who has, say, 10 first and second degree relatives and only 2 have cancer.

Or, conversely, if a fertile family of 20 girls sees 2 of them develop breast cancer, are their sisters really at greatly increased risk?

It seems that given the very small or unusually large sample size that some families represent, a hard number of relatives with cancer is a poor rubric, and perhaps a percentage might be of more use.

But IANAStatistician, so maybe I’m all wet. Thoughts?

2

This isn’t about statistics, this is about coming up with a not-too-complicated message for members of the general public that will get them to seek screening if they are likely to need it, without having everybody and her sister seeking screening.

You don’t want to make it so long that people won’t read it, and IMHO it’s already pushing it. Adding more detail would likely backfire.

3

WhyNot, your questions are perfectly legitimate and potentially answerable by the data on which the guidelines are based. The “red flags” are ones where the risk is over a given threshold and applicable to a large, general US population.

You easily could be right, that a woman’s risk isn’t nearly as high if the woman came from a large family.

What you’re referring to is a conditional risk estimate. “Conditioned on the fact that I came from a family with 20 girls and two girls got breast cancer, what are the chances that I will get breast cancer”.

If, in your hypothetical fertile family of 20 girls, 2 get breast cancer, then to estimate risk, you would have to compare that family to other large families with a large number of girls. It’s entirely possible to estimate that risk, but the uncertainty around that risk estimate is going to be large. The risk estimate’s precision (e.g., standard error or standard deviation) will be imprecise simply because there aren’t many comparable large families with a large number of girls.

Just to make up some numbers, let’s say that the risk for cancer in women with first and second degree cancer-having relatives is 30% greater (risk ratio of 1.30) than those without. Maybe if you were to condition on big families, the risk of getting cancer will decrease to between 5-20% (risk ratio 1.05-1.20). The uncertainty is large. Would you still get screened? Maybe, if the risk was 20%, but maybe not if the risk was 5%.

The raw data set will allow the statistician to do any conditional analyses that you want, but the health authorities want to make the guidelines digestible, like **RTFirefly **said.

Also, it’s entirely possible that the statisticians looked at proportions (%) of relatives instead of counts, but maybe counts were a better predictor for risk. That said, counts are much easier to interpret. They wouldn’t want to put out statements like: “you’re at risk if at least 13% of your first and second degree relatives have cancer, unless you’re from a large family, in which case, you’re at risk if 21% or more have cancer.”

4

Thanks, that makes sense. I do understand that sometimes simplicity must be valued over precision.

But I’m also glad I’m not entirely barking mad for thinking of that. :slight_smile:

5

The other thing is that, in many cases, a more complicated model is only slightly better than a simpler one. That’s probably the case here, so there’s no real payoff in even trying to communicate the complicated model.

6

I’d agree, except that Medicare and insurance reimbursements for testing is often based on such models. I’m not sure I want to not be tested simply because I don’t have enough aunts and sisters to meet the model in the first place, y’know?