Choosing the best method

[Mark Ebbert]

Hi,

Please excuse me if I've overlooked existing questions or documentation that explain this, but I would benefit from greater guidance on which tests are appropriate for a given situation. I understand that you cannot provide guidance for every situation, but general guidance would be great. Here's what I've learned from your R vignette and paper (http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3415556/).

• SKAT: best for non-binary outcomes
• SKAT-Binary: obviously best for binary outcomes (e.g., case-control status).
• method="SKAT": a "non-burden" test that "is more powerful when a large fraction of the variants in a region are noncausal or the effects of causal variants are in different directions." Apparently, "burden" tests "are more powerful when most variants in a region are causal and the effects are in the same direction".
• method="SKATO": combines the benefits of both "burden" and "non-burden" tests by maintaining power for both.
• SKAT-CommonRare: let's you "test combined effects of common and rare variants"
  

Specific questions:

1. Which option performs just a burden test? It isn't clear to me in the documentation.
2. It's difficult to know apriori whether "a large fraction of variants" are causal/non-causal or if the effects are in the same/different directions. Or am I missing something? I don't expect you to resolve this unless I've missed something.
3. In which situations would you choose SKAT-CommonRare over the other methods?

Basically, I think it would be super useful to add a chapter to the vignette that provides more information to choose the best test. But I appreciate any guidance you can provide at present.

Respectfully,

Mark

[Seunggeun (Shawn) Lee]

Hi Mark,

Thanks for the good suggestion. I will consider to add a guidance in the next version. 

1. Which option performs just a burden test? It isn't clear to me in the documentation.

If you want to perform a burden test, you can use 

SKAT function: use r.corr=1

SKATBinary function: use method="Burden"

1. It's difficult to know apriori whether "a large fraction of variants" are causal/non-causal or if the effects are in the same/different directions. Or am I missing something? I don't expect you to resolve this unless I've missed something.

So we developed SKAT-O method which adaptively selects between SKAT and Burden test to maximize power. If you don't have any prior assumption on % of causal variants and directions of them, I recommend to use SKAT-O. 

1. In which situations would you choose SKAT-CommonRare over the other methods?

The burden test, SKAT, and SKAT-O methods (implemented in SKAT and SKATBinary functions) are developed for testing for rare-variant associations. If you want to test for both common and rare variants together, you need to use SKAT-CommonRare. The main difference between these two approaches is a different weighting scheme. 

1) SKAT and SKATBinary functions uses a weighting scheme to up-weight rarer variants (ex beta(1, 25)). This weighting scheme effectively remove the common variants from the test, since common variants are severely under-weighted. 

2) SKAT-CommonRare function uses a weighting scheme in which common and rare variants contribute to the test equally, or a weighting parameter between common and rare variants are selected data-adaptively. So you can test for both common and rare variant effects. 

 

Thanks,
Shawn