error.prob rates using sim.geno

44 views
Skip to first unread message

David Gronwall

unread,
Dec 11, 2009, 5:45:24 PM12/11/09
to R/qtl discussion
Hi Dr. Broman,

I have noticed that using different error.prob rates in the sim.geno
function leads to significantly different results using the
stepwiseqtl function. Keeping all things the same except for
error.prob rates, stepwiseqtl found the following two models:

Scenario 1
error.prob=0.0001 (default)
y ~ Q1 + Q2 + Q3 + Q4 + Q5 + Q6 + Q7 + Q8 + Q9 + Q10 + Q11 + Q12 + Q13
+ Q14 + Q15 + Q1:Q2 + Q1:Q7 + Q3:Q12 + Q5:Q6 + Q13:Q15

Scenario 2
error.prob=0.01
y ~ Q1 + Q2 + Q3 + Q4 + Q5 + Q6 + Q7 + Q8 + Q9 + Q10 + Q11 + Q12 +
Q1:Q7 + Q5:Q6 + Q4:Q10

I am not overly concerned because all of the extra QTLs and
interactions found in Scenario 1 are of small effect (explained
variance is 85% in scenario 1 and 78% in scenario 2), but I was hoping
you could help me to understand what error.prob rate is reasonable for
this particular experiment, and why the results vary. I am using 162
RILs (Cvi x Ler) developed by Alonso-Blanco et. al. (1998). 98.6% of
markers are genotyped. I noticed you use error.prob=0.001 for the
hyper data in "A Guide to QTL Mapping in R/qtl", which is not the
default, so I assume there is an accepted error.prob rate for
different types of experiments. Is there an error.prob value that is
generally accepted for use with RILs?

Karl Broman

unread,
Dec 12, 2009, 11:05:10 PM12/12/09
to rqtl...@googlegroups.com, David Gronwall
Dear David,

While different assumed genotyping error probabilities could give different results, in terms of QTL inference, I would study how large is the evidence for the model 1 vs model 2 with each of these error probabilities. That is, fit each model under each defined error.prob and look at how different the penalized LOD scores are. The other major thing to check is whether the number of imputations you're doing is sufficiently large. If you re-run sim.geno and then re-run stepwiseqtl, do you still get the same inferred models?

There can be no generally accepted error probability. The appropriate value depends on the quality of the genotyping data. But in many cases, genotyping data for RILs would be high density and carefully curated, and so error.prob could be quite small (like 0.0001).

karl
> --
>
> You received this message because you are subscribed to the Google Groups "R/qtl discussion" group.
> To post to this group, send email to rqtl...@googlegroups.com.
> To unsubscribe from this group, send email to rqtl-disc+...@googlegroups.com.
> For more options, visit this group at http://groups.google.com/group/rqtl-disc?hl=en.
>
>

Reply all
Reply to author
Forward
0 new messages