h2>1 about leap HOT 4 CLOSED

omerwe commented on September 27, 2024

h2>1

from leap.

Comments (4)

omerwe commented on September 27, 2024

Hi Elisabetta, Thanks for using LEAP! LEAP estimates heritability by using PCGC (Golan, Lander and Rosset, PNAS 2014) <http://www.pnas.org/content/111/49/E5272.short>. Indeed, the PCGC estimatror can sometimes estimate h2>1 for small datasets. This can happen under model misspecification (e.g. incorrect prevalence estimate), but it can also happen simply due to sampling variance, especially with small datasets (i.e., <5000 individuals). Unfortunately, there is no simple solution for this. I would start by double checking the prevalence estimate and all the other input. If you're sure that everything is correct, then a possible solution is to assume that the SNP heritability is truly close to 100%. In this case, you could input 99% or a similar number to calc_h2.py, under the assumption that this is the true underlying SNP heritability. Good luck, and do not hesitate if you have any other questions! Best, Omer

…

On Mon, Jul 2, 2018 at 1:21 PM e-manduchi ***@***.***> wrote: hello, I've been recently using LEAP on two data sets on which it seems to have run without problems. I'm now trying it on a third data set but on this one I get h2 values >1 (as per below log). This consequently generates problems when I run probit.py: /home/manduchi/mypy/lib/python2.7/site-packages/leap/probit.py:57: RuntimeWarning: invalid value encountered in sqrt normPDF = stats.norm(0, np.sqrt(1-h2)) ... Can you advice? Thanks, Elisabetta === calc_h2.py run python calc_h2.py --bfilesim data_autosome --prev 0.1 --pheno leap.phe --eigen leap_noChr1.kin.npz --extractSim noChr1.snps --related leap.rm /home/ manduchi/mypy/lib/python2.7/site-packages/pysnptools/snpreader/bed.py:42: FutureWarning: 'count_A1' was not set. For now it will default to 'False', but in the future it will default to 'True' warnings.warn("'count_A1' was not set. For now it will default to 'False', but in the future it will default to 'True'", FutureWarning) WARNING:root:loadPhen is using default missing value of '-9'. WARNING:root:loadPhen is using default missing value of '-9'. 29 individuals will be removed due to high relatedness Removing the top 10 PCs from the kinship matrix Computing h2 for a binary phenotype h2: 1.241879 — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#5>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AJog8_wptTmR3tpnc511MPmEm9Vmz5Lcks5uClargaJpZM4U_q8d> .

from leap.

e-manduchi commented on September 27, 2024

Hello Omer,
thanks for your prompt reply!
Regarding prevalence, since from the literature I saw that the prevalence of my trait of interest is 10%, I've simply entered '--prev 0.1' in my runs. Is this correct, or should I have entered another value (like liability threshold)?
My population is just short of 5000 individuals, but I've run LEAP on two other data sets (one much smaller) for the same trait (using the same --prev of 0.1) and it didn't throw errors. The h2 estimated by calc_h2 on these data sets was between 0.2 and 0.24 depending on chromosome.
Elisabetta

from leap.

omerwe commented on September 27, 2024

Hi Elisabetta, What you did sounds perfectly fine. Is your population a "natural population"? LEAP isn't necessarily suitable for inbred populations, such as mice crosses. Another possibility is that the h2 estimate in these datasets is noisy. My code estimates h2, but unfortunately it doesn't estimate its 95% confidence interval. One thing you can try to do is to estimate h2 using gcta <http://cnsgenomics.com/software/gcta> instead of my own code (using the --prevalence flag). These estimates are typically down-biased (as shown in the PCGC paper <http://www.pnas.org/content/111/49/E5272.short> by Golan et al.), but it may be that for some (strange) reason it will give more sensible results on this particular data set. If all else fails, I would just trying running calc_probit with h2=0.95. It may be sub-optimal, but I think that you can still gain statistical power in this way over a naive association analysis. However, I should also point out that according to our simulations in the LEAP paper, the advantage of using LEAP over a plain LMM analysis under 10% prevalence is not very large. LEAP becomes significantly more powerful than a standard LMM with prevalence<1%. Hence, it might not make a large difference anyway whether you use a LEAP or a plain LMM analysis... Good luck again, and do not hesitate if you have any other questions! Omer

…

On Tue, Jul 3, 2018 at 9:53 AM e-manduchi ***@***.***> wrote: Hello Omer, thanks for your prompt reply! Regarding prevalence, since from the literature I saw that the prevalence of my trait of interest is 10%, I've simply entered '--prev 0.1' in my runs. Is this correct, or should I have entered another value (like liability threshold)? My population is just short of 5000 individuals, but I've run LEAP on two other data sets (one much smaller) for the same trait (using the same --prev of 0.1) and it didn't throw errors. The h2 estimated by calc_h2 on these data sets was between 0.2 and 0.24 depending on chromosome. Elisabetta — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#5 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AJog8_Vcz0WA4NjH5sCcud20xQl8PuIrks5uC3dVgaJpZM4U_q8d> .

from leap.

e-manduchi commented on September 27, 2024

Thanks for the suggestions and clarifications. Yes, my population is natural, it's a subset of the WTCCC data. I'll try what you indicate and see what happens. Thanks again,
Elisabetta

from leap.

h2>1 about leap HOT 4 CLOSED

Comments (4)

Related Issues (3)

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent