MetaXcan for binary outcome

[Ani Manichaikul]

HI,

I am getting ready to use MetaXcan with results from GWAS of a dichotomous trait. I seem to remember you had fixed a bug related to dichotomous outcomes previously.  Is there anything special I need to do for this case?  Or is it fine that I just input log(odds ratio) for each SNP as the beta to MetaXcan?

Thanks for your input!

Ani

[Alvaro Barbeira]

Hi Ani,

Just inputing OR column should work (something like --or_column odds_ratio if "odds_ratio" is your column name).

You can use logs(odds ratio) as a --beta_column but there are minor numerical errors in the conversion.

Best regards,

Alvaro

--
You received this message because you are subscribed to the Google Groups "PrediXcan/MetaXcan" group.
To unsubscribe from this group and stop receiving emails from it, send an email to predixcanmetaxcan+unsubscribe@googlegroups.com.
To post to this group, send email to predixcanmetaxcan@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/predixcanmetaxcan/4ae8d7ca-3347-499c-8bd6-bcf7d79ddd09%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Ani Manichaikul]

Ok, thanks both for your replies.  I will give a try with the option --or_column.

Ani

[Eva Krapohl]

Hi Alvaro, 

Would you mind elaborating on what specifically you mean by 'there are minor numerical errors in the conversion'? Do you mean that s-predixcan converts values inputted via --or_column to log(or) using numpy.log(or) and log(or) inputted via the --beta_colum are just treated as regular beta coefficients? How does the interpretation of the S-PrediXcan 'effect_size'/'zscore' change in regard to log(or) vs. or?

Many thanks, 

Eva

On Thursday, August 4, 2016 at 9:25:53 PM UTC+1, Alvaro Barbeira wrote:

Hi Ani,

Just inputing OR column should work (something like --or_column odds_ratio if "odds_ratio" is your column name).

You can use logs(odds ratio) as a --beta_column but there are minor numerical errors in the conversion.

Best regards,

Alvaro

On Thu, Aug 4, 2016 at 4:52 PM, Ani Manichaikul <aman...@gmail.com> wrote:

HI,

I am getting ready to use MetaXcan with results from GWAS of a dichotomous trait. I seem to remember you had fixed a bug related to dichotomous outcomes previously.  Is there anything special I need to do for this case?  Or is it fine that I just input log(odds ratio) for each SNP as the beta to MetaXcan?

Thanks for your input!

Ani

--
You received this message because you are subscribed to the Google Groups "PrediXcan/MetaXcan" group.

To unsubscribe from this group and stop receiving emails from it, send an email to predixcanmetax...@googlegroups.com.
To post to this group, send email to predixca...@googlegroups.com.

[Alvaro Barbeira]

Hi Eva,

Allow me to clarify, please bear with me if I sound redundant.

Using --or_column causes S-PrediXcan to take numpy.log() of the column and use that as an effect size. Taking --beta_column of a column that contains log(OR) will use it as effect size. S-PrediXcan uses a linear regression model and ultimately works with effect sizes, so that using Odd Ratios are an approximation.

For completeness sake: if you provide --pvalue_column, an absolute value of the SNP zscore will be calculated from a normal CDF using a two-sided test, and only the sign of the effect size will be used (its actual magnitude is not used). If you provide a --se_column instead, S-PrediXcan will use effect_size/se as z-score.

Hope this helps. Best,

Alvaro

To unsubscribe from this group and stop receiving emails from it, send an email to predixcanmetaxcan+unsubscribe@googlegroups.com.
To post to this group, send email to predixcanmetaxcan@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/predixcanmetaxcan/cf68ecf8-9567-41dd-8c2c-6ab50b3c2853%40googlegroups.com.

[Eva]

Many thanks for your very prompt reply.

Thanks for the elaboration. Just to clarify, your initial comment regarding 'minor numerical errors in the conversion' referred to using log() as an approximation of effect size when dealing with OR in the context of linear regression, not to using log(OR) as --beta_column vs. using OR as --or_column, which is then converted to log(OR) (as these are equivalent)? 

Thank you for (S-)PrediXcan!

Eva

[Alvaro Barbeira]

Hi Eva,

My initial comment might had been unclear. We are using numpy arrays and numpy.log(); numerical differences might arise in some platforms when using standard python log() function.

The approximation of using numpy.log(OR) as effect size of simple linear regression (instead of logistic regression) is a separate issue. This happens regardless of any numerical issue in python's log() vs numpy.log().

If your GWAS contains a log(OR) column you can use it as --beta_column, and from S-PrediXcan's point of view it will be the same as using a OR column with --or_column.

Please let me know if this is more clear than the previous.

Best,

Alvaro

To unsubscribe from this group and stop receiving emails from it, send an email to predixcanmetaxcan+unsubscribe@googlegroups.com.
To post to this group, send email to predixcanmetaxcan@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/predixcanmetaxcan/ebbd3697-001c-4106-9e2c-2a15a316cb22%40googlegroups.com.

[Eva]

Hi Alvaro, 

All clear. Much appreciated, many thanks!

Eva