[R-meta] Welcome to the "R-sig-meta-analysis" mailing list

Dear Anne-Wil,

You are mixing up two things:

1) The unbiased estimate of the variance is

var(x) = sum((x - mean(x))^2) / (n-1).

2) The variance of a mean is

var(x) / n

So, the correct computation is

vi <- sdi^2/ni

(assuming that sdi is the square-root of the unbiased estimate of the variance, but this is pretty much what is always reported).

Best,
Wolfgang

-----Original Message-----
From: R-sig-meta-analysis [mailto:r-sig-meta-analysis-bounces at r-project.org] On Behalf Of Anne-Wil Kruijt
Sent: Monday, 16 October, 2017 17:20
To: r-sig-meta-analysis at r-project.org
Subject: Re: [R-meta] Welcome to the "R-sig-meta-analysis" mailing list

Apologies, where I wrote ?method = ?MN??, it should have been ?measure = ?MN??.

Dear all,
        
    Thank you, professor Viechtbauer, for referring me to this list. 
    
    Delving into metafor?s mechanics I noticed that when ?method = ?MN? ? is specified in escalc(), its calculation of sampling variance (vi) is ?vi <- sdi^2/ni?. I wondered why there is no option to use the(/a?) unbiased estimator ?vi <- sdi^2/ni-1?. I?m considered to bypass the escalc step and ?manually? compute vi as sdi^2/ni-1 ? but I?m hesitant because it isn?t an option in escalc() (when method = ?MN?, i.e. when obtaining the ?raw mean difference? for input as yi). Does anyone have any insights to share on why it is or is not a good idea to use the ?unbiased estimator? in the context of a REML MA on raw mean difference values? 
    
    All best,
    
    Anne-Wil

Dear all,

Thank you for clarifying. I?ll stick with sdi^2/ni for the metafor analysis. Part my confusion has to do with what I *really* am trying to figure out, namely how to best calculate/estimate the population sd, for use in a follow-up analysis, which I fear is beyond the scope of this list: Bayesfactor?s meta.ttest function allows specification of a null-interval defined in ?standardized values?. Yet, I?ve managed to get myself utterly confused here and seemingly unable to figure out the/a calculation for which test-simulations give me full confidence that I am indeed specifying the exact boundary values that I wish to specify. If anyone has any good ideas, they are very welcome to me.  I?ll take this issue back to the BayesFactor people too ? thank you for helping out so quickly on the metafor-related question! 

All best,

Anne-Wil

Dear Anne-Wil,
    
    You are mixing up two things:
    
    1) The unbiased estimate of the variance is
    
    var(x) = sum((x - mean(x))^2) / (n-1).
    
    2) The variance of a mean is
    
    var(x) / n
    
    So, the correct computation is
    
    vi <- sdi^2/ni
    
    (assuming that sdi is the square-root of the unbiased estimate of the variance, but this is pretty much what is always reported).
    
    Best,
    Wolfgang
    
    -----Original Message-----
    From: R-sig-meta-analysis [mailto:r-sig-meta-analysis-bounces at r-project.org] On Behalf Of Anne-Wil Kruijt
    Sent: Monday, 16 October, 2017 17:20
    To: r-sig-meta-analysis at r-project.org
    Subject: Re: [R-meta] Welcome to the "R-sig-meta-analysis" mailing list
    
    Apologies, where I wrote ?method = ?MN??, it should have been ?measure = ?MN??.

Dear all,
            
        Thank you, professor Viechtbauer, for referring me to this list. 
        
        Delving into metafor?s mechanics I noticed that when ?method = ?MN? ? is specified in escalc(), its calculation of sampling variance (vi) is ?vi <- sdi^2/ni?. I wondered why there is no option to use the(/a?) unbiased estimator ?vi <- sdi^2/ni-1?. I?m considered to bypass the escalc step and ?manually? compute vi as sdi^2/ni-1 ? but I?m hesitant because it isn?t an option in escalc() (when method = ?MN?, i.e. when obtaining the ?raw mean difference? for input as yi). Does anyone have any insights to share on why it is or is not a good idea to use the ?unbiased estimator? in the context of a REML MA on raw mean difference values? 
        
        All best,
        
        Anne-Wil