Skip to content

[R-meta] Calculating variances and z transformation for tetrachoric, biserial correlations?

2 messages · Mark White, James Pustejovsky

Mark White
# 
Hello,

I have converted a number of summary statistics (contingency tables, *t-* and
*F*-statistics,* M*s and *SD*s) to tetrachoric and biserial correlations.
The other effect sizes that I directly observed were raw correlations. I
have my model all set up to run, but I am unsure as to what to do about
these effect sizes. I see two options:

1. Submit raw, tetrachoric, and biserial correlations and their variances
to analyses directly (what I have now).

2. Do Fisher's r-to-z transformation and *then *submit those to analyses.
The problem here is: How do I convert tetrachoric and biserial correlations
to Fisher's z? And if I do that, can I just use N to calculate the
variance? Or, do I have to also convert the variances of tetrachoric and
biserial correlations?

In either case, I am not sure how `metafor::escalc` calculates variances
for tetrachoric (`RTET`) and biserial (`RBIS`) correlations. I tried
looking through the code for `metafor::escalc` on GitHub, but could not
figure out the calculations.

I have included a table describing my effect sizes and how I calculated
them/their variances below.

What do you all think would be the best way to handle these data?

*Effect size*

*k*

*Effect size calculation*

*Variance calculation*

Raw correlation

217

Directly observed

Typical large-samples estimation (see Hedges, 1989, Equation 5), using
`metafor::escalc`

Tetrachoric correlation

12
James Pustejovsky
# 
Mark,

I do not know the specifics of how metafor makes these calculations, but I
will offer the following:

   1. The variance of the statistic (whether r or z) depends on the
   sampling design from which it was collected. So for instance, the variance
   of the tetrachoric correlation will typically differ from the variance of
   the Pearson correlation, and both will differ from the variance of the
   point biserial correlation. This is as should be expected because each of
   these statistics uses progressively more detailed information about the
   underlying distribution.
   2. If you have the variance of the r estimates, you can convert them
   using Fisher's z transformation and then calculate the sampling variance of
   z using the delta method. If V^r is the variance of the r estimate, then
   the variance of z(r) is V^z = V^r / [(1 - r^2)^2].
   3. Application of the z transformation to the point biserial correlation
   is complicated by the possibility that r_{pbs} estimates can be larger than
   1 in absolute value. See the reference below for more detailed discussion
   and one possible solution.
   4. In the reference below, I argued that it is important to consider
   whether it makes theoretical sense to combine evidence across different
   sampling designs. For example, I don't think it is sensible to
   combine biserial correlation estimates based on data from between-groups
   randomized experiments with Pearson correlation estimates from purely
   observational studies. This is because the two parameters represent
   different types of relationships: the former being a causal relationship
   arising from manipulating one of the variables, the other being a purely
   descriptive/correlational relationship. Similarly, tetrachoric correlations
   can arise from either an experimental design, in which one variable is
   manipulated, and the other, dichotomous variable observed as the outcome,
   or from an observational design, in which both variables are
   dichotomizations of a continuous latent construct.

Cheers,
James

Pustejovsky, J. E. (2014). Converting from d to r to z when the design uses
extreme groups, dichotomization, or experimental control. *Psychological
Methods*, *19*(1), 92?112. DOI: 10.1037/a0033788
On Sun, Jul 2, 2017 at 1:05 PM, Mark White <markhwhiteii at gmail.com> wrote: