Comparison of Correlation Method 1 and Correlation Method 2

Ordinal Variables:

Correlations are computed pairwise. If both variables are binary, the method of Demirtas et al. (2012) is used to find the tetrachoric correlation (code adapted from BinNonNor::Tetra.Corr.BB). The tetrachoric correlation is an estimate of the binary correlation measured on a continuous scale. The assumptions are that the binary variables arise from latent normal variables, and the actual trait is continuous and not discrete. This method is based on Emrich and Piedmonte (1991)’s work, in which the joint binary distribution is determined from the third and higher moments of a multivariate normal distribution:

Let \(\Large Y_{1}\) and \(\Large Y_{2}\) be binary variables with \(\Large E[Y_{1}] = Pr(Y_{1} = 1) = p_{1}\), \(\Large E[Y_{2}] = Pr(Y_{2} = 1) = p_{2}\), and correlation \(\Large \rho_{y1y2}\).

Let \(\Large \Phi[x_{1}, x_{2}, \rho_{x1x2}]\) be the standard bivariate normal cumulative distribution function, given by: \[\Large \Phi[x_{1}, x_{2}, \rho_{x1x2}] = \int_{-\infty}^{x_{1}} \int_{-\infty}^{x_{2}} f(z_{1}, z_{2}, \rho_{x1x2})\ dz_{1} dz_{2},\] where \[\Large f(z_{1}, z_{2}, \rho_{x1x2}) = [2\pi\sqrt{1 - \rho_{x1x2}^2}]^{-1} * exp[-0.5(z_{1}^2 - 2\rho_{x1x2}z_{1}z_{2} + z_{2}^2)/(1 - \rho_{x1x2}^2)].\] Then solving the equation \[\Large \Phi[z(p_{1}), z(p_{2}), \rho_{x1x2}] = \rho_{y1y2}\sqrt{p_{1}(1 - p_{1})p_{2}(1 - p_{2})} + p_{1}p_{2}\] for \(\Large \rho_{x1x2}\) gives the intermediate correlation of the standard normal variables needed to generate binary variables with correlation \(\Large \rho_{y1y2}\). Here \(\Large z(p)\) indicates the \(\Large p^{th}\) quantile of the standard normal distribution.

To generate the binary variables from the standard normal variables, set \(\Large Y_{1} = 1\) if \(\Large Z_{1} \le z(p_{1})\) and \(\Large Y_{1} = 0\) otherwise. Similarly, set \(\Large Y_{2} = 1\) if \(\Large Z_{2} \le z(p_{2})\) and \(\Large Y_{2} = 0\) otherwise.

This ensures: \[\Large E[Y_{1}] = Pr(Y_{1} = 1) = Pr(Z_{1} \le z(p_{1})) = p_{1},\] \[\Large E[Y_{2}] = Pr(Y_{2} = 1) = Pr(Z_{2} \le z(p_{2})) = p_{2},\] \[\Large Cov(Y_{1}, Y_{2}) = Pr(Y_{1} = 1, Y_{2} = 1) - p_{1}p_{2}\] \[\Large = Pr(Z_{1} \le z(p_{1}), Z_{2} \le z(p_{2})) - p_{1}p_{2}\] \[\Large = \Phi[z(p_{1}), z(p_{2}), \rho_{x1x2}] - p_{1}p_{2}\] \[\Large = \rho_{y1y2}\sqrt{p_{1}(1 - p_{1})p_{2}(1 - p_{2})},\] \[\Large Cor(Y_{1}, Y_{2}) = Cov(Y_{1}, Y_{2})/\sqrt{p_{1}(1 - p_{1})p_{2}(1 - p_{2})} = \rho_{y1y2}.\]

Otherwise, ordnorm is called for each pair. If the resulting intermediate matrix is not positive-definite, this is corrected for later.

Continuous Variables:

Correlations are computed pairwise. findintercorr_cont is called for each pair.

Continuous-Ordinal Pairs:

findintercorr_cont_cat is called to calculate the intermediate MVN correlation for all Continuous and Ordinal combinations.

Now the two methods will be contrasted.

Simulation Process:

The algorithm used in the simulation function rcorrvar that employs correlation method 1 is as follows:

1. Preliminary checks on the distribution parameters and target correlation matrix rho are performed to ensure they are of the correct dimension, format, and/or sign. This function does NOT verify the feasibility of rho, given the distribution parameters. That should be done first using valid_corr, which checks if rho is within the feasible bounds and returns the lower and upper correlation limits.

2. The constants are calculated for the continuous variables using find_constants. If no solutions are found that generate valid power method pdfs, the function will return constants that produce invalid pdfs (or a stop error if no solutions can be found). Errors regarding constant calculation are the most probable cause of function failure. Possible solutions include:

   1. changing the seed, or
   2. using a list Six of sixth cumulant correction values (if method = “Polynomial”).

3. The support is created for the ordinal variables (if no support provided).

4. The intermediate correlation matrix Sigma is calculated using findintercorr. Note that this will return a matrix that is not positive-definite. If so, there will be a message that it may not be possible to produce variables with the desired distributions. Also, the algorithm of Higham (2002) is used (see Matrix::nearPD) to produce the nearest positive-definite matrix and a message is given.

5. k <- k_cat + k_cont + k_pois + k_nb multivariate normal variables (\(\Large X_{nxk}\)) with correlation matrix Sigma are generated using eigen-value and spectral value decompositions on a \(\Large MVN_{nxk}(0,1)\) matrix.

6. The variables are generated from \(\Large X_{nxk}\) using the appropriate transformations (see Variable Types vignette)

7. The final correlation matrix is calculated, and the maximum error (maxerr) from the target correlation matrix is found.

8. If the error loop is specified (error_loop = TRUE), it is used on each variable pair to correct the final correlation until it is within epsilon of the target correlation or the maximum number of iterations has been reached. Additionally, if the extra correction is specified(extra_correct = TRUE), if the maximum error within each variable pair is still greater than 0.1, the intermediate correlation is set equal to the target correlation (with the assumption that the calculated final correlation will be less than 0.1 away from the target).

9. Summary statistics are calculated by variable type.

Simulation Process:

The algorithm used in the simulation function rcorrvar2 that employs correlation method 2 is similar to that described for rcorrvar, with a few modifications:

1. The feasibility of rho, given the distribution parameters, should be checked first using the function valid_corr2, which checks if rho is within the feasible bounds and returns the lower and upper correlation limits.

2. After the support is created for the ordinal variables (if no support is provided), the maximum support for the count variables is determined using max_count_support, given truncation value vector pois_eps for Poisson variables and/or nb_eps for Negative Binomial variables. The cumulative probability truncation value may differ by variable, but a good value is \(0.0001\). The resulting supports and distribution parameters are used to create marginal lists, consisting of the cumulative probabilities for each count variable.

3. The intermediate correlation matrix Sigma is calculated using findintercorr2.