R/cop_pse_auxiliary.R

In jrosen48/konfound: Quantify the Robustness of Causal Inferences

#' Calculate R2yz based on ryxGz and R2
#'
#' @param ryxGz correlation coefficient between Y and X given Z
#' @param R2 coefficient of determination
#' @return R2yz value
#' @importFrom lavaan parameterEstimates
cal_ryz <- function(ryxGz, R2){
    R2yz <- (ryxGz^2 - R2)/(ryxGz^2 - 1)
    if (R2yz >= 0) {
        ryz <- sqrt(R2yz)
    } else {
        stop("The calculated variance in Y explained by Z is less than 0. This can occur if Z\n suppresses the relationship between X and Y. That is, if partialling on Z increases\n the relationship between X and Y. Note: the unconditional ITCV is not conceptualized\n for this scenario.")
    }
    return(ryz)
}

#' Calculate R2xz based on variances and standard error
#'
#' @param var_x variance of X
#' @param var_y variance of Y
#' @param R2 coefficient of determination
#' @param df degrees of freedom
#' @param std_err standard error
#' @return R2xz value
#' @importFrom lavaan parameterEstimates
cal_rxz <- function(var_x, var_y, R2, df, std_err){
    R2xz <- 1 - ((var_y * (1 - R2))/(var_x * df * std_err^2))
    if (R2xz <= 0) {stop("Error! R2xz < 0!
                         Consider adding more significant digits to your input values.")}
    ## Note output the number of R2xz
    rxz <- sqrt(R2xz)
    return(rxz)
}

#' Calculate rxy based on ryxGz, rxz, and ryz
#'
#' @param ryxGz correlation coefficient between Y and X given Z
#' @param rxz correlation coefficient between X and Z
#' @param ryz correlation coefficient between Y and Z
#' @return rxy value
#' @importFrom lavaan parameterEstimates
cal_rxy <- function(ryxGz, rxz, ryz){
    rxy <- ryxGz * sqrt((1 - rxz^2)*(1 - ryz^2)) + rxz * ryz
    return(rxy)
}

#' Calculate delta star for sensitivity analysis
#'
#' @param FR2max maximum R2
#' @param R2 current R2
#' @param R2_uncond unconditional R2
#' @param est_eff estimated effect
#' @param eff_thr effect threshold
#' @param var_x variance of X
#' @param var_y variance of Y
#' @param est_uncond unconditional estimate
#' @param rxz correlation coefficient between X and Z
#' @param n_obs number of observations
#' @return delta star value
#' @importFrom lavaan parameterEstimates
cal_delta_star <- function(FR2max, 
                           R2, R2_uncond, 
                           est_eff, 
                           eff_thr, 
                           var_x, 
                           var_y, 
                           est_uncond, 
                           rxz, 
                           n_obs) {
    if (FR2max > .99) {FR2max <- .99}
    #if (FR2max < R2 + inci) {FR2max = R2 + inci} check with Ken what this means
    if (FR2max > R2) {D <- sqrt(FR2max - R2)}
    
    #elements for computing Oster's delta_star
    bt_m_b <- est_eff - eff_thr
    rt_m_ro_t_syy <- (R2 - R2_uncond) * var_y
    b0_m_b1 <- est_uncond - est_eff
    rm_m_rt_t_syy <- (FR2max - R2) * var_y
    
    t_x <- var_x * (n_obs / (n_obs - 1)) * (1 - rxz^2)
    ## adjust df for var_x 
    ## var_x is population variance, need sample variance from x
    ## this adjustment is to get closer to what robomit generates as they 
    ## run regression using the sample data 
    num1 <- bt_m_b * rt_m_ro_t_syy * t_x
    num2 <- bt_m_b * var_x * t_x * b0_m_b1^2
    num3 <- 2 * bt_m_b^2 * (t_x * b0_m_b1 * var_x)
    num4 <- bt_m_b^3 * (t_x * var_x - t_x^2)
    num <- num1 + num2 + num3 + num4
    den1 <- rm_m_rt_t_syy * b0_m_b1 * var_x
    den2 <- bt_m_b * rm_m_rt_t_syy * (var_x - t_x)
    den3 <- bt_m_b^2 * (t_x * b0_m_b1 * var_x)
    den4 <- bt_m_b^3 * (t_x * var_x - t_x^2)
    den <- den1 + den2 + den3 + den4
    #obtain delta_star which is Oster's delta
    delta_star <- num / den
    return(delta_star)
}

#' Verify regression model with control variable Z
#'
#' @param n_obs number of observations
#' @param sdx standard deviation of X
#' @param sdy standard deviation of Y
#' @param sdz standard deviation of Z
#' @param rxy correlation coefficient between X and Y
#' @param rxz correlation coefficient between X and Z
#' @param rzy correlation coefficient between Z and Y
#' @param sdcv sd between C and V
#' @param rcvy correlation coefficient between V and Y
#' @param rcvx correlation coefficient between V and X
#' @param rcvz correlation coefficient between V and Z
#' @return list of model parameters
#' @importFrom lavaan sem parameterEstimates
verify_reg_Gzcv <- function(n_obs, sdx, sdy, sdz, sdcv, 
                           rxy, rxz, rzy, rcvy, rcvx, rcvz){
    
    model <- 'Y ~ beta1 * X + beta2 * Z + beta3 * CV'
    
    ccvy <- rcvy * sdcv * sdy # cov(cv, y)
    ccvx <- rcvx * sdcv * sdx # cov(cv, x)
    ccvz <- rcvz * sdcv * sdz
    cxy <- rxy * sdx * sdy
    czy <- rzy * sdz * sdy
    cxz <- rxz * sdx * sdz
    
    # set up the covariance matrix
    cov.matrix <- matrix(c(sdy^2, cxy, czy, ccvy,
                           cxy, sdx^2, cxz, ccvx,
                           czy, cxz, sdz^2, ccvz,
                           ccvy, ccvx, ccvz, sdcv^2), 4, 4)
    rownames(cov.matrix) <- colnames(cov.matrix) <- c("Y", "X", "Z", "CV")
    
    # Check if model can be run
    flag_cov <- tryCatch(
        expr = {
            lavaan::sem(model, 
                        sample.cov = cov.matrix, 
                        sample.nobs = n_obs)
        },
        error = function(e){
            flag_cov <- FALSE
            return(flag_cov)
        },
        warning = function(w){
            flag_cov <- FALSE
            return(flag_cov)
        }
    )
    #if model can be run to verify true delta, then run it can save results
    if (inherits(flag_cov, "lavaan")) {
        fit <- lavaan::sem(model,
                           sample.cov = cov.matrix,
                           sample.nobs = n_obs)
        ## the R2 extracted from summary is NOT right, do the calculation below
        R2 <- (sdy^2 - lavaan::parameterEstimates(fit)[4,]$est) / sdy^2
        betaX <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta1',]$est
        seX <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta1',]$se
        betaZ <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta2',]$est
        seZ <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta2',]$se
        betaCV <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta3',]$est
        seCV <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta3',]$se
    }
    
    #get regression based on true delta in terms of standardized coefficent
    cor.matrix <- matrix(c(1,rxy, rzy, rcvy,
                           rxy, 1, rxz, rcvx,
                           rzy, rxz, 1, rcvz,
                           rcvy, rcvx, rcvz, 1), 4, 4)
    rownames(cor.matrix) <- colnames(cor.matrix) <- c("Y", "X", "Z", "CV")
    
    # check to see if model can be run
    flag_cor <- tryCatch(
        expr = {
            lavaan::sem(model, 
                        sample.cov = cor.matrix, 
                        sample.nobs = n_obs)
        },
        error = function(e){
            flag_cor <- FALSE
            return(flag_cor)
        },
        warning = function(w){
            flag_cor <- FALSE
            return(flag_cor)
        }
    )
    
    # if model can be run, then run it
    if (inherits(flag_cov, "lavaan")) {
        fit <- lavaan::sem(model,
                           sample.cov = cor.matrix,
                           sample.nobs = n_obs)
        std_R2 <- 1 - lavaan::parameterEstimates(fit)[4,]$est
        std_betaX <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta1',]$est
        std_seX <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta1',]$se
        std_betaZ <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta2',]$est
        std_seZ <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta2',]$se
        std_betaCV <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta3',]$est
        std_seCV <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta3',]$se
    }
    
    if ((inherits(flag_cor, "lavaan")) && (inherits(flag_cov, "lavaan"))) {
        result <- list(R2, betaX, seX, betaZ, seZ, betaCV, seCV,
                      std_R2, std_betaX, std_seX,
                      cov.matrix, cor.matrix)
        return(result)
    } else {
        stop("Error!")
    }
}

## TO DO 
## need to code the other solutions for pse as well 
## then see how the different solutions perform in 
## different scenarios (run the regression)
cal_pse <- function(thr, kryx){
    # calculations for preserving standard error
    i1 <- complex(real = 1, imaginary = -sqrt(3))
    i2 <- complex(real = 1, imaginary = sqrt(3))
    temp <- -((2 - thr^2 + 2 * thr * kryx - 2 * kryx^2)/(3 * (-1 + kryx^2))) + 
        ((i1)*(-(2 - thr^2 + 2 * thr * kryx - 2 * kryx^2)^2 + 6 * (-1 + kryx^2) * (thr^2 - 2 * thr * kryx + kryx^2)))/
        (3 * 2^(2/3) * (-1 + kryx^2) * (-16 + 15 * thr^2 + 6 * thr^4 + 2 * thr^6 - 30 * thr * kryx - 24 * thr^3 * kryx - 12 * thr^5 * kryx + 39 * kryx^2 + 
                                            12 * thr^2 * kryx^2 + 18 * thr^4 * kryx^2 + 24 * thr * kryx^3 + 8 * thr^3 * kryx^3 - 30 * kryx^4 - 27 * thr^2 * kryx^4 + 
                                            6 * thr * kryx^5 + 7 * kryx^6 + sqrt(as.complex((-16 + 15 * thr^2 + 6 * thr^4 + 2 * thr^6 - 30 * thr * kryx - 24 * thr^3 * kryx - 12 * thr^5 * kryx + 
                                                                                                 39 * kryx^2 + 12 * thr^2 * kryx^2 + 18 * thr^4 * kryx^2 + 24 * thr * kryx^3 + 8 * thr^3 * kryx^3 - 
                                                                                                 30 * kryx^4 - 27 * thr^2 * kryx^4 + 6 * thr * kryx^5 + 7 * kryx^6)^2 + 
                                                                                                4 * (-(2 - thr^2 + 2 * thr * kryx - 2 * kryx^2)^2 + 6 * (-1 + kryx^2) * (thr^2 - 2 * thr * kryx + 
                                                                                                                                                                             kryx^2))^3)))^(1/3)) - 
        1/(6 * 2^(1/3) * (-1 + kryx^2)) * (i2) * (-16 + 15 * thr^2 + 6 * thr^4 + 2 * thr^6 - 30 * thr * kryx - 24 * thr^3 * kryx - 12 * thr^5 * kryx + 39 * kryx^2 + 
                                                      12 * thr^2 * kryx^2 + 18 * thr^4 * kryx^2 + 24 * thr * kryx^3 + 8 * thr^3 * kryx^3 - 30 * kryx^4 - 27 * thr^2 * kryx^4 +
                                                      6 * thr * kryx^5 + 7 * kryx^6 + 
                                                      sqrt(as.complex((-16 + 15 * thr^2 + 6 * thr^4 + 2 * thr^6 - 30 * thr * kryx - 24 * thr^3 * kryx - 12 * thr^5 * kryx + 
                                                                           39 * kryx^2 + 12 * thr^2 * kryx^2 + 18 * thr^4 * kryx^2 + 24 * thr * kryx^3 + 8 * thr^3 * kryx^3 - 
                                                                           30 * kryx^4 - 27 * thr^2 * kryx^4 + 6 * thr * kryx^5 + 7 * kryx^6)^2 + 4 * (-(2 - thr^2 + 2 * thr * kryx - 
                                                                                                                                                             2 * kryx^2)^2 + 
                                                                                                                                                           6 * (-1 + kryx^2) * 
                                                                                                                                                           (thr^2 - 2 * thr * kryx + kryx^2))^3)))^(1/3)
    # calculations generate correlations conditional on Z: rxcvGz and rycvGz
    rxcvGz_sepreserve <- Re(sqrt(temp))
    rycvGz_sepreserve <- (kryx - thr * (1 - rxcvGz_sepreserve^2))/rxcvGz_sepreserve
    
    return(list(rxcvGz_sepreserve, rycvGz_sepreserve))
}

verify_manual <- function(rxy, rxz, rxcv, ryz, rycv, rzcv, sdy, sdx){
    beta <- (rxy + rycv * rxz * rzcv + ryz * rxcv * rzcv - rxy * rzcv^2 - rycv * rxcv - ryz * rxz) /
        (1 + 2 * rxcv * rzcv * rxz - rxcv^2 - rzcv^2 - rxz^2)
    eff <- beta * sdy / sdx
    return(beta)
}

verify_reg_Gz <- function(n_obs, sdx, sdy, sdz, rxy, rxz, rzy){
    
    model <- 'Y ~ beta1 * X + beta2 * Z'
    
    cxy <- rxy * sdx * sdy
    czy <- rzy * sdz * sdy
    cxz <- rxz * sdx * sdz
    
    # set up the covariance matrix
    cov.matrix <- matrix(c(sdy^2, cxy, czy, 
                           cxy, sdx^2, cxz,
                           czy, cxz, sdz^2), 3, 3)
    rownames(cov.matrix) <- colnames(cov.matrix) <- c("Y", "X", "Z")
    
    # Check if model can be run
    flag_cov <- tryCatch(
        expr = {
            lavaan::sem(model, 
                        sample.cov = cov.matrix, 
                        sample.nobs = n_obs)
        },
        error = function(e){
            flag_cov <- FALSE
            return(flag_cov)
        },
        warning = function(w){
            flag_cov <- FALSE
            return(flag_cov)
        }
    )
    #if model can be run to verify true delta, then run it can save results
    if (inherits(flag_cov, "lavaan")) {
        fit <- lavaan::sem(model,
                           sample.cov = cov.matrix,
                           sample.nobs = n_obs)
        ## the R2 extracted from summary is NOT right, do the calculation below
          R2 <- (sdy^2 - lavaan::parameterEstimates(fit)[3,]$est) / sdy^2
        betaX <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta1',]$est
        seX <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta1',]$se
        betaZ <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta2',]$est
        seZ <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta2',]$se
    }
    
    if (inherits(flag_cov, "lavaan")) {
        result <- list(R2, betaX, seX, betaZ, seZ)
        return(result)
    } else {
        stop("Error!")
    }
}

#' Verify unconditional regression model
#'
#' @param n_obs number of observations
#' @param sdx standard deviation of X
#' @param sdy standard deviation of Y
#' @param rxy correlation coefficient between X and Y
#' @return list of model parameters
#' @importFrom lavaan sem parameterEstimates
verify_reg_uncond <- function(n_obs, sdx, sdy, rxy){
    
    model <- 'Y ~ beta1 * X'
    cxy <- rxy * sdx * sdy
    
    # set up the covariance matrix
    cov.matrix <- matrix(c(sdy^2, cxy,  
                           cxy, sdx^2), 2, 2)
    rownames(cov.matrix) <- colnames(cov.matrix) <- c("Y", "X")
    
    # Check if model can be run
    flag_cov <- tryCatch(
        expr = {
            lavaan::sem(model, 
                        sample.cov = cov.matrix, 
                        sample.nobs = n_obs)
        },
        error = function(e){
            flag_cov <- FALSE
            return(flag_cov)
        },
        warning = function(w){
            flag_cov <- FALSE
            return(flag_cov)
        }
    )
    #if model can be run to verify true delta, then run it can save results
    if (inherits(flag_cov, "lavaan")) {
        fit <- lavaan::sem(model,
                           sample.cov = cov.matrix,
                           sample.nobs = n_obs)
        ## the R2 extracted from summary is NOT right, do the calculation below
        R2 <- (sdy^2 - lavaan::parameterEstimates(fit)[2,]$est) / sdy^2
        betaX <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta1',]$est
        seX <- lavaan::parameterEstimates(fit)[
            lavaan::parameterEstimates(fit)$label == 'beta1',]$se
    }
    
    if (inherits(flag_cov, "lavaan")) {
        result <- list(R2, betaX, seX)
        return(result)
    } else {
        stop("Error!")
    }
}