Package 'multisite.accuracy'

Estimate accuracy in multisite machine learning studies

Description

Function to estimate the accuracy of the predictions of a multisite machine-learning model, controlling the effects of the site.

Usage

multisite.accuracy(y, y.pred, site, estimate = c("auc", "bac", "cor", "hr", "mse"),
                   site.method = "covar", mixed = FALSE, min.n = 10, ...)

Arguments

y	actual value of the variable that you want to predict. For "auc" estimates, it must be a binary vector. For "bac" estimates, it must be either a binary vector or a factor. For "cor" and "mse" estimates, it must be a numeric vector. For "hr" estimates, it must be an object of class "Surv".
y.pred	predicted value of the variable. For "auc", "cor", "hr", and "mse" estimates, it must be a numeric vector. For "bac" estimates, it must be either a binary vector or a factor with the same levels as "y".
site	vector with the name of the site for each observation. Ignored if site.method is "none".
estimate	"auc" (area under the ROC curve, when "y" is a binary variable and "y.pred" is a continous variable), "bac" (balanced accuracy, along with sensitivity and specificity when both "y" and "y.pred" are binary variables; or along with the sensitivity to detect each group when both "y" and "y.pred" are factors), "cor" and "mse" (correlation and mean squared error, when both "y" and "y.pred" are numeric variables), "hr" (hazard ratio, when "y" is an object of class "Surv" and "y.pred" is a numeric variable).
site.method	"covar" (site as covariate, e.g., using lm), "meta" (meta-analysis across sites), or "none" (no control of the effects of the site).
mixed	try to fit mixed-effects models, only for site.method = "covar" and estimate = "cor", "mse", or "hr".
min.n	sites below this minimum sample size will be ignored.
...	further arguments for logistf, lm, rma, etc.

Details

We refer the reader to the publication below for details on the calculations.

Value

A data frame with the estimated accuracy, the methods used, and any warning or error.

References

Solanes, A., Palau, P., Fortea, L., Salvador, R., Gonzalez-Navarro, L., Llach, C.D., Valenti, M., Vieta, E., Radua, J. (2021) Biased accuracy in multisite machine-learning studies due to incomplete removal of the effects of the site. Psychiatry Research: Neuroimaging, 314:111313. Solanes, A., Gosling, C.J., Fortea, L., Ortuno, M., Lopez-Soley, E., Llufriu, S., Madero, S., Martinez-Heras, E., Pomarol-Clotet, E., Solana, E., Vieta, E., Radua, J. (2021) Removing the effects of the site in brain imaging machine-learning - Measurement and extendable benchmark. Neuroimage, 265:119800.

Examples

for (real_effects in c(FALSE, TRUE)) {
  for (eos in c(FALSE, TRUE)) {
    
    # Simulate data
    site = rep(c("Site A", "Site B"), 200)
    y = c(scale(rnorm(400)))
    y.pred = c(scale(residuals(lm(rnorm(400) ~ y))))
    # If real effects:
    if (real_effects) {
      y.pred = c(scale(y.pred + y))
    }
    # If effects of the site:
    if (eos) {
      y = c(scale(y + 2 * (site == "Site B")))
      y.pred = c(scale(y.pred + 2 * (site == "Site B")))
    }
    cat("\n=== Real effects:", ifelse(real_effects, "yes", "no"),
        "\n=== Effects of the site:", ifelse(eos, "yes", "no"), "\n\n")
    
    # Numeric: without real effects, mse.pred should not be < mse.mean and cor should be ~0
    print(rbind(
      multisite.accuracy(y, y.pred, site, "mse", site.method = "covar"),
      multisite.accuracy(y, y.pred, site, "mse", site.method = "meta"),
      multisite.accuracy(y, y.pred, site, "mse", site.method = "none"))[,1:5])
    print(rbind(
      multisite.accuracy(y, y.pred, site, "cor", site.method = "covar"),
      multisite.accuracy(y, y.pred, site, "cor", site.method = "meta"),
      multisite.accuracy(y, y.pred, site, "cor", site.method = "none"))[,1:3])

  }
}

---