MultiFit: Multiple fits

In ModelDataComp: Model-Data Comparison

Description

Fits bivariate or multivariate regressions between a response variable and one or several predictor variables based on multiple fitting methods.

Usage

MultiFit(x, y, fits = c("lm", "quantreg", "poly2", "poly3", "spline", 
    "gam"), xout = NULL, excl.quantile = c(0, 1), fit.quantile = NULL, 
    ...)

Arguments

x	predictor variables: a vector for bivariate fits, or a matrix or data.frame for multivariate fits
y	vector of a response variable
fits	One or several fitting methods that should be used, possible options are: lm, quantreg, poly2, poly3, spline, gam, rf, logistic
xout	vector or data.frame of predictor variables for which fits should be returned. If NULL, fits are returned along a sequence of x values. This allows the plotting of 2D surfaces in case of two predictor variables (see examples). In case of xout=x, fits are returned for the same x values that were used for fitting.
excl.quantile	lower and upper quantiles for which x and y values should be excluded to compute fits. For example, if excl.quantile=c(0, 0.9) all x and y values above the quantile 0.9 will be excluded from fitting.
fit.quantile	Perform a fitting to a certain quantile of x? Setting this argument to an value between 0 and 1 allows quantile regression. Therfore SelectQuantiles is first used to select along a range of x only the values that are around the specified quantile.
...	further arguments (not used)

Details

The following fitting methods are implemented:

• "lm": (multiple) linear regression based on lm: lm(y ~ x)

• "quantreg": quantile regression to the median based on rq: rq(y ~ x, tau=0.5)

• "poly2": 2nd-order polynomial regression based on lm: lm(y ~ poly(x, degree=2))

• "poly3": 3rd-order polynomial regression based on lm: lm(y ~ poly(x, degre=3))

• "spline": smoothing spline based on smooth.spline: smooth.spline(x, y). This method only works for bivariate fits.

• "gam": generalized additive models using spline smoothing based on gam: gam(y ~ s(x))

• "rf": random forest based on randomForest: randomForest(y ~ x). This method is not computed by default because it can be computationally expensive.

• "logistic": multiplicative logistic functions based on FitLogistic: FitLogistic(x, y). This method is not computed by default because it can be computationally expensive.

Furthermore, ensemble statistics like the mean, median, standard deviation and percentiles are computed from the results of the choosen fitting methods.

Author(s)

Matthias Forkel <[e-mail]> [aut, cre]

References

No reference.

See Also

FitLogistic

Examples

# bivariate example
x <- runif(1000, -3, 3) # predictor variable
y <- 0.5 * x + 1 / exp(-0.4 * x) * rnorm(1000, 1, 1) # response variable
ScatterPlot(x, y)
fit <- MultiFit(x, y, fits=c("lm", "quantreg", "poly2", "poly3", 
   "spline", "gam", "rf", "logistic"))
summary(fit)
cols <- piratepal("basel")
matplot(fit$x, fit[,2:11], type="l", add=TRUE, lty=1, col=cols, lwd=2)
legend("topleft", colnames(fit)[2:11], lty=1, col=cols, lwd=2)

# same example but exclude very high values (> quantile 0.9) from fitting
fit1 <- MultiFit(x, y, excl.quantile=c(0, 0.9))
lines(fit1$x, fit1$ensMean, type="l",lty=1, col="purple", lwd=3)

# to compare fitted with original values compute 
# fits at original predictor variables (xout=x)
fit <- MultiFit(x, y, fits=c("poly3", "gam"), xout=x) 
df <- data.frame(sim=c(fit$poly3, fit$gam), obs=rep(y, 2), groups=rep(c("poly3", "gam"), each=length(y)))
of <- ObjFct(df$sim, df$obs, df$groups)
plot(of, which="RMSE")
ScatterPlot(df$sim, df$obs, df$groups, objfct=TRUE)
TaylorPlot(df$sim, df$obs, df$groups)

# bivariate example with fit to a certain quantile
ScatterPlot(x, y)
fit <- MultiFit(x, y, fit.quantile=0.9, fits=c("spline", "gam", "poly3", "rf"))
matplot(fit$x, fit[,2:5], type="l", add=TRUE, lty=1, col=cols, lwd=2)
legend("topleft", colnames(fit)[2:5], lty=1, col=cols, lwd=2)

# example with two predictor variables
a <- runif(1000, -3, 3) # 1st predictor variable
b <- runif(1000, 0, 2) # 2nd predictor variable
y <- 1.2 * b + 1 / exp(-0.4 * a) * rnorm(1000, 1, 0.2) # response variable
plot(a, y)
plot(b, y)
fit <- MultiFit(x=data.frame(a, b), y, xout=NULL) 
image(x=unique(fit$a), y=unique(fit$b), 
   z=matrix(fit$lm, sqrt(nrow(fit))), main="ensMean")

## as 3D plot:
#require(rgl)
#with(data.frame(a, b), plot3d(a, b, y))
#with(fit, surface3d(unique(a), unique(b), ensMean, alpha=0.2, col="red"))

# example with three predictor variables
a <- runif(1000, -3, 3) # 1st predictor variable
b <- runif(1000, 0, 2) # 2nd predictor variable
c <- rnorm(1000, 1, 1) # 3rd predictor variable
y <- 1.2 * b + 1 / exp(-0.4 * a) * c # response variable
x <- data.frame(a, b, c)
fit <- MultiFit(x, y, fits=c("poly2", "rf"), xout=x)
ObjFct(fit$rf, y)
ObjFct(fit$poly2, y)

ModelDataComp documentation built on Nov. 22, 2020, 3 a.m.