I'm running a nonlinear lest squares using the minpack.lm package on Microsoft version or R. However, for each replication, with the same data, the output varies. In fact, my interest lies in the second parameter, however, even it suffers some variation. Base nlminb would return the same awkward output.
If I use the CRAN version of R, the results are reproducible.
Reproducible example:
library(minpack.lm)
df_ss <- read.csv("example.csv")
objective <- function(start, data_ss = df_ss) {
beta_const <- start[1]
beta_1 <- start[2:2]
beta_poly <- start[-(1:2)]
poly_df <- with(data_ss, cbind((var5 - var4 * beta_1)))
poly_df <- poly(poly_df, degree = 2, raw = TRUE)
poly_df <- poly_df %*% diag(beta_poly, nrow = length(beta_poly))
lp2 <- qr(cbind(const = 1 * beta_const, poly_df))
lp2 <- with(data_ss, qr.resid(lp2, y_ss - data_ss$var4 * beta_1))
return(lp2)
}
start <- rep(0, 4)
ss_reg <- nls.lm(par = start, fn = objective)
print(ss_reg$par)
#2.227202e-06 -1.660069e-01 4.659955e-07 2.169863e-07
ss_reg <- nls.lm(par = start, fn = objective)
print(ss_reg$par)
#-1.022907e-06 -1.658407e-01 4.678488e-08 9.447334e-07
ss_reg <- nls.lm(par = start, fn = objective)
print(ss_reg$par)
#1.911939e-06 -1.658408e-01 2.513519e-07 -2.140628e-07
ss_reg <- nls.lm(par = start, fn = objective)
print(ss_reg$par)
#-5.765922e-07 -1.658408e-01 3.606956e-07 5.372434e-07
The data used is here.
Am I doing something wrong?
