Nonlinear total least squares / Deming regression in R

Question

I've been using nls() to fit a custom model to my data, but I don't like how the model is fitting and I would like to use an approach that minimizes residuals in both x and y axes.

I've done a lot of searching, and have found solutions for fitting linear models:

via the deming package,
various stackoverflow posts:
- total-least-square-method-using-r,
- how-to-calculate-total-least-squares-in-r-orthogonal-regression,
and this very nice CrossValidated post: How to perform orthogonal regression (total least squares) via PCA?.
I've also found a MATLAB solution: Total least squares curve fit problem,

but these fit a second order polynomial and not a custom, user-defined model.

What I would like is something similar to nls() that does the x and y residual minimization. This would allow me to enter my custom model. Is anyone aware of any solution in R?

Here's an example, but please note that I'm seeking suggestions on a general solution for nonlinear total least squares regression, and not something specific to this dataset (this is just an example data from here):

df <- structure(list(x = c(3, 4, 5, 6, 7, 8, 9, 10, 11), y = c(1.0385, 
1.0195, 1.0176, 1.01, 1.009, 1.0079, 1.0068, 1.0099, 1.0038)), .Names = c("x", 
"y"), row.names = c(NA, -9L), class = "data.frame")

(nlsfit <- nls(y ~ a^b^x, data = df, start = c(a=0.9, b=0.6)))

library(ggplot2)
ggplot(df, aes(x=x, y=y)) + 
    geom_point() + 
    geom_smooth(method="nls", formula = y ~ a^b^x, se=F, start = list(a=0.9, b=0.6))

Does anyone have any suggestion for how I might proceed?

If you are only looking for R code, that would be off-topic here. However, I can see a more general (NLS + TLS) question here, whose answer can be illustrated in R. You may want to reframe your Q. — gung - Reinstate Monica, Mar 14 '15 at 23:55
Thanks gung, do you have suggestions on how i should reframe? I'm not just looking for R code, i'm looking for examples that show how this might be done. thanks. — Thomas, Mar 15 '15 at 00:25

score 7 · Accepted Answer · answered Mar 15 '15 at 01:09

7

There is a technique called "Orthogonal Distance Regression" that does this. An implementation in R was recently released:

http://www.r-bloggers.com/introducing-orthogonal-nonlinear-least-squares-regression-in-r/

answered Mar 15 '15 at 01:09

Brian Borchers

5,015
1
18
27

Thanks Brian Borchers, that is exactly what I was looking for! – Thomas Mar 15 '15 at 02:19
1

Since links can go dead, can you excerpt the key information here? – gung - Reinstate Monica Mar 15 '15 at 14:56
The link points to a blog posting describing a recent implementation of Orthogonal Distance Regression in R. A google (or google scholar) search for "Orthogonal Distance Regression" will lead you to lots of papers on the topic. – Brian Borchers Mar 15 '15 at 15:48

Nonlinear total least squares / Deming regression in R

1 Answers1