Sigmoidal Modeling in R

Question

I am currently trying to model and plot a sigmoidal curve with a low amount of points.

>myExperiment
 V1  N mean  
0.1  9 0.9 
  1  9 0.8 
 10  9 0.1 
  5  9 0.2

I am using the nlsLM function from the minpack.lm package.

> nlsLM(mean2 ~ -a/(1 + exp(-b * (v1-o))))
Nonlinear regression model
  model: mean2 ~ -a/(1 + exp(-b * (v1 - o)))
   data: parent.frame()
     a      b      o 
-1.452 -0.451  1.292 
 residual sum-of-squares: 0.007017

Number of iterations to convergence: 27 
Achieved convergence tolerance: 1.49e-08
Warning message:
In nlsLM(mean2 ~ -a/(1 + exp(-b * (v1 - o)))) :
  No starting values specified for some parameters.
Initializing ‘a’, ‘b’, ‘o’ to '1.'.
Consider specifying 'start' or using a selfStart model

Using those starting values I receive this error.

> nls(mean~-a/(1 + exp(-b * (v1-o))), start=list(a=-1.452, b=-0.451, o=1.292))
Error in nls(mean ~ -a/(1 + exp(-b * (v1 - o))), start = list(a = -1.452,  : 
  step factor 0.000488281 reduced below 'minFactor' of 0.000976562

I am not well studied in stats to know if this is a syntax R error or a stats failure. What am I doing poorly?

-Thanks

You are going to have a problem fitting a model with three parameters to four observations. — Richard Telford, Sep 15 '16 at 16:20
Would using more unaveraged points be better? (those points are averages) there would still be 4 x values though — Nicholas Hayden, Sep 15 '16 at 16:26
Reading this later, my previous comment may be unclear. The original Data.frame is a average of the y values at the particular x value. Would more y values help or is the a problem with the limited x values? — Nicholas Hayden, Sep 15 '16 at 16:56
I'd try with the original data - at least there will be more degrees of freedom available — Richard Telford, Sep 16 '16 at 06:58

score 0 · Answer 1 · answered Oct 19 '16 at 14:28

This looks like binomial dose-response data. In any case, I would propose a simpler model, like the two parameter log-logistic model, with asymptotes at 0 and 1. A lot of sigmoidal models have been coded up in the drc package.

myExperiment = read.table(header = TRUE, text = 
" V1  N mean  
0.1  9 0.9 
  1  9 0.8 
 10  9 0.1 
  5  9 0.2")

library(drc)

m.ll2 <- drm(mean ~ V1,
  data = myExperiment,
  type = "binomial", 
  fct = LL.2(),
  weights = N)

plot(m.ll2, ylim = c(0, 1))

Sigmoidal Modeling in R

1 Answers1

Linked