Nls Troubles: Missing Value or an Infinity Produced When Evaluating the Model

nls troubles: Missing value or an infinity produced when evaluating the model

minpack.lm to the rescue:

library(minpack.lm)
curve.nlslrc = nlsLM(photolrc ~ Am*(1-((1-(Rd/Am))^(1-(PARlrc/LCP)))),
                   start=list(Am=(max(photolrc)-min(photolrc)),
                              Rd=-min(photolrc),
                              LCP= (max(photolrc)-1)),
                   data = curvelrc)
coef(curve.nlslrc)
  #      Am         Rd        LCP 
  #8.011311   1.087484 -20.752957

plot(photolrc ~ PARlrc, data = curvelrc)
lines(0:1300, 
      predict(curve.nlslrc, 
              newdata = data.frame(PARlrc = 0:1300)))

If you pass start = list(Am = 8, Rd = 1, LCP = -20) to nls you also get a successful fit.

I don't know if the parameter values are sensible estimates considering the science behind this. Can LCP be negative?

R - NLS Error 'Missing value or an infinity produced when evaluating the model'

Use the plinear algorithm. In that case the right hand side and starting values should omit case_networks and its value will be reported as .lin.

fm <- nls(cases ~  r0^(t/5), data = us_jh_cases, alg = "plinear", 
  start = list(r0 = 2.5))
fm

giving:

Nonlinear regression model
  model: cases ~ r0^(t/5)
   data: us_jh_cases
       r0      .lin 
1.094e+00 1.513e+05 
 residual sum-of-squares: 9.836e+12

Number of iterations to convergence: 10 
Achieved convergence tolerance: 6.719e-06

Graph

plot(cases ~ t, us_jh_cases, pch = ".", cex = 2)
lines(fitted(fm) ~ t, us_jh_cases, col = "red") # fit shown in red

nls() in R: Missing value or an infinity produced when evaluating the model

There are several issues here.

First off: your lagged values aren't really lagged. Take a look at df_test and you will se that the 1's and 2's are identical.

This will give you lagged values:

set.seed(1)

rh1 <- rnorm(301, 0.75, 0.1)
rh1[rh1 > 1] <- 1
ta1 <- rnorm(301, 302, 3)
y1 <- rnorm(301, 0.2, 0.05)

df_test <- data.frame(
  rh1 = rh1,
  rh2 = c(NA, head(rh1, -1)),
  ta1 = ta1,
  ta2 = c(NA, head(ta1, -1)),
   y1 = y1,
   y2 = c(NA, head(y1, -1))
)
df_test <- df_test[complete.cases(df_test), ]

Next:

Missing value or an infinity produced when evaluating the model

Means just that, and my eyes immediately fix on the logs in your expression. We all know that taking the log of a negative number is undefined, as is the log of 0, although it is often returned as infinity.

Let's take a look at those expressions

ex1 <- with(df_test, log(-8.2 * ta2 * log(rh2) / 18))
ex2 <- with(df_test, log(-8.3 * ta1 * log(rh1) / 18))

If you look at ex1 and ex2 you will see that both contain a -Inf. Now there's your culprit. But how can we fix this? Let's see which rows in your data gives rise to this.

df_test[which(is.infinite(ex1 + ex2)),]
#        rh1     rh2      ta1      ta2      y1      y2
# 274 1.0000 0.66481 304.5453 300.5972 0.20930 0.17474
# 275 0.7656 1.00000 304.9603 304.5453 0.20882 0.20930

Interesting, they are right next to each other, and they both contain a 1. What's log(1)? What happens if you multiply it by something and take the log of the product?

Let's make sure rh1 and rh2 is always less than 1

set.seed(1)

rh1 <- rnorm(301, 0.75, 0.1)
rh1[rh1 > 0.99] <- 0.99
ta1 <- rnorm(301, 302, 3)
y1 <- rnorm(301, 0.2, 0.05)

df_test <- data.frame(
  rh1 = rh1,
  rh2 = c(NA, head(rh1, -1)),
  ta1 = ta1,
  ta2 = c(NA, head(ta1, -1)),
   y1 = y1,
   y2 = c(NA, head(y1, -1))
)
df_test <- df_test[complete.cases(df_test), ]

But we're still not done. If you run your nls() call now you'll get the error

Convergence failure: initial par violates constraints

And the cause is obvious if you look at the values you specify for your coefficients constraints. coef2 and coef3 has lower constraints set to infinity! That doesn't make sense. "initial par violates constraints" usually means that the start values aren't within the constraints, which is definitely the case here. If we change them to negative infinity everything works fine.

nls(y1 ~
  coef1 ^ 2 * y2 +
  coef1 * (1 - coef1) * 
  (coef2 + coef3 * log(-8.3 * ta2 * log(rh2) / 18)) +
  (1 - coef1) * 
  (coef2 + coef3 * log(-8.3 * ta1 * log(rh1) / 18)),
  data = df_test,
  algorithm = "port",
  start = list(coef1 = 0.7,
               coef2 = 0.15,
               coef3 = 0),
  upper = c(exp(-0.00005), Inf, Inf),
  lower = c(exp(-0.5), -Inf, -Inf)
)
# Nonlinear regression model
#    model: y1 ~ coef1^2 * y2 + coef1 * (1 - coef1) * (coef2 + coef3 * log(…
#    data: df_test
#   coef1   coef2   coef3 
#  0.6065  0.2569 -0.0170 
#  residual sum-of-squares: 1.058

# Algorithm "port", convergence message: 
#   both X-convergence and relative convergence (5)

nls missing value or infinity produced

If I understand what nls is doing, it is estimating the values of parameters a and b in your case. Therefore there is no need to have start values for variables Volume and Height. These are simply the values in the dataset. If you remove them from the start argument, the error goes away and you get a model.

model <- function(data){
  nls(Volume~ a * (Height^b), data=data, start=c(a = 1, b = 1))
}
  
df %>% group_by(Site) %>% nest() %>% 
  mutate(mod= map(.x=data, model))

Why does nls return Missing value or an infinity produced when evaluating the model in this instance?

Welcome to SO.

I think this is due to a minor typing error. When you generate your formula with

f=as.formula(paste(as.name(colnames(csvfile)[1]),"~I(a*exp(b*",as.name(colnames(csvfile[2])),"))"))

then you are actually producing the formula

> f
x ~ I(a * exp(b * y))

I think you have swapped the x and the y around and the starting value for a needs to be back-transformed. At least, when you create the starting values from the logged variables then you model log(y) as a function of log(x). If you do the following instead then it appears to work

> f=as.formula(paste(as.name(colnames(csvfile)[2]),"~I(a*exp(b*",as.name(colnames(csvfile[1])),"))"))
> expo<-nls(f,data=csvfile, start=list(a=exp(a_start),b=b_start))
> expo
Nonlinear regression model
  model: y ~ I(a * exp(b * x))
   data: csvfile
     a      b 
2.7854 0.7079 
 residual sum-of-squares: 4.569

Number of iterations to convergence: 7 
Achieved convergence tolerance: 1.099e-08

A few other comments: Why do you log the x's? log(a exp(bx)) = log(a) + bx

Also, you can write it more compact as

initial_model <- lm(log(y) ~ x, data=csvfile)
f <- as.formula(paste(as.name(colnames(csvfile)[2]),"~I(a*exp(b*",as.name(colnames(csvfile[1])),"))"))
expo <- nls(f, data=csvfile, start=list(a=exp(coef(initial_model)[1]),b=coef(initial_model)[2]))

Why does nls give me an error when called from within ggplot?

When you change your scale, the formula also needs to be changed. Here is a possible solution, although I somehow cannot get confidence intervals to work.

myEquation=y ~ min+((max-min)/(1+10^(ec50-(x))))
ggplot(data=myData,aes(x=x,y=y))+geom_point()+scale_x_log10()+
  geom_smooth(method="nls", formula = myEquation, start = startingGuess, se=FALSE) 

UPDATE: Apparently the reason why confidence intervals do not work, is because standard errors are not currently implemented in predict.nls. Therefore ggplot also cannot display confidence intervals.

---

Related Topics