[R] How to calculate standard error of estimate (S) for my non-linear regression model?
Jeff Newmiller
jdnewmil at dcn.davis.ca.us
Sat Sep 26 22:26:34 CEST 2015
You may not be a statistician, but you should at least learn about the
calculations you are making. You cannot expect to convince others that
your calculations are right just because "Peter on the internet said they
were right".
To give you a gentle push in this direction, I have reproduced the
calculations on that reference web page using R, so you can get a head
start on understanding how to perform them on your data. (Hint: about all
you should need to do is give your dta.nls to predict and use your
column names instead of X and Y.) Once you have convinced yourself that
the pre-defined functions are doing what you expect, then you can omit the
do-it-yourself option with confidence in the results.
Please note that no use of attach is included here... that usually ends in
unhappiness at some point, so prefer to use the data argument instead.
###
dta <- data.frame( X=c(1,2,3,4,5), Y=c(1,2,1.3,3.75,2.25) )
nrow( dta )
# linear regression
y.lm <- lm( Y~X, data=dta )
# compute predictions from original data
## Be aware that you can also "predict" using a nonlinear regression fit
## Also be aware that you can compute estimates using different data if you
## specify the "newdata" argument... see the help for predict.lm
## ?predict.lm
dta$Yprime <- predict( y.lm )
# ?with
dta$YlessYprime <- with( dta, Y - Yprime )
dta$YlessYprime2 <- with( dta, YlessYprime^2 )
# confirm sums
# ?sum
sum( dta$X )
sum( dta$Y )
sum( dta$Yprime )
sum( dta$YlessYprime )
sum( dta$YlessYprime2 )
# standard error of the estimate for population data
sigma.est <- sqrt( sum( dta$YlessYprime2 ) / nrow( dta ) )
sigma.est
# sd function assumes sample standard deviation, can correct the result if
# you want
# ?sd
# ?all.equal
all.equal( sigma.est, sd( dta$YlessYprime ) * sqrt( ( nrow( dta ) - 1 ) /
nrow( dta ) ) )
# alternate formulation
SSY <- sum( ( dta$Y - mean( dta$Y ) )^2 )
rho <- with( dta, cor( Y, X ) )
all.equal( sigma.est, sqrt( (1-rho^2)*SSY/nrow(dta) ) )
# when working with a sample...
s.est <- sqrt( sum( dta$YlessYprime2 ) / ( nrow( dta ) - 2 ) )
s.est
####
> dta <- data.frame( X=c(1,2,3,4,5), Y=c(1,2,1.3,3.75,2.25) )
> nrow( dta )
[1] 5
> # linear regression
> y.lm <- lm( Y~X, data=dta )
> # compute predictions from original data
> ## Be aware that you can also "predict" using a nonlinear regression fit
> ## Also be aware that you can compute estimates using different data if you
> ## specify the "newdata" argument... see the help for predict.lm
> ## ?predict.lm
> dta$Yprime <- predict( y.lm )
> # ?with
> dta$YlessYprime <- with( dta, Y - Yprime )
> dta$YlessYprime2 <- with( dta, YlessYprime^2 )
>
> # confirm sums
> # ?sum
> sum( dta$X )
[1] 15
> sum( dta$Y )
[1] 10.3
> sum( dta$Yprime )
[1] 10.3
> sum( dta$YlessYprime )
[1] 2.220446e-16
> sum( dta$YlessYprime2 )
[1] 2.79075
>
> # standard error of the estimate for population data
> sigma.est <- sqrt( sum( dta$YlessYprime2 ) / nrow( dta ) )
> sigma.est
[1] 0.7470944
> # sd function assumes sample standard deviation, can correct the result
> # if you want
> # ?sd
> # ?all.equal
> all.equal( sigma.est, sd( dta$YlessYprime ) * sqrt( ( nrow( dta ) - 1 )
/ nrow( dta ) ) )
[1] TRUE
>
> # alternate formulation
> SSY <- sum( ( dta$Y - mean( dta$Y ) )^2 )
> rho <- with( dta, cor( Y, X ) )
> all.equal( sigma.est, sqrt( (1-rho^2)*SSY/nrow(dta) ) )
[1] TRUE
>
> # when working with a sample...
> s.est <- sqrt( sum( dta$YlessYprime2 ) / ( nrow( dta ) - 2 ) )
> s.est
[1] 0.9644947
>
On Sat, 26 Sep 2015, Michael Eisenring wrote:
> Dear Peter,
> Thank you for your answer.
> If I look at my summary I see there a Residual standard error: 1394 on 53
> degrees of freedom.
> This number is very high (the fit of the curve is pretty bad I know but
> still...). Are you sure the residual standard error given in the summary is
> the same as the one described on this page:
> http://onlinestatbook.com/2/regression/accuracy.html
> I am basically just looking for a value that describes the goodness of fit
> for my non-linear regression model.
>
>
> This is probably a pretty obvious question, but I am not a statistician and
> as you said the terminology is sometimes pretty confusing.
> Thanks mike
>
> -----Urspr?ngliche Nachricht-----
> Von: peter dalgaard [mailto:pdalgd at gmail.com]
> Gesendet: Samstag, 26. September 2015 01:43
> An: Michael Eisenring <michael.eisenring at gmx.ch>
> Cc: r-help at r-project.org
> Betreff: Re: [R] How to calculate standard error of estimate (S) for my
> non-linear regression model?
>
> This is one area in which terminology in (computational) statistics has gone
> a bit crazy. The thing some call "standard error of estimate" is actually
> the residual standard deviation in the regression model, not to be confused
> with the standard errors that are associated with parameter estimates. In
> summary(nls(...)) (and summary(lm()) for that matter), you'll find it as
> "residual standard error", and even that is a bit of a misnomer.
>
> -pd
>
>> On 26 Sep 2015, at 07:08 , Michael Eisenring <michael.eisenring at gmx.ch>
> wrote:
>>
>> Hi all,
>>
>> I am looking for something that indicates the goodness of fit for my
>> non linear regression model (since R2 is not very reliable).
>>
>> I read that the standard error of estimate (also known as standard
>> error of the regression) is a good alternative.
>>
>>
>>
>> The standard error of estimate is described on this page (including
>> the
>> formula) http://onlinestatbook.com/2/regression/accuracy.html
>> <https://3c.gmx.net/mail/client/dereferrer?redirectUrl=http%3A%2F%2Fon
>> linest atbook.com%2F2%2Fregression%2Faccuracy.html>
>>
>> Unfortunately however, I have no clue how to programm it in R. Does
>> anyone know and could help me?
>>
>> Thank you very much.
>>
>>
>>
>> I added an example of my model and a dput() of my data
>>
>> #CODE
>>
>> dta<-read.csv("Regression_exp2.csv",header=T, sep = ",")
>> attach(dta) # tells R to do the following analyses on this dataset
>> head(dta)
>>
>>
>>
>> # loading packages: analysis of mixed effect models
>> library(nls2)#model
>>
>> #Aim: fit equation to data: y~yo+a*(1-b^x) : Two parameter exp. single
>> rise to the maximum # y =Gossypol (from my data set) x= Damage_cm
>> (from my data set) #The other 3 parameters are unknown: yo=Intercept,
>> a= assymptote ans b=slope
>>
>> plot(Gossypol~Damage_cm, dta)
>> # Looking at the plot, 0 is a plausible estimate for y0:
>> # a+y0 is the asymptote, so estimate about 4000; # b is between 0 and
>> 1, so estimate .5 dta.nls <- nls(Gossypol~y0+a*(1-b^Damage_cm), dta,
>> start=list(y0=0, a=4000, b=.5))
>>
>> xval <- seq(0, 10, 0.1)
>> lines(xval, predict(dta.nls, data.frame(Damage_cm=xval)))
>> profile(dta.nls, alpha= .05)
>>
>>
>> summary(dta.nls)
>>
>>
>>
>>
>>
>>
>>
>> #INPUT
>>
>> structure(list(Gossypol = c(948.2418407, 1180.171957, 3589.187889,
>> 450.7205451, 349.0864019, 592.3403778, 723.885643, 2005.919344,
>> 720.9785449, 1247.806111, 1079.846532, 1500.863038, 4198.569251,
>> 3618.448997, 4140.242559, 1036.331811, 1013.807628, 2547.326207,
>> 2508.417927, 2874.651764, 1120.955, 1782.864308, 1517.045807,
>> 2287.228752, 4171.427741, 3130.376482, 1504.491931, 6132.876396,
>> 3350.203452, 5113.942098, 1989.576826, 3470.09352, 4576.787021,
>> 4854.985845, 1414.161257, 2608.716056, 910.8879471, 2228.522959,
>> 2952.931863, 5909.068158, 1247.806111, 6982.035521, 2867.610671,
>> 5629.979049, 6039.995102, 3747.076592, 3743.331903, 4274.324792,
>> 3378.151945, 3736.144027, 5654.858696, 5972.926124, 3723.629772,
>> 3322.115942, 3575.043632, 2818.419785), Treatment = structure(c(5L,
>> 5L, 5L, 5L, 5L, 5L, 5L, 5L, 5L, 5L, 5L, 1L, 1L, 1L, 1L, 1L, 1L, 1L,
>> 1L, 1L, 1L, 1L, 1L, 3L, 3L, 3L, 3L, 2L, 2L, 2L, 3L, 3L, 3L, 3L, 3L,
>> 3L, 3L, 3L, 2L, 2L, 2L, 4L, 2L, 4L, 4L, 2L, 4L, 2L, 2L, 4L, 4L, 4L,
>> 4L, 4L, 4L, 2L), .Label = c("1c_2d", "1c_7d", "3c_2d", "9c_2d", "C"),
>> class = "factor"), Damage_cm = c(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
>> 0.142, 0.4035, 0.4435, 0.491, 0.4955, 0.578, 0.5895, 0.6925, 0.6965,
>> 0.756, 0.8295, 1.0475, 1.313, 1.516, 1.573, 1.62, 1.8115, 1.8185,
>> 1.8595, 1.989, 2.129, 2.171, 2.3035, 2.411, 2.559, 2.966, 2.974,
>> 3.211, 3.2665, 3.474, 3.51, 3.547, 4.023, 4.409, 4.516, 4.7245, 4.809,
>> 4.9835, 5.568, 5.681, 5.683, 7.272, 8.043, 9.437, 9.7455),
>> Damage_groups = c(0.278, 1.616, 2.501, 3.401, 4.577, 5.644, 7.272,
>> 8.043, 9.591, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA,
>> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA,
>> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA),
>> Gossypol_Averaged = c(1783.211, 3244.129, 2866.307, 3991.809,
>> 4468.809, 5121.309, 3723.629772, 3322.115942, 3196.731, NA, NA, NA,
>> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA,
>> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA,
>> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA), Groups = c(42006L, 42038L,
>> 42067L, 42099L, 42130L, 42162L, 42193L, 42225L, 42257L, NA, NA, NA,
>> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA,
>> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA,
>> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA)), .Names = c("Gossypol",
>> "Treatment", "Damage_cm", "Damage_groups", "Gossypol_Averaged",
>> "Groups"), class = "data.frame", row.names = c(NA, -56L))
>>
>>
>>
>>
>> [[alternative HTML version deleted]]
>>
>> ______________________________________________
>> R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
>> https://stat.ethz.ch/mailman/listinfo/r-help
>> PLEASE do read the posting guide
>> http://www.R-project.org/posting-guide.html
>> and provide commented, minimal, self-contained, reproducible code.
>
> --
> Peter Dalgaard, Professor,
> Center for Statistics, Copenhagen Business School Solbjerg Plads 3, 2000
> Frederiksberg, Denmark
> Phone: (+45)38153501
> Email: pd.mes at cbs.dk Priv: PDalgd at gmail.com
>
> ______________________________________________
> R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>
---------------------------------------------------------------------------
Jeff Newmiller The ..... ..... Go Live...
DCN:<jdnewmil at dcn.davis.ca.us> Basics: ##.#. ##.#. Live Go...
Live: OO#.. Dead: OO#.. Playing
Research Engineer (Solar/Batteries O.O#. #.O#. with
/Software/Embedded Controllers) .OO#. .OO#. rocks...1k
More information about the R-help
mailing list