Re: [R] Fwd: Strange results : bootrstrp CIs
On 13/01/2024 8:58 p.m., Rolf Turner wrote:
On Sat, 13 Jan 2024 17:59:16 -0500
Duncan Murdoch wrote:
My guess is that one of the bootstrap samples had a different
selection of countries, so factor(Country) had different levels, and
that would really mess things up.
You'll need to decide how to handle that: If you are trying to
estimate the coefficient for Italy in a sample that contains no data
from Italy, what should the coefficient be?
Perhaps NA? Ben Bolker conjectured that boot() might be able to handle
this. Getting the NAs into the coefficients is a bit of a fag, but. I
tried:
My question was really intended as a statistical question. From a
statistical perspective, if I have a sampling scheme that sometimes
generates sample size 0, should my CI be (-Inf, Inf) for high enough
confidence level?
A Bayesian might say that inference should be entirely based on the
prior in the case of no relevant data. You could get similar numerical
results by adding some fake data to every bootstrap sample, e.g. a
single weighted observation for each country at your prior mean for that
country, with weight chosen to match the strength of the prior. But
Bayesian methods don't give confidence intervals, they give credible
intervals, and those aren't the same thing even if they are sometimes
numerically similar.
Duncan Murdoch
func <- function(data, idx) {
clyde <- coef(lm(Score~ Time + factor(Country),data=data))
ccc <- coef(lm(Score~ Time + factor(Country),data=data[idx,]))
urk <- rep(NA,length(clyde))
names(urk) <-names(clyde)
urk[names(ccc)] <- ccc
urk
}
It produced a result:
set.seed(42)
B= boot(e, func, R=1000)
B
ORDINARY NONPARAMETRIC BOOTSTRAP
Call:
boot(data = e, statistic = func, R = 1000)
Bootstrap Statistics :
original biasstd. error
t1* 609.62500 3.620405295.39452
t2* -54.81250 -1.662470436.32911
t3* -41.3 -2.7337992 100.72113
t4* -96.0 -1.099571899.78864
t5* -126.0 -0.654888663.47076
t6* -26.3 -1.651668387.80483
t7* -15.7 -0.839117091.72467
t8* -21.7 -5.454401383.69211
t9* 18.3 -0.771100185.57278
However I have no idea if the result is correct, or even meaningful. I
have no idea what I'm doing. Just hammering and hoping. 😊️
cheers,
Rolf
__
R-help@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.
Re: [R] Fwd: Strange results : bootrstrp CIs
Dear R-experts,
I really thank you all for your responses.
Best,
Le dimanche 14 janvier 2024 à 10:22:12 UTC+1, Duncan Murdoch
a écrit :
On 13/01/2024 8:58 p.m., Rolf Turner wrote:
> On Sat, 13 Jan 2024 17:59:16 -0500
> Duncan Murdoch wrote:
>
>
>
>> My guess is that one of the bootstrap samples had a different
>> selection of countries, so factor(Country) had different levels, and
>> that would really mess things up.
>>
>> You'll need to decide how to handle that: If you are trying to
>> estimate the coefficient for Italy in a sample that contains no data
>> from Italy, what should the coefficient be?
>
> Perhaps NA? Ben Bolker conjectured that boot() might be able to handle
> this. Getting the NAs into the coefficients is a bit of a fag, but. I
> tried:
My question was really intended as a statistical question. From a
statistical perspective, if I have a sampling scheme that sometimes
generates sample size 0, should my CI be (-Inf, Inf) for high enough
confidence level?
A Bayesian might say that inference should be entirely based on the
prior in the case of no relevant data. You could get similar numerical
results by adding some fake data to every bootstrap sample, e.g. a
single weighted observation for each country at your prior mean for that
country, with weight chosen to match the strength of the prior. But
Bayesian methods don't give confidence intervals, they give credible
intervals, and those aren't the same thing even if they are sometimes
numerically similar.
Duncan Murdoch
> func <- function(data, idx) {
> clyde <- coef(lm(Score~ Time + factor(Country),data=data))
> ccc <- coef(lm(Score~ Time + factor(Country),data=data[idx,]))
> urk <- rep(NA,length(clyde))
> names(urk) <-names(clyde)
> urk[names(ccc)] <- ccc
> urk
> }
>
> It produced a result:
>
>>> set.seed(42)
>>> B= boot(e, func, R=1000)
>> B
>>
>> ORDINARY NONPARAMETRIC BOOTSTRAP
>>
>>
>> Call:
>> boot(data = e, statistic = func, R = 1000)
>>
>>
>> Bootstrap Statistics :
>> original bias std. error
>> t1* 609.62500 3.6204052 95.39452
>> t2* -54.81250 -1.6624704 36.32911
>> t3* -41.3 -2.7337992 100.72113
>> t4* -96.0 -1.0995718 99.78864
>> t5* -126.0 -0.6548886 63.47076
>> t6* -26.3 -1.6516683 87.80483
>> t7* -15.7 -0.8391170 91.72467
>> t8* -21.7 -5.4544013 83.69211
>> t9* 18.3 -0.7711001 85.57278
>
> However I have no idea if the result is correct, or even meaningful. I
> have no idea what I'm doing. Just hammering and hoping. 😊️
>
>
>
> cheers,
>
> Rolf
>
__
R-help@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.
__
R-help@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.
[R] Plotting extrapolation with R like AUTOBOX does
Dear R-experts, I write to you to know if somebody is aware of a R package (or function) able to plot graphs for extrapolation. I need to be clear on what extrapolation really is to me. It is when we use the model for X variables outside the range of X variables that were used to construct the model and estimates. What I am really looking for is that beyond confidence intervals for predictions for the one estimated model, I want intervals that reflect the uncertainty over the models that could have fit the data as well. I know that Clive Granger worried about this a lot, and wrote a few papers (he called them 'thick' confidence intervals, taking into account model uncertainty). I found AUTOBOX software (https://autobox.com/cms/) allowing multiple time series to be modeled as possible inputs without requiring time as a possible predictor thus it can reproduce ordinary regression. More precisely, I am looking for an R package (or function) able to produce the same 4 graphs (done with AUTOBOX) in the answer here : https://stats.stackexchange.com/questions/327454/right-way-to-extrapolate-data Best, __ R-help@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.
