caculate confidence interval from customized pdf

2024 3 19
1 个回答
回答已采纳
更新时间：2024 3 19
5 次查看（30 天）

0 个投票

Hi
I'm wondering How can I caculate the confidence interval of customized pdf e.g. Gaussian mixture distribution?
pdf=@(x) w1*normpdf(x,mu1,sigma1)+w2*normpdf(x,mu2,sigma2);
cdf=@(x) integral(pdf,-Inf,x);
As icdf function only support specified distribution, I'm wondering how to caculate the shortest confidence interval?

请先登录，再回答此问题。

 采纳的回答

David Goodmanson
David Goodmanson 2024-3-19
编辑：David Goodmanson 2024-3-19

1 个投票

Hello SX,
Ordinarily to find the cdfs you would have to use numerical integration. In this case for the normal distributions, the cdf function is available. Then you can interpolate using the cdf as the independent variable. Here is an example. In the plot you get a wide minimum which you might expect.
mu1 = 1;
mu2 = 2;
sig1 = 1;
sig2 = 3;
w1 = .3;
w2 = .7;
c = .9; % confidence span, there is probably a better name for this
x = -20:.00001:20;
cdf = w1*normcdf(x,mu1,sig1) +w2*normcdf(x,mu2,sig2);
cdn = linspace(min(cdf),max(cdf)-c,1e4);
xdn = interp1(cdf,x,cdn);
cup = linspace(min(cdf)+c,max(cdf),1e4);
xup = interp1(cdf,x,cup);
figure(1); grid on
plot(xup-xdn)
[x0 ind] = min(xup-xdn);
xdn(ind) % lower end of confidence interval
xup(ind) % upper end of confidence interval
cdn(ind) % lower cdf value
cup(ind) % upper cdf value
% ans = -2.4087
% ans = 6.3858
% ans = 0.0497
% ans = 0.9497
D = xup(ind)-xdn(ind) % the result
cup(ind)-cdn(ind) % check, should be c = confidence span
% D = 8.7945
% ans = 0.9000
% try a different case, get a larger confidence interval
xtest = interp1(cdf,x,[.07 .97]);
Dtest = diff(xtest)
% xtest = -1.8602 7.1554
% Dtest = 9.0156

3 个评论
显示 1更早的评论 隐藏 1更早的评论

苏越 徐
苏越 徐 2024-3-19
编辑：苏越 徐 2024-3-19
Thank you David for your valuable suggestion! However, as the pdf of gaussian mixture distribution is asymmetrical, [.05 .95] is not the shortest interval for the given confidence value(e.g. 90% in your example). I guess the function fmincon may help to find the shortest interval, but is there any function easy to efficienctly find the shortest interval by numerical methods?
David Goodmanson
David Goodmanson 2024-3-19
See what you think of the modified answer above
苏越 徐
苏越 徐 2024-3-19
Thank you David! It' good idea to narrow the range for ends of confidence span and then search for the shortest.

请先登录，再进行评论。

更多回答（0 个）

类别

在 帮助中心 和 File Exchange 中查找有关 Gaussian Mixture Models 的更多信息

标签

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by