Construction and Selection of Parametric Models Flashcards
Maximum Likelihood Estimators
1=L(theta)=product of f(xi)
2=l(theta)= ln(L(theta))
3= derivative of l(theta)
4=Set 3=0 and isolate
Incomplete Data
1: left truncated at d = f(x)/S(d)
2: Right-censored at u = S(u)
3: Grouped data on interval (a,b) = Pr( a
Special Cases for MLE
Gamma, fixed alpha : theta= average/alpha
Normal : û = average
sigma^2 = sum of xi^2 / n - û^2
LogNormal : û=sum of ln(xi) /n
sigma^2 = sum of ln(xi)^2 /n - û^2
Poisson : lambda= average
Binomial fixed m : q= average/m
Negative Binomial fixed r : Beta = average /r
Zero-Truncated Distribution
Match E(X^t) to average
Zero-Modified Distribution
Match po^m to the proportion of zero observations Match E(X^m) to average
MLE Uniform on (0, theta)
theta = max(x1,x2,…)
Choosing from (a,b,0) class
1: compare u and sigma^2 ****
2: observe the slope of knk/nk-1
Poisson average=sigma^2 : slope=0
Binomial average > sigma^2 : negative slope
Negative Binomial average < sigma^2 : positive slope
Variance of MLE : Fisher Information
One parameter
I(theta)=-Ex(I’‘(theta))
Var(theta) = I(theta)^-1
Two Parameters
information matrix = I(alpha, theta)
information matrix ^-1 = Var(alpha), Var(theta)
diagonale= cov(alpha, theta)
Delta Approximation
one variable
Var(g(theta))= (d/dtheta g(theta))^2 Var(theta)
two variable
Var(g(alpha, theta)) = (g(alpha)’ )^2 Var(alpha) + 2galpha’gtheta’* Cov(alpha, theta) + (g(theta)’)^2 Var(theta)
Confidence Interval
Theta +- Z(1+p)/2 Var(theta)^0.5
Kolmogorov-Smirnov
Test Statistic : D = max(Dj) where
Dj=max(abs(Fn(xj)-F(xj)), abs(Fn(xj-1)-F(xj)))
If data is truncated at d, then
F*(x)= (F(x)-F(d))/(1-F(d)) for x > d
Kolmogorov-Smirnov Properties
Individual data only
Continuous fit only
Lower critical value for censored data
If parameters are estimated , critical value should be adjusted
Lower critical value if sample size is larger
no discretion
Uniform weight on all parts of distribution
D(x) Plot
Graph the difference between empirical CDF and fitter CDF
Peak = Fn(xj)-F*(xj) Valler = Fn(xj-1)-F*(xj)
P-p Plot
Coordinate : (Fn(xj), F*(xj))) where Fn(xj)= j/(n+1)
Hypothesis Tests: Chi-Square Goodness-of-Fit
Test Statistic : X^2 = sum (Ej-Oj)^2 / Ej or sum Oj^2/Ej - n
Degrees of freedom : k(number of group)-1-r(number of estimated parameters)
Chi-Square Properties
Individual and grouped data
continuous and discrete fit
no adjustments to critical value for censored data
If parameters are estimated , critical value is automatically adjusted via degrees of freedom
no change for critical value is sample size is large
more weights on intervals with poor fit
Hypothesis Test: Likelihood Ratio
Test statistic : T=2(l(theta1)-l(theta0))
Degrees of freedom = number of free parameters in H1 - number of free parameters in H0
Score-Based Approaches
SBC : Schwarz Bayesian Criterion = BIC : Bayesian Information Criterion = l-r/2*ln(n)
Akaike Information Criterion : AIC = l-r
l: log-likelihood
r: number of parameters
n: sample size
Select model with the highest AIC or BIC value