: approximate standard error of skew is
; normal distribution has 0 skew
: approximate standard error of kurtosis is
; normal distribution has kurtosis 3
: essentially singular value decomposition with some standard matrix manipulation
: measure of strength of linear correlation between two variables
: a measure of strength of monotonic relationship between
after the linear effects of other parameters
have been removed
: useful for simulating dependent variables with given marginals or modeling dependence separate from marginal models.
: Asymptotic distribution of the Wald statistic, which is based on an estimate and its variance of a single parameter, allows one to test simple hypotheses regarding a single parameter in a parametric model. Prefer LRT.
Likelihood Ratio Test
: A mechanism for testing nested hypotheses about parameters in a parametric model. Relies on asymptotic distribution of the LRT statistic.
: Alternative to Wald test.
Hotelling's T^2 Statistic
: generalization of Student's t.
: for computing CI on ratios of means.
Generalized Additive Model
: flexible models for parametrically or nonparametrically relating covariates to response via sums of functions of covariates.
: method to fit nonparametric generalized additive regression model
: dimension reduction for discrete variables
: A method to take high dimensional objects and order them along axes, so most similar objects are closest to each other.
: factor analysis, observed variables are modeled as linear combinations of a few, unobserved latent variables; PCA is related, but a merely descriptive technique
: extension of correlation to relationships between two sets of variables
Factor Analysis (CFA)?
: like PCA, but seeks factors with highest canonical correlation with the response
Multiple Factor Analysis (MFA)
: factor analysis when the variables are grouped (also see Canonical correlation)
: approximation of an integral
: Lehmann, E. L. (1986) Testing Statistical Hypotheses.
2nd Ed. John Wiley and Sons, Hoboken, NJ, USA, section 10.2 claims that inference may be done conditional on the initial states if the goal is inference of transition probabilities only. I guess this makes sense. You still need the stationary distribution for computing likelihoods of test sequences.