Clément Viel
2006-Jul-04 15:11 UTC
[R] Problems when computing the 1rst derivative of mixtures of densities
Hi everybody, I am currently working on mixtures of two densities ( f(xi,teta)(1-teta)*f1(xi) + teta*f2(xi) ), particularly on the behavior of the variance for teta=0 (so sample only comes from the first distribution). To determine the maximum likelihood estimator I use the Newton-Rapdon Iteration. But when computing the first derivative I get a none linear function (with several asymptotes) which is completely absurd. This is my function to compute the first derivative: phy=function(teta,vect1,vect2){ return( sum(( vect2 - vect1) / (( 1 - teta) * vect1 + teta * vect2))) } note: vect1 and vect2 contains values of the two distributions computed from sample previously extracted. Beside, vect2 - vect1 is constant and ( 1 - teta) * vect1 + teta * vect2) is linear and always defined so I am expecting a linear function for the first derivative. To my mind, it's likely comes from the operation of division but I don't understand why results are skewed. Have you got any suggestions, please? -- Clément Viel Student in engineering school of Polytech-Lille http://www.polytech-lille.fr [[alternative HTML version deleted]]
Spencer Graves
2006-Jul-08 01:35 UTC
[R] Problems when computing the 1rst derivative of mixtures of densities
What problem are you trying to solve? The mixture proportion you are trying to estimate violates the assumptions that provide a normal approximation to the distribution of maximum likelihood estimators. These situations even violate the assumptions for 2*log(likelihood ratio) being approximately chi-square. If you want hypothesis tests or confidence intervals, Pinheiro and Bates (2000) recommended using 2*log(likelihood ratio) modified as suggested by Monte Carlo results. More recently, Bates has incorporated an 'mcmcsamp' function in the 'lme4' and 'Matrix' libraries. The best references I know on testing a parameter at a boundary are the following: Pinheiro and Bates (2000) Mixed-Effects Models in S and S-PLUS (Springer, sec. 2.4) Crainiceanu, Ruppert and Vogelsang (2003) ?Some properties of Likelihood Ratio tests in linear mixed models? Crainiceanu, Ruppert, Claeskens, and Wand (2005) ?Exact Likelihood Ratio Tests for Penalized Splines?, Biometrika, 92(1) Hope this helps. Spencer Graves p.s. See also inline. Cl?ment Viel wrote:> Hi everybody, > > I am currently working on mixtures of two densities ( f(xi,teta)> (1-teta)*f1(xi) + teta*f2(xi) ), > particularly on the behavior of the variance for teta=0 (so sample only > comes from the first distribution). > To determine the maximum likelihood estimator I use the Newton-Rapdon > Iteration. But when > computing the first derivative I get a none linear function (with several > asymptotes) which is completely > absurd.Why do you think this is absurd? Nonlinear estimation can be very difficult. Also, have you reviewed the capabilities in R for working with mixtures of distributions? I just got 167 hits for RSiteSearch("mixtures") and 49 for RSiteSearch("mixture distributions", "functions").> > This is my function to compute the first derivative: > > phy=function(teta,vect1,vect2){ > return( sum(( vect2 - vect1) / (( 1 - teta) * vect1 + teta * vect2))) > } > > note: vect1 and vect2 contains values of the two distributions computed > from sample previously extracted. > > Beside, vect2 - vect1 is constant and ( 1 - teta) * vect1 + teta * vect2) is > linear and always defined so I am expecting > a linear function for the first derivative. To my mind, it's likely comes > from the operation of division but I don't understand > why results are skewed. > > Have you got any suggestions, please? > > > > ------------------------------------------------------------------------ > > ______________________________________________ > R-help at stat.math.ethz.ch mailing list > https://stat.ethz.ch/mailman/listinfo/r-help > PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html