--- title: "Estimation of unknown elements in admixture models" author: "Xavier Milhaud" bibliography: references.bib output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{Estimation of unknown elements in admixture models} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r, include = FALSE} knitr::opts_chunk$set( collapse = TRUE, comment = "#>", warning = FALSE, message = FALSE ) ``` ```{r setup} library(admix) ``` We remind that a random variable $X$ following an admixture distribution has cumulative distribution function (cdf) $L$ given by $$L(x) = pF(x) + (1-p)G(x), \qquad x \in \mathbb{R},$$ where $G$ is a mixture component whose distribution is perfectly known, whereas $p$ and $F$ are unknown. In this setting, if no parametric assumption is made on the unknown component distribution $F$, the mixture is considered as a semiparametric mixture. For an overview on semiparametric extensions of finite mixture models, see [@Xiang19]. # Estimation of the unknown component weight in an admixture model The mixture weight $p$ of the unknown component distribution can be estimated using diverse techniques depending on the assumptions made on the unknown cdf $F$, among which the ones discussed in the sequel: * the estimator provided by Bordes and Vandekerkhove, see [@BordesVandekerkhove2010]; * the estimator provided by Patra and Sen, see [@PatraSen2016]; * the estimator provided by the Inversion - Best Matching method, see [@MilhaudPommeretSalhiVandekerkhove2024a]. All these estimation methods can be performed using one single generic function for estimation with appropriate arguments, the so-called $admix\_estim$ function. ## The one-sample case Many works studied the estimation of the unknown proportion in two-component admixture models. Among them, seminal papers are [@BordesDelmasVandekerkhove2006] and [@BMM06]. These papers are closely connected to the paper by [@BordesVandekerkhove2010], where an asymptotic normal estimator is provided for the unknown component weight. ### Case of symmetric unknown density In this case, we use the Bordes and Vandekerkhove estimator, see [@BordesVandekerkhove2010]. ```{r} ## Simulate mixture data: mixt1 <- twoComp_mixt(n = 400, weight = 0.7, comp.dist = list("norm", "norm"), comp.param = list(c("mean" = 3, "sd" = 0.5), c("mean" = 0, "sd" = 1))) data1 <- getmixtData(mixt1) ## Define the admixture model: admixMod <- admix_model(knownComp_dist = mixt1$comp.dist[[2]], knownComp_param = mixt1$comp.param[[2]]) admix_estim(samples = list(data1), admixMod = list(admixMod), est.method = 'BVdk', sym.f = TRUE) ``` Because this estimation method relies on the symmetry of the unknown component density, the estimator provides both the estimated mixing weight of the unknown component distribution and the estimated location shift parameter. ### Other cases In full generality (no assumptions made on the unknown component distribution), we use the Patra and Sen estimator, see [@PatraSen2016]. ```{r} admix_estim(samples = list(data1), admixMod = list(admixMod), est.method = 'PS') ``` In this case, the only estimated parameter is the mixing proportion related to the unknown component distribution. ## The two-sample case In the two-sample setting, one idea could be to use the Inversion - Best Matching (IBM) approach. The IBM method ensures asymptotically normal estimators of the unknown quantities, which will be very useful in a testing perspective. However, it is important to note that such estimators are mostly biased when $F_1 \neq F_2$, and general one-sample estimation strategies such as [@PatraSen2016] or [@BordesVandekerkhove2010] may be preferred to estimate the unknown component proportion in general settings (despite that this is more time-consuming). In the latter case, one performs twice the estimation method, on each of the two samples under study. ### Under the null hypothesis $H_0: F_1 = F_2$ When we are under the null, @MilhaudPommeretSalhiVandekerkhove2024a show that the estimators is consistent towards the true parameter values. ```{r} ## Simulate mixture data: mixt1 <- twoComp_mixt(n = 450, weight = 0.4, comp.dist = list("norm", "norm"), comp.param = list(list("mean" = -2, "sd" = 0.5), list("mean" = 0, "sd" = 1))) mixt2 <- twoComp_mixt(n = 380, weight = 0.7, comp.dist = list("norm", "norm"), comp.param = list(list("mean" = -2, "sd" = 0.5), list("mean" = 1, "sd" = 1))) data1 <- getmixtData(mixt1) data2 <- getmixtData(mixt2) ## Define the admixture models: admixMod1 <- admix_model(knownComp_dist = mixt1$comp.dist[[2]], knownComp_param = mixt1$comp.param[[2]]) admixMod2 <- admix_model(knownComp_dist = mixt2$comp.dist[[2]], knownComp_param = mixt2$comp.param[[2]]) admix_estim(samples = list(data1, data2), admixMod = list(admixMod1, admixMod2), est.method = 'IBM') ``` Indeed, one can see that the two unknown proportions were consistently estimated. ### Under the alternative hypothesis $H_1: F_1 \neq F_2$ Estimators are also consistent under $H_1$, although they can be (strongly) biased as compared to their true values as illustrated in the following example. ```{r} ## Simulate mixture data: mixt1 <- twoComp_mixt(n = 800, weight = 0.5, comp.dist = list("norm", "norm"), comp.param = list(list("mean" = 1, "sd" = 0.5), list("mean" = 0, "sd" = 1))) mixt2 <- twoComp_mixt(n = 600, weight = 0.7, comp.dist = list("norm", "norm"), comp.param = list(list("mean" = 3, "sd" = 0.5), list("mean" = 5, "sd" = 2))) data1 <- getmixtData(mixt1) data2 <- getmixtData(mixt2) ## Define the admixture models: admixMod1 <- admix_model(knownComp_dist = mixt1$comp.dist[[2]], knownComp_param = mixt1$comp.param[[2]]) admixMod2 <- admix_model(knownComp_dist = mixt2$comp.dist[[2]], knownComp_param = mixt2$comp.param[[2]]) ## Estimate the mixture weights of the two admixture models (provide only hat(theta)_n): admix_estim(samples = list(data1, data2), admixMod = list(admixMod1, admixMod2), est.method = 'IBM') ``` In such a framework, it is therefore better to use the estimator by [@PatraSen2016], which shows better performance: ```{r} admix_estim(samples = list(data1, data2), admixMod = list(admixMod1, admixMod2), est.method = 'PS') ``` # Estimation of the unknown cumulative distribution function Concerning the unknown cdf $F$, one usually estimate it thanks to the inversion formula $$F(x) = \dfrac{L(x) - (1-p)G(x)}{p},$$ once $p$ has been consistenly estimated. This is what is commonly called the decontaminated density of the unknown component. In the following, we propose to compare the two decontaminated densities obtained once the unknown quantities have been consistently estimated by the IBM approach. Note that we are under the null ($F_1=F_2$), and thus that the decontaminated densities should look similar. ```{r} ## Simulate mixture data: mixt1 <- twoComp_mixt(n = 800, weight = 0.4, comp.dist = list("norm", "norm"), comp.param = list(list("mean" = 3, "sd" = 0.5), list("mean" = 0, "sd" = 1))) mixt2 <- twoComp_mixt(n = 700, weight = 0.6, comp.dist = list("norm", "norm"), comp.param = list(list("mean" = 3, "sd" = 0.5), list("mean" = 5, "sd" = 2))) data1 <- getmixtData(mixt1) data2 <- getmixtData(mixt2) ## Define the admixture models: admixMod1 <- admix_model(knownComp_dist = mixt1$comp.dist[[2]], knownComp_param = mixt1$comp.param[[2]]) admixMod2 <- admix_model(knownComp_dist = mixt2$comp.dist[[2]], knownComp_param = mixt2$comp.param[[2]]) ## Estimation: est <- admix_estim(samples = list(data1,data2), admixMod = list(admixMod1,admixMod2), est.method = 'PS') prop <- getmixingWeight(est) ## Determine the decontaminated version of the unknown density by inversion: res1 <- decontaminated_density(sample1 = data1, estim.p = prop[1], admixMod = admixMod1) res2 <- decontaminated_density(sample1 = data2, estim.p = prop[2], admixMod = admixMod2) ## Use appropriate sequence of x values: plot(x = res1, x_val = seq(from = 0, to = 6, length.out = 100), add_plot = FALSE) plot(x = res2, x_val = seq(from = 0, to = 6, length.out = 100), add_plot = TRUE, col = "red") ``` # References