Generalized Cross Entropy Method for estimating joint distribution from incomplete information
| Autor: | Erika Fille Legara, Hai-Yan Xu, Christopher Monterola, Shyh-hao Kuo, Daxuan Zhao, Guoqi Li |
|---|---|
| Rok vydání: | 2016 |
| Předmět: |
Statistics and Probability
Estimation Kullback–Leibler divergence Computer science Principle of maximum entropy Cross-entropy method 02 engineering and technology Condensed Matter Physics 01 natural sciences Joint entropy 010305 fluids & plasmas Joint probability distribution Complete information 0103 physical sciences Statistics 0202 electrical engineering electronic engineering information engineering Econometrics 020201 artificial intelligence & image processing Marginal distribution |
| Zdroj: | Physica A: Statistical Mechanics and its Applications. 453:162-172 |
| ISSN: | 0378-4371 |
| DOI: | 10.1016/j.physa.2016.02.023 |
| Popis: | Obtaining a full joint distribution from individual marginal distributions with incomplete information is a non-trivial task that continues to challenge researchers from various domains including economics, demography, and statistics. In this work, we develop a new methodology referred to as “Generalized Cross Entropy Method” (GCEM) that is aimed at addressing the issue. The objective function is proposed to be a weighted sum of divergences between joint distributions and various references. We show that the solution of the GCEM is unique and global optimal. Furthermore, we illustrate the applicability and validity of the method by utilizing it to recover the joint distribution of a household profile of a given administrative region. In particular, we estimate the joint distribution of the household size, household dwelling type, and household home ownership in Singapore. Results show a high-accuracy estimation of the full joint distribution of the household profile under study. Finally, the impact of constraints and weight on the estimation of joint distribution is explored. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|
Full Text from ScienceDirect