Log-linear imputation / A. C. Singh. : CS11-613/88-29E-PDF

"A method of imputation based on log-linear methodology is proposed. For this purpose, an initial categorical transformation of all variables is made. Like hot deck imputation (HDI) method, the proposed log-linear imputation (LLI) method is applicable to both discrete and continuous variables. The LLI method generalizes HDI in several ways: (i) chi-square type measure of association is used to choose suitable predictors X for forming "optimal" imputation classes, (ii) the categorical distribution of the variable of interest, Z within an imputation class is model-based; and (iii) Z values are imputed under the constraint of proportional allocation to categories according to imputed proportions. As compared to the linear regression imputation (LRI) method, LLI requires a less restrictive framework. Thus LLI can be placed somewhere between HDI and LRL. Furthermore, since LLI uses model-based procedures for imputing counts corresponding to missing data, imputation variance can be assessed in estimating parameters within a certain class. This class of parameters describes characteristics of the population frequency distribution under the categorical framework. A modification of LLI is also proposed for the problem of statistical matching. This approach offers some control when the commonly used assumption of conditional independence is not valid. Use of supplementary information about conditional dependence, if available, can also be incorporated. Finally, some possible generalizations to the cases of general missing patterns and nonignorable nonresponse are indicated"--Abstract.

Lien permanent pour cette publication :
publications.gc.ca/pub?id=9.837181&sl=1

• Format MARC XML
• Format MARC HTML

Demander des formats alternatifs