A comparison of methods for the imputation of human resources information in the modelling of under-five mortality rate outcome
Seminari GRASS - 29 de novembre 2006
DATE: 29 de novembre 2006 a les 12 hores
SPEAKER: Nuria Pérez Álvarez
IDIOM: Català
PLACE: Aula C5016, Edifici C5, Campus Nord, UPC
SUMMARY:
Background
Only a few studies have investigated the relation between human resources for health and health outcomes, and they reached different conclusions, some report no association whereas others find that high densities can be associated with better or worse health outcomes; moreover a possible generalization of the results is hindered by differences on the methodology, units and explanatory variables from one study to another. We investigated this relation via Bayesian methodology. The influence on the results simulating different missingness mechanism was evaluated.
Methods
The relationship between under-five mortality rate and density of human resources for health was modelled controlling for the effects of income and female adult literacy. For the covariate "density of human resources" missing data were simulated under the assumptions of "Missing completely at random" (MCAR), "Missing at random" (MAR) and "Missing not at random" (MNAR). The model was fitted by Complete Case analysis (CC), where the analysis is done only with data from the subjects who have observed outcome and covariates, using all data sets, using the data after MAR imputation (for MAR and MNAR generated dropout) and using the data after MNAR imputation for the set containing dropout generated following a MNAR mechanism in the covariate density of human resources.
Results and conclusions
The models indicated that larger density of human resources for health, available education to women, as well as an improved economy, were found to reduce under five-mortality significantly. Under MCAR, CC analysis can be acceptable if the rate of missingness is small. When the missingness is at random and the rate of missingness is considerably small the MAR-imputation approach provides accuratest estimators than CC analysis. For MNAR, when the rate of missingness is small, the MNAR imputation seems appropriate. In all cases, the increase in missingness percentage spoils all the intents to solve the missingness problem.
SPEAKER: Nuria Pérez Álvarez
IDIOM: Català
PLACE: Aula C5016, Edifici C5, Campus Nord, UPC
SUMMARY:
Background
Only a few studies have investigated the relation between human resources for health and health outcomes, and they reached different conclusions, some report no association whereas others find that high densities can be associated with better or worse health outcomes; moreover a possible generalization of the results is hindered by differences on the methodology, units and explanatory variables from one study to another. We investigated this relation via Bayesian methodology. The influence on the results simulating different missingness mechanism was evaluated.
Methods
The relationship between under-five mortality rate and density of human resources for health was modelled controlling for the effects of income and female adult literacy. For the covariate "density of human resources" missing data were simulated under the assumptions of "Missing completely at random" (MCAR), "Missing at random" (MAR) and "Missing not at random" (MNAR). The model was fitted by Complete Case analysis (CC), where the analysis is done only with data from the subjects who have observed outcome and covariates, using all data sets, using the data after MAR imputation (for MAR and MNAR generated dropout) and using the data after MNAR imputation for the set containing dropout generated following a MNAR mechanism in the covariate density of human resources.
Results and conclusions
The models indicated that larger density of human resources for health, available education to women, as well as an improved economy, were found to reduce under five-mortality significantly. Under MCAR, CC analysis can be acceptable if the rate of missingness is small. When the missingness is at random and the rate of missingness is considerably small the MAR-imputation approach provides accuratest estimators than CC analysis. For MNAR, when the rate of missingness is small, the MNAR imputation seems appropriate. In all cases, the increase in missingness percentage spoils all the intents to solve the missingness problem.
Share: