Page 82 - ISMCON souvenir 2021

ISMSCON - 2021

period. On comparing gender distribution more number of male patients died as compared to
female patients but it was not significant. Age group analysis presented that majority of patients
died in 60-74 age group whereas majority of patients died having length of stay 0-7 days. Bivariate
Logistic regression reflect that there was significant influence of age group and length of stay on the
mortality of Covid19 positive patients (χ2(9) =495.30, P<.001).
Conclusion: Compulsory vaccination in significant Age Group and early detection is essential for
reducing mortality of Covid 19 patients
Keywords: COVID-19, AGE, Gender, Length of stay, Mortality, Demographic Risk Factors

OS54: COMPARISON OF IMPUTATION METHODS USING RE-

SAMPLING TECHNIQUES IN NFHS-4 DATA

Seena Thomas K , K. Thennarasu 2
1
1Assistant Professor, Department of Statistics, Christ (Deemed to be) University, Bengaluru
2Professor and Head, Department of Biostatistics, NIMHANS, Bangalore
seena.thomas@christuniversity.in, kthenna@gmail.com

INTRODUCTION: The field of survey research is an important area where incomplete data occur. After
imputing the missing values in the survey data, resampling technique can be used to check the efficiency
of the imputation methods, since there is no full data to compare the estimates. Resampling is a procedure
which permits to draw samples again and again from a dataset and to refit the model for each sample in
order to get additional information.

METHODOLOGY: NFHS-4 Delhi data for women was considered for the current study of comparison of
imputation methods which had the highest proportion of missing data compared to all other regions across
the country. Delhi region data, consists of 5914 subjects among which 1337 (22.61%) had unobserved
data for at least one of the variables considered in the study. From the complete data with 4577 (77.39%)
subjects, subsamples of varying sample sizes 50, 100, 200, 500, 800 and 1000 were selected randomly
and created missing-ness under varying proportions of missing data (0.1, 0.2, 0.3, 0.4, 0.5) under MAR
mechanism. During this process 1000 datasets of each sample sizes were selected randomly, induced
missing-ness and imputation was done using various imputation methods. Multiple linear regression
model was fitted for to predict BMI from the covariates Age, Hb, Glucose and SBP. The imputation
methods were compared using the point estimates (mean, SD and SE), regression coefficients, R Akaike
2,
information criterion (AIC) and Bayesian information criterion (BIC) values of multiple linear regression.
The p-values produced by the comparison of imputed data sets were considered for the Hotelling T2 test
and Box’s M test.
RESULTS: The model based methods Multiple Imputation (MI), Quantile Regression (QR) imputation
(QR), Regression Imputation (REG) and stochastic Regression Imputation (SREG) performed better
than donor based methods while comparing with the estimates of complete data. Among the donor based
methods Fractional Imputation method (FI) and Propensity score method (PS) were better than the other
donor based methods.
CONCLUSION: All the model based methods performed better than donor based methods under the
MAR mechanism while comparing with the estimates of complete data.
KEYWORDS: Simulation, MAR mechanism, NFHS data, Multiple Imputation, Quantile Regression
Imputation.

80 CONFERENCE SOUVENIR