All time references are in CEST
new developments and applications in survey to survey imputation |
|
| Session Organisers | Dr Federico Crescenzi (University of Tuscia) Professor Gianni Betti (University of Siena) Professor Tiziana Laureti (University of Tuscia) Dr Lorenzo Mori (University of Bologna ) |
| Time | Tuesday 18 July, 09:00 - 10:30 |
| Room | |
Gathering data to produce accurate and timely welfare estimates re-
quires complex and costly surveys. Only a handful of countries can conduct
these surveys annually to estimate welfare and inequality. Most National
Statistical Offices (NSOs) worldwide carry out these surveys every four to
five years, making it challenging to generate reliable annual welfare indices.
Furthermore, even more common surveys do not follow a standard process
from one year to the next, with changing definitions and questions making
them incomparable. To address these issues, researchers have developed
methods to compare welfare indicators over time using surveys that may
not be directly comparable. These methods, known as Survey-to-Survey Im-
putation Techniques (SSITs), have been effective in predicting comparable
welfare indicators but have struggled with predicting comparable inequality
indicators.
SSITs are a derivative of the poverty map literature (Elbers et al.,
2003; Tarozzi and Deaton, 2009). Poverty maps, which involve imputing
income data onto censuses, have been used to generate geographically de-
tailed poverty estimates in many developing countries. Alongside survey-
to-census imputation, there has been a recent trend toward SSIT, which
involves mapping data from a surveys with certain information to those
with other outcomes of interest but lacking standard welfare aggregates.
The aim of this session is to further explore this argument by examin-
ing the pros and cons of SSITs. The global trend of conducting smaller or
non-probabilistic surveys, along with the increase in independent NSOs and
organizations, brings several benefits to global statistics. However, these
trends also risk reducing the comparability of welfare estimates. This re-
duction in comparability can be influenced by factors such as the definitions
used. SSITs can help by standardizing methods, which aids in obtaining
cross-nationally comparable estimates that transcend cultural boundaries.
Dr Lorenzo Mori (University of Bologna) - Presenting Author
Professor Gianni Betti (University of Siena)
This work addresses the challenge of estimating inequality measures in cross-sectional and longitudinal survey with imputed data, focusing on innovative methodologies to improve accuracy. A novel framework, Survey-to-Survey Imputation Techniques (SSIT) based on Generalized Additive Models for Location, Scale, and Shape (GAMLSS), is proposed to better account for the variability in survey data. Unlike traditional models, SSIT-GAMLSS have the possibility to use auxiliary information and random effects in each parameter of the distribution and not only on the location one.
Applications of this approach are demonstrated through case studies. In Morocco, imputed consumption expenditure data integrates demographic and socio-economic covariates, yielding improved measures of inequality over time. A second application focuses on income data from Italian regions, leveraging SSIT-GAMLSS to address issues such as missing continuous distributions and underreporting, with results validated against official data sources like EU-SILC. A third case examines poverty dynamics in Tuscany, creating synthetic panels to trace changes in inequality over multiple years.
The research highlights significant methodological advancements, including the ability to adapt to unstable parameter relationships in developing regions and mitigate back-transformation errors commonly encountered in imputation methods. Key findings suggest that SSIT-GAMLSS reduces biases in inequality measures, enabling more accurate assessments of socio-economic disparities.
This work has implications for National Statistical Offices (NSOs) and policy applications in countries facing data scarcity or comparability challenges. It concludes by proposing further extensions to analyze poverty dynamics across diverse contexts. The methodology aligns with the objectives of improving poverty and inequality measurement for policy formulation and evaluation.
Mr SOMNATH JANA (PHD RESEARCH SCHOLAR) - Presenting Author
Dr Laxmi Kant Dwivedi (PROFESSOR)
Complex non-hierarchical patterns are frequently seen in large-scale survey data, especially in interactions between interviewers and districts, which present significant analytical challenges. Cross-classified multilevel models provide a sophisticated way to deal with these issues, but putting them into practice necessitates carefully weighing a number of methodological factors. A comparative analysis using data from the National Family Health Survey (NFHS) 2019-21 in India, examines interviewer effects on survey outcomes, with home births as the dependent variable.
Two methodological approaches are compared: the standard hierarchical model and the cross-classified multilevel model. While the hierarchical model assumes rigorous nesting of respondents inside PSUs (Primary Sampling Units), districts, and interviewer levels, the cross-classified model takes into account the fact that interviewers frequently work in many PSUs and districts. The analysis includes multi-level sampling weights calculated using a two-stage stratified sampling design, with both rural and urban samples drawn from India's 707 districts.
The findings reveal significant differences in Intraclass Correlation Coefficients (ICC) between the two models. The hierarchical model shows higher ICC at the interviewer level, suggesting potential overestimation of interviewer-related variance. In contrast, the cross-classified model demonstrates reduced interviewer-level ICC, indicating more accurate variance estimation. The integration of sampling weights further enhances the precision of estimates, as evidenced by caterpillar and variance deviation plots.
Results demonstrate that cross-classified models, particularly when combined with appropriate sampling weights, provide more robust and unbiased estimates compared to traditional hierarchical models in complex survey designs, offering important implications for survey methodology and data analysis in social science research.