Validation and Generalizability of Machine Learning Prediction Models on Attrition in Longitudinal Studies

Attrition in longitudinal studies is a major threat to the representativeness of the data and the generalizability of the findings. Typical approaches to address systematic nonresponse are either expensive and unsatisfactory (e.g., oversampling) or rely on the unrealistic assumption of data missing at random (e.g., multiple imputation). Thus, models that effectively predict who most likely drops out in subsequent occasions might offer the opportunity to take countermeasures (e.g., incentives). With the current study, we introduce a longitudinal model validation approach and examine whether attrition in two nationally representative longitudinal panel studies can be predicted accurately. We compare the performance of a basic logistic regression model to a more flexible, data-driven machine learning algorithm––Gradient Boosting Machines. Our results show almost no difference in accuracies for both modeling approaches, which contradicts claims of similar studies on survey attrition. Prediction models could not be generalized across surveys and were less accurate when tested at a later survey wave. We discuss the implications of these findings for survey retention, the use of complex machine learning algorithms, and give some recommendations to deal with study attrition.

In or out? Poverty dynamics among older individuals in the UK

Using the largest household panel surveyUnderstanding Society, this paper investigates low-income dynamics among pensioner households in the UK controlling for biases due to initial conditions and non-random survey attrition. Estimation results indicate there is a correlation between initial and conditional poverty status, specifically, there is regression towards the mean. The results find no evidence of a correlation between initial poverty status, conditional poverty status and survey attrition. The findings show the importance of benefit income in determining poverty status, suggesting that a dichotomous measure such as poverty status may not suitably reflect actual pensioner living standards. Aside from benefit income, receipt of employer and occupational pension, health, education and subjective financial situation are important in determining initial and conditional poverty status. Stylised examples highlight the significant differences in the ‘poverty experience’ which arise due to differences individual and household characteristics.

Estimación indirecta de tasas de ingreso y de retiro de la actividad económica para México / Indirect Estimate of Income Rates and Retirement from Economic Activity for Mexico

Con base en la información recolectada en la Encuesta Nacional de Ocupación y Empleo de México se han calculado las tasas de ingreso y retiro de la actividad para las 32 entidades federativas durante el periodo 2000-2007. Mediante regresiones lineales ordinarias de mínimos cuadrados se obtienen ecuaciones para predecir las tasas de ingreso y de retiro a partir de las tasas de participación en la actividad por edad, dado que es la información más fácil de conseguir en censos de población o encuestas de hogares por muestreo. Los resultados concuerdan satisfactoriamente con las cifras comparables disponibles. El algoritmo se ejemplifica con la ciudad de Zamora en 2000.AbstractOn the basis of information obtained from Mexico’s National Occupation and Employment Survey, attrition to and separation from activity rates have been calculated for the 32 states during 2000-2007 period. Ordinary least squares linear regressions are used to obtain equations to predict attrition and separation rates on the basis of participation rates in activity by age, given that this is the easiest type of information to obtain in population censuses or household sampling surveys. The results tally satisfactorily with comparable available figures. The algorithm is exemplified by the city of Zamora in 2000.