Open Conference Systems, 50th Scientific meeting of the Italian Statistical Society

Font Size: 
Looking beyond averages: quantile regression approach to model olderâ€adult Europeans’ quality of life
Elisa Cisotto, Giulia Cavrini

Last modified: 2018-05-17


Starting from crossâ€sectional European data coming from the Survey on Health,Ageing and Retirement (SHARE), the present study aims to investigate whether andhow several olderâ€adult (50+) individuals’ characteristics are associated with reportedQuality of Life (QoL). In particular, we analyze the variation in the associationsbetween QoL and olderâ€adults’ characteristics taking into consideration the full QoLdistribution. In doing so, we propose the use of quantile regression as alternativemethod of analysis compared to standard Ordinary Least Squares linear regressionmodels (OLS).To examine individuals’ quality of life we use the SHARE CASPâ€12 scale, ameasurement based on 12 Likert scaled items crossing four theoretically deriveddimensions of QoL that are particularly important in early old age: (C)ontrol,(A)utonomy, (S)elfâ€realization and (P)leasure. Each Likert scale question (three foreach dimension) was recoded so that the most positive response scores 4 and themost negative 1. Further, some of the items were reverse coded so that all responsesare in the same direction. Following a nonâ€hierarchical approach, the CASPâ€12 finalscore is the arithmetic sum of the scores of each item, thus ranking from 12 to 48.High values indicates higher QoL. In our models, the 10 percent quantile estimatescorrespond to the lowest life quality 10 percent of the sample (conditional on theexplanatory variables), the 25 percent quantile to the life quality of the lower 25percent of the sample, and so on. Hence, analysis investigate possible heterogeneousassociations between dependent and independent variables at different segments ofthe conditional QoL distribution. The underling hypothesis is that the common leftskewed distribution of QoL might be better analyzed through a semiâ€parametricmodel that focuses on the entire outcome distribution, while meanâ€based modelscould hide potential specific associations.We fit multiple quantile regression models with CASPâ€12 score as dependentvariable. Overall models for the entire sample are run, followed by separatedestimation by gender and age specific groups. Results are compared with OLSestimates. Variables related to health, functioning, social relations and materialcircumstances are considered as predictor variable. Quantile regression are appliedas crossâ€sectional estimators. Consequently, results cannot be interpreted as causaleffects in any strict sense.4 Rimanere su un max di 4.000 caratteri (dueâ€tre pagine, bibliografia compresa)Statistical analysis testing linearity and normality assumptions about OLS estimatesexplain that, generally, meanâ€average coefficients are reliable and give a goodapproximation of the overall associations. However, quantile regression estimatesdemonstrate significant variation over quantiles and significant differences emergewith regard to material circumstances and social relationships. In many cases,reported associations lose their statistical significance, or come closer to zero, forindividuals reporting high QoL.The associations between QoL and perceived income adequacy, working conditions,as well as education level lose magnitude over quantiles, or even their statisticalsignificance. Moreover, even if for some variables the estimates across quantilesremain statistically significant, they do often change their magnitude and thedifferences between coefficients are also statistically significant. As an example, whileunemployment is associated with nearly 2.5 points of decrease in QoL for the 10thquantile, this negative association do not exceed value 1 for the 90th quantile. Thedeclining relationship can be clearly seen in Figure 1, where coefficients for the wholeQoL distribution are plotted: the greyâ€shaded area depicts the 95 percent confidencebands of the estimated quantile regression coefficients, while the horizontal linedepicts the OLS coefficients estimates and dotted black lines OLS 95 percentconfidence intervals. From the graphs one can observe if the quantile regressionestimates lie outside the confidence intervals of the OLS regression, this shows thatassociations with this variable are not constant across the conditional distribution ofthe independent variable. Similar decreasing coefficients estimates over theconditional QoL distribution emerge for variables related to family structures, countryof residence, social participation, home ownership and neighborhood quality.Moreover, in all models the pseudo Râ€squared estimates, whose interpretation issimilar to the Râ€squared in regular regressions (Hao and Naiman 2007), decrease withincreasing quantiles. In order to test the robustness of results we repeated theanalysis on subsamples disaggregated by gender and age groups. In most cases,similar results emerge.

Full Text: PDF