Different methods, same results? Comparing the consequences of alternative methods of data collection and analysis 3 |
|
| Chair | Professor Elmar Schlueter (Justus-Liebig-University Giessen ) |
| Coordinator 1 | Professor Jochen Mayerl (University of Kaiserslautern) |
Although there is a large body of research on the impact of partner loss due to divorce or separation on different health dimensions (e.g., self-rated, physical, and mental health), opportunities for causal inference are limited, mainly because of the paucity of studies based on multi-wave panel data. Using G-SOEP data and applying diverse types of panel regressions, we aim to explore how robust findings on the effects of union dissolution on health are with regard to model specification, choice of the dependent variable and its measurement scale, as well as choice of the link function.
In our analyses on 2,390 marital disruptions observed between 1984 and 2014, we focus primarily on fixed-effects (FE) panel regressions, contrasting them initially – for the sake of completeness – with random-effects (RE) panel regressions. In particular, we estimate FE panel regressions for individuals nested in households with nonlinear (dummy) impact functions over time, covering the time span of up to five years before to five years after the event of union dissolution. Considering effects on both individual average health level and changes in health across time, we take account of time-constant unobserved heterogeneity. The key outcome variables that we compare are: 1) satisfaction with health (rated on an 11-point scale and available yearly since 1984) and 2) self-rated health (rated on a 5-point scale and available yearly since 1992).
We run through different analytical scenarios, starting with simple FE models disregarding both the ordinal scale of measurement and the nested structure of the data. Next, we dichotomize self-rated health in two different ways – as is customary – and run FE analyses applying a binary logit link. Finally, in order to take advantage of the ordinal nature of the outcome variables we use both an ordered logit hybrid model proposed by Allison (2009) and the so called “blow-up and cluster” (BUC) estimator and compare the results to standard linear FE models. The problem of nested data is solved by using hybrid modeling.
The common pattern known from the literature on divorce is that health starts to decline several years prior to union dissolution (“anticipation effect”) and improves continuously afterwards almost to the baseline level. In order to capture this u-shaped trend, nonlinear impact functions need to be used. However, caution is needed with regard to the robustness of the FE findings. We show that the estimated temporal shape differs considerably, depending on how outcome categories are assessed and whether the nested structure of the data is considered. This points to a serious potential bias in part of the public health literature where dichotomization of ordinal variables is common practice.
We conclude that measurement and analytical strategies may impact results on effects of partnership-related events on health. The recommendation is to keep as much information in the outcome variable as possible and use analytical methods taking advantage of intra-individual change over time in order to minimize potential bias from unobserved heterogeneity.
Panel attrition due to nonresponse can lead to a substantial loss in data quality and is therefore studied extensively in survey research. In this context, it is often of specific interest whether non-respondents differ systematically from respondents since selective refusal can bias results which are based solely on survey participants. While a wide range of predictor variables can be utilized in order to investigate nonresponse patterns in longitudinal surveys, different methods can be used in order to relate these features to the outcome of interest.
Specifically, panel nonresponse can be modeled from two different perspectives: On the one hand, parametric regression methods with a logit or probit link function can be used to study panel dropouts. However, these methods require careful model specification in order to provide valid estimates. Initially, a linear and additive relationship between the predictor variables and the response is assumed, which may be relaxed e.g. through the introduction of polynomial and/or interaction terms -- given some prior knowledge on potential non-linear effects and interactions. On the other hand, data-driven classifiers can be applied which inherently allow for complex non-linear and non-additive effects. In the context of classification trees the predictor space is recursively partitioned into subspaces with respect to some loss function such that trees with highly complex interaction structures can result. On this basis, random forests grow a large number of trees using resampling and a restricted set of predictors in order to overcome the instability of a single tree. However, both methods primarily focus on prediction, which comes at the expense of interpretability especially in the context of random forests.
Against this background, this study compares logistic regression, classification trees and random forests for modeling unit nonresponse in the German Socio-Economic Panel Study (GSOEP). Specifically, the results of these methods are compared with respect to the implied effect structure and prediction accuracy. Concerning variable importance and effect direction, both modeling perspectives lead to quite similar results, while the usage of random forests introduces some improvement in terms of prediction. However, it is shown that tree-based methods offer additional insights concerning the effect structure of the included nonresponse predictors in comparison with standard (parametric) logit regression.
We address methodological challenges in estimating the causal effect of children on parents’ well-being. Particularly, we focus on fixed-effects and instrumental variable methods and compare advantages and disadvantages of both estimation strategies. For the IV approach we exploit variation in family size induced by nature as it has been argued that the birth of twins causes (conditionally!) exogenous variation in family size. This estimation strategy has several advantages: Firstly, and most importantly, reverse causality (happiness leads to births) will not bias the estimates, because it is unlikely that happier people experience twin births more often than less happy individuals. Secondly, if for some variables (e.g. income, sexual frequency and satisfaction and health/stress) it is unclear whether they confound or mediate the causal relation of interest, IV methods can easily provide a consistent estimate of the total causal effect. For estimating the total effect of family size in standard regression models, one would have to control for all confounders – but not for mediators. The true total effect is thus difficult to estimate. In a natural experiment, these potentially confounding variables need not be controlled. However, the twin-instrument has also disadvantages: Previous research has questioned its exogeneity, because twins could have effects on happiness not only via family size, but also via other child characteristics (e.g. twins are less healthy and cause more stress than two children born one after the other). If the direction of the causal relationship under investigation seems theoretically unambiguously clear or one is interested in mediation analysis, fixed effects methods are a powerful alternative requiring similar assumptions. We use the concept of Directed Acyclic Graphs (DAGs) to explicate the respective underlying assumptions for causal identification of direct and indirect effects and present some empirical findings based on 7 waves of pairfam.
A common finding in research on social stratification is that social background strongly influences various educational outcomes (Boudon, 1974; Coleman et al., 1966). Typically, inequalities of educational opportunities are analysed by investigating how strong social background variables determine educational outcomes net of relevant covariates (Breen et al., 2010; Cebolla-Boado et al., 2016). Metric educational outcomes like literacy scores are usually analysed using some kind of conditional mean model (e.g. Linear Regression) assuming homoscedastic error variance. Substantively, this means that the effect of social background on literacy is assumed to be constant across the entire educational outcome distribution. While a violation of this assumption (i.e. the presence of heteroscedastic error variance) does not bias the OLS point estimates, it might be indicative of potentially interesting relations between independent and dependent variables beyond the mean that a Linear Regression Model cannot capture. To date, there is – to the best of our knowledge – no study that investigates the research questions addressed in this paper: Does the effect of social background on educational outcomes differ across the distribution of literacy? If so, how large are the differences between the common OLS estimate and the estimates at different quantiles of the unconditional literacy distribution?
We address our research questions by analysing data from the most recent PISA wave, since it offers a sufficiently high number of students, countries, and variables. As for the analytical strategy, we proceed in two steps: First, OLS regressions are conducted country-wise to explain the literacy score. Secondly, unconditional quantile regression (UQR) models (Firpo et al., 2009) are conducted country-wise for the .10, .25, .50, .75, and .90 quantiles to assess the association of social background with literacy at different points of the unconditional competence distribution. We then test whether the coefficients differ significantly between different quantiles and from the OLS coefficients.
While the effect of social background on literacy is stronger at lower quantiles compared to both higher quantiles and the mean, the differences between OLS estimates and UQR estimates are only statistically significant at the .90 quantile. The social background coefficients from the OLS regressions are between 12 and 31 percent larger than the UQR coefficients at the .90 quantile. This pattern is found primarily in countries where the background dependency is relatively high. Hence, by only looking at the OLS estimate, we overestimate the effect for high achievers substantially and, at the same time, underestimate the associations for low achievers in these countries. We conclude with a discussion of possible mechanisms that might explain the patterns found.
This paper studies racism and ageism in health care in Germany, the Netherlands and Indonesia. One might expect that health professionals in particular have the responsibility to maintain health equality. Inequality leading to discrimination in health care can have major consequences like dying earlier. While negative racism and ageism is expected to occur in Germany and the Netherlands, only negative racism is expected to occur in Indonesia. There no negative ageism is expected yet, since the attitudes and stereotypes against the elderly are still positive.
In order to study perceived racism and ageism,patients were taking part in a patient satisfaction survey,. In order to study objective racism and ageism, physicians took part in a vignette study, with videos displaying patients (played by actors) describing their symptoms. Based on this, the physician had to categorize the disease, give recommendations, and estimate the character of the patient. Furthermore, attitudes toward the elderly and immigrants were measured, in order to test whether discrimination occurs due to attitudes and stereotypes.
As a third method economic experiments were used in order to test whether these methods could replicate the results of the vignette study.
Results showed that perceived racism and ageism as well as objective racism and ageism could be detected. However. economic experiments could only replicate strong effects.