5.4 Towards a relationship
5.4.1 Analysis
The data collected for the analysis are organized as time series cross-sectional data (TSCS) and follow a hierarchical form where countries represent time invariant en- tities (level 2) and year (1971-2009) time variant observations (level 1). Taking under consideration the nature of the data an assessment is necessary between the most common methods of analysis, random and fixed effects. Random effects mod- eling is able to estimate between and within effects as well as time invariant (level2) and time variant (level 1) effects of variables. However, random effects suffer from violations of omitted variables bias and the exogeneity assumption (Palta and Se- plaki, 2002). In the first case, bias exists when the within and between effects are different and the between effect is ignored. In this situation the variance, caused by the omitted between (level 2) country effect, appears in the residuals and eventually
Chapter 5. The association - deindustrialization, recession and mortality 101 is correlated with the independent variable (covariate). That leads to a violation of the exogeneity assumption where residuals are assumed to be independent of the covariates (Shin and Raudenbush, 2010; Neuhaus and Kalbfleisch, 1998).
In order to tackle the above violations, fixed effect modeling is commonly used in various fields but also in studies examining the associations between unemploy- ment and mortality during recessionary periods (Granados, 2005; Gerdtham and Ruhm, 2006; Stuckler et al., 2009a; Svensson, 2007). Fixed effects modeling re- strain this bias since it controls for between effects and level-2 variance. However, fixed effects estimate only within effects and level-1 variance. This is achieved by the inclusion of dummy variables that comprise the level 2 variance (in this case country). Nevertheless, there are certain limitations. This type of modeling is more conservative since it estimates only the coefficients of time variant variables. Con- sequently, between effects (time invariant variables) could not be estimated and this leads to the loss of degrees of freedom of level-2 variables (Snijders and Bosker, 2012).
Following previous research and taking into account the above advantages and lim- itations, the process of the analysis adopted a fixed effects modeling and progressed in two stages. In this case, fixed effects analysis is used to remove confounding of between-country variances. In particular, this method takes into account country level differences. The next table (Table 5.3) illustrates the variables and their de- scription included in the model. The first step of the analysis assesses the overall long-run relationship of the economic indicators on all-cause and suicide mortality. In this case, a long-run relationship is defined as whether a rise in the independent indicators is associated with a rise or fall in mortality. Long-run associations are measured in levels, which in this case are the annual rates of the economic indica- tors included in the model together with the dependent variable describing mortal- ity. This occurs gradually by controlling for the independent variables together with additional general year effects, country-specific time trends and country effects. Nevertheless, the analysis in levels does not offer the possibility to assess short-run associations. More precisely, whether changes of economic and industrial progres- sions (overall economic or industrial decline) can be associated with changes (rises or falls) in mortality (Stuckler et al., 2009a). Therefore, the second stage of the analysis is based on annual changes of the dependent variables and economic in- dicators measuring recession and industrial decline. Results are also stratified by
Chapter 5. The association - deindustrialization, recession and mortality 102 gender in order to identify variations between males and females.
In the same context, further analysis explores the continuous (up to three year lags) of the long and short-term (levels and changes) effects of the economic indicators on mortality fluctuations. These fixed effects models include the lags of the economic covariates and take into account year trends and country-specific time trends. Fi- nally, more intense rises of unemployment and contractions in manufacturing (more than 1%) are also investigated. In this case the analysis includes only those years, when unemployment increases more than 1%, and the reduction in manufacturing employment and GDP per capita is more than 1%.
Before any analysis occurs the dataset has been identified as time-series cross- sectional, where country is the time invariant variable (Level-2) and year the time variant (Level-1) variable. However, the assumption of the error term, suggesting independence and identical distribution, is commonly violated in the TSCS. There- fore, the use of various methods of the Huber/White/sandwich estimator relaxes these prerequisites. The implementation of the bootstrapping method deals with the assumption of independence within clusters. The main assumption of this method is that the distribution of the data represents the underlying population (Guan, 2003). The choice of bootstrapping is based on the number of clusters. It has been in- dicated that for clusters less than 50, in this case there are 15, the calculation of bootstrapped standard errors are appropriate (Cameron et al., 2008). The models are estimated using the Stata version 10 software.
Table 5.3Description of variables
Variables Description
All - Cause Mortality Age standardized rates (20-59 age group) Direct method/European Standard Population Suicides Mortality Age standardized rates (20-59 age group)
Direct method/European Standard Population Country A string variable includes15 countries Year An integer variable 1971 - 2009 GDP per capita Level of GDP per head in US dollars Civilian Unemployment Annual values %
Chapter 5. The association - deindustrialization, recession and mortality 103 Modeling
The main aim is to differentiate between long and short-term relationships of eco- nomic variations on all-cause and suicide mortality after controlling for country and time effects as well as country-specific time trends. The year effects control for determinants that vary across countries and over time, whereas country effects account for factors that differ across countries and they are time-invariant. The im- pact of recession and industrial decline on all-cause and cause-specific mortality is characterized by within country differences occurring in independent variables in relation to the variations taking place in other countries. The estimates of these vari- ations account for time constant differences between countries together with factors that change over time and are distributed across countries.
The step by step modeling is described as follows:
Model 1 It is the most basic approach in levels. The first model introduces the first independent variable employment in manufacturing percentage. The model begins with the indicator describing industrial decline, since the main interest is to explore the overall association of industrial contraction and mortality. In the subsequent models the analysis takes into account periods of recession in order to examine any potential changes in the relationship between industrial decline and mortality. The model controls for country fixed effects.
Model 2 The second model introduces the overall unemployment percentage in- cluding the employment in manufacturing.
Model 3 The third model continues to control for employment in manufacturing and unemployment and introduces the next independent variable (GDP per capita). Due to skewness of the variable, transformation is necessary, therefore the log-GDP is inserted in the model. Log-GDP is included only in the model in levels.
Model 4 Model four includes a general time trend (variable year 1971-2009). It assumes that time varying indicators, for example public health, improve linearly over time at a national level (Ruhm, 2000, 2007).
Model 5The final model controls for country-specific time trends and country fixed effects. The country-specific time trends take control for time-varying determinants within states such as demographic characteristics (Ruhm, 2000, 2007).
Chapter 5. The association - deindustrialization, recession and mortality 104