MOLS 2006: Methodology of Longitudinal Surveys

MOLS 2006: Methodology of Longitudinal Surveys

Andreas Hadjar and Frank Schubert

(Department of Sociology of Education, University of Berne)

Education and Subjective Well-Being in Temporal Change,
1984-2002: a Comparison of two Analytical Methods

This analysis of the change in subjective well-being attempts two aims: a methodological and a theoretical one.On the on hand, the paper will provide an exploration of a methodological problem: How useful are conventional regression analysis and Hierarchic-Linear Modelling (HLM) for the analysis of temporal change. The comparison of these analytical methods focuses on data structure, applicability and parameter estimations. On the other hand, the change of subjective well-being (satisfaction with life, satisfaction with health) over the time period from 1984 to 2002 will be considered.

To analyse social change adequately and to prevent ‘ecological fallacy’ (i.e.to draw wrong micro levelconclusionsfrom macro level findings), social change will be understood as a composition of age, period and cohort effects and therefore analysed by using an A-P-C design — meaning a simultaneous modelling of age, period and cohort effects in one regression model. A problem of A-P-C-Analyses is the confounding or multicollinearity problem that occurs when age, period and cohort variables are introduced simultaneously in a regression model, as Period = Age + Cohort. To prevent confounding, the common strategy to substitute one of the temporal variables is deployed here. Unemployment rate will be included as surrogate for the period variable (time of data gathering). Education and educational expansion are considered as social mechanisms behind these processes of change and therefore introduced into the analyses. Beside the educational level, the interaction between cohort (year of birth) and educational level will be included into the model. This interaction effect shows how the gap between less educated people and more highly-educated people developed over time (cohort succession).

Data base is the German Socio-economic Panel data set (SOEP), although only the West German subsample is used for analysis. The sample comprises of data of over 20.000 people over a time period of 18 years (1984-2002, yearly measurement).SOEP is a representative longitudinal study of German private households.The German Socio-economic Panel 1984-2002 data set provides hierarchical data (time level, person level, household level).

First, graphs will be presented that provide a visual impression of the changes over time. Then, OLS regression models will be estimated; explanatory variables including educational level, age, cohort and period (unemployment rate) and the education-cohort-interaction will be introduced hierarchically. In a last step, Hierarchic-linear Models will be estimated (level 1: time, level 2: person, level 3: household).

Results of conventional regression and HLM lead to similar conclusions: relative stability of satisfaction with life, a strong age effect on satisfaction with health, effects of education on both aspects.The life satisfaction and health satisfaction gaps between higher-educated people and less educated people decrease over time (cohort succession). Results of conventional A-P-C-regression and hierarchic-linear modelling only differ a little regarding direction, strength and significance of the effects. As expected, standard errors of the effects in the HLM analyses are slightly higher. HLM turns out to be the more rigid analytical method. However, HLM cannot totally substitute conventional regression models, as it needs a hierarchic data structure and panel data.

Key Words:

social change, education, subjective well-being, Age-Period-Cohort-analysis, analytical methods, multi level analysis

Dr. Andreas Hadjar
University of Bern
Department of Sociology of Education
Muesmattstrasse 27
CH-3012 Bern
Tel.: +41 (0)31-631 53 56
Fax: +41 (0)31-631 53 52
Email: / Frank Schubert
University of Bern
Department of Sociology of Education
Muesmattstrasse 27
CH-3012 Bern
Tel.: +41 (0)31-631 53 59)
Fax: +41 (0)31-631 53 52
Email:

We will deal with the change in subjective well-being in West Germany and the role of education and the educational expansion. In the last years, multi-level approaches became more and more important for the analysis of social change. Therefore our research interests include – beside temporal and educational differences in subjective well-being – a methodological issue: We will compare conventional OLS regression and Hierarchical Linear Modelling as two methods of data analysis.

What the essence of subjective well-being is, and how it can be produced, is a key question of the social sciences. Subjective well-being is an emotional condition. It is expressed in the subjective feeling that needs are satisfied and goals reached. In the present analyses, we will focus on two aspects of subjective well-being: satisfaction with life and satisfaction with health. Satisfaction is a cognitive component of subjective well-being that is based on a consideration of past, present and future conditions.

To begin with, I would like to outline some hypotheses as basis of our analyses. In exploring subjective well-being in dependence on different temporal variables, it will be focused on education and educational expansion as social mechanisms behind the change in subjective well-being. Looking at the relationship between education and subjective well-being on the individual level, according to Bradburn or Oswald, it may be assumed, that a higher educational level leads to an increase in individual skillsto reach goals, to satisfy needs and to live a healthier lifestyle. And therefore, a higher satisfaction with life and health may be expected among more highly educated people.

An increase in both dimensions of subjective well-being – satisfaction with life and health – in the course of educational expansion seems to be plausible due to the process of cognitive mobilisation of the people, as theorised by Inglehart or Baumert.

There is a second – expected but unwanted – consequence of educational expansion that may not be neglected: the heterogenisation of higher education. As the higher schools were opened to all social classes during the educational reform – although in Germany and Switzerland social origin and educational level remained correlated – the high standards of the higher educational institutions could not be maintained due to the new heterogeneity of pupils.

This heterogenisation led to a decreasing distinction between the more highly educated people and the less educated people regarding their skills, values, lifestyle and subjective well-being, respectively life and health satisfaction. However, heterogenisation was not strong enough to be able to reverse the cognitive mobilisation process.

To analyse social change adequately and to prevent– as we call it – ‘temporal fallacy’ (that is the drawing of wrong conclusions from findings on only one temporal level), social change will be understood as a composition of age, period and cohort effects and therefore analysed by using an A-P-C design — meaning a simultaneous modelling of age, period and cohort effects.

A problem of A-P-C-Analyses is the confounding or multicollinearity problem that occurs when age, period and cohort variables are introduced simultaneously in a regression model, as Period = Age + Cohort. To prevent confounding, the common strategy to substitute one of the temporal variables is deployed here. This means to “include measures of the causal variables for which age, period and cohort are surrogates“ (Tuma/Hannan 1984: 192). We will substitute the period variable by the West German rate of unemployment.

Considering age or – as we interpret it – the position within the life cycle, we assume a clear negative age effect on satisfaction with health, since health problems occur in older ages and the aspiration level does not decrease as much as state of health does.We do not expect a statistical effect of age on life satisfaction, because according to Diener life satisfaction does not change much with age, and other empirical evidence suggests that both younger and older people are happy – the younger due to their more exciting lifestyles and the older due to their fulfilled needs and goals.

Regarding cohort differences, we assume an increasing satisfaction with life and satisfaction with health among younger cohorts, since younger cohorts benefited from educational expansion and cognitive mobilisation.

As already mentioned, the period variable is substituted by the unemployment rate of West Germany, since this strategy avoids multicollinearity. There is much theoretical backing and empirical evidence that unemployment both on the societal level and on the individual level influences subjective well-being negatively. To prevent ecological fallacy, the individual unemployment at the time of survey is also included into the models.

Analyses are based on a panel data-set, including data of all yearlysurveys of the German Socio Economic Panel from 1984 to 2002. GSOEPis a representative longitudinal study of German private households – approximately 20’000 people – and provides hierarchical data. In addition to time level and person level, we also include the household level into our HLM analyses to control for sampling routine.

The sample was reduced to control for socialisation experiences, to prevent selection effects, and to include mainly people with a relatively fixed educational level: So the sample for the analyses of subjective well-being consists of people who are West Germans of German citizenship, older than 20 and born between 1919 and 1963.

Next, let’s move on to the measures. Both variables of subjective well-being – Satisfaction with life and satisfaction with health – were measured on a 11-point-rating scale: People were asked if they are satisfied with their life in general and how satisfied they are with their health.

The education variable was created from both the highest general and the highest vocational educational level. Following the CASMIN educational classification by Müller we classified the people into the following categories: 1 stands for inadequately completed or elementary education (CASMIN 1a, 1b), 2 stands for an intermediate vocational or general qualification (CASMIN 1c, 2a, 2b), 3 for a full general or vocational maturity certificate (CASMIN 2c) and 4 for tertiary education at the university or a higher vocational school (CASMIN 3a, 3b).

The unemployment rate in percent is introduced as period variable – and controlled by the binary variable “individual unemployment at the time of survey”.

The birth cohort will be included in various ways into the following analyses. For a better interpretation of different socialisation periods, cohorts will be bundled up into six dummy variables, each consisting of ten birth cohorts that represent particular times.

The table shows the birth cohorts and main events at the time of socialisation around the age of 15, also regarding educational expansion.

The birth cohorts from 1949 onwards benefited from the educational expansion and therefore, according to our hypotheses, may have a higher level in subjective well-being.

To calculate a trend regarding the subjective well-being over all birth cohorts, the year of birth will be included in some models.

First conclusions regarding our hypotheses and the change of subjective well-being may be drawn from visual inspections. These figures show the cohort-specific change in satisfaction with life and satisfaction with health over the time period between 1984 and 2002.

While regarding satisfaction with life, only marginal – over time slightly increasing – cohort differences and almost no periodical change is to be stated, the figure that shows cohort differences in satisfaction with health over time reveals clearer effects. Younger cohorts are at all times more satisfied with their health than older cohorts are. All cohorts share the trend of a gradual decline in satisfaction with health between 1984 and 2002, which seems to be rather an age effect than a period effect.

Let’s have a look at the next figure that shows educational differences in the change of subjective well-being. The education gap in satisfaction with health is much larger than the educational differences in life satisfaction. While there is not much difference in health satisfaction between people with tertiary educational certificates and A-level holders, intermediate and lower educated people seem to be much less satisfied with their health. Regarding satisfaction with health, there are indicators for a slight period effect, as life satisfaction has been slowly decreasing in all cohorts during the mid and late-1980s, and has been soaring again towards the early-1990s. This fits with the actual increase in unemployment during the late-1980s and the relaxation on the West German labour market after 1990. Health satisfaction is declining in all cohorts between 1984 and the new Century. Whether this is a period effect or an age effect will solved in the multivariate analyses.

Turning to the last figures, a comparison between the four educational levels or educational groups regarding their cohort-specific subjective well-being reveals that over cohort succession the very low educated people keep there negative distinction in life satisfaction, while there is a tendency of convergence between the other educational levels. This fits with the notion of heterogenisation of higher education. Regarding satisfaction with health, there are several trends: The group of the people with a very low educational level did become even more distinct considering the 1949-1958 cohorts to slightly converge again with the other educational groups in the 1959-1963 cohorts. The tertiary educational level seems to have lost its distinction over cohort succession.

Again, whether or not the effects that are suggested in these figures are genuine or based on other effects – especially if there is an actual cohort effect regarding satisfaction with health or of this goes back to an age effect – may be only solved by calculating complex A-P-C models and simultaneous testing of all three temporal effects.

Turning now to the results, we first present OLS Regression Models and then multilevel results.

As one can see in the first column, Satisfaction with Life strongly depends, as hypothesised, on the educational level. More highly educatedpeople are significantly more satisfied with life as a whole. The same effect is to been seen with satisfaction with health, which is also increasing with advanced educational certificates.

In the theoretical part we hypothesised, that in the cohort succession satisfaction with both life and health increases. Satisfaction with health follows this pattern – later born people are more satisfied than earlier born ones. Not as clear are the cohort effects regarding satisfaction with life – they are small and only partly significant, although there is a tendency that later born are more satisfied with life.

In Model II interactioneffects for education and cohort are introduced. Clearly higher education – a-level certificate and above – is, over the cohort succession, a less strong predictor of greater satisfaction with life. In other words: more highly educated people, if later born, are less more satisfied with life than earlier born ones. This effect confirms, what has been found elsewhere too: an effect of the educational expansion is a decreasing distinctiveness of more highly educatedpeople. We observe therefore over the cohort succession more heterogeneity among more highly educated.

In Model III finally age and period are included. As expected reported life satisfaction decreases in years of high unemployment. Age has a small but significant effect – older people tend to be less satisfied with life as a whole. Finally, if we look at the very low R square it is obvious, that satisfaction with live does not change much over time or over the life cycle.

Now turning to Satisfaction with health, again an interactioneffect between cohort and education is included in Model II. Results are meager – apart from the slightly more satisfied intermediate educatedpeople – one does not find an increasing heterogeneity.

When including age in Model III, the cohort effect decreases strongly – as expected younger people are more satisfied with their health than older people. Also, the general unemployment rate is a predictor for less reported satisfaction with health.

We now turn to the Multi-Level Analysis. Hierarchical Linear Modellinghas an important advantage to OLS regressions: The dependence of observations is implied by the method –in contrast to the assumption of independent measurements in OLS regression. This is especially true for longitudinal data, but also when stratified sampling routineswere used. Both applies to the SOEP data used in our analyses. Also, when micro variables as well as disaggregated macro variables are analysed simultaneously – as in conventional regressions – macro error terms interfere with micro error terms and therefore hurt the independence assumption of observations. In both cases significance tests are therefore more precise with HLM than in OLS regressions.

Let’s now compare the results.

Higher education has again a strong significant effect on life and health satisfaction. More highly educated people report higher satisfaction scores.

Satisfaction with life is for all cohorts higher than for the oldest cohorts born 1909-1918, but differs less between later cohorts. Overall, it’s the same trend as in the visual inspection and the OLS regression models, although it is more distinct here. Our ad hoc interpretation is that for the oldest cohort satisfaction with health as well as the general well-being is lowest and therefore affects also life-satisfaction.