Volunteering and Social Inclusion pp 169-201 | Cite as

# Methodological approach: cross-national and longitudinal

## Abstract

In the previous chapter I recalled the hypotheses on the interrelation between unemployment and volunteering which I test in the following empirical chapter. Namely, I am first interested in the influence of job loss on changes in people’s volunteering behaviour and second in the influence of volunteering for the reemployment chances of the unemployed. A special focus is given to long-term unemployed women who are suspected to use volunteering as an alternative activity, especially in West Germany where the institutional context supports women’s withdrawal from the labour market. In the present chapter, I argue how these research questions can be approached methodologically. Generally, my methodological approach can be characterised as being both cross-national and longitudinal. As to the cross-national aspect of my study, I have outlined my approach in the introductory chapter. In the following chapter, I proceed to the discussion of the data, methodological problems related to the sample, and the quantitative methods which I use for testing my hypotheses.

## Keywords

Unobserved Heterogeneity Labour Market Status Random Effect European Community Household Panel Proxy Respondent## Preview

Unable to display preview. Download preview PDF.

## References

- 136.Unobserved heterogeneity is a typical concern in panel data analysis. It describes the heterogeneity across individuals which arises as a result of individual characteristics which are commonly unobservable or simply not measurable, such as preferences or personality characteristics. Failure to account for these unobserved factors may result in biased and inconsistent estimates of the parameters of interest (Kyriazidou 1997). With panel data studies, one has the possibility to differentiate between person-specific and idiosyncratic measurement errors and control for the former one (for a more detailed discussion, see the section on multivariate longitudinal data analysis at the end of this chapter).Google Scholar
- 137.Additionally, in each year the basic information in one of these areas is enlarged by detailed questions. For volunteering, this has been the case in the years 1990, 1995, 1998 and 2003. For further details, see the section on the operationalisation of the variables.Google Scholar
- 138.The head of the household in the GSOEP is defined as the person who knows best about the general conditions under which the household acts and is supposed to answer this questionnaire in each given year. This reduces the risk of longitudinal inconsistencies (Haisken-DeNew and Frick 2005).Google Scholar
- 139.One should note that there is a need to differentiate between first time respondents and those with a repeated interview, since some information does not have to be asked every year, unless a change occurs. Those variables have to be imputed to the following waves (Haisken-DeNew and Frick 2005).Google Scholar
- 140.The ADM (
*Arbeitsgemeinschaft Deutscher Marktforschungsinstitute, Working Group of the German Marketing Research Institutes*) master tape from 1982 served as a basis for collecting sample A. For further details on the sampling procedures of each sub-sample, see Haisken-DeNew and Frick (2005).Google Scholar - 142.These variables are the proportion of heads of households in socio-economic groups 1 to 5 and 13 (that is, in professional or managerial positions), the proportion of the population of pensionable age (i.e. females over 60 and males over 65) and (in PSUs in non-metropolitan areas) the proportion of the employed PSU population working in agriculture or (in PSUs in metropolitan areas) the proportion of the PSU population that was both under pensionable age and living in single person households (Taylor, et al. 2006: A4-1f.).Google Scholar
- 143.The strategy of paid-wise deletion is not advisable for data which are only MAR. One of the problems with this method is that the estimated standard errors and test statistics produced by conventional software are biased (Allison 2002).Google Scholar
- 144.The imputation of item-non-response related to missing income data in the GSOEP follows a two step procedure: The general principle is to apply the so-called row and column imputation technique suggested by Little and Su (1989) whenever longitudinal data is available, and to run purely cross-sectional imputation techniques otherwise (Grabka and Frick 2004: 4). Such supplemental crosssectional imputation methods can be: logical imputation (e.g. the receipt of child benefit is assumed for households with children); median substitution (only for income components with less than 10 affected cases) or median substitution for subgroups; median share substitution (e.g. the Christmas bonus in the private sector as a percentage of 35% of the monthly labour earnings) and finally regression-based substitution (used mostly for more complex income constructs like “interests and dividends” or “individual labour income from first job”) (Grabka and Frick 2004: 5f.).Google Scholar
- 145.In the BHPS, two main imputation techniques are used: Random-within-cell hot-deck imputation and regression imputation (predictive mean matching). The first method was used for certain categorical money variables, such as Proxy’s personal income, and a number of cases where regression methods appeared inappropriate (e.g. income from welfare benefits). The second method was used for money amount variables, such as individual labour earnings and household income. Having imputed a number of primary variables, a number of other income related variables were computed from these variables, with some additional small scale random within-cell hot-deck imputation (Taylor, et al. 2006: A5-21). In a panel study, there may also be variables from the same respondent collected at a different wave. In the BHPS, a three-wave imputation strategy was adopted: The model strategy was used either as forward or backward imputation (Taylor, et al. 2006: A5-22).Google Scholar
- 146.One way of dealing with the problem of sample attrition is the use of sample weights. However, in the case of panel data, conventional sample weights are not appropriate. Instead, one needs longitudinal attrition-adjusted sampling weights. One important critique of those weights is that they implicitly assume independent censoring, i.e. that individuals who drop-out behave the same way after their last interview as they did before (for a description of the construction of longitudinal weights in the GSOEP, see Haisken-DeNew and Frick 2005: 169ff.). To the contrary, since it is quite likely that the drop-out is related to a state change, such as a divorce or unemployment, it can be expected that the person’s behaviour will be changed. For this reason, longitudinal weights are not considered appropriate for dealing with the problem of panel attrition.Google Scholar
- 147.In the GSOEP, temporary drop-outs or persons and households which could not be successfully interviewed in a given year are followed until there are two consecutive temporary drop-outs of all household members or a final refusal. In the case of a successful interview after a drop-out, there is also a small questionnaire including questions on central information which is missing for the dropout year (e.g. employment status) (Haisken-DeNew and Frick 2005: 22).Google Scholar
- 148.For the BHPS, much effort is spent on tracking individuals from wave to wave. At each wave, the Centre and National Opinion Poll (the organisation which carries out the fieldwork) undertake a thorough refusal conversion process to attempt to minimise attrition due to refusal and other forms of non-response. This process covers both previous wave refusals, and also new refusals encountered in the current wave. This refusal-conversion prooved to be very successful: 66% of initial refusal were finally converted (Taylor, et al. 2006: A4-12). Another important element of the panel maintenance is to keep contact with the interviewees between the waves (Taylor, et al. 2006: A4-13). As already mentioned, as a last resort interviews are done by proxy or by telephone.Google Scholar
- 149.More recent figures on panel attrition in the GSOEP are provided by Pannenberg (2000).Google Scholar
- 150.A sample is said to be censored if we have not observed the values of the x variables for all sample observations. So, in the simple case of censoring from below, we would have observed the x values for those cases where y > c (and y is therefore observed exactly), and where y is only known to be either equal to or less than c (Breen 1996).Google Scholar
- 151.If we observe the values of the x variables only for those observations where y is recorded exactly, the entire sample is said to be truncated (Breen 1996).Google Scholar
- 158.The CASMIN scheme is constructed as a certificate-oriented classification that distinguishes educational credentials according to two dimensions: (1) The hierarchical level (length, quality and value of education), and (2) the tracking into different educational pathways, distinguishing general and vocationally oriented education (Brauns, et al. 2003: 222f.).Google Scholar
- 164.In waves 1984 until 1986 and 1988, 1990, 1992, 1994 until 1999, 2001 and 2003 respondents were asked questions related to their social life. The question reads (more or less the same in the different waves): “Which of the following activities do you do in your free time? Please tick how often you practice each activity.” The answer categories vary tremendously over the years.Google Scholar
- 186.As already mentioned, I excluded the first two waves in order to be able to include a person’s work experience in the past two years without confronting the problem of left-censoring. I also excluded the last wave (2005) because I do not have the information on whether people find a job in the following waves.Google Scholar
- 188.Another commonly used methodology for analysing longitudinal data is event history analysis or survival analysis. This method is concerned with analyzing the time to the occurrence of an event, such as re-employment over time, i.e. taking into account the probability that an individual has “survived” in a certain state up to a certain time (Allison 1985; Jenkins 2004). However, I decided against using event history models to test my hypotheses. The main reason for this is related to the structure of the available data: While GSOEP and BHPS contain monthly information on a person’s employment status, data on his or her volunteering behaviour are available only annually or biannually. In order to have a close time-link between the two phenomena, it seems advisable to use the annual information on employment status as well as the (bi-) annual information on volunteering, both at the time of interview.Google Scholar
- 189.Dynamic panel models are characterised by a lagged dependent variables on the right-hand side of the equation. These models, as they are discussed for example by Finkel (1995: 1) stress the possibility to use the time dimension of panel data in order to draw causal inferences. In this understanding, for a causal effect to exist from variable X to variable Y, X and Y must covary and X must precede Y in time, and the relationship must not be “spurious” or produced by the common association with a third variable or set of variables.Google Scholar
- 190.There are several problems associated with the estimation of (lagged) panel models: (1) Reciprocal causality which leads the parameter estimates to be biased and inconsistent. (2) Measurement error: Including Y
_{(t−1)}directly in the model introduces bias because of the problems associated with statistical estimation in the presence of error-laden independent variables. (Y_{i,t−1}is necessarily correlated with the error term. The bias often leads to the underestimation of the true effect of Y_{(t−1)}on Y_{(t)}. The most important problem is however the following one: (3) Omitted variables in panel models may lead to autocorrelation in the endogenous variable’s error term over time (Finkel 1995: 21f.).Google Scholar - 191.As pointed out before, unobserved heterogeneity describes the heterogeneity across individuals which arises from unobserved individual characteristics and leads to biased and inconsistent estimates of the parameters of interest (Kyriazidou 1997).Google Scholar