5.4. Analytical approaches for SHARE and ELSA data

Several models were tested to study the impact of health status, risk factors, and health behaviours on employment likelihood.

The first model was a pooled logistic regression, where the impact of lagged health outcomes from waves 1 and 5 on employment outcomes in waves 2 and 6 were assessed. A variable specifying the ‘era’ from which the data originated was included in the model.

Tests for effect modification for each exposure and control variable were performed, as well as for the impact of ‘era’. Only the impact for the fifth and sixth waves was performed given:

  • The effect previous health status (with the exception of hypertension and stroke) has on employment likelihood was not affected by era

  • Significant pension and retirement reforms occureed between the first/second and fifth/sixth waves.

The second model was a Poisson regression (specifying the Incidence Rate Ratio (IRR) options, which approximate Relative Risks), with country level fixed effects estimating the lagged health outcomes of wave 5 on employment variables in wave 6.

This model is mathematically depicted below, and was also implemented as the main model for the employment likelihood module. The modified Poisson regression has the advantage of approximating relative risks, a more interpretable measure than the odds ratio of a logistic regression; by specifying robust standard errors, any potential variance overestimation is addressed. For these reasons, it was the model of choice to study the impact of lagged health effects on employment likelihood.

\[ \begin{align}\begin{aligned}{ln[Y]}_{i.t} = {\beta}_{0} \times {x}_{1,t} + ({\beta}_{2} \times {\vartheta}_{1,t-1} + ... + {\beta}_{n+2} \times {\vartheta}_{n,t-1}) + {\beta}_{n+3} \times {age}_{t}\\\times {sex}_{t} + ({\beta}_{n+4} \times {x}_{2} + ... + {\beta}_{n+4+ix_{ni}}) + c + \epsilon\end{aligned}\end{align} \]

Where

  • \(Y\) = probability of employment in year/wave t

  • \({\beta}_{0}\) = intercept

  • \({x}_{1}\) = time varying covariate

  • \({\vartheta}_{1,t-1}\) to \({\vartheta}_{n,t-1}\) = lagged health exposure variables

  • \(age \times sex\) = interacted age and sex

  • \(\epsilon\) = error term

  • \(c\) = country

  • \({x}_{2}\) to \({x}_{n}\) = time invariant covariates (potential confounders such as marital status and education)

  • \(t\) = time period of conducted SHARE wave.

Fixed effects logistics regressions were alternative models for model number one, and lagged Poisson regressions alternate models for model two. Further, dynamic fixed effect models were explored for waves that did not require adjustment for missingness patterns. However, convergence remained an issue for dynamic fixed effects models.

The absenteeism and hours worked models were estimated using a zero-inflated Poisson (ZIP) regression [Lambert, 1992 [36]], due to right-skewed distribution of days missed among the employed, and both those ill / not ill with a chronic condition in question had a non-zero probability of zero absenteeism. Given over dispersion and a continuous outcome variable, a ZIP regression is recommended [Lambert, 1992 [36]]. The Vuong test confirmed the model choice [Vuong, 1989 [62]].

In this model, the number of illness-related absenteeism was the main regressor, and both the regular and the inflated models had age (categorical), sex, and education as covariates. Chronic disease status was included in the form of a dummy variable, and based on the diseases detailed in Table 5.8. All analyses were conducted in STATA 15.

Table 5.8 List of countries in SHARE 1, 2, 5, and 6

Waves:

1

2

5

6

Austria

Yes

Yes

Yes

Yes

Belgium

Yes

Yes

Yes

Yes

Czech Republic

Yes

Yes

Denmark

Yes

Yes

Yes

Yes

Estonia

Yes

Yes

France

Yes

Yes

Yes

Yes

Germany

Yes

Yes

Yes

Yes

Greece

Yes

Yes

Italy

Yes

Yes

Yes

Yes

Luxembourg

Yes

Yes

Netherlands

Yes

Yes

Yes

Slovenia

Yes

Yes

Spain

Yes

Yes

Yes

Yes

Sweden

Yes

Yes

Yes

Yes

Switzerland

Yes

Yes

Yes

Yes

Note: The sample size for Luxembourg and Slovenia were to small to include in the main analysis.