the main exposure variables were estimated using a generalized
linear model (robust Poisson method) [35] before
and after adjusting for other variables. Statistical interaction,
or effect modification, was assessed quantitatively by
formal statistical tests introducing product terms into the
regression models. The likelihood ratio test (LRT) was
used to examine whether the effect of job strain on psychological
distress was modified by social support at work
(ISO-STRAIN model), occupational status and by night
shift-work. All the analyses were also stratified by gender,
as most studies throughout the world have found that
women report more psychological symptoms than men, but
much less is known about factors contributing to the
increase of risk among women in countries with different
levels of development [25, 26, 39].
We also tested for the effect of the workplace environment
on psychological distress before and after adjusting
for individual level data using multilevel models [37]. Data
were arranged in a two level hierarchical structure in which
workers were nested within work units. We used multilevel
poisson regression models to estimate the proportion of the
total variance of psychological distress that was due to
differences between work units and individual factors. In
order to test for a group-level effect of job strain using
multilevel methods, we split the variables control and
demand using the group mean (group-level variable) and
the within-group deviation variable (individual score minus
group mean). In this random coefficients model, the association
between job strain and psychological distress is
allowed to vary across work units. The model parameters
were estimated using STATA 10.0.