4.1. Study limitations
The sample size in the present study is relatively small in relation to the analyses
performed in terms of complex models and the multi-level analyses treating each group as a
single observation. The fact that we, in spite of this shortcoming, largely received significant
results supports the validity of the results and conclusions.
As Martens and Haase (Martens & Haase, 2006) concluded, statistical methodology
“provides a necessary, but not sufficient condition for interpreting causal relationship among
constructs” (p. 905). Through the analysis procedure in the autoregressive model we were
able to minimize the influence of unmeasured variables that remained stable over time
(background variables) (Zapf, Dormann, & Frese, 1996). However, we cannot rule out that
the causal system may be under the influence of one or more unmeasured variables
asynchronously influencing the measured variables. This is something that deserves further
study. Firm conclusions would however require an experimental study design controlling for
all relevant factors, i.e., randomized control trials. This is seldom, if ever, possible in
29
organisational research and the approach applied here does offer substantial support for the
proposed causal mechanisms.
In the autoregressive model, the analysis was performed solely at the individual level.
Since the observations were clustered in higher level units, this may have resulted in
underestimated standard errors and thus inflated significance levels. However, the results
from the growth curve model where the multi-level structure of the data were accounted for
corroborated the influence of safety climate on safety behaviour.
Our use of the joint significance test approach to establish evidence of significant
mediation effects has the serious flaw of not addressing the overall mediation effect in a
complex, longitudinal model, but to the best of our knowledge, no test that does is available.
Although no firm conclusions regarding causality may be drawn from the present study, its
longitudinal design, the dual analysis strategy applied, and systematic testing of alternative
models rule out the most obvious threats to conclusions concerning causality and mediation.