Short-term effectiveness of face-to-face periodic occupational health screening versus electronic screening with targeted follow-up: results from a quasi-randomized controlled trial in four Belgian hospitals

Objectives In many countries, organisations are legally obliged to have occupational physicians screen employees regularly. However, this system is time-intensive, and there may be more cost-effective alternatives. Our objective is to compare the short-term effectiveness of periodic occupational health screening of hospital employees by an occupational physician with a system of electronic screening with targeted follow-up. Methods A randomized controlled trial was set up among personnel of four Belgian hospitals, with three measurement moments between June 2019 and December 2020, to compare differences in self-assessed health, healthcare use, productivity and intermediate outcomes over 19 months. Mixed effects models were used to assess differences in effectiveness. Superiority and non-inferiority post-hoc tests were used as a robustness check. The experiment coincided with the first two COVID-19 waves during which hospital employees were exposed to an exceptional period of occupational stress. Results In total, 1077 employees (34% of the target population) participated. Although we observed some immediate effects of the intervention (less trust in the physician, absenteeism, and healthcare use), all these effects disappeared over time. After 19 months, including two waves of COVID-19 hospitalizations, no significant differences were observed between employees screened through face-to-face contact and those screened electronically. Conclusions Our study finds no indication that, in the short-term, substituting physician screening of the workforce with a quicker survey-based screening with targeted follow-up has different effects on the studied endpoints. However, as health and disease are often the result of a long-term process, more evidence is needed to determine long-term effects.

If | is normally distributed with mean + , and covariance , Equation (3) collapses to (2). The density ( ) is then the normal distribution, with = , = 2 , and variance function ( ) = 1. If ( ) is the Poisson density with = , = 1, variance function ( ) = , and log link function, a Poisson mixed model can be estimated, or a negative binomial model when an extra dispersion parameter is added. If ( ) is a Bernoulli density, and using a probit or logit link function, a probit or logit normal model can be obtained.
On the first level (L1), the has subscripts (j) to indicate they belong to one particular employee. On the second level, each for a particular employee ( ) can be represented as a combination of a mean estimate for that parameter ( 0 ), and a random effect for that measurement moment ( ). Only the intercept ( 0 ) and time slope ( 2 ) can vary across employees, the other are fixed. with containing age, gender, hospital, and education. In each model, fixed effects were thus included for group (intervention or control) and time (1,2,3), covariate fixed effects for age, gender, educational attainment, and hospital, and random intercept effects for each individual employee.

Approach in our analyses
In our data, some of the measured outcomes are typically left-or right-skewed. We therefore used logarithmic transformations for musculoskeletal complaints (NMQ), general mental health (GHQ), and need for recovery (NFR). We also make use of continuous ordinal regression (41) to analyse the visual analog scale for general health status (EQ-5D), as several authors have argued that it should be analysed as an ordinal instead of a continuous variable ((42-45) as cited in (41)). As this method is not standardly used, we also include the linear mixed model of the EQ-5D visual analog scale.
It is moreover often likely that count variables diverge from the normality assumption of the residuals because of frequent zero-values (e.g. zero days absent, zero consultations, etc.) and dispersed non-zero values (long tails). We therefore estimated these outcomes with binomial and Poisson-type (e.g. generalised Poisson or negative binomial) mixed models. The fit of these models is tested by several tests for simulated residuals (46): a Kolmogorov-Smirnov test (KS) to test distributional fit, a test for the presence of zero-inflation, and a dispersion test to verify whether the observed variance is higher or lower than the theoretical variance of the model. A generalised Poisson or negative binomial model is adopted when it is necessary to specify a separate dispersion parameter to help alleviate over-or under-dispersion. We also test whether a large number of outliers occur, and quantiles are compared to their theoretical value. Finally, the presence of autocorrelation is tested with a Durbin-Watson test on uniformly scaled residuals , and the covariance structure is adjusted (if autocorrelation is present).

Post-hoc tests
After the analyses, we make use of post-hoc Tukey-Kramer tests (also called honest significant difference tests) as a robustness test of our estimates.
where − is the difference between the estimated means, is the Mean Square Within, and n is the number of employees in the group. The means are obtained by making predictions on the estimated mixed models (e.g. the mean stress score for the control and intervention group in time 1, 2, and 3) with their standard deviation. The pairwise differences of these six (3 times * 2 groups) means are then compared. Results are averaged over the levels of the covariates (hospital, gender, education), and a p-value adjustment is used by the Tukey method for comparing a family of 6 estimates. Because the sample sizes are unequal (unbalanced panel data), the Tukey-Kramer adjustment is used to estimate separate standard deviations for each pairwise comparison. A similar approach is used for the non-inferiority tests.

Time
Wave 1 (N=208) Wave 2 (N=208) Wave 3 (N=208) Total (N=624)       Table S9. Regressions mixed methods: group + time + interaction, and covariates. For each variable, estimates are on the first line, and 95% confidence intervals are in parentheses. *=p<0.1; **=p<0.05; ***=p<0.01. ICC=between/(between+within) variance. σ² = within-group variance; τ00 = variance of random intercept ; τ11 = variance of random slope; ρ01= correlation between random intercept and slope.    Figure S1. Predicted effects plots of mixed models with 95% confidence intervals, blue = control group, orange = intervention group. X-axis=measurement moment (1,2,3). The post-hoc linear predicted means of each outcome are based on the regression estimates, with their confidence intervals, for each measurement moment (x-axis) and by group (colour) Supplemental Table S13. Posthoc estimations with contrasts of within (one group over time) and between (same time different groups) group, with p-value of non-inferiority and nonsuperiority tests. The estimates and non-inferiority and non-superiority tests for the post-hoc estimations are in reversed direction (intervention -control) and the relevant differences (delta) above or below which non-inferiority or non-superiority applies are stated. The estimates for which there is no indication for non-inferiority or non-superiority are formatted in bold. For each variable, estimates are on the first line, and p-values on the second.