Rotamap

Work patterns and sickness - preliminary findings

Alexander Bones, April 2025

Summary

We used statistical models to investigate the links between working patterns and sick leave. Our preliminary findings are:

• Working only night shifts in a week makes clinicians 21 % more likely to go on sick leave
• SASS doctors are 64% more likely to go on sick leave compared to consultants
• Senior resident doctors are 126% more likely to go on sick leave compared to consultants
• Junior resident doctors are 349% more likely to go on sick leave compared to consultants

This article demonstrates how the Report Builder can be used to extract data to gain meaningful insights into staff absences.

Introduction

The links between night shift work and poor health outcomes are long established, with prolonged night shift work linked to a range of conditions, including cardiovascular disease, metabolic disorders, and cancers. In healthcare, night shift work is a necessity in many specialties as patients require around the clock care. Oncall work is crucial for this; however it is important to balance the provisions for patient care with clinician wellbeing.

We wanted to investigate reasons for clinicians taking sick leave using data from Rotamap's electronic rostering systems Medirota and CLWRota detail work performed by clinicians as well as sickness information. Rotamap rosters nearly 90% of NHS anaesthetists and 20% of acute Trust clinicians. We are uniquely positioned with a substantial amount of data on this topic, which we can analyse to gain insights into the relationship between work patterns and sick leave.

Published research using electronic rostering data from nurses has studies this phenomenon and found a significant link between proportion of night shift work and sick leave incidence. We wanted to follow this methodology using data from Rotamap's systems to see if parallel relationships can be drawn and if there are other factors impacting sick leave incidence.

Data sources

All of the data in this article were extracted using the Report Builder. This is Rotamap's tool for organisation-wide reporting, returning data from individual Medirota and CLWRota sites with the option to aggregate it for Key Performance Indicators (KPI) reporting. Report Builder users are already creating reports to:

• Report on the number of PAs consultants are delivering
• Track total hours of work lost due to sickness
• Report on all extra shifts to identify pressure points
• Track extra claims forms needing action

In addition to providing a user interface to explore reports, you can also use our Central API to automatically extract data across the entire organisation. This API can be used to power live dashboards; Figure 1 below is an example visualisation created in Microsoft Power BI using a live data extract from the Central API.

Note that this example consists of data from our demonstration sites and is not reflective of a real organisation.

Figure 1. Example visualisation created from data returned by the Central API in Microsoft Power BI. This example shows when sick leave was taken across the year and the reporting category of the impacted work.

In addition to using business intelligence tools, a number of other languages, such as Python and R, can be used to return data from the Central API. Data sourced from the Central API can be combined with a number of other data sources, enabling you to gain insight by reporting across all areas of your organisation. Please see the Contact us section of this article to find out how to arrange a demo of the Report Builder and the Central API.

In this article, we collected data from three NHS Trusts with established links between Rotamap services and the Electronic Staff Record (ESR). We extracted data dating back up to five years, or from the point at which each Trust began using our services. The dataset comprised:

• 2,351 staff members
• 711,625 assignments
• 3,901 cases of sick leave

We analysed the data using generalised linear mixed-effects models with logistic regression. For a brief explanation of these model types, see Box 1.

All data were extracted using the Report Builder specific to each Trust.

Findings

We identified a significant link between the proportion of night shifts in the previous week and the incidence of sick leave. Our initial model only accounted for the proportion of night shifts worked as a fixed effect, including person identifier and department as random effects, and gave an odds ratio of 1.28 (95 % CI: 1.09 - 1.5). For an explainer of odds ratios and predicted probability, please see Box 2. For an explainer of fixed and random effects, please see Box 3.

There are a number of factors that can impact someone's likelihood of sick leave. We wanted to see if, in addition to night shifts, long shifts (those lasting twelve hours or more) were having any impact. We also suspected that the role of the clinician was having an impact on whether or not people were going on sick leave. Using the reporting category for roles in Rotamap systems, we were able to standardise the roles across the organisations in this article. The types of work that clinicians do at night will vary a lot based on their role. By including the role category as a factor in our models, we could assess the impact of general working patterns on the incidence of sick leave.

We ran an extended model that included the proportion of shifts worked as long shifts, as well as the role category of the person, as fixed effects, while including the individual and department as random effects. We checked if there was a substantial overlap between the long shifts and night shifts, since including both factors could be redundant if there is substantial overlap. We found this was not the case; the computed Variance Inflation Factor (VIF) scores were less than two, indicating low multicollinearity. When we included both terms in the model, we found that the long shift proportion did not significantly predict the likelihood of sick leave (OR 0.94, 95% CI: 0.82 - 1.09). The proportion of night shift work was still significant in this model (OR 1.27, 95% CI: 1.04 - 1.55), but the role category was found to have a greater affect on the incidence of sick leave.

As the role grade is a categorical variable, we had to select one as the baseline. Given the large number of consultants present in the dataset, we decided to use them as the baseline. Other role categories were presented as the difference in likelihood of sick leave based on that role against consultants. Note that from a modelling perspective we are assessing the impact of the different fixed effects assuming they are held constant. Therefore, we imagine that when we compare the role categories the figures are describing the same proportion of night and long shifts across the roles e.g. both roles working no night shifts in the previous week. While this might not happen in reality, this is the models estimate of how much each individual factor contributes to sick leave.

The predicted likelihood of SAS doctors going on sick leave compared with consultants was 64% higher.

The predicted likelihood of senior resident doctors going on sick leave compared with consultants was 126% higher.

Finally, the predicted likelihood of junior resident doctors going on sick leave compared with consultants was 349% higher.

We can visualise these differences by comparing 100 model predictions against the actual data from our systems. This can give us a general idea of how closely model predictions align with the real data. In Figure 2 below, we can see that the blue dot is roughly in the centre for the roles. We can be reasonably confident that our model takes into account the relevant interactions and terms.

Figure 2. Scatter plot of model predictions compared to actual data. Each white circle is a prediction from the model given a randomly selected set of inputs. The blue circle is the mean from the data used to generate the model. We would expect the blue circle to be roughly in the middle of the white circles if the model is capturing relevant terms and interactions.

We can also look at the relationship between the proportion of night shift work across each role grade. In Figure 3, we've plotted the predicted probability of each role grade going on sick leave given a certain proportion of night work in the previous week. The left hand limit is zero, where none of the work was night work, while on the right limit all of the work done was night work. As the graph goes from left to right, there is an increase in the predicted probability of someone going on sick leave. Note that this assumes a linear relationship between proportion of night shift work and sick leave. While this is statistically significant, we will investigate if modeling non-linear relationships provides a better fit.

Figure 3. Line graph of the predicted probability of a clinician going on sick leave based on the proportion of night work in the previous week, grouped by the role category. Solid lines indicate the model predicted probability at that proportion of night work. The shaded area represents the 95% confidence interval for the predictions.

Given the preliminary findings, we will continue to investigate the potential differences for the predicted probability of clinicians going on sick leave based on their role grade. We suspect that the nature of night work being performed differs based on being resident or non-resident; however this is not currently captured by our systems. We aim to publish our findings in a peer-reviewed scientific journal once we have investigated further.

Conclusions

This article has covered our preliminary findings concerning the impact of night shift work on sick leave. We have found that sick leave and staff grade have a significant impact on the likelihood of someone taking sick leave. This investigation has also demonstrated the power of the Report Builder and underlying Central API as tools that can provide data to investigate complex phenomena using statistical methodologies.

Disclaimer

All data used within this article were extracted from Rotamap's services and the organisations have been anonymised. Statistical analysis and visualisation were performed using the R programming language. Statistical modelling was performed using the lme4 package; visualisations were produced using the ggplot2 package. This article is preliminary in nature and should not be used to inform design or planning of rosters.

articles data