Participants
The hospital
We analyzed data that were retrieved from a large Norwegian hospital from 2011 to 2016. The hospital employs more than 20,000 workers dispersed at several different locations. The hospital had 1512 work units, employing between one and 168 employees each (20 employees on average). In addition to providing high-quality healthcare, the hospital also engages in research and has a responsibility for medical education and training. The hospital is highly specialized and has local, regional, and nationwide patient-treatment responsibility.
The employees
The data contained 27,468 unique employee ID numbers for the given study period. In the final sample, we included fulltime employees only (> = 80% cumulative employment contracts) who worked in units comprising a minimum of four employees. To compare downsizing to stability, quarters in which employees experienced other organizational changes (i.e. spin-off, merger, upsizing, insourcing, and outsourcing at the unit-level [16]) were dropped from the analyses.
Data and measures
We employed two different sources of merged data drawn from registers in the hospital’s Human Resources (HR) department, in addition to the hospital’s work environment survey. Using HR register data allowed us to track unit-level downsizing and sickness absence over several years for the same individual employees, using objective data with a high level of accuracy. Although the hospital continuously recorded the HR registries, we aggregated the data into employee-quarters to make them suitable for analysis. By using the hospital’s annual work environment survey, aggregated to the unit-level, we could also track each work-unit’s measures for control and commitment for the same period.
In this manner, the data were organized as employee-quarters (i.e. our level of observation), nested within employees, nested within work units. Because we did not have work environment data for 2014 and 2016, registry information from these years was only utilized when relevant for adjacent quarters (e.g. registry data from Q4 2014 provide information on employees who went through “downsizing in the last quarter” in Q1 2015). Similarly, because we did not have information on whether employees in Q1 2011 experienced downsizing in the previous quarter, Q1 2011 was dropped from the analyses.
The final sample consisted of 173,787 observations, nested in 21,085 employees nested in 1167 work-units. Each employee was observed between one and 15 quarters, and eight on average. Each work unit consists of between one and 114 employees each, and 18 on average.
The project was reported to the Norwegian Data Protection Authorities (NSD) and data protection officers at the hospital and complied with the Declaration of Helsinki.
Sickness absence
We used data on employee sickness absence retrieved from hospital records. Within these registers, all absence periods were listed with the beginning and end date as well as the percentage of absence. We included all registered sickness absence, irrespective of percentage, in the analyses. Consecutive absence periods were merged. Registering sickness absence is a system-based prerequisite for the work-units to recruit substitute personnel. These objective records should therefore be regarded as highly accurate and complete compared to alternative measures such as self-reported absence. The dependent variable in this study was whether an employee entered a period of sickness absence during the quarter. Because shorter spells are more likely to be influenced by aspects other than health [18, 19], we aggregated the data to separate short- (<=8 days) and long (> = 9 days) periods of sickness absence. For absences lasting 9 days or longer, the employee is required to obtain a medical sickness absence certificate. Because the present study was carried out in the Norwegian hospital sector context, we decided to separate between short-term (1–8 days) and long-term (> = 9) sickness absence.
Unit-level downsizing
We derived data on downsizing from hospital records of employment contracts. These registers provided historical data regarding employees’ contracts throughout the study period. The particular change was identified by tracking the specific units in which the employees worked each quarter and how they shifted units. We quantitatively operationalized unit-level downsizing by adopting the threshold defined by Røed and Fevang [17]: a major change necessitates a reduction in employees amounting to more than 20%. In this study, this meant that if a unit downsized by > 20% in one quarter (without simultaneously experiencing an outsourcing amounting to > 90% of the unit or a unit-level merger), we coded the employees as being downsized. The change variable was binary, given values of 0 (stability) or 1 (unit-level downsizing). Measured quarterly, we identified whether an employee was anticipating downsizing in the upcoming quarter (i.e. “unit-level downsizing next quarter”), presently experiencing downsizing (i.e. “unit-level downsizing this quarter”), had experienced downsizing in the previous quarter (i.e. “unit-level downsizing previous quarter”) or was having stability (i.e. the control group). In this paper, experiencing stability entailed an absence of unit-level downsizing, in addition to the absence of unit-level spin-off, merging, upsizing, insourcing and outsourcing of units [16].
Temporary contracts
The data on temporary contracts were drawn from HR registers on employment contracts. In the register, employment contract was categorized as being either temporary or permanent. The objective registers could therefore be considered as highly accurate and complete.
Work environment moderators
The final source of data was employees’ experiences as reported in the work environment survey. This was a comprehensive survey that the hospital conducted in 2011 (response rate 70%), 2012 (response rate 72%), 2013 (response rate 80%), and 2015 (response rate 76%).
The measures for control and organizational commitment were developed and validated by the General Nordic questionnaire (QPS Nordic) for psychological and social factors at work [50]. Both variables were measured using self-report on a 5-point Likert scale. For control, the scale ranged from “very seldom or never,” coded as (0), to “very often or always,” coded as (4). For commitment, the response scale ranged from “disagree totally,” coded as (0) to “agree totally,” coded as (4). The sample items used to measure control at work included “Can you set your own work pace?” and “Can you influence the amount of work assigned to you?” [50]. The sample items measuring organizational commitment from this scale included “This organization really inspires me to give my very best job performance” and “To my friends, I praise this organization a great place to work” [50]. The Cronbach’s alpha was 0.795 for control and 0.873 for organizational commitment.
We aggregated the work environment factors to the work-unit level before merging them with the rest of the data. The mean values of each specific work-unit were assigned to all of the employees working in that particular unit. Each employee-quarter was given the value from the employees’ work-unit that specific year. Aggregating the work environment factors to unit-level was necessary. Anonymity concerns only permitted us to identify the specific work-units in which the employees worked, as opposed to identifying each individual respondent. Although aggregation necessarily meant that some data were lost, it was also an advantage because it allowed us to focus on the work-units’ work environment, rather than on individual perceptions. As argued by Hausknecht et al. [44], conceptualizations of sickness absence have often focused on individual-level predictors, thus diminishing the work-unit context that frames sickness absence behavior.
When aggregating the work environment survey, it was also important to consider group agreement. Commitment to the organization demonstrated strong agreement within the work-units with an rWG of 0.73. Control showed only moderate agreement with an rWG of 0.64 [51]. Traditionally, 0.7 has been set as a cut-off point denoting higher interrater agreement and acceptable aggregation. Lebreton and Senter [51] argued that this cut-off point might be too high in some instances (e.g., the measure is not used for decisions involving specific individuals). Because we could only merge the work environment variables with the rest of the data at the work-unit level, we also aggregated control. However, the moderate level of agreement needed to be taken into consideration when interpreting the findings. Expectation-Maximization algorithm (EM) was used for missing values.
Control variables
Several variables likely to influence the dependent variable were drawn from the HR registers on employment contracts, and were included as control variables. In this paper, control variables comprised salary, age, gender, country of origin, job position, and multiple jobs. We operationalized country of origin into Norwegian, other Nordic, other Western and non-Western. We coded job position into physician, nurse, other patient-oriented position, administration/management, kitchen/cleaning/orderly, other operations (e.g. IT), and other. As hospital employees commonly hold more than one job contract, we included a dummy variable to specify when an employee was a multiple job-holder. Age was measured in years and salary comprised contracted remuneration. Gender was dummy coded (female =1, male = 0). In addition, all three potential moderating variables are also included as control variables, because they are also likely to be confounding variables (e.g. if units with a high percentage of temporary contracts are more likely to downsize).
Statistical analysis
We used a multilevel regression model with random intercept at employee- and work-unit level to analyze the relationship between unit-level downsizing, moderators (temporary contract, control and organizational commitment, separately) and sickness absence. Hence, the unit of analysis was employee-quarters nested within employee, and employees nested within work-unit. For the purpose of analysis, we hierarchically nested employees according to the unit in which they most often appeared. In all, 81% of employees appeared in the same unit in all observations. Stress testing the findings by excluding all quarters in which the employees did not work in their main unit did not significantly alter the results.
We ran a multilevel model because such models are suitable to account for dependency in the data (i.e. employees measured at multiple time points are more similar to themselves than to others, and employees within one work-unit are more similar to each other compared to employees at other units). Multilevel models explicitly and properly estimate the degree of dependencies of observations (i.e., non-independence), thereby protecting researchers from an error of inference and spurious significant results [52,53,54]. Moreover, random effects regression offers the most efficient estimator of the relationship between unit-level downsizing and sickness absence by measuring both the variation between employees and within employees, in a longitudinal design [55]. The dependent variable (sickness absence) was examined at the lowest level of analysis (i.e., employee-quarters). The dependent variable, episodes of sickness absence, was binary; therefore, we used multilevel logistic regression.
We ran separate analyses for each of the three moderators (temporary contract, work-unit control and work-unit organizational commitment). We estimated moderating effects by testing for significant interactions between unit-level downsizing and each moderator, respectively [56]. The level of statistical significance was set to P < .05. All analyses were performed using the xtmelogit command in STATA/SE, version 15.1 (StataCorp LP, College Station, TX, USA, http:www.stata.com/company/contract/).