Study setting and design
This study was conducted in Soonchunhyang University Seoul Hospital, a 734-bed acute-care referral hospital in South Korea. It was approved by the Institutional Review Board (approval number: 2019-01-008). Since 2010, we have maintained an HH monitoring team at the hospital, comprising 24 members across various departments; the infection control team comprises five members. Every quarter, approximately 2500 HH opportunities are monitored by the infection control and HH monitoring team members. We trained HH monitor personnel on monitoring methods, precautions on observation, input of results, and practice through monthly meetings. In the case of the existing monitoring team, we maintained the quality of monitoring by conductiong video training and testing the HH monitors at the first meeting of the year. We follow standard HH monitoring methods by directly observing HH per WHO guidelines [1]. The HH monitor was conducted during the observer’s working hours and there were no restrictions during the week days, weekend, day and night. In order to prevent the Hawthorne effect, observations for one HCW were limited to less than four, and the observation time per department was limited to less than 20 min [4]. Observers in each department did not monitor members of their own department. From January to December 2018, we collected data regarding the HH compliance rates of doctors, nurses, and other HCWs (medical technical assistants, dieticians, physiotherapists, and radiological technologists).
Statistical analysis
The HH compliance rate was calculated by dividing the number of observed HH actions by the total number of opportunities. Opportunities were defined based on the WHO’s “5 moments for HH” (before touching a patient, after touching a patient, before clean/aseptic procedures, after body fluid exposure/risk, and after touching patients’ surroundings). Meanwhile, rates of compliance with optimal HH techniques were calculated based on adherence to the six-step technique recommended by the WHO on each opportunity (rub hands palm to palm, right palm over left dorsum with interlaced fingers, and vice versa; palm to palm with fingers interlaced; backs of fingers to opposing palms with fingers interlocked; rotational rubbing of left thumb clasped in the right palm, and vice versa; and rotational rubbing, backward and forwards, with clasped fingers of the right hand in left palm, and vice versa) [1, 5].
The HH compliance/optimal HH compliance values were calculated for each observed person and the data were expressed as mean, median, and interquartile range (IQR) measurements. We used the generalized estimating equation model for logistic regression using an unstructured working correlation matrix to compare HH compliance or optimal HH compliance rates in different job categories (doctors, nurses, and other HCWs) and year quarters.
To calculate the sample size for estimating the population’s HH compliance and optimal HH compliance, the following conditions were considered: (1) the variability in the target population; (2) the desired precision in the estimate; and (3) the desired confidence in the estimate. In this study, the following equation was applied:
$${\text{n}} \ge Z_{\alpha /2}^{2} \times \rho \times \left( {1 - \rho } \right)/d^{2} ,$$
where ρ represents the population proportion, d the absolute difference, and 1-α the confidence interval (CI) [6, 7]. This sample size can be interpreted as the minimum sample size required to get the sample proportion to fall within 100d% of the true proportion with 100(1 − α)% probability. We considered ds of 5%, 10%, 20%, and 30%, with CIs of 99%, 95%, and 90%, respectively. Among the various cases, we focused on 10% for d and 95% for CI. We calculated the number of n using the R package (‘binomSamSize’) and selected three methods to represent them in a Additional file 1: Table S1. The first method approximation is based on the central limit theorem [8] The other two are the Wilson score method [9] and the Agresti-Coull method [10], which can be used even when the data are asymmetric, the sample is small, and the observations are biased [11].