To identify a set of patient care tasks, we surveyed four critical care RNs trained to care for patients with serious communicable diseases, three infectious disease physicians, and two IPs, who were each asked, “What patient care tasks do you (or others) do where it is important to protect you or your patient from a contagious disease?” Respondents produced 148 tasks in total. To reduce the number of tasks, we printed tasks on separate cards, which two critical care RN subject matter experts (SMEs) sorted into groups of redundant tasks or single tasks that could not be grouped with another task(s). The SMEs reviewed their groupings until they reached a consensus. We then carried forward tasks that RNs in the ICU, ED, and medical/surgical floors perform routinely, which resulted in a final list of 25 tasks (Supplemental material). Tasks were rephrased as needed to minimize ambiguity.

We recruited IPs and RNs from EDs, ICUs, and medical/surgical floors from adult inpatient hospitals across a large academic hospital network. Participants were invited to take an hour-long online survey about infection prevention.

Participants reviewed the 25 tasks and then judged the similarity of all possible pairs of tasks using a scale from 1 (“Not at all similar,”) to 9 (“Extremely similar,”). When rating each task-pair, participants were told to take an “…infection prevention perspective (i.e., preventing the spread of pathogens)” and to assume that the patient is “…average-sized, not bed-bound, able to cooperate (e.g., understand and follow commands), but is weak. No known special precautions are required.” The order of the 300 task-pairs was randomly generated for each participant. After rating all task-pairs, participants described the reasoning behind their ratings. Each reason given by a participant was later printed on a separate card and then sorted into groups of recurring reasons.

Multidimensional scaling (MDS) takes a measure of psychological similarity (e.g., ratings) between all possible pairs of a set of “objects” (e.g., patient care tasks), places each object as a point in an n-dimensional space, and then iteratively configures the objects in space until the distance between each pair of objects reflects the corresponding measure of psychological similarity as closely as possible. Objects perceived as similar are close together in space whereas objects perceived as dissimilar are far apart. The overall congruence between the distances between objects in the multidimensional space and their corresponding measure of similarity is quantified as “stress,” which ranges from 0 to 1 with lower values indicating better fit. The dimensionality of the space is set by the analyst, who strives to find the smallest number of dimensions that best reproduces the psychological similarity of the objects. Each dimension may then be interpreted in terms of known characteristics of the objects, which are reflected in how objects are ordered along a dimension. We used a variant of MDS called Individual Differences Scaling (INDSCAL),¹¹ which also measures how salient (i.e., important) each dimension is to each participant, called a “weight.” If a participant(s) does not perceive a certain dimension, then that dimension will have a weight of zero for that participant(s). The higher a weight is above zero, the more salient that dimension is to that participant(s). We implemented the INDSCAL model using the PROXSCAL algorithm, treating similarity ratings as ordinal data and using the primary approach to ties.¹² We scaled data from all raters in 1 to 6 dimensions. For each dimensionality, we retained the best-fitting final configuration from 1,000 randomly generated starting configurations,^15,16 according to recommended convergence criteria.¹⁵

We identified the dimensions underlying the task-pair ratings using three criteria: model fit ^11–14and the interpretability^11–14 and replicability^12,14 of dimensions. Regarding model fit, we plotted the stress of the best-fitting final configuration as a function of dimensionality to identify the point at which positing additional dimensions failed to improve fit substantially.^11–14

We took a multiple regression approach to interpret dimensions,^12,13 in which we first hypothesized 12 characteristics of patient care tasks based on the input of a SME panel (three infectious disease physicians and four critical care RNs) and recurring reasons participants gave for their similarity judgments. We translated each characteristic into a statement that could be rated on a 5-point scale (“None,” “Low,” “Moderate,” “High,” and “Very High”). A new sample of HCWs independently rated the 25 tasks in terms of the hypothesized characteristics. For each HCW, the order of the 25 tasks and the order of the characteristics for each task were randomly generated. One hypothesized characteristic of tasks, the “Relative importance for preventing versus controlling (e.g., diagnosing or investigating) an infection in a patient,” was calculated from two characteristics as the arithmetic difference between the “Relevance for preventing an infection in a patient,” and the “Relevance for controlling (e.g., diagnosing or investigating) an infection in a patient.” For each of the final 11 hypothesized characteristics (Table 2), we calculated the mean rating for each task, which were predicted from tasks’ coordinate value on each dimension using multiple regression.^12,13 Interpretation of dimensions followed recommended guidelines¹³; as a minimal requirement, the characteristic should have a multiple correlation that is significant at the α = 0.01 level. Additionally, the characteristic should have a high multiple correlation (R ≥ 0.70) and a high regression weight on one dimension only.

To assess the replicability of the best-fitting and most interpretable configuration, we scaled each HCW population’s rating data separately.¹⁴ For each population, we used the coordinate values of the 25 tasks from the full-sample analysis as a reasonable starting configuration for the PROXSCAL algorithm to optimize.¹² We assessed the similarity of dimensions between populations by calculating Pearson correlation coefficients between task coordinates on each pair of dimensions. We then averaged correlations between corresponding dimensions across populations using Fisher’s r to z transformation.

To determine whether the dimensions differed in their salience, and whether such patterns differed between populations of HCWs, we analyzed weights using a Greenhouse-Geisser corrected mixed model analysis of variance (ANOVA) with Dimension as a within-subjects factor and HCW Population as a between-subjects factor.¹¹ Post-hoc pairwise comparisons were performed using Bonferroni’s correction for multiple comparisons. A power analysis in G*Power 3.1.9.2¹⁵ suggested a total sample size of 40 HCWs was required to detect a large interaction (f = 0.4)¹⁶ between Dimension and HCW Population with at least 80% power and α = 0.05 across a range of dimensions (2 – 6). All statistical analyses were performed in SPSS version 26. A p-value less than 0.05 was considered significant.

We received task-pair rating data from 11 IPs and 10 ED, 10 medical/surgical, and 9 ICU (5 Medical ICU, 4 Surgical ICU) RNs. Most (83%) RNs worked in a large quaternary academic medical center. RNs had a median of 8 years (Range: 1 – 38 years) of nursing experience. 16 (40%) RNs reported working or having worked in a COVID-focused unit. 54% of IP respondents were from the same academic medical center. IPs had a median of 4.5 years (Range: 2 – 25 years) of experience as an IP. Nearly half (45%) of the IPs had a background in nursing.

An additional 6 ICU RNs with a median of 8.5 years (Range: 7 – 43 years) of nursing experience rated the 25 tasks on 12 scales. All 6 RNs were also trained in a hospital biocontainment unit, which involves annual exercises, skills training, and online course work related to IPC. Additionally, all 6 ICU RNs reported working or having worked in a COVID-focused.

Figure 1 shows the stress of the best-fitting configuration obtained in 1 through 6 dimensions. Inspection of Figure 1 suggests that positing more than 4 dimensions does not substantially improve fit. However, the best-fitting configuration in 4 dimensions was only marginally better than that obtained in 3 dimensions. Using the multiple regression approach for interpreting dimensions described earlier,¹³ we found that only the 3-dimensional configuration was fully interpretable, which we then retained for further analyses.

To assess the replicability of the 3-dimensional configuration of the full sample, we scaled each HCW population’s data separately in 3 dimensions.¹⁴ The stress for IPs (Stress = 0.03) and ED (Stress = 0.03) and ICU (Stress = 0.03) RNs were comparable, whereas the stress for medical/surgical RNs (Stress = 0.05) was somewhat higher. Table 1 shows all pairwise correlations between each populations’ dimensions. Corresponding dimensions were strongly and significantly correlated between each population (Dimension 1: r¯z' = 0.94, Dimension 2: r¯z' = 0.85, and Dimension 3: r¯z' = 0.79) whereas none of the cross-dimension correlations were significant. Thus, a replicable set of dimensions was found across populations.¹⁴

Following recommended guidelines for interpreting dimensions using multiple regression (Table 2),¹³ we interpreted Dimension 1 (Figure 2) as the perceived magnitude of infection risk to the patient in a task. This dimension was strongly and exclusively related to the characteristics “Infection risk to patient,” and “Infection risk to patient if task performed improperly,” Dimension 2 (Figure 2) reflects the perceived “dirtiness” of a task and the perceived risk of HCW exposure to body fluids. This dimension was related to the “‘Dirtiness’ of task,” “Risk of healthcare worker exposure to body fluids,” and to a lesser extent, the “Amount that task varies.” Although failing to meet recommended guidelines,¹³ “Infection risk to healthcare worker,” “Infection risk to healthcare worker if task performed improperly,” and “Amount of PPE needed,” were weakly related to Dimension 2 as well. Lastly, Dimension 3 (Figure 3) reflects a task’s perceived relative importance for preventing versus controlling an infection in a patient. On this dimension, the arithmetic difference between “Relevance for preventing an infection in a patient,” and “Relevance for controlling (e.g., diagnosing or investigating) an infection in a patient,” was related.

Analysis of weights revealed a significant main effect of Dimension (p < .001), a non-significant main effect of HCW Population (p = .90), and a non-significant interaction between Dimension and HCW Population (p = .72). Post-hoc pairwise comparisons of the main effect of Dimension revealed that HCWs weighted Dimension 1 (M = 0.45, SEM = 0.01) significantly more heavily than Dimension 2 (M = 0.38, SEM = 0.01; p < .001) and Dimension 3 (M = 0.33, SEM = 0.01; p < .001). HCWs also weighted Dimension 2 significantly more heavily than Dimension 3 (p = .001).