Estimates and confidence intervals for mean postoperative health-care-associated infections per week among operating rooms and other anesthetizing locations

Abstract

Purpose

For anesthesia practitioners to contribute to reduced postoperative health-care-associated infections, they need to know which anesthetizing locations at their facilities have infections per week with confidence intervals (CI) exceeding relevant thresholds. We sought the development of more precise estimates and CIs.

Methods

We conducted a retrospective cohort study of the 2024 fiscal year of surgical cases at a large teaching hospital in Iowa (University of Iowa Health Care Medical Center, Iowa City, IA, USA), with 90-day postoperative infection codes from the International Classification of Diseases, 10th Revision, as diagnosed by surgical teams. Following earlier simulations, we used Poisson regression with heteroscedasticity-consistent robust variance estimation. We compared CI widths for new postoperative health-care-associated infections per week per room to the corresponding method of batch means, quantified as costs of Type I errors (e.g., unnecessary vs appropriate Staphylococcus aureus transmission monitoring) and Type II errors (e.g., failure to mitigate vs successful reductions in excess infections).

Results

The 75 operating rooms and other anesthetizing locations (“rooms”) studied had 1,095/42,978 (2.6%) cases with patients who developed postoperative infections. The 75 rooms × 366 days equaled 27,450 room days, with 96.2% having no patient who developed a postoperative infection (26,401), 3.7% having one infection, 0.2% having two infections, and 0.01% having three. There was no significant serial correlation between days by room or between rooms by day. The preceding case of a patient who developed a postoperative infection did not significantly increase the probability of the next case in the room being of a patient who developed an infection, all 75 Šidák-corrected P ≥ 0.32. Counts of infections per day were consistent with Poisson distributions for 73/75 rooms. The other two rooms had ratios of sample variances to sample means of 1.32 and 1.78, respectively. Among the 15 rooms with a mean of ≥ 0.50 infections per week, the 99% CIs with Poisson regression were narrower than using the method of batch means by approximately 35%. Pooling all rooms, the CI width was approximately 38% less. These reductions were equivalent to increasing sample sizes from 1.00 year to 2.35 and 2.60 years or reducing costs by 88% or 92%, respectively.

Conclusions

Poisson regression with robust variance estimation resulted in more precise estimates than did the method of batch means, which will lower health care costs. We recommend reporting counts of postoperative health-care-associated infections per week for each anesthetizing location by using this method.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12630-026-03141-3.

Postoperative health-care-associated infections, including surgical site infections, are a public health challenge, including in the context of anesthesia-related procedures. Around one third (34%) of hospital inpatients and outpatients undergoing care with an anesthesia clinician receive a parenteral antibiotic, and most (64%) hospital patients receiving a parenteral antibiotic undergo an anesthetic.¹ Consider the “anesthesia work area” to include the patient’s nose and skin, the hands of anesthesiologists and anesthesia providers, the anesthesia machine and cart, and intravenous lumens for administering anesthetic drugs.² Transmission of Staphylococcus aureus in the anesthesia work area was associated with a greater risk of postoperative surgical site infection, from 2% to 11%.³ An example of S. aureus transmission in the anesthesia work area is from the nasopharynx of one patient at the start of their case to hands of the anesthesiologist at the end of the case to the nasopharynx of the next patient in the room that day, as well as to the anesthesia machine agent dial.⁴ When the transmitted S. aureus were resistant to the prophylactic antibiotic employed for surgery, the associated risk of developing postoperative infection increased to 18% (P < 0.001 for ordered association).³ Using six categories of infection prevention products designed to prevent pathogen transmission—nasal antisepsis, skin antisepsis, hand sanitizers at the bedside, disinfecting intravenous caps, intraoperative decontamination, and feedback—anesthesiologists reduced S. aureus transmission (relative risk, 0.56; P < 0.001) and consequent surgical site infections by more than half (hazard ratio, 0.12; P = 0.004).⁵ These anesthesiologist-led trial results were validated in a follow-up observational study by another anesthesiology department across its entire surgical suite.⁶

Nevertheless, current statistical analysis methods limit the understanding of how anesthesia practitioners can apply these findings to reduce postoperative health-care-associated infections at individual facilities. Consider examples from two earlier studies of the distribution of postoperative infections among operating rooms and other anesthetizing locations.^7,8 A hospital-based anesthesia group had 112 anesthetizing locations.⁸ The weekly mean was approximately 21 infections among patients uninfected on the day of surgery, diagnosed over 90 postoperative days by surgeons.⁹ The 50% of locations with the fewest cases among patients who developed postoperative infections accounted for 5% of infections (99% confidence interval [CI], 1 to 10).⁸ The 10% of locations with the most cases among patients who developed a postoperative infection accounted for 40% of infections (99% CI, 31 to 49).⁸ Because the anesthetizing locations with the most postoperative infections were operating rooms, we subsequently refer to them as “rooms.” For the 10% of rooms with the most infections, there would be net cost savings from the aforementioned multiple-component bundle of interventions to reduce infections caused by pathogen transmission through the anesthesia work areas: improved environmental cleaning, hand hygiene, vascular care, and preoperative patient decolonization.^7,10–12 For the other 60% of infections from 90% of rooms, using just one or two components of the bundle of interventions as a “compromise” approach is unsupported because a complete patient data meta-analysis reported lower efficacy at preventing surgical site infections.¹³ Estimating suitable, narrow CIs for each room exceeding a high threshold rate of infections per week would facilitate decision-making regarding the anesthetizing locations in which to apply monitoring of S. aureus transmission and optimize the above, basic preventive measures.^2–6,12,13 Prompt recognition of rooms with many infections is important because the sample size for checking operating rooms for intraoperative S. aureus pathogen transmission¹⁴ was based on first detecting the rooms with substantial numbers of cases of patients with postoperative infections.^7,8

In October 2024, Shi et al. published simulations comparing accuracy of multiple methods to estimate means from count data with overdispersion (i.e., variance > mean, unlike for Poisson distribution).¹⁵ They reported nominal coverage (i.e., accurate CIs) by performing simple Poisson regression using robust variance estimation.¹⁵ In statistics packages such as Stata (StataCorp LLC, College Station, TX, USA), this is done simply by selecting a menu option. Their novel results suggest that the mean weekly incidence proportions of postoperative infections after anesthesia might be estimated with precision by relying on variants of Poisson distributions. We tested this hypothesis and examined correlations among days, rooms, and cases, evaluating whether observational studies and randomized trials can reasonably treat the room–day combinations as statistically independent.

Methods

The University of Iowa institutional review board determined that this project (#202505294) did not meet the regulatory definition of human subjects research. The deidentified data used were 42,955 anesthetics administered at a large teaching hospital in Iowa (University of Iowa Health Care Medical Center, Iowa City, IA, USA) from 1 October 2023 through 30 September 2024—N = 366 days because this period included a leap year. The fields were “room,” date/time of patient entrance, date/time of patient exit, specialty (e.g., vascular surgery), and new infection or no infection within 90 postoperative days. The 42,995 cases included 85 operating room or procedure locations (e.g., pediatric hospital surgical suite room). We excluded the 10 rooms each with fewer than 5 cases over the year, leaving 42,978 cases in 75 rooms.

An anesthetizing location is a room with an anesthesia machine. All operating rooms are anesthetizing locations. A cardiac electrophysiology room is an anesthetizing location if it has an anesthesia machine. Our study was of anesthetizing locations, but the locations with postoperative infections were almost entirely operating rooms.^7,8 Regardless, the anesthetizing locations recorded manually in the electronic health record were inaccurate for some rooms.^16,17 Therefore, for accuracy, the fields that we used for anesthetizing locations were instead locations known from the metadata of the monitors transmitting the pulse oximetry measurements.^16,17 For > 20 years, operating room management studies have relied on sensor data for room identification, occupancy, and turnover time information.¹⁸

We obtained data on postoperative health-care-associated infections (surgical site, respiratory, and bloodstream) using codes from the International Classification of Diseases, Tenth Revision (ICD-10), of any encounters that included extensions to all subsequent sixth, seventh, and eighth digits: O86.0, T81.4, T82.6, T82.7, T83.5, T83.6, T84.5, T84.6, T84.7, and T85.7.^7–9,19,20 Each ICD-10 code’s associated fields include presence/no presence on admission, date of the encounter from which the coder identified or confirmed the code, type of encounter (e.g., clinic visit), and practitioner, from which was determined their department (e.g., gynecology).^7–9 Postoperative infections were considered those among patients without preoperative infection who had a health-care-associated infection over 90 postoperative days based on ICD-10 codes abstracted from surgeons’ clinical notes.⁹ We used 90 days because, for the anesthesia studies described at the beginning of the manuscript, follow-up for at least 60 days was associated with approximately 2.61-fold more surgical site infections detected relative to 30 days (P < 0.001).¹³ Initiation of antibiotics by a primary care or emergency medicine department without surgical team involvement was not considered relevant infection.⁹ Using this approach, the positive predictive value for postoperative infections obtained at the hospital was approximately 99.3% compared with National Healthcare Safety Network definitions.⁹

Our retrospective cohort data were analyzed using Stata version 19.5. We treated P < 0.01 as significant. With time series equally spaced at 1-day intervals, counts of infections per day in each room were the same as rates for the room. Serial correlation in the count of infections between successive days for the same room was evaluated using a fixed-effects autoregressive model, treating the room as a fixed effect (command: xtreg, fe). A random effects model was not used because of the large variability among rooms in the mean infections per week. The Wooldridge test for lag 1 autocorrelation was used for inference, and was carried out using Stata’s xtreg and xtserial commands. The lag 1 Pearson correlation was estimated using Stata’s xtregar command. We hypothesized no significant correlation from day to day based on earlier findings for Enterococcus species, which, of vancomycin-resistant Enterococcus spp., S. aureus, Klebsiella pneumoniae, Acinetobacter baumannii, Pseudomonas aeruginosa, and Enterobacter spp. (ESKAPE), is the ESKAPE pathogen most often cultured from anesthesiologists’ hands at the start and end of cases.²¹ Among 389 Enterococcal isolates from 274 pairs of successive cases from three hospitals (in Iowa, New Hampshire, and Massachusetts), 54% (210/389) were subsequently detected on future dates.²² Nevertheless, vancomycin-resistant Enterococcus isolates were less often associated with spread daily in the same room than sensitive isolates (P = 0.002).²² Also, transmitted isolates had no association with postoperative infections (P = 0.23).²² Similarly, few (6%; 84/1,448) gram-negative isolates, including few P. aeruginosa and Enterobacter spp. were detected on future dates, 0% of which were piperacillin-tazobactam resistant.²² None (0%) of those (few) transmission events were associated with postoperative infection.²²

Associations of infections pairwise by day between rooms was evaluated by performing 2,775 Fisher’s exact tests, where $\mathrm{2,\; 775}=75\times \left(75-1\right)/2$. We used Fisher’s exact tests because these data have many zeros (cf. “Results” section). For the multiple comparisons, we used Holm–Bonferroni adjustment as the P values were correlated from using the same rooms repeatedly.

We used the batch means method to calculate, for each of the 75 rooms, a 99% two-sided CI for the mean infections per week. The six periods (i.e., batches) each had 61 successive workdays, where $6\times 61=366$ (i.e., the dates studied). For each of the six periods, we calculated the mean among the 61 observations of infections per day. These were then multiplied by seven to have units of infections per week. We then used Student’s one-group t tests with five degrees of freedom to calculate the CIs. The rooms with the largest infections per week would be targeted for intervention; see the first paragraph of the manuscript and the second paragraph of the “Discussion” section.

The other approach to calculating the 99% two-sided CIs was to assume a Poisson distribution with small to moderate overdispersion (i.e., variance > mean). For each of the 75 rooms, we assessed Pearson goodness of fit to a Poisson distribution using the Chi square test, followed by Šidák correction for the 75 comparisons. We used Šidák correction because the tests were statistically independent, with each room used once. We also summarized among rooms the pairwise ratios of estimated variances to estimated means (all N = 366). We then repeated calculations using deviance goodness of fit tests.

We aimed to determine the mechanisms for underdispersion or overdispersion. Poisson distributions are based on probabilities of postoperative infection being statistically independent among cases. We used Fisher’s exact test for each room to evaluate whether one patient having a postoperative infection changed the probability of the next patient having an infection. Šidák correction was planned for the multiple comparisons. Another potential mechanism for failure to follow a Poisson distribution would be that the first cases of the day in each room had a different probability of infection. These were, again, evaluated with Fisher exact tests for each room, with Šidák correction. The focus on the first was suitable because most postoperative infections were among cases of relatively long duration but sufficiently brief that there was one other case in the operating room.⁷

Shi et al. had multiple simulations for estimating means of count data with overdispersion.¹⁵ Consider the heterogeneity-consistent sandwich estimator:

poisson Infections, vce(robust)

nlcom (mean_HC3: exp(_b[_cons])) ///

(lower: exp(_b[_cons] - 2.5828864 * _se[_cons])) ///

(upper: exp(_b[_cons] + 2.5828864 * _se[_cons]))The first line applies Poisson regression with intercept only, robust variance estimation. The second line calculates the confidence limits. Stata calculates the HC1 heteroscedasticity-consistent estimator. Shi et al. studied the HC3 estimator.¹⁵ For the intercept-only model, the HC3 estimate equals the HC1 estimate multiplied by $N/\left(N-1\right).$²³ The value of 2.5828864 equals the inverse of the standard normal for $1-0.01/2$, which is 2.5758, multiplied by $N/\left(N-1\right)$, which is 1.0027. Multiplication of each value by seven gives the results in units of infections per week. Shi et al. found that the heteroscedasticity-consistent estimator is unbiased for the Poisson distribution and the over-dispersed Poisson, with variance 2.00 × mean.¹⁵ The bootstrap variance estimator was also unbiased with accurate coverage,¹³ but was impractical for our use because some bootstrap samples randomly had all zero infections.

We considered statistical power analysis models to value above estimated percentage reductions in CI widths. Envision the identification of a room with a true Poisson mean rate λ ≥ 0.50 per week. A Type I (alpha) error would be falsely reporting a room because its sample mean λ³ 0.50 per week, despite its true mean rate < 0.50 infections per week. The consequence would be unnecessary costs for pathogen transmission monitoring and/or empirical use of extra infection prevention interventions (e.g., use of chlorhexidine wipes²⁴). We expect use of the typical Type I error rate of alpha = 0.05. A Type II (beta) error would be not reporting a room because its sample mean would be < 0.50 per week, although its true mean would be ≥ 0.50 per week. The cost of postoperative infections per room day is at least as large as the cost of transmission monitoring or intervention (i.e., beta ≥ alpha).^7,10 For simplicity, we set alpha = beta = 0.05. Algebra and suitability of study sample size are in the Electronic Supplementary Material eAppendix.

Results

Among the 42,978 cases in 75 rooms, the mean (standard deviation [SD]) operating or procedure room time was 2.53 (1.97) hr, with a median [interquartile range (IQR)] of 2.03 [1.18–3.32] hr. There were 24 distinct specialties. There were 2.55% of cases in which patients developed postoperative infection (1,095/42,978).

The room with the most cases (> 1,500) was the electroconvulsive therapy suite. There were zero postoperative surgical infections. The room with the fewest cases was a cardiology procedural suite (10). There also were zero postoperative infections. Both would be equally unimportant for anesthesiologists’ strategies to reduce postoperative infections. Nevertheless, while for the former that could be based on estimating the incidence proportion of postoperative infection ≈ 0%, for the latter it was because there were so few cases. The exact upper 99% confidence limit for the incidence proportion equaled 37%. The implications being the same, we analyzed daily counts of cases of patients who developed postoperative infections by combination of room and day: N = 366 observations per room.

The 75 rooms × 366 days equaled 27,450 room days, with 96.2% having no patients who developed an infection (26,401), 3.7% having one patient who developed an infection (1,005), 0.2% having two cases of patients who developed infections (42), and 0.01% having three cases of patients with postoperative infections (2). These were longitudinal data by room, with one daily observation each. Consequently, from these counts of infections per day, we can estimate CIs for the mean infections per week. As motivation for our work on reducing CI widths, we consider a room that had 313 days of zero infections, 49 days with one infection, and 4 days with two infections. The sample mean was 0.156 infections per day and the sample variance 0.154. The product of 7 and 0.156 equaled 1.09 per week. The point estimate from Poisson regression matched, as it was also 1.09 per week. Its 99% CI was 0.77 to 1.53 per week. An alternative approach not considered further would be to apply dichotomization and logistic regression. The point estimate was biased (1.19 per week), and the CI was wider (0.81 to 1.74 per week).

Continuing with Poisson regression, there was tiny or no serial correlation between days (Wooldridge test, P = 0.07; Pearson correlation, 0.027). There were no detected pairwise associations between rooms by day, with the smallest of the 2,775 Holm–Bonferroni-adjusted P = 0.17. Three data suggested the corresponding unadjusted P < 0.001 was a false discovery. First, neither of those two operating rooms had any days with > 1 patient having a postoperative infection (i.e., those rooms’ data were fully binary, zero infections or one infection). Second, the 5/366 days for which both rooms had patients with postoperative infections were days during which the specialties differed. Third, one room was in the adult inpatient surgical suite and the other at the pediatric surgical suite.

Counts of infections per room among the 366 days were consistent with Poisson distributions for 73/75 rooms, each with an unadjusted P > 0.01. Consistent with Poisson distributions’ properties of variances equal to means, the means (SDs) of the ratios of the sample variances to sample means were 1.00 (0.05), with a median [IQR] of 0.99 [0.97–1.00]. The other two rooms’ infections per day may not have followed Poisson distributions, based on the Pearson goodness of fit adjusted P ≤ 0.003. Nevertheless, all 75 rooms had deviance goodness of fit unadjusted P > 0.01. Still, investigating the two rooms, the ratios of sample variances to sample means were 1.32 and 1.78. The mechanism for over-dispersion was not excess zeros from no performed cases for many days, as both rooms had at least one case on every workday. The room with the largest estimated overdispersion had a mean (SD) of 3.58 (1.12) cases per day per workday while the room with the second largest had a mean (SD) of 2.29 (1.38) cases per day.

Matching Poisson assumptions, the preceding case having postoperative infection did not significantly increase the probability of the next case in the room having infection, all 75 Šidák-corrected P ≥ 0.32. Nevertheless, the one room with an unadjusted P < 0.01, specifically P = 0.005, had the largest estimated overdispersion. The first cases of the day did not have a greater probability of postoperative infection. Specifically, the 75 Šidák-adjusted Fisher exact tests P ≥ 0.94.

We used Poisson regression with heteroscedasticity-consistent robust variance estimation. The point estimates of weekly infections were identical to those of batch means. The 99% CIs were narrower using the Poisson regression, by a mean (SD) of 13.8% (33.2%), median [IQR] 20.3% [−2.3% to 38.7%]. There were 15 rooms with a mean of ≥ 0.50 infections per week. Among those rooms of particular importance, the 99% CIs were narrower using Poisson regression by a mean (SD) of 34.8% (28.8%), with a median [IQR] of 47.1% [12.5%–54.1%]. From equation (1) in the ESM eAppendix, a 34.8% reduction in CI widths would be equivalent to increasing the functional sample size 2.35-fold (e.g., getting 2.35 years of information from 1.00 years of data). Applying equations (2) and (3) from the ESM eAppendix, in units of costs, instead of alpha = beta =0.05, alpha = beta = 0.0058, a reduction ≈ 88.4%.

Tests of corollary prediction

Data by room followed Poisson distributions, as would the probability distribution of the total postoperative infections among all rooms, because the sum of Poisson variables with unequal means is a Poisson variable with a mean equal to the sum of the individual means. Nevertheless, the zero infection observations would now be conflated between days with no cases. Among the 252 workdays, each with many cases, a Poisson model was suitable (Chi square test, P = 0.66), with observed ratio of variance to mean of 0.96, and absence of serial correlation (r = −0.022; P = 0.73). Among the 114 weekends and holidays, with many fewer cases, a Poisson model was also suitable (Chi square test, P = 0.10), with a ratio of variance to mean of 1.17, and absence of significant serial correlation (r = 0.085, P = 0.37). Using a Poisson regression model with N = 366 independent days and the binary variable of weekend or holiday, or not, the estimated marginal mean (standard error) was 20.94 (0.63) postoperative health-care-associated infections per week. In contrast, the standard error estimated using batch means was 1.01 infections per week. The Poisson model’s smaller standard error functionally increased the sample size from 1 year to 2.60 years and reduced the costs of misclassification from alpha = beta = 0.05 to alpha = beta = 0.0040, a reduction of 92.0%. The Table presents the totals with rooms put into four categories of frequencies of postoperative infections.

Discussion

Last year, Shi et al. published simulation results reporting accurate CI coverage for rates (e.g., infections per week) using Poisson intercept-only regression with robust variance estimation.¹⁵ In the current article, we applied their findings.¹⁵ Our multiple test results (e.g., probability distribution fits) showed that postoperative infection data categorized by anesthetizing location matched their simulated conditions. The resulting CIs and standard errors for infections per week are sufficiently narrower to expect an 88–92% associated cost reduction. The savings are manifested either by more rapid detection of change in infections per week or by the prevention of unnecessary S. aureus transmission monitoring to evaluate the anesthesia workspaces’ contribution to, or lack thereof, the postoperative infections.^2,10–14

More precise estimates for finding anesthetizing locations with greater than threshold postoperative infections, or, equivalently, quicker detection of change, would not be useful if there were no interventions for anesthesiologists to make, but that is not so. For example, chlorhexidine-impregnated wipes used in the preoperative holding area significantly reduces bacterial contamination of the patient’s groin and axillary skin sites.²⁴ For example, after long anesthetics, syringe tips regularly have bacterial contamination.²⁵ Disinfection caps can be scrubbed to address such contamination because, with 10 sec of drying time, they are more effective against hyper-transmissible S. aureus than alcohol pads and non-scrubbing alcohol disinfection caps (both P ≤ 0.0087).²⁶ For example, even after routine surface disinfection, operating room environmental sites routinely have thousands of bacterial colony-forming units per surface area sampled, which anesthesiologists reduced >100-fold with 2 min of ultraviolet-C disinfection.²⁷ Take for example the previously considered problem of how many rooms can be disinfected nightly by an ultraviolet light disinfection robotic system.²⁸ Because the time for disinfection of each room has a large coefficient of variation,²⁹ the answer needs to be determined separately for each surgical suite.²⁸ The same applies if the ultraviolet disinfection system were ceiling mounted because the rooms for installation must be chosen.^30,31 Conceptually, the anesthetizing rooms that should be disinfected at least daily would be those with an upper confidence limit of infections per week exceeding some threshold (e.g., 0.25 per week). An effective approach to increase the value from each disinfection system would be to increase the precision of the estimates of infections per week such that the upper confidence limits are not sufficiently greater than the point estimates of the means to result in false detection of rooms as needing daily ultraviolet disinfection. This has been done.

Conceptually, our results could be limited by being the study of one large teaching hospital in Iowa. Nevertheless, as the analyses involved scores of operating rooms, our studying of only one hospital is unlikely to have affected our results for comparisons of confidence interval widths. In contrast, the study’s usefulness is limited to organizations that carry out interventions by room. Effective anesthesia machine and equipment disinfection between patients and during cases, hand hygiene by anesthesiologists, vascular care by anesthesiologists, and preoperative patient decolonization are actions chosen to be done by room and are ones that achieve substantial reduction in S. aureus transmission among anesthesia workspace reservoirs between successive patients in the room.^{5,6,12,21,22,24–27} If there were no pathogen transmission from patient to patient with a contribution or role of the anesthesia team,¹² then the monitoring would be unnecessary after that initial period of optimization. Put another way, if additional and deliberately placed hand hygiene dispensers are made available for anesthesiologists in specific rooms, then our paper would be useful.¹² In contrast, if chlorhexidine wipes²⁴ were used for every patient at a facility, comparing infection rates among rooms would be of no value. Finally, the current study’s scientific advance is in showing how to analyze infections per week categorized by room, by which we mean anesthetizing locations. Individual hospitals can have rooms’ infections per week reported using broad categories to emphasize the importance of anesthesiologists’ targeting of infection prevention (Table). Quantifying overall variability among rooms was addressed previously using the Gini index^7,8 because it permitted decomposition to understand the causes of variability; longer case durations are associated with greater risk of infection but fewer cases per day.⁷

Table.

Incremental risk ratios of infections per week among operating rooms and other anesthetizing locations in categories

Category of postoperative infections per week	Rooms, n/total N (%)	Marginal mean estimate	Standard error of the mean	Incremental risk ratio	95% Bonferroni-adjusted one-sided lower confidence limit
< 0.10	28/75 (37%)	0.044	0.006	(Base)
≥ 0.10 to < 0.25	12/75 (16%)	0.169	0.017	3.86	2.79
≥ 0.25 to < 0.50	19/75 (25%)	0.346	0.021	7.92	5.92
≥ 0.50	16/75 (21%)	0.695	0.032	15.89	12.01

Poisson regression was performed with robust clustered variance, the cluster being the day (i.e., N = 366 clusters). The one independent variable was the category in the first column of the table. The second column (Rooms) reports the raw counts of rooms’ cases that contribute to the category, but room is not a variable in the calculations. The Bonferroni-adjusted one-sided lower confidence limits are 98.83% intervals, where $98.83\%=100\cdot \left(1-0.05/3\right)$. The marginal mean estimate has units of postoperative health-care-associated infections per week (e.g., incremental risk ratio 15.89 = 0.694502/ 0.043716).

Supplementary Information

Below is the link to the electronic supplementary material.