The effects of publishing emergency department wait time on patient utilization patterns in a community with two emergency department sites: a retrospective, quasi-experiment design

The study setting is a middle-sized community with a population of half a million with two adult ED sites in southwest Ontario, Canada. These two sites are: University Hospital (UH) and Victoria Hospital (VH). The two sites are 8.5 km apart (travel distance by road), and in normal traffic conditions it takes approximately 15 min to travel from one site to another by car, and about 30 min by public transportation. The two sites are staffed by the same group of physicians.

Starting from 19 February 2009, the average wait time information of these two sites has been available to the public (updated daily) via a website [8]. The number of "hits" to the website ranged from 18 to 39 per day during the month of June 2010 (historical data for number of "hits" per day are not available), suggesting that around 10% to 20% of the visitors were using this website.

Administrative ED records for these two hospitals were used in the data analysis. Such ED records include demographics (age, gender, postal code) for the patients and timing (time of triage, admission, initial physician assessment, and discharge) and clinical information (triage level, main reason for the visit, discharge) of the visits. We used ED records during the period from 1 August 2008 to 31 August 2009, excluding the transition period of February 2009. The period between 1 August 2008 and 31 January 2009 will be referred to as the "before" period hereafter, and the period between 1 March 2009 and 31 August 2009 the "after" period. A dummy variable, "period," was created to represent the period (0 as the "before" period and 1 "after").

Patients aged 19 and above who self-arrived at the ED (i.e., not by ambulance), whose triage levels were not "Resuscitation" or "Emergency," who provided a valid residential postal code, and who resided within 50 km from either site were included in the analysis, as these patients are the most likely to use the published wait time information to select which site to visit. Those who resided more than 50 km away were excluded as they were most likely visiting the region rather than traveling from their home to visit the ED, making their travel distances impossible to calculate. Follow-up visits were also excluded as such visits did not involve a choice of site. Independent visits from the same patient were treated as separate visits as the same patient may visit different sites at different times. Please note that ability to pay was not a factor because of Canada's universal healthcare system.

Patient age was grouped into three categories: 19-40, 41-60, and 61 and older. Age was not treated as a continuous variable as we believe that age is likely to have a non-linear effect [20]. We chose 40 and 60 as the two cutoff points as we believed that people below and above these two ages may have important differences in their reactions to information on the Internet. Travel distances between the patients' resident addresses and the two ED sites were calculated using ArcGIS 9.3.1, a Geographical Information Systems (GIS) software product from Environmental Systems Research Institute, Inc. (ESRI). The latest Ontario road map files were used, with longitude and latitude values of the patients' postal codes obtained from Statistics Canada Post Code Conversion File (PCCF) [21]. The accuracy of PCCF in measuring distances between two postal codes was validated elsewhere [22]. Differences in travel distances between the patient's residence and the two ED sites were grouped into two categories: UH closer and VH closer. Another patient level variable, gender, was also included in the model.

Time of the day was grouped into three categories based on triage time: daytime (8 a.m.-4 p.m.), early evening (4 p.m.-12 a.m.), and midnight (12 a.m.-8 a.m.). Such categorization corresponds to the time shifts of the ED sites, with changes in staffing and other resources. Days of the week were grouped into two categories: weekdays (Monday through Friday) and weekends (Saturday and Sunday) based on triage time.

Main reason for visit was categorized into five categories using ICD 10 codes: "Mental and behavioral disorders" (ICD codes starting with F), "Pregnancy, childbirth, and the puerperium" (ICD codes starting with O), "Diseases of the circulatory system" (ICD codes starting with I), "Injury, poisoning, or external causes of morbidity and mortality" (ICD codes starting with S, T, V, or Y), and other reasons.

Level of emergency was determined at triage (triage level). There are five possible levels: Resuscitation, Emergency, Urgent, Less Urgent, and Non Urgent. As mentioned above, those with triage levels Resuscitation or Emergency were excluded.

Wait time was defined as the time between registration and discharge ("door to door"). Another measure of ED wait time, namely the time between registration and initial assessment by a physician ("door to doctor"), was also used in the literature; however, since such wait time was not published in the website, we did not use it in the analysis. Differences in wait times were grouped into three categories: No difference; UH shorter; VH shorter.

Rate of wait time exceeding 4 h and the mean, standard deviation, and 95th percentile of wait times in the two sites were used as indicators for the likelihood of lengthy wait times in the descriptive analysis. The patients' selection of ED sites was used as the outcome variable in the logistic regression. Independent variables include the patients' demographic and clinical variables, the characteristics of the timing of the visit, and the distances between the patients' residences to the two ED sites.

Univariate analyses were performed on all covariates, and those with a p value less than 0.10 were included in the multivariate model. An interaction variable between the difference in wait time and the dummy variable "period" was included in the multivariate model to test the hypothesis that the publication of wait time information led to more patients selecting the site with shorter wait time. Given the likely lack of independence among the visits of the same patient, we used the Huber-White method to correct for heteroscedasticity and for correlated responses from cluster samples [23].

Data analyses were carried out using R 2.11.0 [24].