Design and settings
This exploratory study was conducted at 4 hospitals in the Midwestern United States and used archived hospital data and reports. Twenty-eight adult medical, surgical, and medical-surgical inpatient acute care units provided the data. Due to the difference across study hospitals in backing up archived data, the covered data periods varied between hospitals.
The study hospitals included the following: Hospital 1, academic medical center, bed size about 900, 14 participating units, data from January 2004 to December 2008; Hospital 2, community hospital, bed size about 300, 4 participating units, data from February 2007 to December 2008; Hospital 3, teaching hospital, bed size about 900, 4 participating units, data from April 2008 to May 2009; and Hospital 4, teaching hospital, bed size about 700, 6 participating units, data from January 2006 to December 2008. Data were included from different time periods to increase the sample size.
The unit of analysis was the patient care unit-month (abbreviated as unit-month) defined as data aggregated by month for each patient care unit. Some interdependence for the data points from a single unit and for the data points from other units in the same hospital existed. For statistical analyses and result interpretation, each data point for a study unit was assumed to be independent from all others. The study was approved by each hospital's institutional review board and the corresponding author's employer university. There was no conflict of interest.
Data sources and collection
In each study hospital, a designated site coordinator (a hospital staff or administrator) retrieved the archived hospital data and facilitated chart reviews. Each site coordinator was instructed by the corresponding author about the desired hospital data to be used to ensure the consistency and reliability of the data across the 4 study hospitals. Under the corresponding author's supervision, the retrieved data were entered by a trained research assistant and verified by another trained research assistant for accuracy. Detailed information about the study variables are described in Table .
Study variables and definitions
The 2 dependent variables were the fall rate and the injurious fall rate. The fall rate was defined as the rate at which patients fell during their hospital stays/1000 patient-days [15
]. A fall was defined as an unplanned descent to the floor with or without injury. All falls types were included, whether falls resulted from physiologic or environmental causes [28
]. The operational definition of the fall rate was (number of total falls × 1000)/(total patient-days). The injurious fall rate was defined as the fall rate/1000 inpatient-days during which physical injury occurred, regardless of severity (including minor, moderate, major injury and death) [15
]. The operational definition of the injurious fall rate was (number of injury falls × 1000)/(total patient-days).
The predictor was the average response time to call lights. These data were retrieved from the call light tracking system at each hospital. Patient/family-initiated calls made from the pillow speaker or call cord were categorized as normal calls but calls initiated in the bathrooms were not included in the analysis. The response time was defined as the time that elapsed between a normal call activation to its cancellation from the patient room. The response times for "staff response" on the reports generated from the call light tracking system were aggregated at the unit level for each month and calculated by (call light response time in seconds for all the calls made for the unit and month)/(total number of calls for the unit and month). The operational definition of this variable was (sum of the call light response time for the calls in seconds)/(total call light use).
As for covariates, the data on the percentages of patients with altered mental status and hearing problems came from chart review. Due to constrained resources, one data point by quarter for each patient care unit was collected. The percentages of patients hospitalized at the study unit on the 15th of the first month of each quarter, who had cognitive impairment or altered mental status at admission, were calculated. As for the chart review procedure, the charts of 10 randomly sampled patients per study unit were reviewed by a trained research assistant. If altered mental status was identified at admission in the chart, the patient was coded as Yes (1); otherwise, No (0) was coded (Table ).
For each study hospital, the patient management database was used to generate the total patient-days per unit-month. The daily count of total patient-days was the midnight census. The daily counts for a unit for a specified month were added up to indicate the total patient-days for that unit and month. The designated site coordinators calculated this variable (the total patient-days per unit-month) before sending the data to the corresponding author. Total patient-days per unit-month were used to compute the call light use rate per patient-day and fall and injurious fall rates.
Data were entered into the Statistical Package for the Social Sciences (SPSS; 18.0 Window version; SPSS Inc., Chicago, IL, USA). All data points were matched by patient care unit as well as by year and month. Only unit-month data with valid fall rate and injurious fall rate data were included in the analysis.
In the course of data management, means and standard deviations were calculated for the continuous variables, and the skewness and kurtosis values of these variables were examined. The call light use rate per patient-day (skewness value = 11.57; kurtosis value = 235.82) and the call light response time (skewness value = 11.00; kurtosis value = 137.61) had high skewness and kurtosis values. The log transformation was done on both variables, but the log transformation left the distributions still skewed. As a result, the continuous variable of the patient call light use rate per patient-day was recorded to fall within 1 of 10 groups: 10 = low to 0.95; 20 = more than 0.95 and up to 3.52; 30 = more than 3.52 and up to 4.66; 40 = more than 4.66 and up to 5.83; 50 = more than 5.83 and up to 6.91; 60 = more than 6.91 and up to 7.56; 70 = more than 7.56 and up to 8.12; 80 = more than 8.12 and up to 8.65; 90 = more than 8.65 and up to 9.65; and 100 = more than 9.65. In addition, the continuous variable of the call light response time (in seconds) was recorded to fall within 1 of 10 groups: 10 = low to 128.10; 20 = more than 128.10 and up to 153.00; 30 = more than 153.00 and up to 167.00; 40 = more than 167.00 and up to 179.40; 50 = more than 179.40 and up to 193.00; 60 = more than 193.00 and up to 207.00; 70 = more than 207.00 and up to 221.00; 80 = more than 221.00 and up to 241.00; 90 = more than 8241.00 and up to 730.40; and 100 = more than 730.40. These 2 recoded variables were used to test the hypotheses.
SPSS was also used for data analyses. The sample (1063 unit-months) was the total number of months with available data for each patient care unit. One-way analysis of variance (ANOVA) tests were conducted to elucidate differences in the study variable means across the 4 study hospitals and 3 unit types. Separate hierarchical multiple regression analyses were used to test the 2 hypotheses. Hierarchical regression is also called sequential regression; predictors are entered into the equation in the order specified by the researcher. Predictors are entered in steps or blocks with each predictor or a set of predictors being assess in terms of what it/they add(s) to the prediction of the dependent variable, after the previous variables have been controlled for [29
Missing values for the covariates and predictor were replaced by mean values because data were missing at random. Before entering the categorical covariates into the regression models, 2 sets of dummy variables were created to capture 4 hospitals and 3 unit types. Collinearity among the predictor and covariates was a possible concern and was checked [29
]; we found that it is not to be a problem. All predictors were included in the analyses.
The covariates were entered into the multiple regression equation first. Then, the average call light response time was entered as a predictor into each model. Key outcomes of the analyses were the significance tests and estimates of regression coefficients for the average call light response time in the final regression models. Alpha was set at 0.05 for the analyses.
The power analysis was used to compute the required sample size. For the power analysis for the multiple linear regression analysis, making the assumption of including up to 13 predictor variables to explain a medium-sized squared multiple correlation (R2
= 0.13) with alpha of 0.05 (2-tailed) and desired statistical power of 0.80 requires a sample size of 149. The total sample size of 1063 unit-months provided more than 99% power; that is, it was more than sufficient. Thus, power was fully adequate for the proposed project [30