Interprofessional collaboration and patient-reported outcomes: a secondary data analysis based on large scale survey data

The dataset was comprised of the entire patient and employee reported data that the Picker Institute Germany collected in 342 cross-sectional surveys between 2003 and 2016, which was the entire time period in which the Picker Institute conducted surveys in Germany. They were conducted on the occasion of quality improvement and monitoring processes in hospitals (i.e., the timing of the surveys was based on single occasions and the dataset does not represent systematic repeated cross-sections or panel data). For this purpose, validated employee (“Picker Employee Questionnaires”) as well as patient questionnaires (“Picker Inpatient Questionnaires”) were applied in different departments within hospitals. Good psychometric properties have been shown in several investigations [9‐13]. The paper-pencil questionnaires were introduced to employees respective patients along with up to two reminders according to the Total Design Method by Dillman [14]. Initial information on voluntariness in participation and pseudonymized analysis was given (see declarations for further information on ethics approval and consent to participate). Only those patients who did not respond within two (1st reminder) or 4 weeks (2nd reminder) received an additional reminder with a copy of the questionnaire and a return envelope. As employee surveys generally result in a lower response rate, particular importance was attached to ensuring a high level of commitment and willingness to participate in these surveys. For this reason, employee surveys were conducted anonymously, which meant that all those initially contacted also received all additional reminder letters. Employees who have already answered were advised to dispose the questionnaire in the reminder letter.

General information on the composition of Picker questionnaires can be found in Additional file 1. For our analyses, we used all questions assessing PROs as main dependent variables. Items assessing employee-rated IPC were included in Principal Component Analysis (PCA, see statistical analysis). The resulting principal components served as independent variables in further analyses. Demographic variables (gender, age, and education) were used as control variables for employees as well as patients. Items come with different response options, ranging from binary item formats to items with up to seven on a Likert scale (higher scores represent better IPC respective PRO). The specific items can be found in the Additional file 2.

All calculations have been performed using Stata MP 17.0. The level of significance was defined as 0.05. As some hospitals decided to survey their patients and employees in two consecutive years for organizational reasons (instead of conducting the two types of surveys simultaneously), data from hospitals that conducted both an employee and patient survey within a two-year period were analysed. In addition, the departments surveyed had to be identical. As we were able to create unique hospital-department-codes, consisting of a unique hospital- plus a unique code for a given department, irrespective of questionnaire type (patient or employee questionnaire) or survey year, we were able to match the respective department-level IPC score to the individual-level patient data.

A PCA on items assessing IPC from the employee’s point of view was used to derive main components underlying these scales. The overview of all items included in PCA can be found, along with the prespecified PCA model, in Additional file 3. Items were identified as suitable for main components if the Kaiser-Meyer-Olkin (KMO) measure of sampling adequacy coefficient was > 0.50, and the Bartlett’s test of sphericity significant (i.e., < 0.05). Only factors with eigenvalues ≥1 were considered (Guttmann-Kaiser Criterion) and results were rotated using both promax and varimax rotation [15‐17] in order to check the sensitivity of results to these alternative interpretations. Reassuringly, the results mirrored the unrotated principal solution. Thus, we rely on the unrotated solution to predict standardized component scores for the employees and subsequently aggregate these at the department-year-level (i.e., the department-year-IPC-score for each component was the mean of the employee component scores for each component).

These mean component scores were z-standardized to allow the statement that an increase of 1 SD on IPC score at department-year-level was associated with a change of β₁ (OR) in PRO outcomes at the individual-patient-level. The unit of analysis was individual patient data from departments within hospitals which have surveyed employees from the identical departments in the same year, the year before or after.

Due to the ordinal scaling of the dependent variables (i.e., PROs), the regression model was implemented as an ordered logistic regression model, such that estimated coefficients (logged odds) are subsequently transformed and interpreted as odds ratios (OR). The formal model takes the general form:where y_icj denotes PRO of patient i treated in department c. As the outcome variable y_j consists of ordered outcome categories, the score was estimated as a linear function of the independent variables (department-level IPC and department- and individual-level covariates) and cutpoints defined by the number of categories of the dependent variable. The probability of observing outcome l was then defined as the probability of the function and error term being within the range of the estimated cutpoints. β₁ was the coefficient of interest denoting the change in the probability of outcomes in response to a one-unit increase in the department-year-level IPC score (IPC_c). X_kij and X_gcj denote vectors of individual and cluster-level control variables, respectively. θ_h denotes hospital fixed effects and ε_icj denotes the random error term. Standard errors are cluster-robust at the department-year-level. Additionally, we adjust the reported p-values for the false discovery rate (FDR) using Andersons’ q values for multiple hypothesis testing, and examine of the association between patient- and employee-rated IPC to assure validity of the IPC measure.

In order to answer the secondary research questions, exploratory subgroup analyses were undertaken afterwards. Results were adjusted for multiple hypothesis testing based on the false discovery rate [18].

In total, 43 Hospitals including 519 departments and 88,243 individual-level observations (25,470 employees and 62,773 patients) surveyed both employees and patients within 2 years in the period from 2003 to 2016 (n = 132 surveys). After PCA, matching of employee-based department-level IPC scores to patient-level PRO data, and limitation to those patients for whom demographic (control) variables were available (see below) there were 6154 patient observations (unit of analysis) from 19 hospitals respective 103 departments and 49 surveys included in the analyses. Table 1 presents employees’ and patients’ descriptive statistics. Survey characteristics, including information regarding the specific survey years, department types and number of surveyed patients and employees from each included hospital can be found in Additional file 4.

The PCA revealed three principal components, explaining 67% of the total variance:

Both promax and varimax rotation revealed the same composition of components. Thus, we rely on the original (unrotated) unrotated principal solution to predict the component scores. The KMO measure of sampling adequacy was .830 and Bartlett’s test of sphericity significant (p < 0.001). The final set of IPC items can be found in Additional file 2.

All items assessing PROs from patients which were included in the final data set were used for regression analyses (see Additional file 2 for main dependent variables).

When neither controlling for employees’ nor patients’ demographic characteristics (gender, age, education) a statistically significant association was found between the PRO complications and department-specific IPC, meaning that an increase of 1 SD in the first IPC component leads to a statistically significant chance of a higher (i.e., better) PRO-rating regarding the question “Did complications occur after your/ your child’s discharge from hospital?” (OR 1.06, 95% CI 1.00–1.11, p = 0.046). This result was consistent after adding control variables to the ordered logistic regression model (OR 1.07, 95% CI 1.00–1.13, p = 0.029) (see Table 2).

No statistically significant results between IPC components and PROs emerge neither with nor without adding control variables in the remaining analyses (see Table 2).

To account for the problem of multiple inference (i.e., 15 tests due to 5 outcomes and three IPC variables), we also report sharpened q-values for the adjusted regression estimates based on the method developed by Benjamini et al. [19], i.e., the expected proportion of false rejections of the null hypothesis [18]. This approach allows a straightforward comparison with the original p-values and has more power than Bonferroni correction: With a large number of tests, Bonferroni correction (which controls for the familywise error rate, i.e.., the probability of any false rejections of the null hypothesis), may be considered too conservative. In our case, both sharpened q-values and Bonferroni corrected p-values (i.e., p-values multiplied by the number of tests) are severely inflated, indicating no significant relationship between IPC and PROs in any of our analyses when accounting for multiple inference.

Focusing on the association between patient- and employee-rated IPC, we estimated OR (using ordered logistic regression) regarding the question “How well does the communication and coordination between doctors and nurses work?” which was asked of patients as well as employees. Here, a statistically significant result (OR 1.10, 95% CI 1.02–1.18, p = 0.011) was found (n = 5345 patients), meaning that patients as well as employees of this sample have comparable ratings of IPC.

Regarding exploratory analysis, we decided to focus on heterogenous associations between IPC and PRO by department-type. Given the fact, that there are 13 department types left in the final data set, we decided to analyze the three departments with the largest sample size. As a result, we analyzed “Internal medicine” (n = 2100), “Surgery” (n = 1698), and “Pediatrics” (n = 442). As stated in the pre-analysis plan (see above), results generated in these exploratory analyses are intended to be viewed as hypothesis generating rather than hypotheses testing. Nevertheless, we investigated Andersons’ q values to account for the multitude of analyses.

Results are presented in the Additional file 5. Significant Andersons’ q values confirm the significance of all p values emerged in these analyses. Statistically significant results for “Surgery” and “Internal medicine” are found only regarding one single IPC-component (department-specific IPC). In accordance with the above-described results, an increase of 1 SD in department-specific IPC leads to higher odds for a better PRO-rating of patients which were treated within these departments, not only in relation to the PRO complications, but also to the other investigated PROs.

To the contrary, results for pediatric patients are more ambiguous. Whereas the chance of a better PRO-rating regarding complications was higher with an increase in department-specific IPC (OR 1.15, 95% CI 1.07–1.23, p < 0.001), it was lower with an increase in interprofessional organization (second IPC component) (OR 0.59, 95% CI 0.44–0.79, p < 0.001). The opposite picture arises when looking at the PROs overall satisfaction, less discomforts, and treatment success. Regarding the third IPC component (overall IPC) an increase of 1 SD was associated with a worse PRO-rating in all investigated outcomes.