Statistical analysis

Descriptive statistics about the presence and components of the discharge letter were analysed using Stata 13.0 [12]. The relation between the presence and quality of the discharge letter and admission and patient characteristics was analysed with logistic multilevel regression analysis. Multilevel analysis was used because the data had a hierarchical structure: patients (level 1) were clustered within hospital departments (level 2) and hospital departments were clustered within hospitals (level 3) [13]. The 2^nd order PQL estimation procedure was used.

The presence of the discharge letter and the presence of discharge letter components were set as dichotomous (no/yes) outcome variables. For each component, separate regression analyses were run. The independent admission and patient characteristics were age, discharge status (deceased or discharged), admission urgency (elective/urgent), unplanned readmission (no/yes), one or more positive screening criteria for adverse events (no/yes), adverse event during admission (no/yes), preventable adverse event during admission (no/yes), and hospital type (university/tertiary teaching/general).

The variances of the model were tested for statistical significance using a one-sided Wald test.

ICCs were calculated for the hospitals and departments. The ICC indicates the relative influence of that level on the total variance of the outcome in a year. A high ICC at the hospital level means that there is less heterogeneity within hospitals and a high variation in the presence of the discharge letter between hospitals [13]. The variance at the patient level, the lowest level, is approximated by π23 to calculate the total variance [13].

The multilevel analyses were carried out using MLwiN 2.30 and the application “runmlwin” in Stata 13.0 [14, 15].