Statistical analyses

For the inclusion criteria, descriptive statistics were computed based on the total number of questionnaires distributed and, for the investigated variables, based on the questionnaires which met the inclusion criteria. Percentages were computed for dichotomous variables, and means, standard deviations, minima and maxima for continuous variables. This was undertaken for both the total sample and separately for the six region-specific samples. Regional differences regarding dichotomous variables were statistically tested using chi-square tests, and regional differences regarding continuous variables using Kruskal-Wallis tests.

The relationship of travel distance, travel time, in-practice waiting time and quality of communication with visits to the primary diabetes care provider was investigated using log-linear Poisson regression models [26] with frequency of visits as the criterion. Poisson regression models are specially designed for predicting count variables as for example the frequency of visits. The de-logarithmised coefficients estimated by these models reflect the proportion with which the predicted variable increases or, respectively, decreases when the corresponding predictor variable increases by one unit.

All focal predictor variables and all control variables were applied as predictor variables. All focal predictors were entered together into the same model as they were considered to reflect different aspects. This also applies to travel distance and travel time because travel time is not solely determined by travel distance, but also by factors such as the method of transport available to the patient and road congestion. Hence, both variables might have an independent effect on the visits to the diabetes care provider. Dummy-coded study regions were included to control for region-specific tendencies regarding provider visits. The status of diabetes and the five secondary complications of diabetes were included to control for the effect of health status because health status usually affects the number of health care provider visits [2–4, 8, 9, 15]. The three socio-demographic variables (age, gender and education) were included because these variables are known to be related to health status [27–30] and thus might help to control for those aspects of health status that are not reflected by diabetes status and secondary complications. Taken together, the demographic variables and the variables directly addressing health status can be regarded as variables covering the need for visits to the primary diabetes care provider.

The coefficients for all predictor variables were estimated and tested for deviation from zero. Fit of the models was investigated by comparing the complete model via a likelihood ratio test with only the criterion mean as predictor and by using the index proposed by Cameron and Windmeijer [31] (CW-index).¹ The CW-index serves a similar purpose as the multiple R square in multivariate linear regression. This index is 0 when the model predicts the criterion as poorly as the criterion mean, and it is 1 when the model predicts the criterion optimally. The extent to which a specific predictor variable contributes to predicting visit frequency was assessed by subtracting the CW-index for a model without this variable from the CW-index for the complete model. To obtain a statistic that can be interpreted analogously to the percentage of variance exclusively explained by a specific predictor variable in a multivariate regression model, the resulting difference was multiplied by 100. In contrast to the coefficient belonging to a predictor variable, the statistic just described is independent of the unit with which the predictor variable is measured. Therefore, by means of this statistic, the contributions of different predictor variables can be compared.

The correlations between the focal predictors were computed and, to investigate the influence of these correlations on the results, four additional models were computed in each of which a different focal predictor was omitted.

To investigate the stability of the results across the six study regions, the interactions between the focal predictor variables and the study regions were analysed. The statistical significance for the interactions with a specific focal predictor was tested via a likelihood ratio test by comparing a model containing all predictor variables and all investigated interactions with a model from which the interaction terms for the respective focal predictor variable had been removed. In addition, the contribution of the interaction for explaining the visit frequency was investigated using CW-indices in an analogous manner as described above for the main effects. To investigate the extent to which possible interactions can be attributed to the countries’ health care systems, analogue computations were performed after grouping the study regions according to the Reibling categories assigned to the corresponding country [23, 24] (see above).

To check the generalisability of the results, all analyses were performed in two variants: 1) for all included participants (i.e. persons who had up to two missing data items for the control variables), and 2) for participants with complete data for all investigated variables. For the analysis referring to participants with missing values, the missing values were imputed by the study region-specific means. As data possessing less than interval scale level were dummy coded, the study region-specific relative frequencies of persons belonging to the respective category were taken for imputing missing values of these data.