140 Chapter 6 it would be interesting to combine repeated measurements of baseline characteristics with measurement of adherence. Combined with investigating the patient-therapist interactions during treatment sessions and their effects on patient adherence, this can help to further understand patient adherence. Strengths and limitations Our study has a number of strengths. The application of multiple imputation by chained equations for missing data helped to reduce bias introduced by missing data. Since a large number of cases had at least 1 missing data point due to illegible reporting by the therapist, no reporting by the therapist or other problems not related to the patient, performing a complete case analysis would have significantly reduced the number of cases available for the analysis. Imputing missing data allowed optimal use of the available data and therefore provide more robust results. Another strength of the study is the use of the EXAS for the measurement of adherence during every treatment session. The detailed information on patient adherence provided by the EXAS allowed the use of LCGA to determine different groups of patients with distinct adherence trajectories. Limitations of the study should also be discussed. The first limitation is the introduction of missing data through the way data on adherence was collected. To keep the added workload for the physical therapists participating in the study low, we chose a method that allowed the physical therapists to write down the data on a form they could keep on their desk. Although this methodology requires little effort from the therapist, it introduced more room for errors in reporting (illegible handwriting, forgetting to complete part of the form, etc.) than for example digital reporting through a web–based application. Although imputation was used to minimize the effects of missing data on the results, the best way to handle missing data is to prevent it. A second limitation is that there are currently no existing rules or conventions for the pooling of estimates from LCGA on imputed datasets. Imputation of missing data and analysis of the imputed data generally consists of 3 steps (37). First, a number of different datasets with imputed data are created. Then, the parameters of interest are estimated from each imputed dataset. The last step is the pooling of the parameter estimates and estimating the variance of the pooled estimate. Although the mice package from R provides the tools to pool estimates for linear models, these tools are not available for LCGA in the mice package. Although manually pooling and estimating the parameters of interest would have been possible, similar procedures for the Kruskal-Wallis test used to compare baseline characteristics of the identified trajectory groups do not exist. Instead, we decided to calculate the average of all variables with imputed data over all imputed datasets and find the dataset with the smallest mean deviation from the overall mean to perform the analyses on. This allows the use of imputation to maximize the data available for analysis at the cost of precision
RkJQdWJsaXNoZXIy MTk4NDMw