Marieke van Rosmalen

Chapter 4 60 Statistical analysis IBM SPSS Statistics (Version 25, Chicago, Illinois, United States) was used for statistical analysis. To compare patient characteristics between cases and controls, we used one-way analysis of variance (ANOVA) for numerical data and Chi-squared test for categorical data. To evaluate the feasibility of our method we compared numbers of successfully performed measurements between the coronal and sagittal plane and between G 0 and G 1 using an independent samples t test. To determine mean nerve root size we also used an independent samples t test. Results with a p value < 0.05 were considered significant. To evaluate intra- and interrater reliability we used the intraclass correlation coefficient (ICC). We calculated a mean ICC of the right and left side per measurement site. We considered an ICC < 0.50 as poor reliability, 0.50 – 0.75 as moderate, 0.75 – 0.90 as good and > 0.90 as excellent reliability. 17 ROC analysis and development of risk chart We used receiver operating characteristic (ROC) analysis to determine area under the curve (AUC) per nerve root (C5, C6, C7) and for two different combinations of measurement: 1. mean of all three nerve roots bilaterally next to the ganglion (3 variables) and 2. mean of all three nerve roots 1 cm distal from the ganglion (3 variables). We then used a multivariate binary logistic model for both combinations separately with measurement sites as covariates. With the results of this model we calculated the log odds for having an inflammatory neuropathy using the following equation (Eq. 1): log ! " $ ! ! " = % + " 5 + & 6 + ' 7 (Eq. 1) Where β 0 is the constant, b 1 , b 2 and b 3 the logistic regression coefficients of nerve roots C5, C6 and C7 respectively and C5, C6 and C7 the diameters of the nerve roots in millimetres. Subsequently, we took the inverse logit to obtain p, i.e. the absolute probability of having an inflammatory neuropathy, using the following equation (Eq. 2): = 1 1 + !(# ! % # " &' % # # &( % # $ &)) (Eq. 2) To develop a risk chart, we calculated p for different combinations of C5, C6 and C7 and for both combinations of measurement sites. Finally, we obtained a cut-off value for p obtaining 95% specificity, i.e. we determined at which p we considered MRI to be abnormal.