Multiple imputation Sample Clauses

Multiple imputation. Imputation is a method to estimate data where it is missing using information gleaned from other, complete cases in the dataset (Xxxxx, 1987). Multiple imputation describes when this process is repeated to generate multiple possible datasets by introduction of an element of randomness in generating dummy values in place of missing values. Multiple imputation is preferred to mean or median value imputation, because mean and median imputation decrease the variance in the dataset. In the MND Register dataset, there were many cases of missingness of date of diagnosis. Although date of onset of symptoms could arguably be used instead, due to lag times between onset and diagnosis, it is possible we would miss new cases. Instead, we decided to impute diagnostic delay from the other values in the sample and add that to date of onset to estimate date of diagnosis. This was repeated 20 times, as there were 20% of date of diagnosis values missing (Xxxxxx, 2008; Xxxxx, Xxxxxxx, & Xxxx, 2011). Predictive mean matching was used to estimate diagnostic delay values. Predictive mean matching is where a regression is used to estimate values for a particular record and then records with full data that closely match the predicted value are used to pick an imputed value from (Xxxxx et al., 2011). Pooled parameter and standard error estimates are calculated using Xxxxx’x rules (Xxxxx, 1987). To calculate the pooled parameter estimate, the mean of all imputed incidence calculations were taken (Xxxxxx, 2010). To calculate the pooled standard error, the within imputation variance (squared standard error of each dataset divided by the number of datasets) and the between imputation variance (variance of each imputed incidence calculation) are combined (Xxxxxx, 2010). The variance derived from these estimates are not directly comparable to historical estimates but confidence intervals from unimputed data were provided as well as pooled incidence estimates to give an indication of data spread. Multiple imputation was performed in R, using the ‘mice’ package and incidence estimates were calculated in MS Excel (Xxxxxxxxx-Xxxxxxxxx, 2011; Team, 2018).

Multiple imputation. To estimate date and age of diagnosis for those records with date of onset only we used predictive mean matching to generate 22 datasets (as 22% of cases were missing diagnostic delay) and calculated standardised incidence for all datasets (Xxxxxx, 2008; Xxxxx et al., 2011). The model for predictive mean matching included data collection centre, age of onset, gender, diagnosis subtype, site of onset and family history. We calculated pooled estimates of incidence and 95% confidence intervals using Xxxxx’x rules (Xxxxx, 1987). Imputation datasets were generated using the R package ‘mice’. (Xxxxxxxxx-Xxxxxxxxx, 2011; Team, 2018).