PMC:2846953 / 13251-14865 JSONTXT

{"target":"http://pubannotation.org/docs/sourcedb/PMC/sourceid/2846953","sourcedb":"PMC","sourceid":"2846953","source_url":"https://www.ncbi.nlm.nih.gov/pmc/2846953","text":"Handling of missing data and BLLOQ values\nIn general, missing data were present in the data set, but rare. In the data mining analysis, it was deemed prudent to impute missing data. Missing continuous variables were imputed in MARS and Random Forest Analyses by replacing missing observations with the median of non-missing cases for each variable: this was done separately for smokers and non-smokers, and for validation and analysis datasets. Imputation was carried out on predictor variables only, cases with missing dependant variables were dropped, not imputed, in the MARS and Random Forest analyses.\nFinal regression models were fit to data sets in which no imputation was used on any variable, i.e. cases with missing dependant or predictor variables were dropped from the analysis and the number of cases dropped was noted in the regression output. In general, final regression results seemed little changed between imputed and unimputed datasets.\nFor categorical variables, missing data were handled by adding an additional category as indicated in the following section. This approach does not depend on any assumptions about the sources of missing data and we do not regard it as a form of imputation.\nBelow the lower limit of quantification (BLLOQ) flags were present in the data set for 9 variables of exposure (NECEQ, MHBMA, TOTNN, DHBMA, OHP, HPMA3, COTIN, COHB, and 4-ABP). If a BLLOQ flag is positive for any of these variables and if the variable itself has a missing value, the value of the affected variable is set to 0. Non-missing value variables BLLOQ values with flags were left unchanged.","divisions":[{"label":"title","span":{"begin":0,"end":41}},{"label":"p","span":{"begin":42,"end":606}},{"label":"p","span":{"begin":607,"end":956}},{"label":"p","span":{"begin":957,"end":1213}}],"tracks":[]}