Net probleem der waarnemingsfouten bij een massaal onderzoek

Errors of observation in a large‐scale statistical research. In an investigation carried out on behalf of one of the large department stores in the Netherlands, 15 different body measurement were taken on a sample consisting of 5000 women, with a view to establishing, on a sound statistical basis, a sizing system for ready‐made clothes. The present paper describes the error analysis carried out as a part of this investigation. The errors are classified into (1) clerical errors, (2) systematic errors, and (3) random errors. Clerical errors—that is ‘slips of the pen’ in recording, re‐writing or card punching—were succesfully traced by plotting scatter diagrams of highly correlated dimensions; compare fig. 4 where the cross indicates an outlying observation probably resulting from such an error. Out of 75000 observations 264 clerical errors have been detected; of which 233 could be corrected, while 31 were discarded, with their concomitant records, since the source of error could not be determined with sufficient accuracy. Systematic errors were traced by inter comparison of regressions in samples of 50 records of each of the 18 female students who performed the measurements. A significant difference found for one particular dimension and for one observer appeared to be due to a faulty method of observation, for which a correction could afterwards be applied. Lastly by a series of replications the random errors were determined for each of the 15 measurements taken. The standard deviations, β1‘sβ2’s, and correlations coefficients computed from the data, have all been corrected so as to eliminate the systematic errors resulting from these random errors in the observations. The formulae needed for this purpose were separately derived. In some cases these corrections turned out to be by no means trivial; their omission would have led to serious errors in the final conclusions.