UNECE Conference Work Session on Statistical Data Editing

1. An R package for selective editing based on a latent class model M.T. Buglielli, M. Di Zio, U. Guarnera and F. R. Pogelli Istat – Italy UNECE Conference Work Session on Statistical Data Editing Ljubljana, Slovenia, 9-11 May 2011

2. Introduction UNECE Work Session on Statistical Data Editing Selective editing looks for units affected by important errors in order to limit accurate reviewing. Error quantification - Observations are prioritised according to the values of a score function that expresses the impact of their potential error on the estimates of interest. Accuracy level - Units above a given threshold are selected since they potentially represent the observations affected by important errors. Ljubljana, Slovenia, 9-11 May 2011

3. Problems UNECE Work Session on Statistical Data Editing The score function is generally based on the difference between observed and “anticipated” values. The problem is that differences are due to both errors and to the natural variability of the phenomenon. Score values cannot be interpreted as a direct evaluation of the accuracy of estimates. Without historical (true and contaminated) information it is not possible to select the most influential units such that a prefixed level of accuracy for the target estimates is attained. Ljubljana, Slovenia, 9-11 May 2011

4. A latent model approach – The contamination model UNECE Work Session on Statistical Data Editing • The use of a latent model for true data and errors, allows to • distinguish the error and the variability component of the residuals • the score value of an observation is directly interpreted as the expected error of the units. • The method estimates the probability of being in error and the • error impact, that suitably combined determine the • conditional expected error • In this framework we can select units by estimating the expected • error left in data once they are restored (also without hist info) Ljubljana, Slovenia, 9-11 May 2011

5. SeleMix a software for selective editing UNECE Work Session on Statistical Data Editing • SeleMix is a package in R for the selection of influential errors according to the contamination model. • Implements the ECM-algorithm developed to estimate model parameters • computes local and global scores • returns the set of observations affected by influential errors with respect to a certain prefixed level of accuracy of the target estimates. • Moreover, it provides anticipated values (predictions) for each unit for both observed and non observed variables. The imputation can be considered “robust” in that the model used to compute the “anticipated” values takes into account the presence of errors in data. Ljubljana, Slovenia, 9-11 May 2011

6. SeleMix functions UNECE Work Session on Statistical Data Editing The package is composed of three functions ml.est, pred.y, sel.edit. ml.est - This function estimates the parameters of the model by using an ECM-algorithm suitably developed. The output is a list of: model parameters, anticipated values, BIC and AIC scores, outlier flags, and posterior probabilities pred.y - makes a prediction of the true values for the variables Y through their expected value conditional on all the available information. It returns, for each unit, a "prediction" for both observed and missing items of each Y variable, the outlier flag and the posterior probability. Ljubljana, Slovenia, 9-11 May 2011

7. SeleMix UNECE Work Session on Statistical Data Editing sel.edit - This function prioritises observations according to the score function values and flags the units to be edited so that the expected residual error is below a prefixed level of accuracy. The output of sel.edit is a matrix containing the flag of influential units, the observed and anticipated values ordered by the global score, the local scores. Ljubljana, Slovenia, 9-11 May 2011

8. Warning UNECE Work Session on Statistical Data Editing • 1) Model assumptions • True data are log-normal/normal • Error is Gaussian and it inflates the covariance matrix • However: • The Gaussian or log-normal assumption is frequently adopted • Some experiments show that it can be usefully applied to cases when data depart form the assumptions • 2) The accuracy level is for estimates of totals (means). Ljubljana, Slovenia, 9-11 May 2011

9. Warning - edits UNECE Work Session on Statistical Data Editing 1) it is generally difficult to incorporate fatal edits in the model 2) On the other hand, soft edits (when the values are anomalous but plausible) are implicitly considered since the units are classified as erroneous with a certain probability, and this probability is explicitly considered in the computation of the score. Ljubljana, Slovenia, 9-11 May 2011

10. If you are interested… UNECE Work Session on Statistical Data Editing The software can be freely downloaded from www.osor.eu the Open Source Observatory and Repository for European public administrations (OSOR). In future it will be made available on the Cran library (R website) Ljubljana, Slovenia, 9-11 May 2011