Créer une présentation
Télécharger la présentation

Download

Download Presentation

Score Functions under the Optimization Approach Work Session on Statistical Data Editing

142 Vues
Download Presentation

Télécharger la présentation
## Score Functions under the Optimization Approach Work Session on Statistical Data Editing

- - - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - - -

**Score Functions under the Optimization Approach**Work Session on Statistical Data Editing Paris, 28-30 April 2014 Ignacio Arbués and Pedro Revilla INE Spain**Outline**Selective editing as an optimization problem Score functions obtained from the optimization approach Case studies**Selective editing as an optimization problem** Based on techniques of optimization We will determine a selection strategy that allows editing the minimum number of units, while obtaining certain accuracy requirements in the aggregates Score functions can be obtained References Arbués I., Revilla, P. and Salgado D. (2013). “An optimization approach to selective editing”. Journal of Official Statistics (JOS) Arbués I., González M. and Revilla, P. (2012). “A class of stochastic optimization problems with application to selective data editing". Optimization**General approach**True (yk0) observed (ykobs) and edited (ykedit) values The ultimate variables are the selection strategy vector RT = (R1, R2,…, Rn) for the sample units s = {1,...,n}, where Rk= 0 if the unit k is selected for interactive editing and Rk = 1 otherwise Since the measurement error ek = ykobs-yk0 is random, conditional on the realized sample s, and on the available information Z chosen to make the selection of units, this selection can vary depending on the realized ykobs, yk0 and Z**Selection strategy**• R(w) = r is a realized selection (ii)Em[R/Z] is the vector of probabilities of non-selection under the specific model m given the chosen information Z relation the true, observed and edited values ykedit(r) = (1 − rk) yk0 + rk ykobs, we have made explicit the dependence of the edited values upon the selection strategy r we are implicitly assuming that the editing work drives us from the observed to the true values We can also write ykedit(r) = yk0+ rk ek**Objective function to maximize**Em[ΣRi/Z] Em[1TR/Z] where 1 stands for a vector of ones whose maximization amounts to minimizing the number of selected units**Constraints**Derive from the application of a loss function to the survey estimators Two loss functions most used in practice absolute loss L=L(1) (a,b) = │a-b│ squared loss L=L(2) (a,b) = (a-b)2 Each constraint controls the loss of accuracy in terms of the chosen loss function L due to non-selected units For these loss functions, each constraint can always be written as a bound on a quadratic form, denoted by Em [RTRZ]**Loss matrix**The n x n matrix of conditional moments of the estimated measurement errors specifies the potential losses at the unit level Measures of bias and/or MSE seem natural in practice and they stem from the choice of the absolute or the squared loss function respectively. In particular, non-zero off-diagonal terms of allow for cross-unit terms to be included in the “overall” loss. The choice of the matrix is naturally linked to the choice of the loss function L, hence the term loss matrix**Entries of the matrices of moments**The entries of the corresponding matrices of moments of the estimated measurement errors are denoted by Δkl = ωkωl(Ykobs – Yk0) (Ylobs – Yl0)**Generic optimization problem**[Po] max Em [1T R| Z] s.t. Em [RT Δ(q)R| Z] ≤ q , q = 1, 2,…, Q, R Ω0 - 0 denotes the admissible outcome space of R - q refers to the different constraints The different constraints q may arise from the fact that there are multiple variables of interest inside the questionnaire**Stochastic version**Choosing the auxiliary information Z and the subset S0of sought selection strategies in the general problem P0, we end up with different optimization versions If no auxiliary information is used and the sought selection strategies are of the form: kis uniform (0,1) random variable and Qk = Qk (X,Yobs, S) is a continuous (0,1) random variable, we have the stochastic version of the optimization problem**Combinatorial version**We use both all available auxiliary information and the observed values found in the sample and do not restrict the form of the sought selection strategies S0 = S [Pco] max 1´r s.t. rt M(q) r ≤ mq2 , q = 1,…, Q r Bn r stands for the realized selection M(q) condenses the modelization of the measurement error mq are bounds chosen by the statistician B = {0,1}.**Score functions obtained from the optimization approach**Additional assumption is neglecting the cross-unit terms in each constraint Then these constraints can be rewritten as Em [RTR/Z] = Em [RT diag()/Z]**Solution resulting from this linear problem**The solution resulting from this linear problem is given in terms of matrices M (q) = Em [(q)/ Z] Since this selection scheme is to be applied unit by unit upon receipt of each questionnaire, and no cross-sectional information except that regarding each unit k separately will be actually used, the formal conditioning reduces effectively to conditioning upon the information of each unit Thus we write M (q) = Em [/ Z ] = diag (Em [kk(q)/Zk]) = diag (Mkk(q)) On the other hand, in order to obtain the optimal Lagrange multipliers λ* involved in the dual problem, a historic double-data set with raw and edited values is necessary**Final solution**Putting it all together we arrive at the final solution, which only requires the diagonal entries of the matrices M (q) :**Score function**One constrain (Q = 1) Unit k is selected provided Mkk > 1/* Mkk can be regarded as a single score and 1/* as the threshold value. * Mkk can be considered as a “standardized” score, in the sense that the threshold value is generically set to 1. Multiple constraints (Q >1) each q* Mkk(q) is a standardized local score, and qq* Mkk(q) is the standardized global score, with the generic global threshold value 1.**Case study 1: Periodic survey with a simple questionnaire** Application to Turnover/New Orders survey Monthly data from 13.500 units Only two of the variables requested are considered (total turnover and total new orders) Data from January 2002 to September 2006**Model to obtain the conditional moments**We need a model for the data in order to obtain the conditional moments of the estimated measurement errors Since the variables are distributed in a strongly asymmetric way, we use their logarithm transform, ytij = log (xtij + m), We assume that variables xtij are independent across (i, j) and for any pair (i, j), and we choose among the following simple univariate ARIMA models (1 - B) ytij = at (1 – B12) ytij = at (1 – B12) (1 - B) ytij= at We select the model which produces lesser mean of squared residuals**Computing the conditional moments**With this model, we compute the prediction ŷtij and the prediction standard deviation vij. The a priori standard deviation of the observation errors and the error probability are considered constant across units (that is possible because of the logarithm transformation). We denote them by j and pj with j = 1,2 and they are estimated using historical data of the survey.**Evaluation**We intend to compare the performance of the score obtained under the optimization approach to that of the score-function described, for example, in Hedlin (2003) i = i | xiobs - xipre|, where xipre is a prediction of x according to some criterion. - i0 = i | xiobs - xipre|,prediction data t-1 -1 = i | xiobs - xipre|,predictionARIMAmodel - 2 score function computed under optimization approach**Effectiveness of the score functions**E1j (n) = i n (ij)2 (xijobs - xij0)2 E2j(n) = [i nij (xijobs - xij0)]2 Units arranged in descending order according to the corresponding score function These measures can be interpreted as estimates of the remaining error after editing the n first units The difference is that E1j(n) is the aggregate squared error and E2j(n) is the squared aggregate error. Thus, E2j(n) is the one that has practical relevance, but we also include the values of E2j(n) because in the linear problem, it is the aggregate squared error which appears in the left side of the expectation constraints**Case study 2: Cross-sectional survey with a complex**questionnaire The results are related to a sample of 7215 questionnaires and 186 quantitative variables extracted from the 1999 Spanish Agricultural Census We need a model for the data in order to obtain the conditional moments of the estimated measurement errors Since we cannot use past information to make the predictions, we will use regression models instead of time series models first model a classical linear regression. For each of the variables in the questionnaire, we build a model in which the log-transformed variable under study yj = log (1+xj ) is regressed against a subset of the remaining ones (yrem) yj = β0 + βrem yrem + or yj = X´β +. The most difficult and time-consuming task is the selection of the regressors, which is done by an automatic method (a kind of stepwise algorithm)**Figure 1: Histograms of the R2 of the linear regressions,**unweighted and weighted.**Alternative models**• Most variables (land area, livestock, etc) are distributed as a mixture of a degenerate distribution in zero and some other continuous distribution in the positive semi-axis. Some models have been proposed in the literature to deal with this kind of data (see Schafer, 1999). • Second model is a two-part model in which a logit is used to predict whether the variable will be equal to zero or not, and a linear regression model conditional to the event {yj > 0}. The model has the form • yj = (1-Z) (X`β +), with N (0, v2 ) • and where Z {0, 1}. We establish a logistic regression model for the dichotomous event of having zero or positive values. • P(Z = 0) = e x’/1+e x’’**Figure 2. Histogram of the variable “land in property”,**transformed.**Figure 3. Remaining quadratic error (relative to total**error) as a function of the number of questionnaires edited in logarithmic scale. Each curve represents the quantile indicated by is colour: blue=50%, green=75%, red=90%, cyan=95%, purple=99%. Strong error**Remaining error as a function of the edited questionnaires**(N~9000).**Final remarks**•We have introduced a theoretical framework to deal with the problem of selective editing. We consider the search for an adequate selection strategy as a generic optimization problem with an stochastic and a combinatorial version. We have shown that a certain score function provides the solution to the problem with linear constraints •Our experiments with real data suggest that the method provides good selection strategies. The results, although still preliminary, are encouraging. The selection obtained outperforms that of traditional score functions in general •Much work is still under progress. More methodological research is needed to find generic multivariate models. From the practical point of view, all the applications have been carried out using data from traditional surveys. It would be convenient to test the procedures using other sources of information such as administrative and big data