Official statistics are intended to support decision makers by providing reliable information on different population groups, identifying what their needs are and where they are located. This allows, for example, to better guide public policies and focus resources on the population most in need. Statistical information must have some characteristics to be useful for this purpose. This data must be reliable, up-to-date and also disaggregated at different domain levels, e.g., geographically or by sociodemographic groups (Eurostat, 2017). Statistical data producers (e.g., national statistical offices) face great challenges in delivering statistics with these three characteristics, mainly due to lack of resources. Population censuses collect data on demographic, economic and social aspects of all persons in a country which makes information at all domains of interest available. They quickly become outdated since they are carried out only every 10 years, especially in developing countries. Furthermore, administrative data sources in many countries have not enough quality to produce statistics that are reliable and comparable with other relevant sources. On the contrary, national surveys are conducted more frequently than censuses and offer the possibility of studying more complex topics. Due to their sample sizes, direct estimates are only published based on domains where the estimates reach a specific level of precision. These domains are called planned domains or large areas in this thesis, and the domains in which direct estimates cannot be produced due to lack of sample size or low precision will be called small areas or domains. Small area estimation (SAE) methods have been proposed as a solution to produce reliable estimates in small domains. These methods allow improving the precision of direct estimates, as well as providing reliable information in domains where the sample size is zero or where direct estimates cannot be obtained by combining data from censuses and surveys (Rao and Molina, 2015). Thereby, the variables obtained from both data sources are assumed to be highly correlated but the census actually may be outdated. In these cases, structure preservation estimation (SPREE) methods offer a solution when the target indicator is a categorical variable, with at least two categories (for example, the labor market status of an individual can be categorised as: ‘employed’, ‘unemployed’, and ‘out of labor force’). The population counts are arranged in contingency tables: by rows (domains of interest) and columns (the categories of the variable of interest) (Purcell and Kish, 1980). These types of estimators are studied in Part I of this work. In Chapter 1, SPREE methods are applied to produce postcensal population counts for the indicators that make up the ‘health’ dimension of the multidimensional poverty index (MPI) defined by Costa Rica. This case study is also used to illustrate the functionalities of the R spree package. It is a user-friendly tool designed to produce updated point and uncertainty estimates based on three different approaches: SPREE (Purcell and Kish, 1980), generalised SPREE (GSPREE) (Zhang and Chambers, 2004), and multivariate SPREE (MSPREE) (Luna-Hernández, 2016). SPREE-type estimators help to update population counts by preserving the census structure and relying on new and updated totals that are usually provided by recent survey data. However, two scenarios can jeopardise the use of standard SPREE methods: a) the indicator of interest is not available in the census data e.g., income or expenditure information to estimate monetary based poverty indicators, and b) the total margins are not reliable, for instance, when changes in the population distribution between areas are not captured correctly by the surveys or when some domains are not selected in the sample. Chapters 2 and 3 offer a solution for these cases, respectively. Chapter 2 presents a two-step procedure that allows obtaining reliable and updated estimates for small areas when the variable of interest is not available in the census. The first step is to obtain the population counts for the census year using a well-known small-area estimation approach: the empirical best prediction (EBP) (Molina and Rao, 2010) method. Then, the result of this procedure is used as input to proceed with the update for postcensal years by implementing the MSPREE (Luna-Hernández, 2016) method. This methodology is applied to the case of local areas in Costa Rica, where incidence of poverty (based on income) is estimated and updated for postcensal years (2012-2017). Chapter 3 deals with the second scenario where the population totals in local areas provided by the survey data are strengthened by including satellite imagery as an auxiliary source. These new margins are used as input in the SPREE procedure. In the case study in this paper, annual updates of the MPI for female-headed households in Senegal are produced. While the use of satellite imagery and other big data sources can improve the reliability of small-area estimates, access to survey data that can be matched with these novel sources is restricted for confidentiality reasons. Therefore, a data dissemination strategy for micro-level survey data is proposed in the paper presented in Part II. This strategy aims to help statistical data producers to improve the trade-off between privacy risk and utility of the data that they release for research purposes.
