Demografska statistika ali statistika prebivalstva je eno najbolj tradicionalnih in hkrati tudi eno najstarejših področij statistike. Tudi na Statističnem uradu Republike Slovenije je zbirka podatkov o prebivalcih in njihovih dogodkih ena najobsežnejših in najstarejših. Ti podatki so tudi med najbolj iskanimi podatki med uporabniki, saj so osnova za izračun različnih socialnih, ekonomskih, okoljskih in drugih kazalnikov.
Na tokratnem predavanju bom iz prakse predstavila, kako demografske podatke zagotavljamo danes. Katere podatke o prebivalcih in njihovih življenjskih dogodkih (rojstvo, smrt, poroka, razveza, selitev) zbiramo, kako jih med seboj povezujemo, prikazujemo, kje jih objavljamo in kateri izmed njih so najbolj iskani. Ker pa v zadnjih letih demografski podatki postajajo središčna tema razprav o naši prihodnosti - zaradi staranja prebivalstva, selitvenih tokov iz manj razvitih v bolj razvita območja, zaradi manjše rodnosti, zmanjševanja delovno sposobnega prebivalstva, zaradi pričakovane daljše življenjske dobe, bom predavanje zaključila s kratko analizo zadnjih demografskih trendov v Sloveniji. Slides: PPTPhotos: 1, 2, 3, 4, 5, 6, 7, 8.
Sequential statistical methods have been around since the 1950's but are not widely known or used. We discuss the adaptation of the binomial sequential probability ratio test (SPRT) statistic to a very sensitive problem in quality. The situation was that approximately 750,000 homes and businesses in a major metropolitan area needed to be inspected for mercury contamination as quickly as possible and with a high degree of reliability in the results. Because of the nature of the situation, the quality of the inspection process needed to be demonstrated to meet certain standards as quickly as possible and monitored thereafter.
Slides: PDF
Quality is a widespread concept among people of all professions. Therefore, the definition of its meaning should not be a self-given right of experts in the area. Also, the notion of quality should be freed from its traditional ties to products and presence or absence of defects. The usual notions of "quality" are: good, bad, excellent. And this clear view should not be distorted. What we can do is to expand this notions in terms of the object and in terms of several quality levels of the individual object.
Expansion in terms of the object leads us from the quality of results to the quality of sources (product / service process, resulting in the product / service working environment, which results in the quality process, management, resulting in a quality of work environment). Of course, the improvement of quality in each of these four areas demands different knowledge and different approaches.
Expansion in the direction of several quality levels leads to a statistical, that is distributional view on the problem. If there are enough levels of quality, we usually end up with a Gaussian-like distribution. Thus we see quality of the majority, what is the quality of the negative minority and what is the quality of the positive minority. In this way, we work on improving on three levels:
High-throughput technologies such as microarrays simultaneously measure thousands or even ten thousands of variables. Usually the number of subjects included in such studies is limited and much smaller than the number of variables; therefore these data are called high-dimensional. High-dimensional data are increasingly often used for class prediction; for example, we would like to use the gene expression data to classify a sample in one of the predefined and distinct groups (classes).
In the talk I will present how high-dimensionality affects class prediction for class-imbalanced data, i.e. data where the number of samples in each group is not the same. I will also present some popular strategies that were proposed to solve the class-imbalance problem and explain their advantages and disadvantages when used on high-dimensional data.
The primary task of MR imaging is tissue delineation and then analysis of tissue deviations from normality. Despite the apparent simplicity of the delineation, the development of segmentation algorithms is very active area due to different scanning errors and differences in normal brain structures. Segmentation approaches using anatomical atlases, that provide a priori information about the spatial distribution of tissue types, yield the best results. But these methods require the alignment of the atlases, which are calculated as average of some healthy brains. The alignment process is again sensitive to scanner artifacts and to similarity between the atlas brain structures and the brain structures to be studied.
I will describe an approach for generating tissue probability atlases with the use of local descriptive statistics, where the alignment process is not required. Slides: PDFPhotos: 1, 2, 3, 4, 5, 6, 7, 8, 9.
Different ways how to deal with the demand for the process to be within tolerance will be described. Also, some of the process capability indices will be presented.