Triangulación de un estudio cualitativo mediante regresión logística / Triangulation of a qualitative research by logistic regression

Fundamento y Objetivo: La investigación cualitativa complementa su visión de la realidad mediante la triangulación. La regresión logística binaria es un instrumento de predicción de riesgo en epidemiología analítica. Nuestro objetivo ha sido triangular una investigación cualitativa de tipo pedagógico con modelos de regresión logística. Material y Método: Sobre la información recogida por un grupo focal, organizamos los datos en tres variables: Aforismo / Frase corta (variable dependiente), Profesor y Tipo (variables predictoras) y construimos dos modelos con regresión logística binaria. El error alfa fue del 5 y del 10%. El tamaño muestral venía impuesto por el grupo focal anterior (saturación cualitativa). Se diseñaron rutinas para trabajar con los datos en el programa R. Resultados: Con 127 elementos (44 aforismos y 83 frases cortas) se obtuvieron significaciones crudas del 10% para dos de los diez profesores con información relevante para el grupo focal (odds ratios de 0.42 y 2.33 respectivamente; índice de Brier escalado = 0.06 y área bajo curva ROC = 0.63) y significaciones menores del 5% para cuatro de los cinco epígrafes en que habíamos dividido la variable tipo (epidemiológicos, epistemológicos, estadísticos y pragmáticos o heurísticos). El epígrafe "Estadístico" fue significativo con respecto a "Epistemológico" (OR=5,00; IC al 95% = 14.431-1.743) y con respecto a "Pragmático" (OR=4.80; IC al 95%=14.602-1.577). El epígrafe "Difusión Científica" no resultó significativo. Conclusiones: En un entorno de investigación cualitativo-pedagógica sobre aforismos y frases cortas, la regresión logística binaria se ha mostrado eficaz, dentro de una estrategia de triangulación, para identificar docentes originales para el grupo focal (p<0.10) y señalar epígrafes con interés clasificatorio (p<0.05). La capacidad predictiva de los modelos ha sido baja y la capacidad discriminativa aceptable

Background and objectives: Qualitative research seeks to enrich its vision of reality through triangulation. Binary logistic regression is a prediction tool in analytical epidemiology. Our aim was to complement a qualitative study by logistic regression models. Methods: On gathered information by a previous focus group, we organized the data into three variables: Aphorism / short phrase (dependent), Professor and Type (predictive) and built two models with binary logistic regression. The alpha error was 5 and 10%. The sample size was imposed by the previous focus group task (qualitative saturation). Routines were implemented to work with the program R. Results: With 127 elements (44 aphorisms and 83 short sentences) we obtained a 10% raw signification for two of the ten teachers with relevant information for the focus group (odds ratios of 0.42 and 2.33 respectively; Brier scaled =0.06 and area under ROC curve = 0.63) and significations less than 5% for four the five sections in which we divided the variable "Type" (epidemiological, epistemological, statistical, pragmatic or heuristic). The heading "Statistics" was significant with respect to "Epistemological" (OR = 5.00, CI 95% = 14.431-1.743) and with respect to "Pragmatic" (OR = 4.80, CI 95% = 14.602-1.577). The label "Scientific Spread" was not significant. Conclusions: In an environment of qualitative and pedagogical research on aphorisms and short phrases, binary logistic regression has been shown effective in identifying original teachers for focus group (p<0.1) and to identify qualifying entries with interest (p<0.05). The predictive capability of models has been low and acceptable the discriminative capacity

Missing value imputation on missing completely at random data using multilayer perceptrons.

Data mining is based on data files which usually contain errors in the form of missing values. This paper focuses on a methodological framework for the development of an automated data imputation model based on artificial neural networks. Fifteen real and simulated data sets are exposed to a perturbation experiment, based on the random generation of missing values. These data set sizes range from 47 to 1389 records. A perturbation experiment was performed for each data set where the probability of missing value was set to 0.05. Several architectures and learning algorithms for the multilayer perceptron are tested and compared with three classic imputation procedures: mean/mode imputation, regression and hot-deck. The obtained results, considering different performance measures, not only suggest this approach improves the quality of a database with missing values, but also the best results are clearly obtained using the Multilayer Perceptron model in data sets with categorical variables. Three learning rules (Levenberg-Marquardt, BFGS Quasi-Newton and Conjugate Gradient Fletcher-Reeves Update) and a small number of hidden nodes are recommended.