| dc.description.abstract | The present research had as objective verify the probability of Civil Society Organizations of Public Interest (OSCIPs) having their donations impacted by information from the accounting reports. To reach the proposed objective, a predominantly quantitative approach was adopted - with use of logistic regression and artificial neural network. The data, object of analysis of the study, were obtained by means of 463 accountability provided by the Institute of Applied Economic Research (IPEA). Therefore, from of a logistic regression, we tested a model to donation forecast. Initially, through the correlation matrix, It was found a low correlation between the explanatory variables. By means of the logistic regression estimation, we obtained the likelihood ratio test statistic with a value of 63.42, what allows us to state that the coefficients were jointly significant to explain the probability of OSCIPs having their levels of donations affected by accounting information - being possible reject to 5% the hypothesis that the coefficients together are equal to zero. Also, based on the estimated logit model, it was found that the coefficients of the variables government revenues (GOV), total assets (ASSETS) and short-term debt (ECP) are statically different from zero to 5% of significance. It was evidenced that the variables "administrative expenses (ADM), general indebtedness (EG) and long-term debt (ELP)" are not significant. Using the stepwise criterion, the logistic regression was again estimated to verify fit of the model and to observe the capacity of the logit to classify the event of interest of the study. Through the ROC curve, with an area of 0.7077, it was verified that the estimated model to classify the donations to the OSCIPs presents a good predictive capacity. The fit of the model was still tested by means of three cut-off points (0.4, 0.5 and 0.6), being found a improvement significant improvement in the adjustment of the model by means of a cut-off point of 0.4, since the sensitivity ranged from 62.56% (stepwise regression) to 81.06%. Additionally, the performance of the logistic regression and a structured artificial neural network with 1 input layer, a hidden layer (tested with three and five neurons) and an output layer were compared. In the comparison of the performance of the two methodologies, it was verified that the best performance of both occurred in the cutoff point of 0.4 and in the artificial neural network configured with three neurons in the hidden layer. The percentage of accuracy of the logistic regression for category 1 (with donations) was 81.06%, while the neural network presented 74.45% of correct answers in the classification of event 1. Finally, it was verified that the logit model had better performance In the classification of the event under study. Based on the results found, it can be concluded that there is evidence on the relevance of the accounting information - which signals the organizational reputation, evidences size and short-term indebtedness - to explain the probability of occurrence of donations. | |
