O desbalanceamento de dados em experimentos está muitas vezes presente em diversas pesquisas nas mais variadas áreas do conhecimento. Embora existam muitas maneiras de análise de tais dados, além de diversos recursos computacionais já implementados em diversos softwares estatísticos, ainda perdura dúvidas entre os pesquisadores a respeito da opção de análise mais eficiente. A literatura fornece ao pesquisador direção na escolha da metodologia de análise a obter maior eficácia nos resultados de sua pesquisa, mas o número elevado de opções pode tornar a escolha difícil. Em se tratando de testes estatísticos, algumas das opções para se trabalhar com dados desbalanceados são os testes t e Wald-F, mas ainda resta ao pesquisador decidir entre as várias opções disponíveis nos pacotes, pois nem sempre as opções padrões são as mais indicadas. No presente trabalho foram realizadas simulações com diferentes cenários experimentais, utilizando-se o delineamento casualizado em blocos com um fator de tratamento em uma situação e o esquema de tratamentos em parcelas subdividas em outra, sendo comparados quatro métodos de cálculo do número aproximado de graus de liberdade (Containment, Residual, Satterthwaite e Kenward-Roger). Verificou-se que o método de Kenward-Roger controla de maneira mais eficiente a taxa de erro tipo I e não é inferior aos outros métodos com respeito ao poder do teste Wald-F. / The data imbalance in experiments is often present in several researches in various fields of knowledge. While there are many ways to analyze these data in addition to various computer resources already implemented in many statistical software, doubt still lingers among researchers about the most efficient analysis option. The literature provides the researcher direction in choosing the analysis methodology to get better in your search results, but the large number of options can make the difficult choice. When it comes to statistical tests, some of the options for working with unbalanced data are the tests t and Wald-F, but there is still the researcher to decide between the various options available in the packages because the defaults are not always the most suitable. This experiment was carried out simulations with different experimental scenarios, using the randomized block design with one factor in a situation treatment and treatment regimen subdivided parcels in another, and compared four methods of calculating the approximate number of degrees of freedom (Containment, Residual, Satterthwaite and Kenward-Roger). It has been found that the method of Kenward-Roger controls more efficiently the type I error rate and is not inferior to other methods with respect to the power of the test Wald-F.

O desbalanceamento de dados em experimentos está muitas vezes presente em diversas pesquisas nas mais variadas áreas do conhecimento. Embora existam muitas maneiras de análise de tais dados, além de diversos recursos computacionais já implementados em diversos softwares estatísticos, ainda perdura dúvidas entre os pesquisadores a respeito da opção de análise mais eficiente. A literatura fornece ao pesquisador direção na escolha da metodologia de análise a obter maior eficácia nos resultados de sua pesquisa, mas o número elevado de opções pode tornar a escolha difícil. Em se tratando de testes estatísticos, algumas das opções para se trabalhar com dados desbalanceados são os testes t e Wald-F, mas ainda resta ao pesquisador decidir entre as várias opções disponíveis nos pacotes, pois nem sempre as opções padrões são as mais indicadas. No presente trabalho foram realizadas simulações com diferentes cenários experimentais, utilizando-se o delineamento casualizado em blocos com um fator de tratamento em uma situação e o esquema de tratamentos em parcelas subdividas em outra, sendo comparados quatro métodos de cálculo do número aproximado de graus de liberdade (Containment, Residual, Satterthwaite e Kenward-Roger). Verificou-se que o método de Kenward-Roger controla de maneira mais eficiente a taxa de erro tipo I e não é inferior aos outros métodos com respeito ao poder do teste Wald-F. / The data imbalance in experiments is often present in several researches in various fields of knowledge. While there are many ways to analyze these data in addition to various computer resources already implemented in many statistical software, doubt still lingers among researchers about the most efficient analysis option. The literature provides the researcher direction in choosing the analysis methodology to get better in your search results, but the large number of options can make the difficult choice. When it comes to statistical tests, some of the options for working with unbalanced data are the tests t and Wald-F, but there is still the researcher to decide between the various options available in the packages because the defaults are not always the most suitable. This experiment was carried out simulations with different experimental scenarios, using the randomized block design with one factor in a situation treatment and treatment regimen subdivided parcels in another, and compared four methods of calculating the approximate number of degrees of freedom (Containment, Residual, Satterthwaite and Kenward-Roger). It has been found that the method of Kenward-Roger controls more efficiently the type I error rate and is not inferior to other methods with respect to the power of the test Wald-F.

Linear mixed models are frequently used to analyze data with random effects and/or repeated measures. A common approach to such analyses requires choosing a covariance structure. Information criteria, such as AIC and BIC, are often used by statisticians to help with this task. However, these criteria do not always point to the true covariance structure and therefore the wrong covariance structure is sometimes chosen. Once this step is complete, Wald statistics are used to test fixed effects. Degrees of freedom for these statistics are not known. However, there are approximation methods, such as Kenward and Roger (KR) and Satterthwaite (SW) that have been shown to work well in some situations. Schaalje et al. (2002) concluded that the KR method would perform at least as well as or better than the SW method in many cases assuming that the covariance structure was known. On the other hand, Keselman et al. (1999) concluded that the performance of the SW method when the covariance structure was selected using AIC was poor for negative pairings of treatment sizes and covariance matrices and small sample sizes. Our study used simulations to investigate Type I error rates in test of fixed effects using Wald statistics with the KR adjustment method, incorporating the selection of the covariance structure using AIC and BIC. Performance of the AIC and BIC criteria in selecting the true covariance structure was also studied. The MIXED procedure (SAS v. 9) was used to analyze each simulated data set. Type I error rates from the best AIC and BIC models were always higher than target values. However, Type I error rates obtained by using the BIC criterion were better than those obtained by using the AIC criterion. Type I error rates for the correct models were often adequate depending on the sample size and complexity of covariance structure. Performance of AIC and BIC was poor. This could be a consequence of small sample sizes and the high number of covariance structures these criteria had to choose from.