OSCAR OLVERA: Hello.My name is Oscar Olvera.I am a first year PhD student in the Measurement, Evaluation,and Research Methodology Program at the Universityof British Columbia, UBC, here in Vancouver, Canada.I am the first author, alongside with my advisor, Dr. BrunoZumbo, of the article "A Cautionary Note

OSCAR OLVERA [continued]: on the Use of the Vale and Maurelli Methodto Generate Multivariate, Nonnormal Data for SimulationPurposes," which was published-- which came outlast year in September of 2014 in The Journal of Educationaland Psychological Measurement.The reason why I wrote this article

OSCAR OLVERA [continued]: and I started this research was rightafter I took a seminar on population distributionswith Dr. Harry Joe in the Department of Statistics.I wanted to see what types of multivariate nonnormalwere regularly used in simulations including,you know, psychometrics or [INAUDIBLE] social sciences.I ended up finding that the majority of the articles,

OSCAR OLVERA [continued]: not to say-- probably-- well, not all of them,but a majority of the articles useda method that's the multivariate extension of the Fleishmanthird-order polynomial.It's called the Vale and Maurelli method.And if you go to any journal in the social sciencesthat [INAUDIBLE] with these types of distributions,they use that method.

OSCAR OLVERA [continued]: And I started wondering, well, whatabout the other distributions that could be used?Like, I learned one of the thingswith populous that I think would bevery applicable to psychometrics,but I wasn't seeing that in the journals that I was reading.So we set up two [INAUDIBLE] simulation studies.The first one was just to evaluate the algorithm itself

OSCAR OLVERA [continued]: and to see whether, you know, if you ask the Vale Maurellimethod to get you marginal populations, [INAUDIBLE] of twoand kurtosis of seven, whether the average across simulationswas actually two or seven.What we found first is that it is not true, especiallyfor kurtosis.The estimates are severely downward biased,

OSCAR OLVERA [continued]: so most of the simulation conditions that peopleuse when they use this method are actually not reflectedin the data being generated.The other aspect that is also very worrisomeis that there is a very high variabilityin the estimates in the data generated by the Vale Maurellimethod.So for instance if you specify a kurtosis population kurtosis

OSCAR OLVERA [continued]: of 21, you might end up with a lot of kurtosisare in the 200s, in the 500s, and whatnot.So you look at the pictures in the article,you would end up seeing that those two regions havea very, very long tail.The second study that we set out wasto choose a different method and see

OSCAR OLVERA [continued]: whether if you graph, if you did a simulation study where youused both the different methods, wouldyou end up getting different sets of conclusions thatwere solely dependent on the methodthat you used to generate your data.The method that we used was the Headrick 2002 methodof the fifth-order polynomial.

OSCAR OLVERA [continued]: It's within the same idea of taking powersof normal distributions, the second powerwould be the variance, the third power would be the [INAUDIBLE],and the fourth power would be the kurtosis.And the Headrick method adds the fifth and sixth powerto control better for the other two.The first thing that we found was

OSCAR OLVERA [continued]: that indeed although the Headrick method alsogenerates estimates that are downward bias,they are not as downward bias as the ones generateby Vale Maurelli.In a few of the conditions they actuallymatched exactly what you would expect in the population.And, much more importantly, there are a lot less variables.

OSCAR OLVERA [continued]: So if see the tails of the distributions generatedby the Headrick method, they're a lot shorts.So they're a lot more precise in the [INAUDIBLE] kurtosisthat they're getting.The really interesting part thoughwas that we took an article wherethey are very [INAUDIBLE] cited--

OSCAR OLVERA [continued]: it's been cited more than a thousand times--on the impact of normality in the chi-square test of fitwhere the authors compared the [INAUDIBLE] theorychi-squared test of fit, the [INAUDIBLE] correction,the asymptotic distribution [INAUDIBLE] of chisquared that was derived by Brown.

OSCAR OLVERA [continued]: The authors used the Vale Maurelli method,and the general conclusion was that across the boardthe [INAUDIBLE] correction outperformedthe asymptotic distribution free methodand it was much better than the just normal theorychi-squared distribution.When you use the Headrick method,

OSCAR OLVERA [continued]: even within the same simulation conditions,the results are not the same or that they're not as clear.There are various moments in which [INAUDIBLE] distribution[INAUDIBLE] chi squared outperforms [INAUDIBLE]correction.The place where we want to take this kind of research idea

OSCAR OLVERA [continued]: was to explore other algorithms to simulate dataand see which types of distributionshave not been considered [INAUDIBLE]nonnormal multivariate distributions have notbeen considered by people within the social sciencesand psychometrics and quantitative psychologybecause they are either-- they're not awarethat these methods exist.

OSCAR OLVERA [continued]: We're taking forward to the simulation of this experimentcorrelation.We're taking forward to the use of [INAUDIBLE] distributions.And we're taking forward to this ideathat people who do simulations should not onlyworry about the methods that they're feedingand the conditions they're doing,but they should also be interested in knowingthe quality of the data that the software is using

OSCAR OLVERA [continued]: and how the data is being generated.That would be my main goal for this research.Thank you.