Assumptions and Conditions for Using Statistical Tests Essay Example

indicating the direction of arrows connecting the variables in the model. When a misspecification error is corrected by changing the model, all parameter estimates in the model are subject to change, not only in magnitude, but sometimes even in direction. There is no statistical test for misspecification. A good literature review is important in identifying variables which need to be specified.

As a rule of thumb, the lower the overall effect (ex. , R2 in multiple regression, goodness of fit in logistic regression), the more likely it is that important variables have been omitted from the model and that existing interpretations of the model will change when the model is correctly specified. The specification problem is lessened when the research task is simply to compare models to see which has a better fit to the data, as opposed to the purpose being to justify one model and then assess the relative importance of the independent variables.

Adequate Cell Size Adequate cell count is an assumption of any procedure which uses Pearson chi-square or model likelihood chi-square (deviance chi-square) in significance testing when categorical predictors are present. This includes but is not limited to chi-square testing of crosstabulation, loglinear analysis, binomial logistic regression, multinomial logistic regression, ordinal regression, and general or generalized linear models of the same.

Factor Space

Factor space is the set of cells which are generated by a crosstabulation of the categorical dependent with all the categorical factors but not the continuous covariates.

 Data Level Requirements Measurement level requirements vary by statistical procedure but most procedures require an interval or ratio level of measurement. It is common in

social science, however, to utilize dichotomies and ordinal data, such as Likert scale data, even in procedures which technically require interval-level data.

Dichotomies are often included in statistical models (ex. , regression models) provided the split is less than 90:10. Ordinal variables are often included in statistical models provided the normality of error terms may be demonstrated, as discussed below. Some researchers require the ordinal scale to have five or more values. Violations of data level assumptions mean that actual standard error will be greater than the computed standard error and significance is overestimated (that is, the chance of Type I error is greater than computed).