Scott H. Koeneman,Joseph E. Cavanaugh
Abstract:In this work, the distributional properties of the goodness-of-fit term in likelihood-based information criteria are explored. These properties are then leveraged to construct a novel goodness-of-fit test for normal linear regression models that relies on a non-parametric bootstrap. Several simulation studies are performed to investigate the properties and efficacy of the developed procedure, with these studies demonstrating that the bootstrap test offers distinct advantages as compared to other methods of assessing the goodness-of-fit of a normal linear regression model.
What problem does this paper attempt to address?
The problem this paper attempts to address is the development of a new goodness-of-fit test method based on the non-parametric bootstrap approach for evaluating the goodness-of-fit of normal linear regression models. Specifically, the authors explore the distribution characteristics of the goodness-of-fit term in the likelihood information criterion and use these characteristics to construct a new goodness-of-fit test method. This method aims to overcome the limitations of existing methods in detecting whether model assumptions are met, especially when there are unobserved covariates causing heteroscedasticity, which existing methods may fail to effectively detect.
### Main Issues:
1. **Limitations of existing methods**: Existing goodness-of-fit test methods, such as the Breusch-Pagan test and the White test, mainly rely on observed covariates to detect heteroscedasticity. When heteroscedasticity is caused by unobserved covariates, these methods may fail to effectively detect model misspecification.
2. **Comprehensive evaluation of model assumptions**: Existing goodness-of-fit test methods usually only test specific assumptions, such as normality and homoscedasticity, and cannot comprehensively evaluate all assumptions. Therefore, a method that can comprehensively test multiple assumptions is needed.
3. **Performance in small samples**: Existing goodness-of-fit test methods may perform poorly in small samples, especially when the sample size is small, their Type I error rate may deviate from the expected value.
### Solution:
The authors propose a goodness-of-fit test method based on the non-parametric bootstrap approach, which has the following features:
1. **Does not rely on correct model specification**: This method uses the White robust sandwich variance estimator, which can provide reliable variance estimates even if the model is misspecified.
2. **Comprehensive evaluation of all assumptions**: This method not only tests specific assumptions, such as normality and homoscedasticity, but also detects other forms of model misspecification, such as incorrect mean structure.
3. **Applicable to different sample sizes**: Through simulation studies, this method shows good performance under different sample sizes, especially as the sample size increases, its detection ability gradually improves.
### Simulation Study Results:
The authors conducted four simulation studies to evaluate the effectiveness of the proposed method:
1. **Correct model specification**: When the model is correctly specified, the bootstrap goodness-of-fit test's Type I error rate is slightly higher than the expected value, but it gradually approaches the expected value as the sample size increases.
2. **Incorrect mean structure**: When the model omits an important covariate, the bootstrap goodness-of-fit test shows high detection ability, especially in large sample sizes.
3. **Heteroscedasticity caused by unobserved covariates**: In this case, the bootstrap goodness-of-fit test shows high detection ability, while traditional Breusch-Pagan and White tests fail to effectively detect this form of heteroscedasticity.
4. **Heteroscedasticity caused by covariates in the model**: In this case, traditional Breusch-Pagan and White tests show high detection ability, while the bootstrap goodness-of-fit test also achieves similar detection ability in large sample sizes.
### Conclusion:
This paper proposes a new goodness-of-fit test method based on the non-parametric bootstrap approach, which can effectively evaluate the goodness-of-fit of normal linear regression models under different conditions, especially when there are unobserved covariates causing heteroscedasticity. This method not only detects specific assumptions but also comprehensively evaluates all model assumptions, showing broad application prospects.