卡方检验结合蒙特卡洛模拟的合理性及权威引用咨询

1. Is using Monte Carlo simulation the correct approach here? #

Absolutely—this is a standard, widely accepted workaround when the expected frequency assumption for chi-square tests isn't met. Traditional Pearson's chi-square test relies on the asymptotic approximation that the test statistic follows a chi-square distribution when sample sizes are large. When too many cells have expected counts <5, this approximation breaks down, leading to inaccurate p-values. Monte Carlo simulation bypasses this by generating thousands (or millions) of simulated contingency tables from the null hypothesis, calculating the test statistic for each, and then comparing your observed statistic to this simulated distribution to get a more reliable p-value.

2. Does this solve the problem of violating the expected frequency assumption? #

Yes, it addresses the core issue. The problem with low expected counts is that the chi-square distribution doesn't accurately reflect the true sampling distribution of the test statistic. Monte Carlo simulation directly estimates this true distribution by resampling, so it produces p-values that are far more trustworthy in sparse or small-sample contingency tables.

That said, there are edge cases where even Monte Carlo might struggle (e.g., if most cells have expected counts of 0), but for your scenario where only the 80% threshold is missed, it’s a robust solution. For context: Fisher’s exact test is another option, but it’s computationally infeasible for tables larger than 2x2—Monte Carlo is the practical alternative for bigger contingency tables.

3. What authoritative sources support this method? #

Here are some key references you can cite:

• Categorical Data Analysis by Alan Agresti (3rd Edition): This is the definitive textbook for categorical data methods. In Chapter 2, Agresti explicitly recommends Monte Carlo simulation as a valid approach when expected frequencies are too small for asymptotic chi-square tests, noting that it provides accurate p-values by estimating the exact sampling distribution.
• Nonparametric Statistical Methods by Hollander & Wolfe (3rd Edition): This nonparametrics classic covers Monte Carlo resampling techniques for hypothesis testing, including their application to contingency tables when asymptotic assumptions fail.
• SAS/STAT Documentation: The official documentation for SAS’s PROC FREQ (which includes Monte Carlo options for chi-square tests) states that simulation-based p-values are preferred over asymptotic ones when expected counts are low, as they reduce Type I error rates that come from poor chi-square approximations.
• NCSS Statistical Software Manual: The NCSS guide for contingency table tests highlights Monte Carlo simulation as a reliable method for handling sparse data, emphasizing its ability to produce valid inference when asymptotic conditions aren’t met.

内容的提问来源于stack exchange，提问作者Leandra