使用R语言计算Cohen's d为何结果始终不同？

Hey there, I totally get how frustrating it is when your Cohen's d results keep shifting every time you run your R code—let’s walk through the most likely culprits and how to fix them:

• Your function uses random sampling under the hood
  Some Cohen's d implementations (especially those tied to bootstrapped confidence intervals or permutation tests) rely on random processes. For example, if you’re calculating d alongside bootstrap CIs, each run will pull different random samples unless you lock in the random number generator.
  Fix: Drop set.seed(123) (or any integer you like) at the very start of your script. This forces R to generate the same sequence of random numbers every time, making your results 100% reproducible.

• Missing values are being handled unpredictably
  If your dataset has NA values, some functions might use random imputation or exclude cases in a non-deterministic way (though this is less common). It’s easy to overlook how your Cohen's d function deals with missing data.
  Fix: Take control of NA handling first. Use na.omit() to strip out rows with missing values, or use a consistent imputation method (like mice::mice()) with a set seed before calculating d.

• You’re accidentally using different data subsets each run
  Maybe your script includes a sampling step (like sample_data <- df[sample(nrow(df), 50), ]) without a seed, so you’re working with a different slice of data every time. Even a tiny change in the dataset can alter Cohen's d.
  Fix: Double-check your code to ensure you’re using the exact same dataset each run. If you do need to sample, add set.seed() right before the sampling line.

• You’re mixing different packages/functions
  Different R packages calculate Cohen's d slightly differently. For example, effsize::cohen.d() uses pooled SD with n-1 degrees of freedom, while some functions in lsr or psych might use n instead, or handle unequal sample sizes differently. Switching between them without noticing will lead to varying results.
  Fix: Pick one package and stick with it. effsize is a solid, consistent choice—read its documentation to understand exactly how it computes d, so you know what to expect every time.

• Your original data is being modified between runs
  If your script alters the dataset (e.g., adding calculated columns, filtering rows) in a way that depends on previous runs or randomness, your input data changes each time you run the code.
  Fix: Load your raw data fresh at the start of every script, or use rm(list = ls()) to clear your environment before running—this avoids leftover variables messing with your results.

内容的提问来源于stack exchange，提问作者user195214