Getting Rejected for Statistics? You're Probably Making One of These 3 Mistakes

Every year, thousands of medical manuscripts are rejected. One of the most common reviewer comments reads:

"The choice of statistical methods is inappropriate, which undermines the reliability of the conclusions."

Behind that sentence, there are usually three very specific mistakes.

Mistake 1: Reporting Skewed Data with Mean ± SD

Have you ever written something like this?

"Length of hospital stay was 8.3 ± 12.7 days."

Mean 8.3, standard deviation 12.7. The SD is larger than the mean—which almost always signals a right-skewed distribution.

Hospital stay, medical costs, survival time—these variables have a natural lower bound of zero and a long right tail. Reporting them with a mean is like describing average income in a classroom where half the students sat next to a billionaire: technically accurate, completely misleading.

What to do instead:

• Report as Median (IQR) instead of Mean ± SD
• Replace t-tests with Mann-Whitney U (two groups) or Kruskal-Wallis (multiple groups)

Many researchers know this rule in theory. But when they ask a general AI for help with analysis, the AI defaults to mean ± SD and t-tests without ever asking about data distribution.

---

Mistake 2: Running Multiple Comparisons Without Correction

You compare four treatment arms with six pairwise tests, each using p < 0.05 as the threshold.

The problem: with six tests, the probability of getting at least one false positive exceeds 26%.

This is the Multiple Comparisons Problem. Every test is a coin flip for a false positive. The more tests you run, the more likely you are to find something that isn't real.

Experienced reviewers catch this immediately. It's one of the top reasons manuscripts get sent back for major revision.

What to do instead:

• Use Bonferroni correction (conservative; good for confirmatory research)
• Or FDR correction (Benjamini-Hochberg) (higher power; better for discovery research)

---

Mistake 3: Using Stepwise Regression for Variable Selection

"I used stepwise regression to select which variables entered the final model."

This sentence is a red flag for statistical reviewers—and has been for decades.

The problems with stepwise regression:

1. Unstable results — swap in a different sample, and the selected variables may be completely different
2. Distorted p-values — after multiple rounds of filtering, the p-values in the final model can no longer be correctly interpreted
3. Overfitting — the model performs well on your data but generalizes poorly to new patients

This method became popular in the 1970s due to computational constraints. Those constraints are gone. The method's flaws remain.

What to do instead:

• LASSO regression (L1 regularization; automatically shrinks unimportant coefficients to zero)
• Elastic Net (combines LASSO and Ridge; handles multicollinearity better)
• Or base variable selection on prior clinical knowledge, not data-driven automation

---

Why AI Can't Catch These Errors Without Rules

Ask ChatGPT "help me choose statistical methods" and you'll typically receive:

• A list of methods that looks reasonable
• Qualified with phrases like "depending on your data" or "usually recommended"

What you won't get: an AI that asks whether your data is skewed, checks whether you're doing multiple comparisons, or refuses to recommend stepwise regression.

These checks require explicit rules—the kind a senior biostatistician runs through automatically when reviewing a methods section.

Our Statistical Method Advisor encodes these rules directly into the prompt template:

• Detects skewed data → recommends non-parametric methods by default
• Flags multiple comparisons → prompts for correction strategy
• Prohibits stepwise regression → redirects to LASSO or Elastic Net

Fill in your data description. Get a statistical plan that reviewers won't send back.

---

Quick Reference

Statistical methods aren't complicated—but keeping every rule in mind under deadline pressure is. A structured prompt template automates that checklist for you.

Try the Statistical Method Advisor → · Browse all prompt templates →