G*Power Alternative: AI Power Analysis

Recompute a study's statistical power without installing G*Power. Paste your numbers or a screenshot and MoClaw checks whether the claim holds.

Computers
AI Computer
Connectors
Telegram
Slack
Files
scripts/
power_analysis.py
power-analysis-report.md
sensitivity-table.csv
Schedules
No scheduled tasks
I'm checking a published study for my thesis. It reports a two-tailed independent-samples t-test with 45 participants per group, Cohen's d of 0.5, alpha 0.05, and claims a statistical power of 0.84. Can you verify that?
MoClaw Replay

This is a post hoc power analysis. I'll: 1. Recompute power from n=45/group, d=0.5, alpha=0.05 using the non-central t distribution (the method G*Power uses) 2. Compare the result against the claimed 0.84 3. If it disagrees, solve for the effect size and sample size that would actually produce 0.84

Used 3 tools
Bash python3 power_analysis.py --test two-sample-t --n 45 --d 0.5 --alpha 0.05 --tails 2
Post-hoc power analysis (two-tailed independent t-test)

Noncentrality (delta) = d * sqrt(n/2) = 0.5 * sqrt(22.5) = 2.371
df = 2(45) - 2 = 88
Critical t (alpha=.05, two-tailed) = 1.987

Actual power (1 - beta) = 0.6502

Paper claims: 0.84
Computed:     0.65
Gap:          +0.19 (19 percentage points overstated)

At 65% power the study has roughly a 1-in-3 chance of a Type II error.
Bash python3 power_analysis.py --solve-for power=0.84 --n 45 --d 0.5 --alpha 0.05
To actually reach power = 0.84, one input must change:

Keep n=45/group  ->  need d = 0.63 (not the reported 0.50)
Keep d=0.50      ->  need n = 72/group (144 total, about 60% more participants)

Sensitivity (n=45, alpha=.05, two-tailed):
  d=0.30   power=0.29
  d=0.50   power=0.65   <- paper's stated d
  d=0.63   power=0.84   <- needed for the claim
  d=0.80   power=0.96
Write power-analysis-report.md
power-analysis-report.md created:
├── Inputs as reported
├── Recomputed power (non-central t)
├── Claim vs computed (19 pp gap)
├── Required d and n for 0.84
└── Write-up sentence + APA note
power-analysis-report.md
Used 3 tools

Verdict: the claimed power of 0.84 is not reproducible from the parameters reported. With n=45 per group, d=0.50, two-tailed at alpha=.05, the correct post-hoc power is 0.65, overstated by 19 points.

For your write-up:
"A post hoc power analysis (two-tailed independent t-test, d=0.50, n=45 per group, alpha=.05) yields 1 - beta = .65, well below the authors' reported .84; reaching .84 would require d=.63 or roughly 72 participants per group (N=144), suggesting either a computational error or an undisclosed difference in the parameters used."

Want me to check the rest of the paper's stats the same way?

Yeah. Also, why would a researcher report a higher effect size than the real one?
MoClaw Replay

Conceptual question, no computation needed. Two buckets: honest errors and motivated distortion. I'll keep them separate so the reader can judge which one applies.

First, the direction: a smaller true effect means lower power, not higher. If the real d were 0.30 instead of 0.50, power at n=45 drops to about 29%, so an inflated effect size makes a study look far better powered than it really is.

Common reasons it happens:

Honest errors

  • Basing d on a tiny pilot study, which systematically overestimates effects
  • Borrowing an effect size from published work without adjusting for publication bias (the winner's curse)
  • Confusing effect-size metrics (Cohen's d vs partial eta squared vs r)
  • Dividing by the wrong standard deviation, which inflates the standardized effect

Motivated distortion

  • Reverse-engineering: picking the d that justifies a sample size already chosen
  • HARKing: forming the hypothesis after seeing results, then reporting post-hoc power as if it were planned
  • Defaulting to d=0.5 ("medium") with no real basis for it

From the outside you usually cannot separate error from intent, which is why power claims are best pre-registered. Want me to draft a short reviewer comment flagging this for the paper?

scripts/
G*Power Alternative: AI Power Analysis
  • First, the direction: a smaller true effect means lower power, not higher. If the real d were 0.30 instead of 0.50, power at n=45 drops to about 29%, so an inflated effect size makes a study look far better powered than it really is.
  • Common reasons it happens:
  • *Honest errors**
  • Basing d on a tiny pilot study, which systematically overestimates effects
scripts/ power-analysis-report.md sensitivity-table.csv

Try follow-up prompts

What MoClaw tracks

  • AI recomputes power with the non-central t distribution, the same method G*Power uses
  • Catches a paper's claimed 0.84 power that is actually 0.65
  • Solves for the effect size and sample size you would actually need

How G*Power Alternative: AI Power Analysis Works with MoClaw

1

Paste Your Study's Numbers

Give MoClaw the test type, sample size, effect size, and alpha, or just upload a screenshot of your G*Power setup. No formulas or software to install, describe it in plain English.

2

AI Recomputes the Real Power

MoClaw runs the post hoc power analysis with the non-central t or F distribution, the same method G*Power uses, then compares the result against any power the study claims and flags the gap.

3

Get a Verdict and a Write-Up Line

You get the exact power, a sensitivity table, the effect size or sample size you would actually need, and a ready-to-paste sentence for your thesis or paper.

Ways to Extend This Workflow

🔍

Verify a Published Paper's Power Claim

Recompute the power a study reports and find out whether the number holds up before you cite it or build on it.

📐

Justify Sample Size for a Grant or IRB

Run an a priori power analysis to show exactly how many participants you need for your target power, including unequal group sizes.

🧪

Catch Effect-Size Inflation in Peer Review

Spot when an author's reported effect size is too optimistic for the sample they ran, and quantify how underpowered the study really is.

🎓

Check Your Thesis Stats Section

Run every comparison in your results through a power analysis at once and find the underpowered tests before your committee does.

Power Analysis: By Hand vs G*Power vs MoClaw

See how MoClaw's AI-powered approach differs from traditional tools.

FeatureBy HandG*PowerMoClaw
Setup Formulas and tables Install software, learn the UI Paste the numbers or a screenshot
Reads your inputs You do it all You enter each field manually Reads a screenshot and checks the setup
Judges the result Up to you Computes only, no judgment Flags the gap vs the claimed power
Required sample size Separate calculation A separate a priori analysis Solves n and d for the target in one pass
Output A single number A number and a plot Verdict, sensitivity table, write-up line
Price Free (your time) Free (desktop install) Free to try, then $20/mo

Why AI-Powered Power Analysis?

Traditional calculators give you a number. They never tell you whether a study's reported power is actually true.

Nothing to Install

G*Power is a desktop download with a dense interface. MoClaw runs in the browser, reads your numbers or screenshot, and computes power the same rigorous way.

Judgment, Not Just a Number

MoClaw recomputes the power, then tells you whether a claimed 0.84 actually holds, catches a 19-point overstatement, and explains which effect-size assumption broke.

Decision-Ready Output

Instead of a lone figure, you get the required sample size, a sensitivity table, and a sentence you can paste straight into your paper or reviewer comment.

G*Power Alternative: AI Power Analysis FAQ

Is MoClaw a free online G*Power alternative?

Yes. MoClaw recomputes statistical power the same way G*Power does, using the non-central t or F distribution, but it runs in your browser with nothing to install. You paste your inputs (effect size, sample size, alpha, and test type) or a screenshot in plain language, and it returns the exact power plus a check against any value the study claims.

Can I paste a G*Power screenshot instead of typing the numbers?

Yes. You can upload a screenshot of your G*Power setup or a results table, and MoClaw reads the inputs, checks whether they are configured correctly for a post hoc power analysis, and recomputes the power. If a field looks wrong, like a one-tailed test where it should be two-tailed, it flags it.

What does the power analysis output look like?

You get the exact power value, a clear verdict on whether it matches the claimed power, a sensitivity table showing power across different effect sizes, and a write-up sentence you can drop into a thesis or paper. MoClaw can also save it as a Markdown or PDF report.

Can I check more than one test or study at a time?

Yes. You can hand MoClaw several tests at once, such as every comparison in a results section, and it runs a separate power analysis for each and summarizes which ones are underpowered. This is useful when reviewing a full paper or a multi-experiment thesis.

Can MoClaw calculate the sample size I need before I run a study?

Yes. MoClaw does both post hoc and a priori power analysis. Give it a target power like 0.80, an expected effect size, and alpha, and it solves for the sample size you need before collecting any data, including unequal group allocations.

Is this better than using G*Power for power analysis?

G*Power is excellent and free, but it only computes, it does not judge. MoClaw recomputes power the same way, then tells you whether a claimed power holds, what effect size or sample size would be required, and why a number might be off. It runs in the browser with nothing to install.

How much does power analysis with MoClaw cost?

You can try MoClaw for free, which covers individual power analyses. Paid plans start at $20/month and include your own cloud computer, Claude Opus, and the ability to batch-check full papers and combine power analysis with other work.

Can I combine post hoc power analysis with a full stats review of my paper?

Absolutely. MoClaw can run a post hoc power analysis as part of a larger review, checking your assumptions, recomputing test statistics and p-values, and flagging effect-size inflation across the whole results section. Everything runs on the same cloud computer, so the outputs stay together.

Related Use Cases

Try MoClaw Free