HolmesCo: Report Card

Effect sizes vs p-values — Week 7 handout with instructor notes

HolmesCo Technical Summary — Three Results to Evaluate

This is a fictional document for teaching purposes.



QUARTERLY TECHNICAL SUMMARY — Q1 2026

HolmesCo is pleased to present three key findings from our recent project portfolio. All results have been subjected to rigorous statistical testing.


Result 1: Wind farm output and elevation

HolmesCo compared annual energy output (GWh) for 15 wind farms on hilltop sites vs 15 wind farms in lowland areas across Northern England.

Hilltop Lowland
Mean output 48.2 GWh 31.5 GWh
SD 11.3 10.8

t-test: t(28) = 4.12, p < 0.001

95% CI for difference: 8.4 to 25.0 GWh

Cohen’s d = 1.51

“Hilltop wind farms produce significantly more energy. This confirms what we’ve always known — elevation matters.”


Result 2: Borehole temperature and county

HolmesCo measured temperatures at 12 m depth in 1,000 boreholes in County Durham and 1,000 boreholes in North Yorkshire.

Durham N. Yorkshire
Mean temperature 11.52°C 11.50°C
SD 0.8 0.9

t-test: t(1998) = 0.53, p = 0.001

95% CI for difference: 0.008 to 0.032°C

Cohen’s d = 0.02

“A highly significant difference in geothermal potential between the two counties. HolmesCo recommends prioritising County Durham for geothermal development.”


Result 3: Solar panel degradation rate

HolmesCo compared annual power output degradation rates for 20 monocrystalline vs 20 polycrystalline solar panels over 10 years.

Mono Poly
Mean degradation 0.45%/yr 0.58%/yr
SD 0.22 0.25

t-test: t(38) = 1.74, p = 0.09

95% CI for difference: −0.02 to 0.28 percentage points/yr

Cohen’s d = 0.55

“No significant difference between panel types. Both technologies degrade at essentially the same rate.”


Your task

For each result, answer: Should we act on this? Why or why not?

Consider:

  • Is the result statistically significant?
  • Is the effect big enough to matter for policy or engineering decisions?
  • Does the confidence interval tell a different story from the p-value?
  • Would you trust HolmesCo’s interpretation?

Instructor notes

Result 1: Significant with a LARGE effect — act on this

This is the genuine article. The effect is large (d = 1.51), the CI is entirely above zero and practically meaningful (8–25 GWh), and the p-value is very small. The conclusion is well-supported.

HolmesCo’s interpretation is unusually reasonable here. The only quibble: “confirms what we’ve always known” suggests they didn’t need the test — confirmation bias, but the evidence genuinely supports it.

Result 2: Significant with a TINY effect — don’t act

This is the key teaching example. The p-value is 0.001 (“highly significant!”), but Cohen’s d is 0.02 — negligibly small. The actual difference is 0.02°C. No one can detect, measure, or use a 0.02°C difference for geothermal planning.

The large sample size (n = 2000) made a trivially small difference “significant.” This is the statistical equivalent of measuring a flea and calling it a horse because your ruler is very precise.

HolmesCo’s recommendation to “prioritise County Durham” based on a 0.02°C difference is absurd — but this is exactly the kind of reasoning that appears in real consultancy reports when people don’t look past the p-value.

Key message: “Significant” ≠ “important.” Always report effect sizes.

Result 3: Non-significant but SUGGESTIVE — investigate further

The p-value (0.09) is above the conventional 0.05 threshold, so HolmesCo says “no difference.” But look at the effect size: d = 0.55 is a medium effect. The CI (−0.02 to 0.28) nearly excludes zero — most of the interval is above zero, suggesting mono panels probably do degrade less.

With only 20 panels per group, the study is likely underpowered — it may lack the statistical power to detect a real difference of this size. A larger study might well find significance.

HolmesCo’s “no difference” conclusion is premature. The honest conclusion is: “We can’t rule out an important difference; more data is needed.”

Key message: Non-significant ≠ no effect. The absence of evidence is not evidence of absence.