←Back to the bookChapter 59 · Clinical Trial Efficacy (Student's t)

Chapter 59 · Case Study

Clinical Trial Efficacy: Student's t 💊

Does a blood-pressure drug actually work? The t-distribution is the tool for testing means when sigma is unknown, but the answer here carries a warning: comparing to zero can fool you, and only the placebo-controlled test tells the truth.

Statistics, Data Science and AI: A Visual Handbook · John Fisher · 2026

⚙️ Setup & data

import numpy as np, pandas as pd, matplotlib.pyplot as plt
from scipy import stats
AMBER="#d97706"; TEAL="#0d9488"; INK="#1a2138"; GRID="#e6e9f2"; PINK="#db2777"
plt.rcParams.update({"figure.facecolor":"white","axes.facecolor":"white","figure.dpi":110,"font.size":11,
   "axes.edgecolor":GRID,"axes.grid":True,"grid.color":GRID,"axes.axisbelow":True,"axes.spines.top":False,
   "axes.spines.right":False,"axes.titlesize":12,"axes.titleweight":"bold","legend.frameon":False})
BASE="https://raw.githubusercontent.com/johnfisher-ai/Statistics-Data-Science-AI-Visual-Book/main/data/"
try:    df = pd.read_csv("../../data/biomedical_clinical_trials.csv")
except FileNotFoundError: df = pd.read_csv(BASE+"biomedical_clinical_trials.csv")
print("loaded:", df.shape)
df.head()

loaded: (1000, 5)

	patient_id	treatment_group	baseline_systolic_bp	post_treatment_systolic_bp	bp_change_mmHg
0	PT_1000	Placebo	140.5	131.3	-9.2
1	PT_1001	Placebo	140.4	142.7	2.3
2	PT_1002	Placebo	119.4	117.3	-2.1
3	PT_1003	Placebo	129.6	134.6	5.0
4	PT_1004	Placebo	131.4	133.3	1.9

BEAT 1 · THE ONE-SAMPLE t-TEST

Did blood pressure drop in the treatment arm?

The t-distribution tests a mean when the population sigma is unknown, which is always. We test H0: mean bp_change = 0 in the Active arm, the obvious first question: did patients improve?

active = df[df.treatment_group=="Active_Compound_X"]["bp_change_mmHg"]
placebo = df[df.treatment_group=="Placebo"]["bp_change_mmHg"]
print(f"treatment groups: {df.treatment_group.value_counts().to_dict()}")
tstat, pval = stats.ttest_1samp(active, popmean=0)
print(f"\nActive arm (n={len(active)}): mean bp change = {active.mean():.3f} mmHg, sd = {active.std():.3f}")
print(f"t = (xbar - 0)/(s/sqrt(n)) = {tstat:.2f},  df = {len(active)-1},  p-value = {pval:.2e}")
print("-> reject H0: BP dropped significantly in the treatment arm" if pval<0.05 else "-> not significant")

treatment groups: {'Placebo': 515, 'Active_Compound_X': 485}

Active arm (n=485): mean bp change = -2.580 mmHg, sd = 4.010
t = (xbar - 0)/(s/sqrt(n)) = -14.17,  df = 484,  p-value = 2.31e-38
-> reject H0: BP dropped significantly in the treatment arm

Blood pressure fell by about 2.6 mmHg in the treatment arm, and the one-sample t-test is overwhelmingly significant (p far below 0.05). The t-statistic, (x̄ − 0)/(s/√n), uses the sample standard deviation s, which is exactly why the t-distribution is the right reference. Case closed, the drug works? Not so fast.

BEAT 2 · THE PLACEBO TRAP

Comparing to zero is not enough

Efficacy means beating PLACEBO, not just changing from baseline. Blood pressure drifts on its own (regression to the mean, the placebo effect). The two-sample t-test compares the arms directly.

tstat2, pval2 = stats.ttest_ind(active, placebo, equal_var=False)
print(f"Active  arm: mean bp change = {active.mean():.3f} mmHg")
print(f"Placebo arm: mean bp change = {placebo.mean():.3f} mmHg")
print(f"difference (Active - Placebo) = {active.mean()-placebo.mean():+.3f} mmHg")
print(f"two-sample t = {tstat2:.3f}, p-value = {pval2:.3f}")
na, npl = len(active), len(placebo)
sp = np.sqrt(((na-1)*active.std(ddof=1)**2 + (npl-1)*placebo.std(ddof=1)**2)/(na+npl-2))   # pooled SD
print(f"Cohen d = {(active.mean()-placebo.mean())/sp:+.2f}  (effect size; |0.2| small, |0.5| medium, |0.8| large)")
print("-> Active significantly beats Placebo" if pval2<0.05 else "-> NO significant difference: the drug adds nothing beyond placebo")

Active  arm: mean bp change = -2.580 mmHg
Placebo arm: mean bp change = -2.737 mmHg
difference (Active - Placebo) = +0.157 mmHg
two-sample t = 0.640, p-value = 0.522
Cohen d = +0.04  (effect size; |0.2| small, |0.5| medium, |0.8| large)
-> NO significant difference: the drug adds nothing beyond placebo

fig,ax=plt.subplots(figsize=(6.5,3.2))
ax.boxplot([placebo,active],tick_labels=["Placebo","Active_Compound_X"],patch_artist=True,
           boxprops=dict(facecolor="#fde68a"),medianprops=dict(color=PINK,linewidth=2))
ax.axhline(0,color=INK,ls=":",lw=1); ax.set_ylabel("BP change (mmHg)"); ax.set_title("BP change: Placebo vs Active (nearly identical)")
plt.tight_layout(); plt.show()

Here is the twist: the placebo arm dropped just as much (−2.74 vs −2.58 mmHg), and the two-sample t-test gives p ≈ 0.52, no significant difference. The apparent improvement in Beat 1 was real but not caused by the drug; blood pressure fell in both groups. This is the entire reason trials are placebo-controlled: testing against zero confuses a real-but-spurious drop with genuine efficacy, and only the comparison reveals the truth.

BEAT 3 · WHY t, NOT z?

Heavier tails for small samples

The t looks like the normal but with heavier tails that fatten as the sample shrinks. Its critical value exceeds 1.96, demanding stronger evidence when sigma is estimated from few data.

for n in [5,10,15,30,100]:
    tcrit=stats.t.ppf(0.975,n-1); print(f"n={n:3d} (df={n-1:3d}): 95% t critical = {tcrit:.3f}  vs normal z = 1.960")

n=  5 (df=  4): 95% t critical = 2.776  vs normal z = 1.960
n= 10 (df=  9): 95% t critical = 2.262  vs normal z = 1.960
n= 15 (df= 14): 95% t critical = 2.145  vs normal z = 1.960
n= 30 (df= 29): 95% t critical = 2.045  vs normal z = 1.960
n=100 (df= 99): 95% t critical = 1.984  vs normal z = 1.960

At n = 5 the 95% critical value is 2.78, far above the normal's 1.96, the t's way of demanding more evidence when σ is estimated from few points. As n grows the gap closes, and by n = 100 the t is essentially the normal. Using z on a tiny trial would overstate significance; the t keeps small-sample conclusions honest.

---

✅ Case closed

A one-sample t-test showed BP dropped in the treatment arm, but the two-sample test against placebo (p ≈ 0.52) revealed the drug adds nothing: both arms fell alike. The t-distribution is the right tool when sigma is unknown, and the placebo control is what separates a real effect from a hopeful one.

Statistics, Data Science and AI: A Visual Handbook · © 2026 John Fisher

←Back to the bookChapter 59 · Clinical Trial Efficacy (Student's t)