Compare independent means without assuming equal variances

Try

Two independently sampled groups have means 25 and 20 units. Their sample sizes are 10 and 15, and both sample SDs happen to be 3. Equal observed SDs do not prove equal population variances. Start with a method that estimates each variance contribution separately rather than imposing equality unnecessarily.

Predict

Predict how each sample size affects uncertainty in the difference. A noisy small group can dominate the standard error. Also predict why n1+n2−2 is not automatically the correct degrees of freedom when separate variances are used. The reference must reflect uncertainty in both estimated variance contributions.

Experiment

The estimated difference is 5. Let a=s1²/n1=9/10=0.9 and b=s2²/n2=9/15=0.6. Then SE=√(a+b)=√1.5≈1.224745. Welch degrees of freedom are (a+b)²/[a²/(n1−1)+b²/(n2−1)]=2.25/(0.81/9+0.36/14)=175/9≈19.444.

For illustration use a supplied conservative critical value t*=2.093 from df=19. The margin is about 2.56339 and the interval is approximately [2.43661,7.56339]. State that reference choice explicitly. A calculator using the fractional df directly would give slightly different endpoints. For a zero-difference null, t=5/√1.5≈4.082, exceeding the supplied two-sided 5% cutoff 2.093. Reject that null under the stated model; this critical-value comparison does not require inventing an exact p-value.

Understand

Welch inference estimates μ1−μ2 for independent quantitative groups without requiring equal population variances. Its standard error is √(s1²/n1+s2²/n2). The test statistic subtracts a specified null difference from x̄1−x̄2 and divides by that SE. The interval uses the same separate-variance estimate with an appropriate t critical value.

The Welch–Satterthwaite degrees-of-freedom approximation is ν=(a+b)²/[a²/(n1−1)+b²/(n2−1)], where a and b are the two estimated variance contributions. It generally is not an integer. If a printed table requires an integer row, using a lower df can be conservative; document the convention instead of rounding silently. Degenerate zero-variance cases need separate judgment rather than automatic division by zero.

Classical equal-variance pooled t inference uses a different variance estimator and df=n1+n2−2 under its own assumptions. Pooling should not be chosen merely because sample sizes are equal or an informal variance comparison appears nonsignificant. Welch is a useful default for many independent-group comparisons, but it still needs appropriate design and distributional conditions.

Check independence within and between groups, representative sampling or relevant randomized assignment, and suitable population shapes. Small groups with severe skew or extreme observations can make approximate t inference unreliable. Unequal sample sizes and spreads can amplify problems. Examine plots and provenance before relying on summary statistics alone.

The same people measured twice do not satisfy the independent-group setup. Their means are correlated, and variance of their difference includes covariance. Paired analysis works with individual differences and their observed spread. Treating a paired study as two independent samples may misrepresent uncertainty even when the group sizes match.

Interpret the interval in the chosen orientation and units. A positive difference is not automatically causal in observational data, nor practically useful merely because zero is excluded. Include an effect threshold and design limitations when supporting a decision.

Apply

Use the practice samples to compute separate variance contributions before taking a square root. Verify the supplied df using the formula, then calculate the interval. Finally, explain to a colleague why a low pooled-variance estimate cannot justify assuming equal population variances without considering the design and model.

Continue

Next choose among methods without being told which formula to use. Begin with outcome type, target and dependence rather than searching for a familiar number pattern.