How Contrast Signs Determine Dependence

Why Contrast Direction Matters

In a multi-arm trial, two estimated treatment contrasts can use some of the same group means. Their estimates are then generally dependent, even when the groups themselves are sampled independently. The direction in which a shared group enters each contrast matters: if both contrasts subtract the group, its sampling variation moves both estimates in the same direction; if one adds it and the other subtracts it, that variation moves the estimates in opposite directions.

The central task here is to read that dependence from the contrast coefficients. A shared group does not guarantee positive covariance: its contribution can be positive or negative, and contributions from several shared groups can offset one another. This is why checking only whether contrasts share an arm is not enough.

Assumptions and the Shared-Group Contribution

Use the following sampling model. Group \(g\) contains \(n_g\) observations \(Y_{gi}\), each with variance \(\sigma_g^2\). Observations are independent within each group, and the groups are independent of one another. The within-group observations have a common variance, though that variance may differ across groups. These assumptions imply \(\operatorname{Var}(\bar Y_g)=v_g=\sigma_g^2/n_g\), and the group means are uncorrelated across distinct groups. Normal outcomes are not required for these covariance calculations.

For two contrasts \(L_1=\sum_g c_g\bar Y_g\) and \(L_2=\sum_g d_g\bar Y_g\), expand their covariance by bilinearity. Terms involving different group means vanish because those means are uncorrelated. For a group included in both contrasts, the remaining term is the product of its two coefficients times its group-mean variance:

This expression is useful here as a group-by-group sign check. If only one group is shared, the covariance is \(c_gd_gv_g\) for that group. Since \(v_g>0\), the sign is determined by whether its two coefficients have the same or opposite signs. If several groups are shared, add their contributions: same-sign coefficient pairs contribute positively, opposite-sign pairs negatively, and exact cancellation can give zero covariance.

The condition \(\sum_g c_g=0\) makes a contrast insensitive to adding the same constant to every group mean. It does not determine the sign of covariance between two contrasts. That sign depends on their shared coefficients and the associated variances \(v_g\).

Opposite Signs and Negative Covariance

Reversing the direction of one comparison reverses the sign of its coefficient on the shared group. The same sampling fluctuation can then push the two estimates in opposite directions. The sign describes dependence between the estimates, not whether the underlying treatment effects are positive or negative.

When Contributions Cancel

With more than one shared group, coefficient signs give the sign of each contribution, not necessarily the sign of the total covariance. Contributions are weighted by the relevant group-mean variances, so equal and opposite coefficient products need not cancel if the group variances differ.

How to Read the Signs Reliably

A reliable calculation keeps the coefficient signs attached to the contrast as written. For each shared group, multiply the coefficient in the first contrast by the coefficient in the second, then weight that product by the group-mean variance. Add the resulting contributions. A group appearing in just one contrast has coefficient zero in the other and contributes nothing.

A useful comparison is between sign inspection and a direct calculation from the group means. Direct expansion can verify the result, but the coefficient approach makes the source of each term visible and scales more cleanly when many arms are involved. Neither method can repair a wrong sampling model: if group means are correlated across groups, the cross-group covariance terms that were set to zero must be included.

Diagnostics, Limits, and a Common Pitfall

Before interpreting a computed sign, check that the model supports the calculation. Within each group, independent observations with common variance give \(\operatorname{Var}(\bar Y_g)=\sigma_g^2/n_g\); independence across groups gives zero covariance between distinct group means. If observations within a group are dependent, this variance formula need not hold. If groups are related, the omitted cross-group covariance terms can change the result.

Also check whether you are discussing covariance or correlation. Covariance magnitude depends on the measurement scale and the variances of shared group means. Correlation is standardized, but it still reflects the same coefficient-driven direction. Reversing the sign of an entire contrast reverses its covariance and correlation with another contrast; it does not change the underlying data.

• Common mistake: “These contrasts share the control group, so their covariance must be positive.”
• Correction: Inspect the control coefficients. Two negative coefficients give a positive control contribution; opposite signs give a negative contribution. With multiple shared groups, sum all weighted contributions before deciding the total sign.
• Another pitfall: Treating zero covariance as proof of independence without a joint-normality condition.

These calculations concern sampling dependence between contrast estimates. They do not imply that the observed patient outcomes are correlated across independently assigned groups, nor do they establish a causal relation between treatment effects. They describe how shared estimated group means contribute uncertainty to the contrasts.