When Conventional Adjustment Is Adequate

When Conventional Adjustment Can Work

With repeated treatment decisions, a covariate measured between decisions may help determine later treatment and predict the outcome. If earlier treatment also changes that covariate, directly adjusting for it can alter the causal contrast or introduce bias. But this difficulty is not inevitable: conventional adjustment can be adequate when the covariates used for adjustment are unaffected by earlier treatment, and when the treatment decisions are adequately controlled by those covariates.

The key is to separate two questions. First, does the covariate’s timing and causal role make ordinary adjustment appropriate for the contrast of interest? Second, are the treatment groups comparable after adjustment, and is there enough treatment variation within the adjustment strata? A “yes” to the first question does not guarantee a “yes” to the second.

Definition: Treatment-confounder feedback occurs when a prior treatment affects a later covariate that, in turn, predicts a subsequent treatment and the outcome. For the settings considered here, there is no treatment-confounder feedback when every covariate used for adjustment is unaffected by prior treatment. In particular, if \(L_1\) is measured after \(A_0\), no feedback requires that its value would be the same under either value of \(A_0\).

Use \(L_0\) for baseline covariates, \(A_0\) and \(A_1\) for treatments at two successive decisions, and \(L_1\) for covariates measured between those decisions. Let \(Y(a_0,a_1)\) denote the outcome that would occur under treatment sequence \((a_0,a_1)\). A common target is the mean outcome difference between two sequences, such as \((1,1)\) and \((0,0)\). The notation \(L_1\) is useful even when that covariate is treatment-invariant: it records when it was measured, not whether it was caused by prior treatment.

Assumptions That Make Adjustment Valid

Conventional adjustment requires more than an absence of feedback. The covariates must control confounding at the treatment decisions, the treatment sequences being compared must be possible within the relevant covariate strata, and the observed outcome must correspond to the potential outcome under the treatment actually received.

Conditions for the two-decision adjustment formula: Let \(L=(L_0,L_1)\), and suppose \(L_1\) is unaffected by \(A_0\). For each treatment sequence of interest, assume:

Consistency: a person who receives \((A_0,A_1)=(a_0,a_1)\) has observed outcome \(Y=Y(a_0,a_1)\).
Exchangeability at the first decision: conditional on \(L_0,L_1\), the first treatment is independent of the potential outcomes for the sequences being compared.
Exchangeability at the second decision: among people with \(A_0=a_0\), conditional on \(L_0,L_1\), the second treatment is independent of the potential outcome under \((a_0,a_1)\).
Positivity: within every covariate stratum given positive probability in the target population, each treatment decision needed for the sequence has positive probability of taking the required value.

The first exchangeability condition is stated conditional on both \(L_0\) and \(L_1\), not just on baseline \(L_0\). This detail matters if the adjustment formula conditions on \(L_1\). Independence conditional on \(L_0\) alone does not, in general, remain true after further conditioning on \(L_1\). The fact that \(L_1\) is unaffected by treatment makes ordinary adjustment conceptually available; it does not by itself guarantee exchangeability within the \(L_1\) strata.

Under these assumptions, the mean outcome under a specified sequence can be obtained by estimating the outcome mean in each covariate stratum among people who received that sequence, then averaging those means over the target population’s covariate distribution:

\mu(a_0,a_1) = \sum_{l_0,l_1} E\!\left[Y\mid A_0=a_0,A_1=a_1,L_0=l_0,L_1=l_1\right] P(L_0=l_0,L_1=l_1).

For continuous covariates, the sum is replaced by an integral. The formula is a standardization: each treatment sequence is evaluated against the same covariate distribution. A contrast such as \(\mu(1,1)-\mu(0,0)\) then compares sequence-specific mean outcomes in the target population, rather than comparing the covariate distributions that happened to occur in the observed treatment groups.

Why the Formula Follows

The argument is a sequence of conditional-mean substitutions. Consistency links the observed outcome to the relevant potential outcome for people who received the specified sequence. Exchangeability at each decision allows the corresponding observed treatment group to represent that potential outcome within the covariate stratum. Positivity ensures that those observed groups exist in every stratum that contributes to the target mean.

\begin{aligned} E[Y(a_0,a_1)] &= \sum_{l_0,l_1} E[Y(a_0,a_1)\mid L_0=l_0,L_1=l_1]P(l_0,l_1)\\ &= \sum_{l_0,l_1} E[Y(a_0,a_1)\mid A_0=a_0,L_0=l_0,L_1=l_1]P(l_0,l_1)\\ &= \sum_{l_0,l_1} E[Y(a_0,a_1)\mid A_0=a_0,A_1=a_1,L_0=l_0,L_1=l_1]P(l_0,l_1)\\ &= \sum_{l_0,l_1} E[Y\mid A_0=a_0,A_1=a_1,L_0=l_0,L_1=l_1]P(l_0,l_1). \end{aligned}

The first equality averages over the target population’s covariate distribution. The second uses first-decision exchangeability conditional on \(L_0,L_1\). The third uses second-decision exchangeability among those with \(A_0=a_0\). The final equality uses consistency. This derivation also shows why the first exchangeability condition cannot be weakened to conditioning on \(L_0\) alone when the displayed formula conditions on \(L_1\).

Worked Examples

Worked Example: Baseline Adjustment for a Treatment Contrast

Suppose a hypothetical study compares a treatment with control. A baseline risk category \(L_0\) is low for half the target population and high for the other half. Assume, for illustration, that the study’s conditional outcome means are as follows.

Baseline risk	Mean outcome, treated	Mean outcome, control
Low	8 units	6 units
High	12 units	10 units

Assume treatment is exchangeable with respect to the potential outcomes within each baseline-risk category, both treatments are possible in each category, and consistency holds. There are no treatment-affected covariates in this one-decision example. Standardize both treatment means to the target population, which has proportions \(0.5\) and \(0.5\) in the two categories:

\begin{aligned} \mu(1)&=(0.5)(8)+(0.5)(12)=4+6=10\text{ units},\\ \mu(0)&=(0.5)(6)+(0.5)(10)=3+5=8\text{ units},\\ \mu(1)-\mu(0)&=10-8=2\text{ units}. \end{aligned}

Under the stated assumptions, the estimated treatment contrast is an average increase of 2 outcome units in this target population. The calculation does not establish exchangeability or positivity; those are assumptions about the study and its data-generating process, not conclusions obtained from these four means.

Worked Example: Two Decisions With a Treatment-Invariant Intermediate Covariate

Consider a hypothetical two-decision study with sequence \((1,1)\) compared with \((0,0)\). The intermediate covariate \(L_1\) is low for 60% of the target population and high for 40%. It is measured after \(A_0\), but the study’s substantive assumptions say that \(A_0\) does not affect \(L_1\). Suppose the conditional mean outcomes for the two sequences are:

\(L_1\) stratum	Mean for sequence (1,1)	Mean for sequence (0,0)
Low	14 units	10 units
High	20 units	15 units

Assume consistency, positivity for both decisions in each stratum, first-decision exchangeability conditional on \(L_1\), and second-decision exchangeability conditional on \(L_1,A_0\). In particular, do not infer the first condition merely from exchangeability conditional on some other baseline covariates. Standardize to the stated target distribution:

\begin{aligned} \mu(1,1)&=(0.6)(14)+(0.4)(20)=8.4+8=16.4\text{ units},\\ \mu(0,0)&=(0.6)(10)+(0.4)(15)=6+6=12\text{ units},\\ \mu(1,1)-\mu(0,0)&=16.4-12=4.4\text{ units}. \end{aligned}

The standardized contrast is 4.4 units. Although \(L_1\) was measured between treatment decisions, it can be used in this conventional adjustment because it is unaffected by \(A_0\) and the required exchangeability conditions hold. The timing of measurement alone is not sufficient: if \(L_1\) were affected by \(A_0\), this calculation would not generally estimate the total contrast between the two treatment sequences.

Worked Example: No Feedback Does Not Eliminate Confounding

In a hypothetical one-decision study, a prognostic covariate \(L\) is low or high, with half the target population in each category. Treatment does not affect \(L\). Within the low category, 20% receive treatment; within the high category, 80% receive it. The conditional outcome means are 6 units for treated and 4 for untreated people with low \(L\), and 12 units for treated and 10 for untreated people with high \(L\). Suppose the means are otherwise comparable within each \(L\) stratum.

The treated group’s population share is \((0.5)(0.2)+(0.5)(0.8)=0.5\). Of the full population, \((0.5)(0.2)=0.1\) is treated with low \(L\), and \((0.5)(0.8)=0.4\) is treated with high \(L\). Its crude mean is therefore:

\frac{(0.1)(6)+(0.4)(12)}{0.5} = \frac{0.6+4.8}{0.5} = 10.8\text{ units}.

The untreated group also has population share \(0.5\): 0.4 of the full population is untreated with low \(L\), and 0.1 is untreated with high \(L\). Its crude mean is:

\frac{(0.4)(4)+(0.1)(10)}{0.5} = \frac{1.6+1.0}{0.5} = 5.2\text{ units}.

The crude difference is \(10.8-5.2=5.6\) units. But the adjusted means use the same half-low, half-high target distribution: \((0.5)(6)+(0.5)(12)=9\) units treated and \((0.5)(4)+(0.5)(10)=7\) units untreated. The standardized contrast is \(9-7=2\) units. Thus, treatment-invariant covariates still need adjustment when treatment allocation differs across prognostic strata.

Regression Is a Tool, Not the Assumption

A conventional outcome regression can implement adjustment by including the relevant covariates and predicting outcomes under each treatment sequence. For example, one may fit a model for \(E[Y\mid A_0,A_1,L_0,L_1]\), set the treatments to each sequence in turn, predict for the same target-population covariate values, and average those predictions. This approach relies on adequate model specification as well as the causal assumptions above.

A regression coefficient is not automatically the standardized contrast. If the model includes treatment-by-covariate interactions, treatment effects can differ across strata, and a single treatment coefficient may not equal the target-population average contrast. Even without interactions, the coefficient’s interpretation depends on the model’s scale and coding. State the estimand first, then use predictions or another adjustment method that actually computes it.

Common Pitfalls and Practical Checks

Assuming no feedback means no confounding. Treatment may still be associated with prognosis through baseline or treatment-invariant covariates. Check exchangeability and adjust for the relevant confounders.
Using only baseline exchangeability when conditioning on an intermediate covariate. If the formula conditions on \(L_1\), require first-decision exchangeability conditional on \(L_0,L_1\), or provide a separate valid argument for the conditional-mean substitution. Conditioning on \(L_1\) does not preserve an independence statement known only given \(L_0\).
Ignoring positivity. If a treatment sequence is never observed in some target-population stratum, its conditional outcome mean cannot be learned from those data by ordinary adjustment.
Calling every post-baseline covariate a feedback variable. A covariate measured after treatment is not necessarily affected by it. The causal ordering and whether earlier treatment changes the covariate must be established, not inferred from the date alone.
Interpreting a fitted coefficient without specifying the contrast. Report which treatment sequences are compared, the target population, the outcome scale, and how predictions were averaged.

Key takeaway: Conventional adjustment is adequate for a stated longitudinal treatment contrast when the covariates used for adjustment are unaffected by prior treatment, treatment assignment is exchangeable conditional on the appropriate covariate history at each decision, positivity holds, and the estimation method correctly implements the target contrast. No feedback is an enabling condition, not a substitute for these other requirements.

Check Your Understanding

For each question, identify the assumption or calculation that matters most.

A covariate \(L_1\) is measured after \(A_0\) but is known to be unaffected by it. What additional assumptions are needed before adjusting for \(L_1\) in a two-decision sequence contrast?
Why does exchangeability conditional on \(L_0\) alone not necessarily justify a formula that compares treatment groups within \(L_0,L_1\) strata?
In a target population with 30% low risk and 70% high risk, conditional means are 5 and 9 under treatment and 3 and 8 under control, respectively. Calculate the standardized contrast.
What does positivity require for a treatment sequence in a covariate stratum that occurs in the target population?
When a regression includes treatment-by-covariate interactions, why might its treatment coefficient differ from a standardized average contrast?