Why a Missing Covariate Can Change the Design
An adaptive allocation rule uses information available before assignment to choose a probability of receiving each treatment. If a baseline covariate used by that rule is missing, the rule cannot use the participant’s true covariate profile as intended. The practical choice is not simply how to fill a blank in a data file: it is how the design will act when that blank is encountered. Different choices can produce different assignment probabilities, different balance patterns, and different records of the randomization process.
Here the focus is on covariates used for allocation. Handling missing covariates later in an outcome analysis is a related but distinct task. A strategy that is acceptable for making an assignment need not be adequate for estimating a treatment effect, and vice versa.
The rule must be implementable using information available at the decision time. It should be specified in advance, applied consistently, and recorded together with the information and probability used for each assignment. In particular, a value learned only after assignment cannot be used retrospectively to describe the probability that governed that assignment.
Three Strategies and Their Assumptions
Complete-case handling for allocation means using the covariate contribution only when that value is observed. For a participant whose value is missing, the allocation rule omits that covariate’s contribution; it does not omit the participant from randomization. This preserves the opportunity to use other observed baseline information, but it provides no balance control for the missing factor on that assignment.
Missing-category handling treats “missing” as an additional level of the covariate. For example, a binary risk factor with levels low and high gains a third allocation level, missing. The rule can then balance participants whose values are unavailable against other participants whose values are also unavailable. This does not identify their true risk level or balance them with the low- and high-risk groups.
Imputation handling uses a model or other prespecified rule to represent the unobserved covariate value or its distribution. It may use a single imputed value, a random draw from a predictive distribution, or probabilities over possible values. For allocation, the method must specify how uncertainty in the missing value affects the assignment probability.
In missing-data terminology, missing completely at random means the chance of a value being missing does not depend on observed or unobserved data. Missing at random means that, conditional on specified observed information, missingness does not additionally depend on the unobserved value. Missing not at random allows such a dependence to remain. These labels do not mechanically select an allocation strategy; they help assess whether an imputation model and the resulting use of covariate information are credible.
Worked Example: Omitting a Missing Covariate Contribution
Worked Example: Omitting a Missing Covariate Contribution
Suppose a hypothetical two-arm trial uses a minimization rule to balance a binary baseline risk factor. Among participants with observed values, the current counts are low risk: 5 assigned to A and 3 to B; high risk: 2 assigned to A and 4 to B. A new participant’s risk value is missing. For illustration, define the imbalance score as the sum, over levels, of the absolute differences in treatment counts. The rule assigns probability 0.5 to each arm when the candidate scores are tied, and otherwise favors the arm with the lower score.
Under complete-case handling, this participant contributes to neither the low-risk nor high-risk count. If assigned to A, the score remains \( |5-3|+|2-4|=2+2=4 \). If assigned to B, the score also remains \( |5-3|+|2-4|=2+2=4 \).
The scores tie, so this covariate provides no reason to favor either arm: the specified rule assigns \(P(A)=0.5\) and \(P(B)=0.5\). The participant is still randomized. The limitation is that the assignment does not use the risk factor to balance this participant against others.
Worked Example: Creating a Missing Category
Worked Example: Creating a Missing Category
Consider a second hypothetical trial using the same general minimization idea, but now with a separate missing level. Before the next assignment, the missing-level counts are 1 on A and 3 on B. Suppose all other covariate contributions are tied between candidate assignments, and the prespecified rule assigns probability 0.8 to the arm with the lower resulting imbalance and 0.2 to the other arm.
If the new participant is assigned to A, the missing-level counts become 2 and 3, giving imbalance \( |2-3|=1 \). If assigned to B, they become 1 and 4, giving imbalance \( |1-4|=3 \). The rule therefore assigns \(P(A)=0.8\) and \(P(B)=0.2\).
The missing category has changed the assignment probabilities even though the participant’s underlying risk is unknown. What it balances is the pattern of recorded missingness, not the unobserved risk values. For example, a group with many missing values could still contain more high-risk participants than the other treatment group.
Worked Example: Averaging Over Imputed Profiles
Worked Example: Averaging Over Imputed Profiles
Suppose a third hypothetical design has prespecified profile-specific allocation probabilities. For a participant with the observed information available, the rule would assign to A with probability 0.8 if the missing risk value were high and with probability 0.2 if it were low. An imputation model, fitted using information available before this assignment, gives probability 0.75 that the value is high and 0.25 that it is low.
If the design averages over the two possible profiles, the probability of assigning A is the weighted average \(0.75(0.8)+0.25(0.2)=0.60+0.05=0.65\). Thus \(P(A)=0.65\) and \(P(B)=1-0.65=0.35\). This probability accounts for uncertainty about the profile rather than acting as if one profile were known with certainty.
The calculation assumes the profile probabilities and profile-specific allocation probabilities are defined for the same participant and history. If the design instead draws one imputed value and then applies the rule, it must also define whether the assignment probability recorded is conditional on that draw or averaged over the imputation randomness. Those are different descriptions of the randomization process.
Why Probability Averaging Is Justified
The averaging calculation follows from conditioning on the unknown profile. Let \(X_i\) take possible values \(x\), and suppose the imputation model assigns probabilities \(q(x)\) that sum to 1. If the assignment rule would use probability \(p_A(x)\) for A under profile \(x\), then the probability of A before learning \(X_i\), under the imputation model, is
The first equality is the law of total probability: assignment to A can occur under each possible profile, and each profile is weighted by its probability. In the preceding example, the sum has two terms, \(0.8(0.75)+0.2(0.25)=0.65\). This derivation does not establish that the imputation model is correct. It only shows how to average the allocation probabilities once that model and the allocation rule have been specified.
Comparing the Approaches in Practice
| Approach | What the allocation rule uses | Potential benefit | Main limitation |
|---|---|---|---|
| Complete-case handling | The covariate only when observed | Simple; does not invent a value | No balancing contribution for missing values; observed-value balance may not represent everyone |
| Missing category | A separate level for unavailable values | Explicitly tracks and may balance missingness | Does not balance the unknown true covariate levels |
| Imputation | A predicted value or distribution over values | Can use other observed information to inform allocation | Depends on model quality and careful handling of uncertainty |
These are alternatives, not a universal ranking. Complete-case handling is transparent when missing values are rare, but systematic missingness can leave the very profiles of interest less controlled. A missing category preserves a distinct record of unavailable values, but it can create a small, unstable level and only balances the fact of missingness. Imputation may make more use of observed covariates, but a misspecified model can direct assignments using misleading profile probabilities. No method guarantees exact balance.
The choice also affects the meaning of the allocation history. If an imputed profile determines the assignment probability, the design record should retain the observed inputs, the imputation method and result (or distribution), the allocation probability actually used, and the assignment. If a missing category is used, record that the category was missing at decision time rather than replacing it later with a value and rewriting the earlier decision.
Diagnostics, Failure Modes, and a Common Mistake
Useful checks focus on whether the procedure is behaving as specified. Tabulate missingness over time, by observed covariate profiles, and by assigned treatment. Recalculate a sample of assignments from the recorded inputs and the prespecified rule. For imputation, inspect whether predicted profiles are plausible and whether the model is being applied to participant types it can reasonably predict. Look for unexpected changes in allocation probabilities when the frequency of missingness changes.
A common mistake is to say, “The missing value was imputed as high risk, so this participant was high risk when randomized.” That treats an uncertain prediction as an observed fact. The correction is to state exactly what the design did: it used a single imputed value, sampled an imputed value, or averaged over a predictive distribution. The imputation assumption and the probability used for assignment should be recorded, and later analysis should not mistake the imputed value for a measurement.
Another operational failure is to use a value obtained after assignment to reconstruct the allocation decision as though it had been available beforehand. The assignment probability is determined by the inputs and randomization procedure at the time of assignment. Later information may be relevant to data analysis, but it does not change that historical probability.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Under complete-case handling, what happens to a participant whose allocation covariate is missing? Are they excluded from randomization?
- A missing-category level has counts 2 on A and 5 on B. A new participant with a missing value is considered for assignment. What are the resulting missing-level imbalances under assignment to A and to B?
- A profile-specific rule gives \(P(A)=0.7\) for a high value and \(P(A)=0.3\) for a low value. The imputation model assigns probabilities 0.4 and 0.6 to high and low. What is the averaged probability of A?
- What does balancing a missing category establish, and what does it not establish?
- Why should the allocation record retain the probability and information used at the time of assignment?
Solutions. (1) The rule omits that covariate’s contribution, but the participant is still randomized using the remaining specified procedure. (2) Assignment to A gives counts 3 and 5, so imbalance \( |3-5|=2 \); assignment to B gives counts 2 and 6, so imbalance \( |2-6|=4 \). (3) The averaged probability is \(0.4(0.7)+0.6(0.3)=0.28+0.18=0.46\), so \(P(A)=0.46\). (4) It can balance how many participants have unavailable values across arms; it does not establish balance of their unknown underlying covariate values. (5) The record allows the actual assignment mechanism to be checked and described using only information available when the decision was made; later data must not be substituted for the inputs that governed randomization.