-
Notifications
You must be signed in to change notification settings - Fork 0
Comparison with Other Methods
This page compares urn randomization with other common randomization methods used in clinical trials. Each claim is referenced to the supporting literature.
| Feature | Simple | Permuted Block | Stratified Block | Minimization | Urn (Wei 1978) |
|---|---|---|---|---|---|
| Balance guarantee | None | Within blocks | Within strata | Strong | Adaptive |
| Allocation predictability | None | High | High | High (deterministic component) | Low |
| Many covariates (>3) | N/A | N/A | Poor | Good | Good |
| Small trials (n<100) | Poor balance | Good | Limited strata | Good | Good |
| Large trials (n>200) | Good | Good | Good | Good | Good |
| Pre-generated lists needed | No | Yes | Yes | No | No |
| Implementation | Trivial | Simple | Moderate | Moderate | Moderate |
Method: Each participant is assigned to a treatment with fixed probability (e.g., 50/50 for two arms), independent of all previous assignments.
Pros:
- Maximum allocation randomness — correct guess probability is exactly 0.50 (Zhao, 2014)
- Simplest to implement and audit
- Eliminates any possibility of selection bias
Cons:
- No balance guarantee — may result in unequal numbers of individuals assigned to each arm, especially in small samples (Efron, 1971)
- No mechanism to balance prognostic factors
When to choose: Rosenberger & Lachin (2016) note that simple randomization is appropriate for large trials (n > 100) where natural convergence is sufficient.
Urn comparison: Wei & Lachin (1988) showed that the urn design is "a compromise between designs that yield perfect balance in treatment assignments and complete randomization" and that as trial size increases, the urn design "approaches complete randomization." This means urn randomization offers balance when it matters most (early in the trial) without sacrificing much randomness overall.
Method: Participants are assigned within fixed-size blocks where each block contains a pre-specified number of each treatment in random order. For example, a block of 4 for a two-arm trial contains exactly 2 assignments to each arm.
Pros:
- Guarantees exact balance at every block boundary
- The entire allocation sequence can be pre-generated
- Widely accepted by regulators
Cons:
- Allocation predictability: "If the block size is known to trial personnel and the intervention group is revealed after assignment, then the last allocation within each block can always be predicted" — this is a unique property of permuted blocks where assignments become deterministic (Berger et al., 2003; NCI Randomization Tool)
- Selection bias risk: Schulz et al. (1995) demonstrated that trials with inadequate allocation concealment overestimated treatment effects by approximately 41%
- Variable block sizes mitigate but do not eliminate predictability (Matts & Lachin, 1988)
- Zhao et al. (2012) found that permuted block designs have the highest correct guess probability among common designs, making them the most vulnerable to selection bias
When to choose: Double-blinded trials where investigators cannot observe assignments, or when regulatory requirements specifically mandate block randomization.
Urn comparison: Wei & Lachin (1988) showed the urn design "is not as vulnerable to experimental bias as are other restricted randomization procedures." Unlike permuted blocks, the urn design has no deterministic assignments — the allocation probability shifts continuously but never reaches 0% or 100% for any arm (Zhao et al., 2012). Note: The NCI Clinical Trial Randomization Tool describes a related but distinct method, the "block urn design" (Zhao, 2011), as "analogous to permuted blocks but much improved in that the blocks 'reset' under certain conditions, thereby reducing overall predictability."
Method: Separate permuted block randomization within each stratum defined by combinations of prognostic factor levels.
Pros:
- Achieves balance within each prognostic subgroup
- Combines the strengths of stratification and blocking
Cons:
- Combinatorial explosion: With k factors having l levels each, there are l^k strata. Kalish & Begg (1985) noted that stratified randomization begins to fail when the number of strata approaches half the sample size. With 4 factors having 3 levels each there are 81 strata — most will have very few or zero participants.
- Incomplete blocks: Sparse strata will end mid-block, negating the balance guarantee
- Practical limit: Typically feasible for only 2–3 stratification factors (Kalish & Begg, 1985)
- Inherits the predictability problem from the underlying block design
When to choose: When a small number of critical factors (1–3) must be perfectly balanced within subgroups.
Urn comparison: Covariate-adaptive methods like urn randomization can better handle the problem of increasing numbers of covariates compared to stratified approaches (Rosenberger & Lachin, 2016). The urn design maintains per-factor urns rather than per-stratum blocks, avoiding the combinatorial problem entirely.
Method: For each new participant, calculate how overall factor balance would change under each possible treatment assignment. Assign the treatment that minimizes imbalance, typically with a random component (e.g., p=0.75 for the best arm).
Pros:
- Achieves the strongest balance across multiple factors simultaneously
- Handles many factors well without combinatorial explosion
- Can be started mid-trial
Cons:
- Allocation predictability: Treatment assignments can become highly predictable (Scott et al., 2002). With p=1 (fully deterministic), the next assignment can be predicted with certainty by anyone who knows the current marginal totals. Even with p<1, the favored arm is often apparent.
- Scott et al. (2002) noted that predictability concerns also apply to other methods, including stratified randomization, but the debate remains unresolved.
- The theoretical properties of test statistics under minimization are less well-established than under urn randomization (Rosenberger & Lachin, 2016)
When to choose: When factor balance is the highest priority and the trial is double-blinded (making selection bias from predictability less of a concern).
Urn comparison: Both methods handle many factors and adapt in real time. The key difference is the balance–randomness tradeoff. Minimization achieves better balance but at the cost of higher predictability. Urn randomization provides "tight control of balance in the early phase" (Wei & Lachin, 1988) with lower predictability. Wei & Lachin (1988) developed formal permutational test statistics for the urn design, giving it stronger theoretical grounding for inference.
Method: When groups are balanced, assign with equal probability. When one group is smaller, use a biased probability (e.g., 2/3) favoring the smaller group.
Pros:
- Simple adaptive mechanism with a single tuning parameter
- Less predictable than block randomization
- Well-studied theoretical properties (Efron, 1971)
Cons:
- Balances only overall treatment counts — does not address prognostic factor balance
- The bias probability is constant regardless of the degree of imbalance
When to choose: Two-arm trials without important prognostic factors where better-than-simple balance is desired.
Urn comparison: Wei's (1978) urn design can be viewed as a generalization of Efron's biased coin to multiple treatments and multiple factors. The urn mechanism provides a smooth, graded response to imbalance rather than a binary switch, and it naturally extends to factor-level balance through per-factor urns.
Based on the literature, urn randomization is well-suited when:
- You have 3+ prognostic factors — stratified methods become impractical due to the combinatorial explosion of strata (Kalish & Begg, 1985)
- Unpredictability matters — especially in open-label or single-blinded trials where selection bias is a concern (Wei & Lachin, 1988)
- The trial is small to moderate (n < 200) — where the urn provides tight control of balance in the early phase (Wei & Lachin, 1988) while simple randomization may produce significant imbalance
- Subgroup analyses are planned — per-factor urn balance improves power for subgroup comparisons (Wei & Lachin, 1988)
- You want valid permutational inference — Wei & Lachin (1988) derived large-sample test statistics specifically for the urn design
- Regulatory requirement for blocks: Some IRBs or regulators specifically require permuted block randomization. Use variable block sizes and ensure allocation concealment.
- Perfect balance is paramount and the trial is blinded: Minimization achieves stronger balance. In double-blinded trials, the higher predictability is less of a concern.
- Very large trial with few factors: Simple randomization (n > 200) is the least controversial and may suffice when balance isn't a major concern.
- Berger, V.W., Ivanova, A., & Deloria-Knoll, M. (2003). Minimizing predictability while retaining balance through the use of less restrictive randomization procedures. Statistics in Medicine, 22(19), 3017–3028.
- Efron, B. (1971). Forcing a Sequential Experiment to be Balanced. Biometrika, 58(3), 403–417.
- Kalish, L.A. & Begg, C.B. (1985). Treatment allocation methods in clinical trials: a review. Statistics in Medicine, 4(2), 129–144.
- Matts, J.P. & Lachin, J.M. (1988). Properties of permuted-block randomization in clinical trials. Controlled Clinical Trials, 9(4), 327–344.
- NCI Division of Cancer Prevention. Clinical Trial Randomization Tool. https://prevention.cancer.gov/ctrandomization/about/
- Pocock, S.J. & Simon, R. (1975). Sequential Treatment Assignment with Balancing for Prognostic Factors in the Controlled Clinical Trial. Biometrics, 31(1), 103–115.
- Rosenberger, W.F. & Lachin, J.M. (2016). Randomization in Clinical Trials: Theory and Practice. 2nd ed. Wiley.
- Schulz, K.F., Chalmers, I., Hayes, R.J., & Altman, D.G. (1995). Empirical evidence of bias: dimensions of methodological quality associated with estimates of treatment effects in controlled trials. JAMA, 273(5), 408–412.
- Scott, N.W., McPherson, G.C., Ramsay, C.R., & Campbell, M.K. (2002). The method of minimization for allocation to clinical trials: a review. Controlled Clinical Trials, 23(6), 662–674.
- Wei, L.J. (1978). An Application of an Urn Model to the Design of Sequential Controlled Clinical Trials. Journal of the American Statistical Association, 73(363), 559–563.
- Wei, L.J. & Lachin, J.M. (1988). Properties of the urn randomization in clinical trials. Controlled Clinical Trials, 9(4), 345–364.
- Zhao, W. & Weng, Y. (2011). Block urn design - a new randomization algorithm for sequential trials with two or more treatments and balanced or unbalanced allocation. Contemporary Clinical Trials, 32(6), 953–961.
- Zhao, W. (2014). A better alternative to stratified permuted block design for subject randomization in clinical trials. Statistics in Medicine, 33(30), 5239–5248.
- Zhao, W., Weng, Y., Wu, Q., & Palesch, Y. (2012). Quantitative comparison of randomization designs in sequential clinical trials based on treatment balance and allocation randomness. Pharmaceutical Statistics, 11(1), 39–48.