Completely Randomized Design (CRD)
Analyze single-factor experiments where treatments are assigned to experimental units entirely at random. Build the ANOVA table, test treatment effects with the F-statistic, and interpret results for balanced or unbalanced designs.
Run CRD Analysis →What is a Completely Randomized Design?
A Completely Randomized Design (CRD) is the simplest design of experiments layout. A single factor is studied at two or more levels (treatments), and experimental units are allocated to those treatments completely at random, with no blocking or other restriction on randomization.
Because randomization is unrestricted, CRD is appropriate when the experimental units are homogeneous, that is, when no major nuisance source of variation is known in advance. Any remaining unit-to-unit differences are absorbed into the experimental error term.
CRD is analyzed with one-way ANOVA. The total variation in the response is partitioned into variation between treatments and variation within treatments (error), and their mean squares are compared through an F-test.
In plain terms: You have one thing you're changing (the factor), several settings for it (treatments), and you randomly decide which test unit gets which setting. Then you check whether the settings actually produced different average results, or whether the differences are just noise.
Design Structure
Single Factor, k Treatments
One controllable factor is tested at k levels. Example: curing temperature at 150, 175 and 200 °C. Each treatment is applied to one or more experimental units chosen at random.
Balanced vs Unbalanced
A balanced CRD has equal replications per treatment (equal n). Unequal replications give an unbalanced CRD, which is still valid but has lower power and is more sensitive to unequal variances.
Replication & Randomization
Replication provides an estimate of pure error; randomization protects against bias from lurking variables. Both are essential for valid CRD inference.
Key Formulas
Interpreting the F-test
The null hypothesis is that all treatment means are equal (τ1 = τ2 = … = 0). A large F relative to the critical value means between-treatment variation exceeds random error, so at least one treatment mean differs.
A significant F is an omnibus result: it does not say which treatments differ. Follow a significant CRD with post-hoc comparisons (Tukey HSD) to locate the specific differences.
Assumptions & Validation
Independence
Errors must be independent, guaranteed by random assignment of units to treatments.
If violated: Re-examine the randomization; correlated units require a mixed or repeated-measures model.
Normality
Residuals should be approximately normal within each treatment.
If violated: For small samples use the non-parametric Kruskal-Wallis test.
Homogeneity of Variance
Treatment groups should have similar variance (homoscedasticity); check with Levene's test.
If violated: Use a variance-stabilizing transformation or Welch's ANOVA.
Homogeneous Units
Experimental units should be similar; if a strong nuisance factor exists, CRD is the wrong design.
If violated: Switch to a Randomized Block Design to remove the nuisance factor.
⚠️ Check assumptions first
CRD is only efficient when experimental units are homogeneous. If a known nuisance variable (batch, operator, day) affects the response, an unblocked CRD inflates the error term and can hide real treatment effects. In that situation, block the nuisance factor with an RBD instead.
When NOT to Use Completely Randomized Design (CRD)
Known Nuisance Factor
When a systematic source of variation (material batch, machine, shift) is present, CRD wastes power. Use a Randomized Block Design to isolate it.
Two Nuisance Factors
With two crossed nuisance factors (e.g. operator and day), a Latin Square Design controls both simultaneously; CRD cannot.
Multiple Factors of Interest
To study several factors and their interactions at once, use a factorial design rather than separate CRDs.
Industry Applications
Material Testing
Compare tensile strength across several supplier materials when test coupons are essentially identical.
Agriculture & Lab Trials
Compare yields across fertilizer types on uniform plots, or reaction yield across catalysts in identical vessels.
Process Screening
Quickly compare a handful of process settings when the equipment and raw material are uniform across runs.
Clinical & Product Trials
Compare outcomes across treatment groups when subjects are randomized and reasonably homogeneous.
Frequently Asked Questions
What is the difference between CRD and a Randomized Block Design?
CRD assigns treatments to units completely at random and assumes homogeneous units. An RBD first groups similar units into blocks, then randomizes treatments within each block. RBD removes the block-to-block variation from the error term, giving more power whenever a known nuisance factor exists. Use CRD only when units are uniform.
How many replications do I need for a CRD?
More replication increases power and gives a better error estimate. There is no universal minimum, but at least three to five units per treatment is common for detecting moderate effects. Formal sample-size determination uses the desired power, the effect size you want to detect, and the expected error variance.
Can a CRD be unbalanced?
Yes. Equal replication per treatment (balanced) is preferred because it maximizes power and is robust to mild variance inequality, but a CRD with unequal group sizes is still valid. Unbalanced designs are more sensitive to violations of the equal-variance assumption.
What analysis does a CRD use?
A single-factor CRD is analyzed with one-way ANOVA. The total sum of squares is split into a treatment component and an error component, and their mean squares are compared with an F-test to judge whether treatment means differ.
What if the ANOVA assumptions are violated?
If residuals are non-normal or variances are unequal, try a transformation (log or square root), use Welch's ANOVA for unequal variances, or switch to the non-parametric Kruskal-Wallis test, which compares treatments using ranks and does not assume normality.
Does a significant CRD tell me which treatment is best?
No. A significant F only shows that at least one treatment mean differs. To identify which specific treatments differ, follow up with post-hoc pairwise comparisons such as Tukey's HSD, which control the family-wise error rate across all pairs.
Analyze Your Single-Factor Experiment
Build the CRD ANOVA table and test treatment effects in seconds. Free during Beta.
Run CRD Analysis →