Cochran-Mantel-Haenszel: A Comprehensive Guide to the CMH Method in Stratified Analysis

12. April 2025 By Webmasterr Off

The Cochran-Mantel-Haenszel method, commonly abbreviated as CMH, is a cornerstone technique in biostatistics for analysing associations in stratified data. When researchers deal with observational studies or epidemiological data, confounding factors can blur the true relationship between exposure and outcome. The Cochran-Mantel-Haenszel approach offers a principled way to adjust for a categorical stratification variable, providing a pooled measure of association and a formal test of significance. In this guide, we will explore the Cochran-Mantel-Haenszel method in depth, including its history, mathematical foundations, practical implementation, and common pitfalls. Whether you encounter this technique under its full name, as CMH, or through the related Mantel-Haenszel nomenclature, you will gain a clear understanding of how to apply it in the real world.

Origins and naming: the Cochran-Mantel-Haenszel method in context

The CMH method is named after three prominent statisticians: Thomas Louis Cochran, Norman Mantel, and William Haenszel. Each contributed to the development of stratified analysis for binary outcomes. The method is sometimes referred to as the Cochran-Mantel-Haenszel test or simply the Mantel-Haenszel test, depending on the emphasis placed on the contributing researchers. In practice, “Cochran-Mantel-Haenszel” is the most widely recognised terminology, but readers may see “Mantel-Haenszel” or “Haenszel-Mantel-Cochran” in older texts or software documentation. Regardless of the ordering, the underlying idea remains the same: combine evidence across strata to obtain a stable, adjusted measure of association between exposure and outcome.

When to use the CMH method: controlling confounding by stratification

In observational studies, confounding variables can distort the relationship between an exposure and an outcome. The CMH approach is particularly suited to situations where you have multiple 2×2 contingency tables, one for each level of a stratification variable (for example, age group, hospital, or sex). By analysing these 2x2xK tables collectively, CMH provides two key outputs:

A pooled, common odds ratio (the CMH odds ratio) that represents the average association across strata, adjusting for the stratification factor.
A chi-squared test statistic (CMH chi-square) that tests the null hypothesis of no association after adjustment, with one degree of freedom in the standard 2x2xK setting.

The practical advantage of CMH is that it can reduce bias due to confounding without resorting to more complex modelling in some contexts. It is particularly valuable when data are sparse within individual strata, as the stratified approach borrows strength across strata to stabilise the estimate.

Structural data: 2x2xK tables and how to organise your data

To apply the Cochran-Mantel-Haenszel method, your data must be arranged as a series of 2×2 contingency tables, one per stratum. Each stratum i contains counts for the four cells of a binary outcome and a binary exposure:

a_i: Disease Yes and Exposed Yes
b_i: Disease No and Exposed Yes
c_i: Disease Yes and Exposed No
d_i: Disease No and Exposed No

Across strata, the numbers can differ, but within each stratum the table remains 2×2. For illustration, consider a scenario with three strata (K = 3). The counts might look like this:

A worked partial example: a simple 2x2xK data structure

Stratum 1:

a₁ = 8
b₁ = 4
c₁ = 2
d₁ = 16

Stratum 2:

a₂ = 5
b₂ = 7
c₂ = 3
d₂ = 15

Stratum 3:

a₃ = 3
b₃ = 5
c₃ = 4
d₃ = 18

For each stratum you can verify the total N_i = a_i + b_i + c_i + d_i. In this example, every stratum contains 30 observations, which simplifies the arithmetic and makes interpretation straightforward.

Key quantities: the CMH odds ratio and the CMH chi-square statistic

The two central outputs of the Cochran-Mantel-Haenszel method are the pooled odds ratio and the associated chi-square test. Here is a concise description of how these are formed in 2x2xK data.

The CMH odds ratio

The CMH estimate of the common odds ratio (OR_MH) across all strata is given by the ratio of two sums that blend information from each stratum:

OR_MH = [sum over i of (a_i d_i / N_i)] / [sum over i of (b_i c_i / N_i)], where N_i = a_i + b_i + c_i + d_i.

A value greater than 1 indicates a positive association between exposure and disease after adjustment for the stratification variable; a value less than 1 indicates a negative association. In our worked example, the sums yield an OR_MH of about 5.25, suggesting a strong positive association after accounting for strata. It is important to emphasise that OR_MH is an odds ratio, not a risk ratio, and its interpretation should reflect this distinction.

The CMH chi-square statistic

The CMH chi-square test assesses whether the observed association across strata differs from what would be expected under the null hypothesis of no association after adjustment. In the standard 2x2xK framework, the CMH test statistic is a chi-square with 1 degree of freedom. Conceptually, it aggregates the cross-stratum deviations of the observed a_i counts from their stratum-specific expectations, weighting them by the corresponding variances. When the CMH chi-square statistic is large and the p-value is small, you have evidence against the null hypothesis: there is a detectable association between exposure and disease that persists after stratification.

In practical terms, the computation is implemented in statistical software, but understanding the logic helps with interpretation. The CMH statistic and the OR_MH are complementary: OR_MH communicates the magnitude of the association, while the CMH chi-square tests whether that association is unlikely to arise by chance under the null model that the exposure and disease are independent within strata.

CMH in practice: software, workflows, and interpretation

Many statistical packages implement the Cochran-Mantel-Haenszel approach, reflecting its enduring relevance in epidemiology and clinical research. Here are common routes to applying CMH in real data analyses, with notes on syntax and common pitfalls.

R: Mantel-Haenszel and CMH in R

In R, the Mantel-Haenszel test is typically performed with a dedicated function that handles stratified 2×2 tables. The most frequent usage is via the mantelhaen.test() function, which returns both the CMH chi-square statistic and the common odds ratio estimate. Data are commonly arranged as a multi-dimensional array or as a list of 2×2 tables across strata. The command resembles:

# Example syntax (conceptual)
result <- mantelhaen.test(x, y, z)
print(result)

Where x, y, and z represent ingredients of your data structure. While the exact input format depends on your data layout, the key outcome is the CMH chi-square statistic and OR_MH. When reporting results, include the CMH odds ratio and the corresponding 95% confidence interval, if available, along with the p-value of the CMH chi-square test.

Python: Statsmodels and Mantel-Haenszel in Python workflows

In Python, advanced users can apply the Mantel-Haenszel method through the statsmodels package. The typical approach involves stratified contingency tables constructed via pandas, and then invoking the Mantel-Haenszel test function. The syntax resembles:

# Conceptual example (not runnable as-is)
from statsmodels.stats.contingency_tables import StratifiedTable

# Build a list of 2x2 arrays, one per stratum
strata = [
    [[a1, b1], [c1, d1]],
    [[a2, b2], [c2, d2]],
    [[a3, b3], [c3, d3]],
]

table = StratifiedTable(strata)
# CMH statistic
cmh = table.test_null_pvalue()  # or an equivalent CMH test accessor
# Common odds ratio
or_mh = table.odds_ratio_pooled
print(or_mh, cmh)

As with R, it is prudent to report the CMH odds ratio, its confidence interval if available, and the p-value associated with the CMH chi-square test when presenting results.

Other software: SAS, Stata, and specialised packages

Stata has the “cmh” procedure, SAS provides PROC FREQ with options for CMH, and specialised biostatistics software often packages the CMH test under the Mantel-Haenszel umbrella. Regardless of the software, the interpretation remains consistent: a pooled measure of association across strata, plus a test of significance that is robust to stratification effects.

Interpreting CMH results: what the numbers mean for decision making

A good CMH analysis yields two key takeaways: the magnitude of the association and the statistical significance after controlling for the stratification variable. Consider the following interpretive guidance:

The CMH odds ratio (OR_MH) indicates how strongly exposure is associated with disease when the data are aggregated across strata. An OR_MH greater than 1 suggests increased odds of disease with exposure after adjustment; an OR_MH less than 1 suggests decreased odds.
The CMH chi-square test assesses whether that association could plausibly arise by chance under the null hypothesis of no association within strata. A small p-value (commonly < 0.05) supports the notion of a statistically significant association after adjustment for stratification.
Confidence intervals are crucial for understanding precision. A wide interval that includes 1 indicates uncertainty about the direction or magnitude of the association, whereas a narrow interval wholly above or below 1 strengthens the conclusion.
Assumptions matter. CMH assumes that the strata are properly defined and that within each stratum the data arise from a binomial process with a consistent effect. Large heterogeneity of effects across strata can complicate interpretation, and tests for homogeneity (see Breslow-Day) may be appropriate.

Common pitfalls and misinterpretations to avoid

While the Cochran-Mantel-Haenszel approach is robust in many scenarios, several caveats deserve attention:

Strata definition matters. If the stratification variable is misclassified or if important effect modifiers are omitted, the CMH analysis may still be biased or misrepresent the true relationship.
Heterogeneity across strata. If the odds ratios vary considerably across strata, a single pooled OR_MH may be misleading. In such cases, investigators should assess homogeneity (e.g., Breslow-Day test) and consider stratified reporting or meta-analytic approaches.
Sparse data issues. In strata with very small cell counts, estimates can become unstable. Consider collapsing strata or using exact methods if appropriate.
Interpretation of odds ratios. The CMH odds ratio is an odds ratio, not a risk ratio. In common observational contexts, especially when outcomes are not rare, these measures can diverge in interpretation.
Confounding versus interaction. CMH helps adjust for confounding by stratification, but it does not inherently reveal interaction effects between exposure and the stratification variable. If interactions are suspected, additional modelling or stratified analyses are warranted.

Extensions and related concepts: homogeneity, stratified models, and beyond

Beyond the standard CMH framework, researchers often explore extensions and complementary tests to obtain a fuller picture of the data. Notable concepts include:

Breslow-Day test for homogeneity of odds ratios. This test assesses whether the odds ratios across strata are consistent enough to justify a common odds ratio estimate. A significant Breslow-Day test suggests substantial heterogeneity and prompts further investigation.
Stratified logistic regression. When there are multiple strata, a logistic regression model with stratum indicators (fixed effects) can adjust for the stratification variable while allowing the exposure effect to be estimated directly. This approach can also handle additional covariates.
Mantel-Haenszel pooled analyses in meta-analytic contexts. In meta-analyses, a Mantel-Haenszel estimator can pool study-specific odds ratios across diverse settings, provided that the studies are sufficiently homogeneous in design and outcome definition.

Practical considerations: data preparation, reporting, and reproducibility

To ensure a robust CMH analysis, follow best practices in data preparation and reporting:

Ensure strata are meaningful and mutually exclusive. Each stratum should represent a well-defined category of the stratification variable, with sufficient observations in each 2×2 table.
Check for zero cells and sparse data. If a stratum contains zero in critical cells, consider continuity corrections or exact methods where appropriate, noting any limitations.
Document your data structure and analysis steps. Reproducibility is key: provide a clear description of how the 2x2xK data were assembled, how OR_MH was computed, and how the CMH statistic was derived.
Report both the CMH odds ratio and the CMH chi-square test results, including p-values and confidence intervals where available. When possible, present stratum-specific results alongside the pooled estimate to convey the picture of heterogeneity.

A practical, illustrative example: applying the CMH method to stratified data

To bring these ideas to life, consider a hypothetical study examining whether exposure to a particular environmental factor is associated with a respiratory condition, while controlling for age group as the stratification variable. The data are organised into three age strata, each with a 2×2 table for disease status by exposure.

Stratum 1 data (age group 18–34):

a₁ = 8
b₁ = 4
c₁ = 2
d₁ = 16

Stratum 2 data (age group 35–54):

a₂ = 5
b₂ = 7
c₂ = 3
d₂ = 15

Stratum 3 data (age group 55+):

a₃ = 3
b₃ = 5
c₃ = 4
d₃ = 18

With these numbers, compute N_i for each stratum (they all sum to 30). The CMH odds ratio is obtained by combining across strata as:

OR_MH = [sum_i (a_i d_i / N_i)] / [sum_i (b_i c_i / N_i)]

Plugging in our numbers:

Sum of a_i d_i / N_i = (8×16/30) + (5×15/30) + (3×18/30) = 4.2667 + 2.5 + 1.8 ≈ 8.5667
Sum of b_i c_i / N_i = (4×2/30) + (7×3/30) + (5×4/30) = 0.2667 + 0.7 + 0.6667 ≈ 1.6334
OR_MH ≈ 8.5667 / 1.6334 ≈ 5.25

Interpretation: after adjusting for age group, the odds of disease among those exposed are approximately 5.25 times those without exposure. This magnitude, subject to precision and confidence intervals, signals a meaningful association in the data. The CMH chi-square test would then indicate whether this observed association is statistically significant after stratification. While we have not shown the full p-value calculation here, the framework allows statisticians to quantify both the direction and the reliability of the association.

Interpreting results responsibly: limitations and context

Like any statistical method, the Cochran-Mantel-Haenszel approach has limitations that researchers should acknowledge in their reporting:

It assumes homogeneous effects across strata. If there is substantial heterogeneity in the effect of exposure across different strata, the pooled OR_MH may be misleading. A Breslow-Day test for homogeneity can help diagnose this issue.
The method adjusts for observed stratification variables only. Unmeasured or misclassified confounders within strata can still bias results.
It is most straightforward for binary outcomes and binary exposures within each stratum. More complex outcome types or multi-valued exposures require extensions or alternative modelling approaches.

Connecting CMH to broader analytic strategies

The Cochran-Mantel-Haenszel method sits at the intersection of stratified analysis and classic contingency table methods. It complements, rather than replaces, other approaches such as:

Multivariable logistic regression with stratum indicators as fixed effects to adjust for stratification while permitting the inclusion of multiple covariates and potential interactions.
Propensity score stratification or matching, where CMH-like pooling can help summarise effects within propensity-score strata.
Meta-analytic techniques that combine study-specific odds ratios across diverse sources, using Mantel-Haenszel weighting in homogeneous settings.

Key takeaways: summarising the Cochran-Mantel-Haenszel method

The Cochran-Mantel-Haenszel method provides a robust, interpretable framework for analyzing stratified 2×2 data. Its strengths include the ability to pool information across strata while adjusting for a categorical confounder, and to offer a formal test of association via the CMH chi-square statistic. The CMH odds ratio gives a single, interpretable measure of effect size, while the test results indicate whether that effect is unlikely to be due to random variation under the null hypothesis.

In practice, CMH is widely used in epidemiology, clinical research, and public health. It offers a balance between simplicity and rigor, enabling researchers to draw meaningful conclusions from stratified observational data without immediately resorting to more complex modelling. By understanding both the OR_MH and the CMH chi-square test, practitioners can communicate findings clearly, transparently, and in a way that is accessible to a broad audience.

Glossary of terms to help navigate the literature

Cochran-Mantel-Haenszel (CMH) method: A stratified analysis technique for binary outcomes that yields a pooled odds ratio and a chi-square test of association.
CMH odds ratio (OR_MH): The combined estimate of the odds ratio across strata, adjusting for the stratification variable.
CMH chi-square statistic: The test statistic used to assess the null hypothesis of no association after stratification, following a chi-square distribution with 1 degree of freedom in the standard 2x2xK setup.
Stratum (plural: strata): A category or level of the stratification variable (e.g., age group, hospital location) in which the data are organised into 2×2 tables.
Breslow-Day test: A test for homogeneity of the odds ratios across strata, used to assess whether a common odds ratio is appropriate.

Final thoughts: adopting the Cochran-Mantel-Haenszel approach with confidence

The Cochran-Mantel-Haenszel method remains a practical and powerful tool for analysts dealing with stratified binary data. Its balanced blend of mathematical clarity and interpretability makes it a mainstay in both teaching and applied research. By structuring data into 2x2xK tables, computing the CMH odds ratio, and testing for overall association with the CMH chi-square statistic, researchers can obtain a nuanced understanding of exposure effects that accounts for confounding through stratification. Whether you encounter the term Cochran-Mantel-Haenszel, Mantel-Haenszel, or a hyphenated variant of the same name, the underlying principles are consistent and accessible. As you apply this method, remember to consider potential heterogeneity across strata, report both pooled and stratum-specific results, and complement CMH with additional analyses when appropriate to build a robust evidence base.

CategoryMisc