Explain Mantel - Haenszel Method please.

Searching the Web

Mantel-Haenszel method epidemiology statistics explained

Finding Sources

Reading Web Page

Mantel Haenszel method explained stratification confounding

Reading File

Here is a thorough explanation of the Mantel-Haenszel method.

The Mantel-Haenszel (MH) Method

Background and Origin

Introduced by Nathan Mantel and William Haenszel in their landmark 1959 paper "Statistical Aspects of the Analysis of Data from Retrospective Studies of Disease" (Journal of the National Cancer Institute), the method was originally designed to analyze case-control (retrospective) data. It has since become a standard tool in epidemiology, clinical trials, and meta-analysis.

The Core Problem it Solves: Confounding

Suppose you want to measure the association between an exposure (e.g., smoking) and a disease (e.g., lung cancer). A simple 2x2 table gives you a crude odds ratio or risk ratio. But a third variable - say, age - might be associated with both smoking AND lung cancer. This is confounding: the crude estimate is distorted because the groups being compared differ on age, not just smoking.

The solution is stratification: split your data into strata (age groups), compute the association within each stratum (where the confounder is held constant), then combine those stratum-specific estimates into a single adjusted summary. The MH method is the standard algorithm for doing this.

The Setup: The 2x2xK Table

For each stratum i (i = 1, 2, ..., K), you have a 2x2 table:

	Disease +	Disease -	Row Total
Exposed	a_i	b_i	n₁ᵢ
Unexposed	c_i	d_i	n₀ᵢ
Col Total	m₁ᵢ	m₀ᵢ	Nᵢ

The MH Pooled Odds Ratio

Each stratum contributes a weighted estimate of the odds ratio. The weight for each stratum is:

wᵢ = (bᵢ × cᵢ) / Nᵢ

The Mantel-Haenszel pooled odds ratio is:

OR_MH = Σ(aᵢdᵢ / Nᵢ) / Σ(bᵢcᵢ / Nᵢ)

That is: sum the "concordant" products weighted by stratum size, divided by the sum of "discordant" products weighted by stratum size. Strata with larger samples and more informative data receive more weight automatically.

The MH Pooled Risk Ratio (for cohort studies)

For cohort or cross-sectional data:

RR_MH = Σ(aᵢ × n₀ᵢ / Nᵢ) / Σ(cᵢ × n₁ᵢ / Nᵢ)

Again, each stratum is weighted by its contribution to the total, ensuring larger strata carry more influence.

The Mantel-Haenszel Chi-Square Test

Beyond just estimating the adjusted association, MH also provides a hypothesis test of whether the adjusted association is statistically significant (H₀: OR_MH = 1 or RR_MH = 1):

χ²_MH = [|Σ(aᵢ - Eᵢ)| - 0.5]² / Σ Var(aᵢ)

Where:

Eᵢ = n₁ᵢ × m₁ᵢ / Nᵢ (expected value of aᵢ under H₀)
Var(aᵢ) = n₁ᵢ × n₀ᵢ × m₁ᵢ × m₀ᵢ / (Nᵢ² × (Nᵢ - 1))
The 0.5 is a continuity correction (Yates correction)
This statistic follows a chi-square distribution with 1 degree of freedom, regardless of how many strata K you have

Worked Conceptual Example

(From Creasy & Resnik's Maternal-Fetal Medicine)

A case-control study examines alcohol use and preeclampsia. The crude OR is 2.26, suggesting alcohol raises risk. But parity (first vs. later pregnancy) could confound this. Stratifying:

Nulliparous stratum: OR = 1.0
Multiparous stratum: OR = 1.0
MH Summary OR = 1.0 (95% CI: 0.42 to 2.34)

The adjusted OR of 1.0 shows no true association - parity was confounding the original estimate of 2.26. The crude OR was inflated because parous women are more likely to drink alcohol AND less likely to get preeclampsia.

Assessing Confounding

To decide if confounding is present, compare:

Crude measure vs. MH adjusted measure
A common rule of thumb: if they differ by >10-15%, confounding is likely

There is no formal statistical test for confounding - it is a judgment call based on the magnitude of the difference.

Effect Modification (Interaction)

Before computing a pooled MH estimate, check whether the stratum-specific OR/RR values are homogeneous (similar across strata). If they differ substantially, the effect of exposure varies by level of the third variable - this is effect modification (also called interaction), NOT confounding.

To test this formally, use the Breslow-Day test:

H₀: All stratum-specific odds ratios are equal
If p < 0.05, effect modification is likely present
In that case, report stratum-specific estimates rather than a pooled MH estimate - pooling is misleading when there is true heterogeneity

The Cochran-Mantel-Haenszel (CMH) Extension

The CMH statistic generalizes the method to 2x2xK tables where K is any number of strata (not just 2). It:

Assumes a common odds ratio across all K strata
Tests the null hypothesis that the two main variables are conditionally independent after controlling for the stratifying variable
Still yields 1 degree of freedom regardless of K, giving it strong power

(Kaplan & Sadock's Comprehensive Textbook of Psychiatry)

Applications

Context	Use
Epidemiology	Controlling confounders in case-control / cohort studies
Clinical trials	Adjusting treatment comparisons across center or baseline severity strata
Meta-analysis	Combining odds ratios or risk ratios from multiple 2x2 tables (one per study)
Psychometrics	Differential item functioning (DIF) analysis in test development
Survival analysis	Mantel-Cox (log-rank) test for comparing survival curves

Assumptions and Limitations

Common effect assumption: MH assumes the OR/RR is the same (or approximately the same) in each stratum. Violation = effect modification.
No sparse data: Works best when each stratum has enough observations. Strata with very small counts contribute little but can distort results.
Only one confounder at a time: Stratification handles one (or at most two) confounders simultaneously. For multiple confounders, use logistic regression (which generalizes MH).
Binary exposure and outcome: Standard MH applies to 2x2 tables. Ordinal or continuous data require extensions.

MH vs. Other Methods

Method	When to Use
MH	Simple confounder adjustment, one variable at a time; transparent and non-parametric
Logistic regression	Multiple confounders simultaneously; continuous covariates
Woolf method	Alternative pooling using log(OR) weights; less stable with sparse data
Random-effects meta-analysis	When heterogeneity between strata/studies is substantial

Key Takeaway

The Mantel-Haenszel method produces a weighted average of stratum-specific effect estimates, where larger, more informative strata receive more weight. This makes it a simple, robust, and transparent way to control for a single confounding variable and obtain an adjusted measure of association - whether an OR, RR, or rate ratio - along with a valid significance test, all within a 1-df chi-square framework.

Please explain in simple and understandable terminology

The Mantel-Haenszel Method - Simply Explained

Start With a Story

Imagine you want to answer this question:

"Does drinking coffee cause heart disease?"

You collect data on 1,000 people - some drink coffee, some don't - and you check who got heart disease. You find that coffee drinkers get heart disease MORE often. Alarming, right?

But wait. Coffee drinkers also tend to smoke more. And smoking causes heart disease. So is it really the coffee... or is it the smoking sneaking into your results?

This "sneaky third variable" is called a confounder. The Mantel-Haenszel method is a tool to remove the confounder's influence and see the true picture.

The Simple Idea: Divide, Then Combine

The MH method works in two steps:

Step 1 - Divide (Stratify)

Split your data into groups based on the confounder.

Group 1: Smokers only
Group 2: Non-smokers only

Now compare coffee vs. no coffee within each group separately. Within a group of all-smokers, everyone smokes the same amount - so smoking can't distort the coffee comparison anymore. You've neutralized the confounder.

Step 2 - Combine (Pool)

You now have two separate results (one per group). The MH method combines them into one overall answer, giving more weight to the groups with more people (because bigger groups give more reliable estimates).

Think of it like averaging exam scores from two classes - but if one class has 100 students and the other has 10, you trust the bigger class more. MH does exactly this mathematically.

What Does "Weight" Mean Here?

Each group (stratum) gets a weight based on how many people are in it. A group with 500 people contributes more to the final answer than a group with 20 people. This makes the combined result more accurate and stable.

What Number Does It Produce?

The MH method produces an adjusted Odds Ratio (OR) or adjusted Risk Ratio (RR) - a single number that tells you how strongly the exposure is linked to the disease, after removing the confounder's effect.

Result	Meaning
OR = 1.0	No association (exposure doesn't matter)
OR > 1.0	Exposure increases the risk
OR < 1.0	Exposure decreases the risk (protective)

A Concrete Example

Going back to the preeclampsia example from earlier:

Crude (unadjusted) OR = 2.26 - looks like alcohol doubles the risk of preeclampsia
You suspect parity (whether a woman has had children before) is confounding this
You split the data: first-time moms vs. women who've had babies before
In each group separately, the OR = 1.0
MH adjusted OR = 1.0 - no real association!

The original scary result of 2.26 was entirely due to parity sneaking in. Alcohol wasn't actually the problem - parity was. The MH method revealed this.

What About the Statistical Test?

After computing the adjusted OR, you want to know: is this result real or just due to chance?

The MH method includes a chi-square test that answers exactly this. It produces a p-value:

p < 0.05 = the association is statistically significant
p ≥ 0.05 = could be due to chance

The clever part: this test gives you just 1 single p-value, no matter how many groups you split into. So splitting into 10 age groups still gives one clean answer.

One Important Warning: Effect Modification

Before you combine the groups, you should check: are the results similar in all groups?

If coffee has OR = 3.0 in smokers but OR = 0.8 in non-smokers, that's very different.
This means the effect of coffee changes depending on smoking status - this is called effect modification.
In this case, you should NOT combine the groups into one number - you should report them separately, because they tell different stories.

You only use MH to combine results when the groups tell roughly the same story.

In a Nutshell

Step	What You Do	Why
1	Split data by confounder	Neutralize its effect
2	Calculate OR/RR in each group	Get a "clean" estimate per group
3	Check groups are similar	Make sure combining makes sense
4	Combine using MH weights	Get one adjusted overall estimate
5	Run the chi-square test	Check if the result is statistically real

When Is It Used?

Epidemiology: Controlling for confounders in disease studies
Clinical trials: Comparing treatments across different hospitals or patient subgroups
Meta-analysis: Combining results from multiple studies into one pooled estimate
Any situation where a third variable might be distorting a two-variable comparison

The Bottom Line

The Mantel-Haenszel method is like a fairness referee. It makes sure you're comparing apples to apples by holding confounders constant, then combines the fair comparisons into one trustworthy answer - giving more say to the groups with more data.

Can you explain with an example question?

Mantel-Haenszel Method: Step-by-Step Example

The Question

"Does smoking cause lung cancer?" But we suspect age might be a confounder (older people smoke more AND are more likely to get cancer anyway).

We collected data on 200 people and divided them into two age groups:

Young (under 50)
Old (50 and above)

Step 1: Look at the Raw (Crude) Data First

Here's the overall data, ignoring age:

	Cancer +	Cancer -	Total
Smoker	40	60	100
Non-smoker	20	80	100
Total	60	140	200

Crude Odds Ratio = (40 × 80) / (60 × 20) = 3200 / 1200 = 2.67

This says smokers have 2.67 times the odds of getting lung cancer. But is age distorting this?

Step 2: Stratify by Age (The Confounder)

Stratum 1 - Young People (Under 50)

	Cancer +	Cancer -	Total
Smoker	10	40	50
Non-smoker	5	45	50
Total	15	85	100

OR (Young) = (10 × 45) / (40 × 5) = 450 / 200 = 2.25

Stratum 2 - Old People (50 and above)

	Cancer +	Cancer -	Total
Smoker	30	20	50
Non-smoker	15	35	50
Total	45	55	100

OR (Old) = (30 × 35) / (20 × 15) = 1050 / 300 = 3.50

Step 3: Check - Are the Two ORs Similar?

Young group: OR = 2.25
Old group: OR = 3.50

They're in the same direction (both above 1, both suggesting smoking increases cancer risk). They're not wildly different, so it's reasonable to combine them. ✅

(If one was 2.25 and the other was 0.4, we'd say there's effect modification and we should NOT combine them.)

Step 4: Calculate the MH Weights

The MH formula weights each stratum by how much data it contributes. The weight for each stratum is:

Weight = (b × c) / N

Where:

b = exposed without disease
c = unexposed with disease
N = total in that stratum

Weight for Young group: = (40 × 5) / 100 = 200 / 100 = 2.0

Weight for Old group: = (20 × 15) / 100 = 300 / 100 = 3.0

The old group gets more weight because its numbers are more informative (the disease is more common there, so there's more statistical signal).

Step 5: Calculate the MH Adjusted Odds Ratio

The formula is:

OR_MH = Σ(a × d / N) / Σ(b × c / N)

Numerator (top of the fraction):

Young: (10 × 45) / 100 = 450/100 = 4.5
Old: (30 × 35) / 100 = 1050/100 = 10.5
Total numerator = 4.5 + 10.5 = 15.0

Denominator (bottom of the fraction):

Young: (40 × 5) / 100 = 200/100 = 2.0
Old: (20 × 15) / 100 = 300/100 = 3.0
Total denominator = 2.0 + 3.0 = 5.0

OR_MH = 15.0 / 5.0 = 3.0

Step 6: Interpret the Results

Estimate	Value	Meaning
Crude OR	2.67	Raw association before adjustment
MH Adjusted OR	3.00	Association after controlling for age

The adjusted OR (3.0) is slightly higher than the crude OR (2.67). This means age was actually masking some of the smoking effect (a phenomenon called negative confounding). After removing age's influence, the true association between smoking and lung cancer is a bit stronger.

Conclusion: Smokers have 3 times the odds of developing lung cancer compared to non-smokers, after adjusting for age.

Quick Visual Summary of What Just Happened

RAW DATA
(All mixed together)
Crude OR = 2.67  ← possibly distorted by age
        |
        ▼
SPLIT BY AGE
  Young group → OR = 2.25
  Old group   → OR = 3.50
        |
        ▼
COMBINE WITH WEIGHTS
  (Old group gets more weight = 3.0)
  (Young group gets less weight = 2.0)
        |
        ▼
MH ADJUSTED OR = 3.0  ← the honest, age-controlled answer

The Key Lesson From This Example

The crude OR was 2.67 - it looked like age was inflating the result
After controlling for age using MH, the true OR was actually 3.0 - slightly stronger
This showed age was suppressing the real effect, not inflating it
The MH method peeled away age's influence and revealed the cleaner truth

This is exactly why researchers never rely on a crude estimate alone in observational studies - confounders can hide, inflate, or reverse the real association.

What's the difference between confounding and effect modification?

Confounding vs. Effect Modification

The One-Line Difference

Confounding = a third variable is distorting your result. Fix it, remove it, adjust for it. Effect modification = a third variable is changing the true effect. Don't remove it — report it, because it's telling you something real and important.

Think of It This Way

Imagine you're trying to hear a song (the true exposure-disease relationship):

Confounding = background noise. It's interfering with what you're trying to hear. You want to remove it.
Effect modification = the song genuinely sounds different in different rooms. That difference is the finding. You want to describe it.

Confounding - Explained Simply

A confounder is a variable that:

Is associated with the exposure
Is associated with the outcome
Is NOT on the causal pathway between them

It creates a fake or distorted association. The confounder is like a shadow - it makes things look different from what they really are.

Example of Confounding

Question: Does carrying a lighter cause lung cancer?

You find that lighter-carriers get lung cancer more often. Shocking? Not really.

Lighter-carriers tend to smoke
Smoking causes lung cancer
Carrying a lighter does NOT cause cancer - smoking does

Smoking is the confounder. Once you control for smoking, the lighter-cancer association disappears completely. The lighter was just a proxy for the real culprit.

Lighter ──────────────────────────── Cancer
    ↘                              ↗
      Smoking (the real cause)

The lighter-cancer association was fake - created entirely by the confounder.

Effect Modification - Explained Simply

Effect modification (also called interaction) is when the effect of an exposure on an outcome is genuinely different depending on a third variable.

It is NOT a bias or an error. It is a real biological or clinical phenomenon that you want to know about.

Example of Effect Modification

Question: Does aspirin prevent heart attacks?

You study two groups - men and women:

In men: Aspirin reduces heart attack risk significantly (OR = 0.6)
In women: Aspirin has almost no effect on heart attacks (OR = 0.97)

Sex is an effect modifier. The drug works differently in men vs. women. This is a real and important finding - not a statistical nuisance to remove.

If you pooled the data and computed one average OR, you'd miss this entirely and give wrong clinical advice to half the population.

Side-by-Side Comparison

Feature	Confounding	Effect Modification
What it is	A distortion / bias	A real biological difference
Is it a problem?	Yes - it misleads you	No - it informs you
What to do with it	Adjust for it, control it, remove it	Report it, stratify by it
MH adjusted OR vs crude OR	They differ significantly	Stratum-specific ORs differ from each other
Should you pool the ORs?	Yes, after adjustment	No - report separately
Example	Age confounding coffee-cancer link	Sex modifying aspirin's effect on heart attacks
Clinical implication	"The crude result was misleading"	"The treatment works differently in subgroups"

How to Tell Them Apart in Practice

Here's the decision process when you stratify your data:

Stratify by the third variable
            │
            ▼
Are the stratum-specific ORs
    similar to each other?
        /           \
      YES             NO
       │               │
       ▼               ▼
Are they different    EFFECT MODIFICATION
from the crude OR?    → Report each stratum
    /       \           separately
  YES        NO
   │          │
   ▼          ▼
CONFOUNDING  No confounding,
→ Use the    no effect modification
  MH adjusted
  OR

A Worked Example Showing Both

Question: Does a new drug reduce mortality? Potential third variable: Disease severity (mild vs. severe)

Scenario A - Confounding

Group	Drug OR
Mild disease	0.5
Severe disease	0.5
Crude OR	0.7

Both strata show the same OR (0.5). The crude was 0.7 - distorted by the fact that sicker people got the drug more often. The true effect is 0.5 in both groups. Severity was confounding the result.

✅ Use the MH adjusted OR = 0.5 and report that.

Scenario B - Effect Modification

Group	Drug OR
Mild disease	0.9 (barely works)
Severe disease	0.3 (works very well)
Crude OR	0.6

The ORs are very different between strata. The drug works much better in severe disease. This is effect modification.

❌ Do NOT report a single pooled OR of 0.6 - it hides the truth. ✅ Report: "The drug reduces mortality by 70% in severe disease but has little effect in mild disease."

The Most Important Practical Point

Situation	Wrong Action	Right Action
Confounding	Report the crude OR	Adjust using MH or regression
Effect modification	Pool into one number	Stratify and report each group

Confounding is something you correct. Effect modification is something you communicate.

Memory Hook

Confounder = a liar in your data. Catch it, remove it. Effect modifier = a storyteller in your data. Listen to it, report it.

This is a shared conversation. Sign in to Orris to start your own chat.