Risk Difference Model Overview

Risk difference is the exposed-group probability minus the reference-group probability. It is also called the absolute risk difference, risk difference measure, rate difference in some incidence settings, or absolute percentage-point effect when both probabilities refer to the same follow-up period. Unlike a ratio, risk difference remains on the original probability scale.

The worked design defines paid classes = yes as group 1 and paid classes = no as group 0. The event is a final-grade pass, coded as G3 ≥ 10. The source data contain 39 paid-class students and 610 students without paid classes. The observed pass counts are 30 and 519, respectively.

The Exact Analytical Question

How many more or fewer students pass per 100 when the paid-class group is compared with the no-paid-class group in this dataset?

Quick Answer: Risk Difference Result

The absolute risk difference is obtained by subtracting the no-paid-class pass probability from the paid-class pass probability. The sign is negative because 76.92% is below 85.08%.

Table of Contents

1. What risk difference means
2. Research question and hypotheses
3. When to use risk difference
4. Risk-difference variants and scales
5. Variables and data dictionary
6. Formula and manual calculation
7. Observed data and G3 profiles
8. Complete results and diagnostics
9. Assumptions and design checks
10. Risk difference versus related measures
11. Calculator workflow
12. Python chart interpretations
13. Paired R chart interpretations
14. Python, R, SPSS and Excel workflows
15. Expandable code
16. Interpretation and APA reporting
17. Common mistakes
18. Practice questions
19. Reports and worked files
20. Related guides
21. Frequently asked questions
22. Conclusion

What Is Risk Difference?

Risk difference is an additive effect measure for a binary outcome. If p₁ is the risk in group 1 and p₀ is the risk in group 0, then RD = p₁ − p₀. A value of zero means equal observed risks. Positive values indicate more events in group 1; negative values indicate fewer events in group 1.

Why the Additive Scale Matters

Decision-makers often need the absolute number of outcomes associated with an exposure or intervention. A risk ratio says how many times as likely an event is; a risk difference says how many additional or fewer events occur per 100 or per 1,000 people. In this dataset, RD = −0.081589 corresponds to approximately 81.59 fewer passes per 1,000 paid-class students compared with the no-paid-class risk, if the sample estimate were transported without adjustment.

Direction Depends on Coding

What Risk Difference Does Not Prove

Risk difference describes an observed association. It does not prove causation, eliminate confounding, adjust for baseline differences, or guarantee that the same absolute effect applies in another population. The difference is especially vulnerable to selection when exposure is self-chosen rather than randomized. Review experimental versus quasi-experimental design and sampling bias before making causal claims.

Research Question, Hypotheses and Data Design

Research Question

In the observed student data, how different is the G3 pass probability for students reporting paid classes compared with students reporting no paid classes?

Null and Alternative Hypotheses

Design Facts That Control Interpretation

• Exposure is observational and self-reported.
• The paid-class group is small: 39 of 649 students, or 6.01%.
• The reference group contains 610 students.
• The event threshold is fixed at G3 ≥ 10 in the workbook.
• Every student contributes once to one exposure group and one outcome.

When to Use Risk Difference

Decision Flow

For binary cross-tabulations and expected-count checks, see cross-tabulation, chi-square test of independence and Fisher’s exact test.

Risk Difference Variants and Interpretation Scales

The phrase “risk difference” can describe several closely related estimands. The formula must match the design and outcome timing.

Negative Risk Difference

Yes, risk difference can be negative. Under the current coding, a negative value means the event is less common in group 1. Whether that is beneficial or harmful depends on the event. Here the event is passing, so a negative value is adverse for the paid-class group. If the event were failure, the same underlying table would produce a positive risk difference.

Variables, Coding and Data Dictionary

Group Composition

Raw G3 Descriptive Statistics

The raw-grade comparison supports the direction of the binary result: the paid-class mean is 11.21 versus 11.95, and the upper quartile is 13 versus 14. The medians are both 12, showing that the binary difference is not a complete summary of the two distributions.

Risk Difference Formula and Manual Calculation

Arrange the counts so group 1 is the first row and the event is the first outcome column.

Step 1: Calculate the Group Risks

Step 2: Subtract in the Declared Direction

Multiplying by 100 gives -8.1589 percentage points.

Step 3: Wald Standard Error

The paid-class component contributes 95.63% of the Wald variance because n = 39 is much smaller than n = 610. Increasing the reference group further would have limited effect unless the paid-class group also increased.

Step 4: Wald 95% Confidence Interval

Step 5: Newcombe 95% Confidence Interval

The workbook uses Wilson intervals for each group and then combines the one-sided distances. This avoids relying entirely on the symmetric Wald approximation.

Observed Data, Outcome Counts and G3 Profiles

Observed 2 × 2 Table

Outcome Count Reconciliation

Complete G3 Distribution by Paid-Class Status

Zeros are retained because the table is a full audit of all integer G3 values observed from 0 through 19.

Threshold Sensitivity

The primary event is G3 ≥ 10. The following table shows how the absolute difference changes when the pass threshold is moved. This is a descriptive sensitivity analysis, not permission to choose the threshold that produces the most favorable result.

Complete Risk Difference Results and Data Diagnostics

Primary Results Table

Wald and Newcombe Interval Comparison

Both intervals cross zero. The Newcombe interval is wider and asymmetric around the estimate because it inherits the unequal Wilson uncertainty of the two group risks.

Expected Counts Under Equal Group Risk

The largest Pearson residual is +1.22 in the paid-class failure cell. None exceeds ±2, which is consistent with the confidence interval crossing zero. These residuals describe the table composition; the Newcombe interval remains the primary inferential summary.

Supplementary Scale Translation

Risk Difference Assumptions and Diagnostic Checks

Risk difference itself is a descriptive contrast, but interval interpretation depends on the design and sampling process. The following checks are more important than mechanically applying one formula.

Observed Diagnostic Summary

Confounding and Selection

Paid classes are not randomly assigned. Students who are struggling may be more likely to seek paid instruction, creating confounding by indication. Socioeconomic access, prior grades, school, study time and family support may also influence both exposure and outcome. An adjusted analysis could use binary logistic regression followed by standardized predicted probabilities, but the simple workbook deliberately reports the unadjusted absolute contrast.

Generalizability

The estimate is sample-specific. Transporting −8.16 points to another school system assumes comparable exposure definitions, outcome thresholds, student characteristics and selection mechanisms. Those assumptions are not established by the 2 × 2 table.

Risk Difference Versus Related Measures

Risk Difference Versus Relative Measures

The same table produces an 8.16-point absolute decrease, a risk ratio of 0.9041 and an odds ratio of 0.5845. These numbers are not interchangeable. The odds ratio appears more extreme because odds and probabilities diverge when the event is common. With pass risks near 77% and 85%, the odds ratio should not be described as a risk ratio.

Risk Difference Versus Percentage Change

A decrease from 85.08% to 76.92% is an 8.16 percentage-point decrease. Relative to the 85.08% reference risk, it is a 9.59% relative decrease. Writing “8.16% lower” is ambiguous unless the scale is identified.

Risk Difference Versus Chi-Square or Fisher’s Exact Test

A significance test asks whether the table is compatible with independence; risk difference asks how large the absolute contrast is. A strong report may include both, but effect magnitude and confidence limits should remain central. See chi-square test, Fisher’s exact test and odds ratio.

Risk Difference Calculator: Step-by-Step Workflow

A reliable risk difference calculator should require four counts and make the group direction visible. The following workflow reproduces the workbook.

Calculator Audit Table

Excel Cell Logic

Python Risk Difference Chart Interpretations

The six Python figures translate the 2 × 2 table into risks, intervals, component uncertainty, outcome counts and the final absolute-effect summary. Every chart is interpreted with the exact workbook values.

Python Chart 1: Risk Difference Observed 2 × 2 Table

Pattern

The no-paid group dominates the table numerically, with 610 students versus 39 in the paid group.

Key values

Paid: 30 pass and 9 fail. No paid: 519 pass and 91 fail.

Interpretation

Counts alone are not comparable because denominators differ; row risks must be calculated.

Why it matters

A transposed table or reversed event column would change the meaning and potentially the sign.

Next step: Reconcile row and column totals to 649 before calculating any effect measure.

Python Chart 2: Risk Difference Group Risks

Pattern

The paid-class risk bar is lower than the no-paid-class bar.

Key values

Paid = 0.769231; no paid = 0.850820; vertical gap = 0.081589.

Interpretation

The observed absolute pass-probability difference is −8.16 percentage points.

Why it matters

Showing both risks prevents an isolated effect estimate from hiding the baseline risk.

Next step: State the event definition and group order beside the plotted risks.

Python Chart 3: Risk Difference Confidence Intervals

Pattern

The paid-group interval is much wider because n = 39.

Key values

Paid Wilson: 0.6166–0.8735; no-paid Wilson: 0.8204–0.8769; RD Newcombe: −0.2364–0.0271.

Interpretation

The risk-difference interval crosses zero, so the observed negative direction is not resolved at 95% confidence.

Why it matters

Precision is dominated by the small paid-class group rather than the large reference group.

Next step: Use the Newcombe limits for the main report and show the Wald interval only as a comparison.

Python Chart 4: Risk Difference Effect Components

Pattern

The absolute gap is modest compared with the uncertainty attached to group 1.

Key values

RD = -0.081589; SE = 0.068991; paid variance share = 95.63%.

Interpretation

Most interval width originates from the paid-group binomial variance.

Why it matters

A much larger reference sample cannot compensate fully for a small exposed sample.

Next step: Prioritize recruitment or data quality in the smaller group when planning a more precise study.

Python Chart 5: Risk Difference Outcome Counts

Pattern

Passes are the majority in both groups; failures are relatively more common in the paid group.

Key values

Paid failure risk = 9/39 = 23.08%; no-paid failure risk = 91/610 = 14.92%.

Interpretation

The failure-risk difference is +8.16 points, the sign-reversed counterpart of the pass-risk difference.

Why it matters

Outcome coding changes the sign but not the underlying information in a binary table.

Next step: Keep the pass and failure analyses consistent rather than presenting both as independent findings.

Python Chart 6: Risk Difference Result Summary

Pattern

The point estimate is below zero and the upper confidence limit is slightly above zero.

Key values

RD = −0.081589; 95% CI = −0.236394 to 0.027080.

Interpretation

The most compatible values include a meaningful decrease, no effect and a small increase.

Why it matters

The confidence interval conveys both direction and uncertainty more completely than a binary significance label.

Next step: Report the point estimate and both limits in probability and percentage-point units.

R Risk Difference Charts in Paired Rows

The R figures are arranged in three horizontal pairs on desktop and collapse to one column on mobile. The pair structure keeps each visual beside the analytical check needed to interpret it.

Pair 1 — Counts and Risks

R Chart 1

Observed Table

The observed table confirms 30/9 in the paid group and 519/91 in the no-paid group. The denominators differ sharply, so raw bar heights must not be interpreted as risks.

Report check: Counts sum to N = 649.

R Chart 2

Group Risks

The R risk plot reproduces 76.92% and 85.08%. The 8.16-point vertical separation is the absolute effect estimate.

Report check: Risk gap equals p₁ − p₀.

Pair 2 — Absolute Effect and Interval

R Chart 3

Absolute Difference

The point at −0.081589 lies below the zero reference line. Its sign reflects paid yes minus paid no, not a universal direction.

Report check: Zero is the no-difference reference.

R Chart 4

Confidence Intervals

The Newcombe interval extends from −0.236394 to 0.027080 and crosses zero. The wider paid-group uncertainty is visible beside the narrow no-paid interval.

Report check: The upper limit is +0.027080.

Pair 3 — Percentage Points and Final Summary

R Chart 5

Percentage-Point Scale

The effect is displayed as −8.1589 points, an intuitive absolute scale. The compatible range is −23.6394 to +2.7080 points.

Report check: Use “points,” not relative percent.

R Chart 6

Result Summary

The compact summary aligns with Python and Excel. It should be read as an observational association with unresolved direction at the 95% level.

Report check: Keep causal language out of the report.

Risk Difference in Python, R, SPSS and Excel

Expandable Risk Difference Code

from math import sqrt
from statistics import NormalDist

a, b = 30, 9       # paid: pass, fail
c, d = 519, 91     # no paid: pass, fail
z = NormalDist().inv_cdf(0.975)

p1 = a / (a + b)
p0 = c / (c + d)
rd = p1 - p0
se = sqrt(p1*(1-p1)/(a+b) + p0*(1-p0)/(c+d))
wald = (rd - z*se, rd + z*se)

def wilson(x, n, z=1.959963984540054):
    p = x / n
    den = 1 + z*z/n
    center = (p + z*z/(2*n)) / den
    half = z*sqrt(p*(1-p)/n + z*z/(4*n*n)) / den
    return center-half, center+half

l1, u1 = wilson(a, a+b)
l0, u0 = wilson(c, c+d)
newcombe = (
    rd - sqrt((p1-l1)**2 + (u0-p0)**2),
    rd + sqrt((u1-p1)**2 + (p0-l0)**2),
)
print(p1, p0, rd, se, wald, newcombe)

a <- 30; b <- 9
c <- 519; d <- 91
p1 <- a/(a+b)
p0 <- c/(c+d)
rd <- p1-p0
se <- sqrt(p1*(1-p1)/(a+b) + p0*(1-p0)/(c+d))
wald <- rd + c(-1,1)*qnorm(.975)*se

wilson <- function(x,n,z=qnorm(.975)) {
  p <- x/n
  den <- 1 + z^2/n
  ctr <- (p + z^2/(2*n))/den
  half <- z*sqrt(p*(1-p)/n + z^2/(4*n^2))/den
  c(ctr-half, ctr+half)
}
w1 <- wilson(a,a+b); w0 <- wilson(c,c+d)
newcombe <- c(
  rd - sqrt((p1-w1[1])^2 + (w0[2]-p0)^2),
  rd + sqrt((w1[2]-p1)^2 + (p0-w0[1])^2)
)
list(paid=p1, no_paid=p0, RD=rd, Wald=wald, Newcombe=newcombe)

COMPUTE G3_pass = (G3 >= 10).
VALUE LABELS G3_pass 0 'Fail' 1 'Pass'.
EXECUTE.

CROSSTABS
  /TABLES=paid BY G3_pass
  /FORMAT=AVALUE TABLES
  /CELLS=COUNT ROW EXPECTED
  /STATISTICS=CHISQ RISK.

* Confirm paid=yes: 30 pass, 9 fail; paid=no: 519 pass, 91 fail.
* Calculate the reported RD as 30/39 - 519/610.

Paid risk      = 30/(30+9)
No-paid risk   = 519/(519+91)
Risk difference= Paid risk - No-paid risk
Wald SE        = SQRT(PaidRisk*(1-PaidRisk)/39 + NoPaidRisk*(1-NoPaidRisk)/610)
Wald lower     = RD - 1.96*SE
Wald upper     = RD + 1.96*SE

How to Interpret and Report Risk Difference

Worked APA-Style Report

Data-Discussion Paragraph

Reusable Reporting Templates

Required Elements in a Complete Report

• Exact group 1 and group 0 definitions.
• Event and non-event coding.
• Four 2 × 2 cell counts.
• Both group risks.
• Risk difference in probability and percentage-point units.
• Confidence-interval method and limits.
• Study-design limitations and non-causal wording.
• Supplementary ratio measures only when clearly labeled.

For broader guidance on magnitude, review effect size and confidence interval formulas.

Common Risk Difference Mistakes and Corrections

Result-Specific Corrections

{tbl(['Mistaken statement','Why it is wrong','Correct statement'],[
['Paid classes lower passing by 8.16%.','Ambiguous scale and causal wording.','The observed pass probability was 8.16 percentage points lower in the paid-class group.'],
['The result proves no effect.','An interval including zero does not prove equality.','The 95% interval includes zero and also permits sizeable negative values.'],
['The paid group is 9.59% lower, therefore RD = −9.59%.','9.59% is a relative decrease, not the absolute difference.','RD = −8.16 points; RR = 0.9041.'],
['The odds ratio is 0.5845, so risk is 41.55% lower.','Odds ratio is not a risk ratio when the event is common.','Report OR separately and retain RD as the absolute measure.'],
['A larger no-paid sample will solve precision.','The paid group already drives 95.63% of variance.','Precision requires more information in the paid group.']],1100)}

Risk Difference Practice Questions

Risk Difference Reports and Worked Files

Use the complete files to reproduce the counts, formulas, software output and reporting checks.

Related Salar Cafe Guides

Risk Difference Frequently Asked Questions

Risk Difference Conclusion

The worked risk difference is built from a transparent 2 × 2 table: 30 passes and 9 failures among 39 students reporting paid classes, compared with 519 passes and 91 failures among 610 students reporting no paid classes. These counts produce pass risks of 76.92% and 85.08%.

The absolute difference is −0.081589, or −8.16 percentage points. The source workbook’s Newcombe 95% confidence interval is −0.236394 to 0.027080. Because zero lies inside that range, the observed negative direction is not statistically resolved at the 95% level.

The data discussion adds important context: the paid-class group contains only 39 students, supplies 95.63% of the estimated variance, has a lower mean G3, and shows negative risk differences across all displayed threshold definitions. Even so, paid classes were not randomized, and the result cannot distinguish a treatment effect from self-selection, prior difficulty or other confounding.