Outcome Scale And Patient-Benefit Conclusions

Why the Outcome Scale Changes the Conclusion

For a binary clinical outcome, a patient either experiences the favorable outcome by a specified time or does not. The same trial data can be summarized as an absolute difference in outcome probabilities, a relative comparison of those probabilities, or a count of favorable outcomes. Each description answers a different question. In particular, the number of favorable outcomes observed in an arm reflects both its outcome rate and the number of patients assigned to it.

The distinction matters when comparing designs with different allocation ratios. A larger treatment arm can produce more favorable outcomes even if its patients have the same outcome probability as patients in a smaller control arm. A count alone therefore cannot establish how much treatment changes an individual patient’s chance of a favorable outcome.

Definitions and Assumptions

Let \(Y=1\) denote that a patient has the prespecified favorable outcome by the assessment time, and \(Y=0\) otherwise. Let \(p_T\) and \(p_C\) be the probabilities of that outcome under assignment to treatment and control, respectively. Their sample estimates are \(\hat{p}_T=x_T/n_T\) and \(\hat{p}_C=x_C/n_C\), where \(x_T,x_C\) are the observed numbers of favorable outcomes and \(n_T,n_C\) are the numbers assigned to the corresponding arms.

Definition: The risk difference is \(p_T-p_C\), estimated by \(\hat{p}_T-\hat{p}_C\). It expresses the absolute change in outcome probability. The risk ratio is \(p_T/p_C\), estimated by \(\hat{p}_T/\hat{p}_C\), when \(p_C>0\). It expresses the treatment risk relative to the control risk. The favorable-outcome count in an arm is the number of patients in that arm with \(Y=1\); its expected value for \(n\) patients with risk \(p\) is \(np\).

For a causal interpretation, suppose treatment assignment is randomized, treatment and outcome are well defined, the endpoint is assessed in the same way in both arms, and each patient’s outcome is not affected by other patients’ assignments. Assume also that outcomes are observed for all assigned patients, or that missing outcomes are handled by a justified analysis. These conditions support comparing outcome probabilities under assignment. The numerical illustration below uses prespecified 3:2 randomization and independent patient outcomes; this keeps the arithmetic transparent.

The risk difference is measured in probability units. A difference of \(0.10\) is a 10-percentage-point increase, not a 10% relative increase. The risk ratio is unitless: a ratio of \(1.33\) means the treatment risk is about 1.33 times the control risk, or 33% higher relative to that control risk. Neither scale by itself describes every aspect of clinical importance; interpretation depends on the endpoint and the patients to whom the risks apply.

One Binary-Outcome Trial, Three Descriptions

Worked Example: Absolute and Relative Benefit With Unequal Arm Sizes

Hypothetical data. In a randomized trial, 120 patients are assigned to treatment and 80 to control. By the fixed assessment time, 48 treatment patients and 24 control patients have the favorable outcome. Assume the endpoint is completely observed and the trial’s prespecified randomization and follow-up procedures apply equally to both arms.

The estimated treatment risk is \(48/120=0.400\), or 40.0%. The estimated control risk is \(24/80=0.300\), or 30.0%.

Risk difference. Substituting the observed risks gives

\hat{p}_T-\hat{p}_C =\frac{48}{120}-\frac{24}{80} =0.400-0.300 =0.100.

The estimated absolute increase is 0.100, or 10.0 percentage points. In this example, the estimated favorable-outcome probability is 10 percentage points higher under treatment assignment.

Risk ratio. The estimated relative risk is

\frac{\hat{p}_T}{\hat{p}_C} =\frac{0.400}{0.300} =1.333.

The estimated treatment risk is about 1.33 times the control risk, a relative increase of about \(1.333-1=0.333\), or 33.3%. This is not a 33.3-percentage-point increase; the absolute increase is 10.0 percentage points.

Observed counts. There are 48 favorable outcomes in the treatment arm and 24 in the control arm, a difference of 24 observed outcomes. But the treatment arm has 40 more patients. The count difference is not a direct estimate of the difference in an individual patient’s outcome probability.

Why Raw Counts Depend on Allocation

For any arm with \(n\) patients and outcome probability \(p\), the expected favorable-outcome count is \(np\). Thus the count is tied to both the risk and the arm size. The observed count difference can be decomposed using the control risk as a reference:

n_T\hat{p}_T-n_C\hat{p}_C =n_T(\hat{p}_T-\hat{p}_C)+(n_T-n_C)\hat{p}_C.

In the example, this gives \(120(0.100)+40(0.300)=12+12=24\). The identity separates a component based on the estimated risk difference from a component reflecting the unequal arm sizes at the control risk. This decomposition is an arithmetic explanation, not a claim that the observed count difference identifies a causal number of patients helped.

A useful comparison instead places the risks on the same number of patients. The next example does that directly.

Worked Example: Standardizing Counts to the Same Number of Patients

Use the risks from the hypothetical trial: 40.0% under treatment and 30.0% under control. Ask how many favorable outcomes would be expected per 100 patients assigned to each arm if these estimated risks applied to that common group size.

For 100 treatment patients, the expected count is \(100(0.400)=40\). For 100 control patients, it is \(100(0.300)=30\). The standardized difference is \(40-30=10\) favorable outcomes per 100 patients, corresponding to the 10-percentage-point risk difference.

For a hypothetical group of 200 patients, the estimated counts under assignment of all 200 to treatment or all 200 to control would be \(200(0.400)=80\) and \(200(0.300)=60\), respectively. The difference is 20 expected favorable outcomes in that group size. These are estimates based on the trial risks, not outcomes observed in two alternative versions of the same patients.

The standardization changes the comparison from the trial’s unequal observed arm sizes to a common size. It does not eliminate sampling uncertainty or guarantee that the estimated risks will apply to a different patient population.

Uncertainty and Interpretation

Point estimates summarize the observed data, but they do not show how much the estimates might vary across comparable trials. Under independent Bernoulli outcomes in each arm, approximate standard errors can be calculated from the observed proportions. For the risk difference, the plug-in standard error is the square root of the sum of the estimated binomial variances divided by their arm sizes. A large-sample interval is the estimate plus or minus 1.96 standard errors for an approximate 95% confidence interval.

Worked Example: Approximate Confidence Intervals on Both Scales

Continue with the same hypothetical data: \(\hat{p}_T=0.400\), \(\hat{p}_C=0.300\), \(n_T=120\), and \(n_C=80\). The independent-binomial plug-in standard error for the risk difference is

SE(\hat{p}_T-\hat{p}_C) =\sqrt{\frac{0.400(0.600)}{120}+\frac{0.300(0.700)}{80}} =\sqrt{0.004625} =0.0680.

The approximate 95% interval is \(0.100\pm1.96(0.0680)\), or \(0.100\pm0.133\). Its limits are approximately \(-0.0333\) and \(0.233\), equivalent to about -3.33 to 23.3 percentage points. Under the assumptions for this interval, the data are compatible with effects ranging from a small absolute decrease to a substantial absolute increase. The interval includes zero.

For the risk ratio, a common large-sample interval is calculated on the logarithmic scale. Its estimated standard error is

SE\{\log(\hat{p}_T/\hat{p}_C)\} =\sqrt{\frac{1}{48}-\frac{1}{120}+\frac{1}{24}-\frac{1}{80}} =\sqrt{0.04167} =0.2041.

Since \(\log(1.333)=0.2877\), the interval on the log scale is \(0.2877\pm1.96(0.2041)\), or approximately \(-0.112\) to \(0.688\). Exponentiating the limits gives an approximate risk-ratio interval of \(0.894\) to \(1.99\). This interval includes 1, the value corresponding to equal risks.

These intervals are illustrative large-sample calculations, not universal formulas for every response-adaptive design. If assignment probabilities change in response to accumulating outcomes, the usual independent-sample variance calculations may not reflect the randomization and sampling process. The analysis must use an inferential method appropriate to that design. Also, approximate intervals can be unreliable with very small counts or risks near zero or one.

Choosing and Comparing the Scales

Summary	Question it answers	Interpretation in the example	Main caution
Risk difference	How much does the outcome probability change absolutely?	10.0 percentage points higher under treatment	Absolute impact depends on the population’s baseline risk.
Risk ratio	How large is the treatment risk relative to control?	1.33 times the control risk, or 33.3% higher	A relative change is not a percentage-point change.
Favorable-outcome count	How many favorable outcomes occurred or are expected in a specified group?	48 versus 24 observed with unequal arm sizes; 40 versus 30 expected per 100	Raw counts depend on the number assigned to each arm.

The risk difference gives an absolute comparison, while the risk ratio gives a relative comparison. Neither makes the other redundant: the same risk ratio can correspond to different absolute differences at different baseline risks, and the same risk difference can correspond to different ratios. A count becomes interpretable as a treatment comparison only when the group sizes and the basis for the comparison are made clear.

In a fixed-size population, the estimated number of favorable outcomes under assignment to one arm is the arm’s size multiplied by its estimated risk. Comparing counts under alternative assignments requires specifying a common population size and treating the risks as estimates for that population. It is not valid to subtract observed outcomes across differently sized randomized arms and call the result the number of patients who benefited.

Common Mistakes and Diagnostics

Calling a 0.10 risk difference a 10% relative increase. It is an increase of 10 percentage points. The relative increase in this example is 33.3%, calculated from the risk ratio.
Comparing raw counts without checking denominators. The treatment arm has more observed favorable outcomes here, but it also has more assigned patients. Calculate the arm-specific risks before drawing a treatment-benefit conclusion.
Presenting a risk ratio without its baseline risk. A ratio alone can obscure the absolute probability difference. Report the underlying risks or the risk difference as well.
Treating standardized counts as observed facts. “40 per 100 under treatment” is an estimate based on the trial risk, not an observed count from a newly assigned group of 100.
Ignoring endpoint and follow-up definitions. The risks are comparable only if “favorable outcome” and the assessment time mean the same thing in both arms, and outcome ascertainment is adequate.

As a diagnostic, always check that each reported risk equals its favorable count divided by its own denominator. Then check the scale: risk differences should lie between -1 and 1, risk ratios should be nonnegative when defined, and expected counts should correspond to an explicitly stated group size. If arm sizes differ, do not treat a difference in raw counts as a difference in treatment effectiveness.

Key takeaway: For a binary outcome, the risk difference describes absolute change, the risk ratio describes relative change, and the favorable-outcome count describes outcomes for a specified number of patients. State the scale and denominators whenever drawing a patient-benefit conclusion.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

In the hypothetical trial, what are the estimated treatment and control risks, and what is their risk difference in percentage points?
What does the estimated risk ratio of 1.333 mean? Why is it not a 33.3-percentage-point increase?
Why can the observed count difference of 24 not be interpreted directly as 24 patients benefiting from treatment?
If the estimated risks are applied to 50 patients in each arm, what are the expected favorable-outcome counts and their difference?
What does it mean that the approximate 95% risk-difference interval includes zero?

Solutions. (1) The risks are \(48/120=0.400\) and \(24/80=0.300\), so the risk difference is \(0.400-0.300=0.100\), or 10.0 percentage points. (2) The treatment risk is about 1.333 times the control risk, a relative increase of about 33.3%. The absolute increase is the difference between probabilities, 0.100 or 10.0 percentage points. (3) The treatment arm has 120 patients and the control arm has 80; the counts reflect both outcome risks and group sizes. The count difference is not a count of individual causal outcomes. (4) Expected counts are \(50(0.400)=20\) and \(50(0.300)=15\), a difference of 5 expected favorable outcomes per 50 patients. (5) The approximate interval is compatible with a range of risk differences that includes no absolute difference, as well as negative and positive differences; it does not establish that the true difference is zero.