Cumulative Probabilities for Ordered Categories

Why Summarize an Ordered Response Cumulatively?

For an ordinal response, the categories have a meaningful order, even when the gaps between categories are not quantitatively meaningful. For example, pain ratings of “none,” “mild,” “moderate,” and “severe” can be ranked, but the difference between none and mild need not equal the difference between moderate and severe. A cumulative probability uses this ordering to answer a direct question: what is the probability that the response is at or below a specified category?

This summary differs from simply listing the probability of each category. A category probability describes one category alone; a cumulative probability combines all categories up through a cutpoint. The cutpoints therefore follow the response order, and their meaning depends on that order.

Definition and Calculation

Suppose \(Y\) is an ordinal response with \(K\) ordered categories, labelled \(1,2,\ldots,K\) in increasing order. Write \(\pi_k=P(Y=k)\) for the probability of category \(k\). The labels represent the order; they do not, by themselves, assert equal spacing between categories.

Definition: For each cutpoint \(j=1,\ldots,K-1\), the cumulative event is \(\{Y\leq j\}\), and its cumulative probability is \(P(Y\leq j)\). Because the categories \(1,\ldots,j\) are mutually exclusive and together form this event, \(P(Y\leq j)=\sum_{k=1}^{j}\pi_k\). We may also use the endpoint conventions \(P(Y\leq 0)=0\) and \(P(Y\leq K)=1\).

To calculate a cumulative probability, add the category probabilities from the lowest category through the cutpoint. For example, \(P(Y\leq 2)\) includes the probabilities of categories 1 and 2, not just category 2. The endpoint at \(K\) includes every possible response and consequently has probability 1.

Worked Example: Cumulative Probabilities for Four Ordered Categories

Suppose a response has categories 1 = low, 2 = moderate, 3 = high, and 4 = very high, with category probabilities \(0.10, 0.25, 0.40,\) and \(0.25\), respectively. First check that the probabilities form a distribution:

$$ 0.10+0.25+0.40+0.25=1.00. $$

At the first cutpoint, only category 1 is included:

P(Y\leq 1)=\pi_1=0.10.

At the second cutpoint, categories 1 and 2 are included:

P(Y\leq 2)=\pi_1+\pi_2=0.10+0.25=0.35.

At the third cutpoint, categories 1, 2, and 3 are included:

P(Y\leq 3)=\pi_1+\pi_2+\pi_3=0.10+0.25+0.40=0.75.

Finally, all four categories are included at the endpoint: \(P(Y\leq 4)=1.00\). Thus the cumulative probabilities at the three interior cutpoints are \(0.10, 0.35,\) and \(0.75\). In context, 35% of responses are moderate or lower, and 75% are high or lower.

What the Cumulative Summary Says

Each cutpoint creates a split of the ordered categories into those at or below the cutpoint and those above it. For the four-category example, \(P(Y\leq 2)\) is the probability of a moderate-or-lower response. The complementary event is \(Y>2\), meaning a high-or-very-high response. These statements describe the same split from opposite sides.

P(Y\leq j)+P(Y>j)=1.

A cumulative probability is not a probability conditional on being in a subset of categories. It is an ordinary probability under the distribution of \(Y\). For instance, \(P(Y\leq 2)\) is the fraction of all responses expected to be in categories 1 or 2, not the fraction in category 1 among responses in categories 1 or 2.

Worked Example: Interpreting Clinical Severity Categories

A clinical assessment classifies patients as 1 = no symptoms, 2 = mild symptoms, 3 = moderate symptoms, or 4 = severe symptoms. Suppose the category probabilities are \(0.15, 0.35, 0.30,\) and \(0.20\). They sum to \(1.00\), so they account for all possible categories.

The probability of no symptoms or mild symptoms is

P(Y\leq 2)=0.15+0.35=0.50.

Thus half of the patients are expected to have mild symptoms or less. The probability of moderate symptoms or less is

P(Y\leq 3)=0.15+0.35+0.30=0.80.

The complementary probability of severe symptoms is \(P(Y>3)=1-0.80=0.20\), agreeing with the category probability for category 4. The calculation makes clear which categories are included in each clinical summary.

Cumulative summaries retain the order because each cutpoint refers to a contiguous lower portion of the scale. They do not treat categories as unrelated labels. This is useful when the scientific question concerns being no worse than a given level, or being above a given level. The specific cutpoint must still be reported: “at or below mild” and “at or below moderate” are different events.

Worked Example: Comparing Two Ordered Distributions

Consider two groups assessed on a three-category scale: 1 = favorable, 2 = intermediate, and 3 = unfavorable. Suppose the category probabilities are:

Group	Favorable (1)	Intermediate (2)	Unfavorable (3)
A	0.30	0.40	0.30
B	0.50	0.30	0.20

For Group A, the cumulative probability at the first cutpoint is \(P(Y\leq 1)=0.30\), and at the second it is \(P(Y\leq 2)=0.30+0.40=0.70\). For Group B, the corresponding values are \(P(Y\leq 1)=0.50\) and \(P(Y\leq 2)=0.50+0.30=0.80\).

At both cutpoints, Group B has the larger probability of a response at or below the stated category: 0.50 versus 0.30 at the favorable cutpoint, and 0.80 versus 0.70 at the intermediate cutpoint. The category probabilities also show the distributions, but cumulative summaries express the comparison in terms of the ordered thresholds. These calculations alone do not establish why the groups differ or whether the differences are statistically meaningful.

Two Basic Properties

The cumulative events are nested: if a response is at or below category \(j\), it is also at or below category \(j+1\). This set relationship gives an immediate ordering of the cumulative probabilities.

Result: For any distribution on ordered categories, \(P(Y\leq j)\leq P(Y\leq j+1)\) for \(j=0,\ldots,K-1\).

Proof. The event \(\{Y\leq j\}\) is contained in \(\{Y\leq j+1\}\). Therefore, \(\{Y\leq j+1\}\) is the disjoint union of \(\{Y\leq j\}\) and \(\{Y=j+1\}\). By additivity of probability,

P(Y\leq j+1)=P(Y\leq j)+P(Y=j+1)\geq P(Y\leq j),

since probabilities are nonnegative. This also covers the endpoints under the conventions \(P(Y\leq 0)=0\) and \(P(Y\leq K)=1\). \(\square\)

A second useful identity connects cumulative probabilities to the mean when the categories are coded \(1,\ldots,K\). This identity uses the numerical coding to form a mean; it does not imply that the category steps represent equal amounts of an underlying quantity.

Result: If \(Y\) takes values \(1,\ldots,K\), then \(E[Y]=K-\sum_{j=1}^{K-1}P(Y\leq j)\).

Proof. For any particular response value \(y\) in \(\{1,\ldots,K\}\), the indicators \(1\{y>j\}\), for \(j=1,\ldots,K-1\), equal 1 exactly when \(j<y\). There are \(y-1\) such integers \(j\). Hence the following identity holds for every possible value of \(Y\):

Y=1+\sum_{j=1}^{K-1}1\{Y>j\}.

Taking expectations and using linearity for this finite sum gives

E[Y]=1+\sum_{j=1}^{K-1}P(Y>j) =1+\sum_{j=1}^{K-1}\bigl(1-P(Y\leq j)\bigr) =K-\sum_{j=1}^{K-1}P(Y\leq j).

The last equality follows because there are \(K-1\) interior cutpoints: \(1+(K-1)=K\). \(\square\)

Worked Example: Checking the Mean From Cumulative Probabilities

For the four-category distribution \(0.10, 0.25, 0.40, 0.25\), the cumulative probabilities at the interior cutpoints are \(0.10, 0.35,\) and \(0.75\). The identity gives

$$ E[Y]=4-(0.10+0.35+0.75)=4-1.20=2.80. $$

Direct calculation from the category probabilities checks the result:

$$ E[Y]=1(0.10)+2(0.25)+3(0.40)+4(0.25) =0.10+0.50+1.20+1.00=2.80. $$

The agreement verifies the arithmetic. The interpretation of this mean depends on treating the category labels as numerical scores; the cumulative probabilities themselves require only the ordering.

Common Mistakes and Practical Interpretation

Using only the probability at the cutpoint. For \(P(Y\leq 2)\), adding \(\pi_1+\pi_2\) is required. Using \(\pi_2\) alone gives \(P(Y=2)\), a different quantity.
Including categories above the cutpoint. \(P(Y\leq j)\) includes categories \(1\) through \(j\), not categories \(j\) through \(K\). The latter categories describe a different event.
Calling the cumulative probability conditional. \(P(Y\leq j)\) uses the full response distribution as its denominator. A conditional probability would require an additional conditioning event.
Ignoring the category order. A statement such as “at or below category 2” is meaningful only after the categories and their ordering have been specified.
Assuming equal category spacing. Cumulative probabilities use the order, not equal distances between categories. A numerical mean may require a separate justification if the category codes are merely ordinal labels.

Key takeaway: For ordered categories, the cumulative probability at cutpoint \(j\) is the sum of the probabilities from category 1 through category \(j\). It answers a threshold question about the ordered response, while retaining a clear distinction between the cutpoint event and any single category.

Check Your Understanding

Use the category order stated in each question and show the probability sums.

A five-category response has probabilities \(0.08, 0.17, 0.30, 0.25,\) and \(0.20\). Calculate \(P(Y\leq 2)\) and \(P(Y\leq 4)\).
In an ordered three-category scale, explain the difference between \(P(Y\leq 2)\) and \(P(Y=2)\).
Prove that \(P(Y\leq j)\leq P(Y\leq j+1)\) using the relationship between the two events.
For categories coded 1 through 3 with cumulative probabilities \(0.20\) and \(0.65\), calculate \(E[Y]\) using the identity derived above.
Why does calculating a cumulative probability require an ordering of categories, even though it does not require equal spacing between them?