Why Count-Based Reasoning Helps
Bayes' Rule answers a reverse question: after observing evidence \(E\), what is the probability that event \(A\) occurred? The information in a problem is often given in the opposite direction, as the probability of seeing \(E\) among cases where \(A\) is true. Natural frequencies make the reversal concrete. Instead of manipulating several probabilities at once, imagine a sufficiently large group, count how many cases have the evidence, and ask what fraction of those cases also belong to \(A\).
The counts are a way of organizing the stated probabilities, not necessarily records from a real group of people or objects. A carefully chosen hypothetical population can turn percentages into whole numbers and make the denominator of the answer visible.
From Probabilities to Counts
Suppose \(A\) and its complement \(A^c\) form two disjoint groups that include every possibility. Let \(E\) be the evidence observed. Assume \(P(A)>0\), \(P(A^c)>0\), and that \(P(E\mid A)\) and \(P(E\mid A^c)\) are defined. To calculate \(P(A\mid E)\), we must also have \(P(E)>0\); otherwise there are no cases matching the evidence and the conditional probability is undefined.
Imagine \(N\) cases. The count in group \(A\) is \(NP(A)\), and the count in group \(A^c\) is \(NP(A^c)\). Among the \(A\) cases, the count with evidence \(E\) is \(NP(A)P(E\mid A)\). Among the \(A^c\) cases, it is \(NP(A^c)P(E\mid A^c)\). Add those two counts to get the total matching \(E\). The fraction of that total that comes from \(A\) is the desired reverse conditional probability.
The condition \(P(E)>0\) is important: by the law of total probability, \(P(E)=P(A)P(E\mid A)+P(A^c)P(E\mid A^c)\). This quantity is exactly the denominator in the count-based expression after cancelling \(N\). If it equals zero, no hypothetical cases match \(E\), so the ratio cannot define \(P(A\mid E)\).
A useful choice for \(N\) makes the first set of counts whole numbers. For example, if the group probabilities are 0.10 and 0.90, \(N=1{,}000\) gives 100 and 900 cases. Conditional probabilities then determine how many cases within each group match the evidence. The same ratio results for any positive \(N\), because \(N\) multiplies every count and cancels.
Worked Example: Messages Flagged by a Filter
Imagine a hypothetical set of 1,000 messages. Ten percent come from a project mailing list, and 80% of those messages are flagged by a filter. Among messages not from the list, 5% are flagged. Given that a message is flagged, what is the probability it came from the project list?
Let \(A\) mean “the message came from the project list” and \(E\) mean “the message was flagged.” The groups \(A\) and \(A^c\) exhaust the messages. The stated probabilities are \(P(A)=0.10\), \(P(A^c)=0.90\), \(P(E\mid A)=0.80\), and \(P(E\mid A^c)=0.05\).
Count the messages from the list that are flagged. There are \(1{,}000(0.10)=100\) messages from the list. Of these, \(100(0.80)=80\) are flagged.
Count flagged messages not from the list. There are \(1{,}000(0.90)=900\) other messages. Of these, \(900(0.05)=45\) are flagged.
There are \(80+45=125\) flagged messages in all, so \(P(E)>0\) and the conditional probability is defined. Of those 125, 80 came from the list:
In this hypothetical population, 64% of flagged messages come from the project list. This is not the same as the 80% probability that a message from the list is flagged: the question changes the group under consideration from list messages to flagged messages.
A Step-by-Step Method
The arithmetic is usually simple once the correct denominator is identified. Use this sequence when probabilities are given for two complementary groups.
Let \(A\) be the group whose probability you want after observing \(E\). State clearly what counts as evidence.
Select \(N\) so that multiplying by the given group probabilities produces simple counts.
For each group, multiply its count by the conditional probability of \(E\) in that group.
Add the evidence counts across groups, then divide the count in the target group by that total. Confirm that the evidence count is positive.
This method also works when the evidence is uncommon. The critical question is not simply how often the evidence occurs in the target group; it is how many evidence-matching cases come from each group. A group can contribute many such cases because it is large, even if the evidence is less likely within that group.
Worked Example: Defective Parts From Two Machines
A hypothetical factory produces 30% of its parts on Machine 1 and 70% on Machine 2. Suppose 1% of Machine 1's parts and 3% of Machine 2's parts are defective. Given that a randomly selected part is defective, find the probability it came from Machine 1.
Let \(A\) mean “the part came from Machine 1” and \(E\) mean “the part is defective.” Imagine 10,000 parts. This size gives 3,000 parts from Machine 1 and 7,000 from Machine 2.
The defective count from Machine 1 is \(3{,}000(0.01)=30\). The defective count from Machine 2 is \(7{,}000(0.03)=210\). Thus, the total number of defective parts is \(30+210=240\), which is positive, so conditioning on defectiveness is defined.
The fraction of defective parts that came from Machine 1 is:
Therefore, 12.5% of defective parts in this hypothetical production mix came from Machine 1. The answer is not 1%: that is the probability of a defect among Machine 1 parts, whereas the question asks about the machine among defective parts.
Why the Count Ratio Gives the Conditional Probability
The count method is not a separate rule from probability; it follows directly from the definition of conditional probability. For events \(A\) and \(E\) with \(P(E)>0\), \(P(A\mid E)\) is the probability of the joint event \(A\cap E\), divided by the probability of \(E\). In a hypothetical population of size \(N\), the joint count is \(NP(A\cap E)\), while the evidence count is \(NP(E)\). Taking their ratio cancels the scale \(N\).
For two complementary groups, \(A\cap E\) has probability \(P(A)P(E\mid A)\), by the multiplication rule. The evidence event \(E\) consists of the disjoint cases \(A\cap E\) and \(A^c\cap E\), so its probability is the sum of their probabilities. These facts produce the count formula shown earlier. They also show why the denominator must include evidence from every possible group, not only evidence from the target group.
Worked Example: Evidence From Three Sources
Suppose a package comes from Source A with probability 0.20, Source B with probability 0.50, and Source C with probability 0.30. A characteristic \(E\) occurs in 60% of Source A packages, 20% of Source B packages, and 10% of Source C packages. Given a package has the characteristic, find the probability it came from Source A.
The three source groups are mutually exclusive and exhaustive. Imagine 1,000 packages. There are 200 from A, 500 from B, and 300 from C. The counts with the characteristic are \(200(0.60)=120\) from A, \(500(0.20)=100\) from B, and \(300(0.10)=30\) from C.
The total count matching the evidence is \(120+100+30=250\), which is positive. Of these, 120 are from A. Thus:
In this hypothetical population, 48% of packages with the characteristic come from Source A. The 60% figure describes how often the characteristic appears among packages from A; 48% answers the reverse question among all packages with the characteristic.
Choosing Counts and Interpreting Them
A hypothetical population is a calculation aid. It does not assert that exactly \(N\) cases were observed. If a problem gives probabilities that do not yield whole-number counts for a convenient \(N\), choose a larger multiple when practical. For instance, a group probability of 0.025 becomes 25 cases in a hypothetical population of 1,000. If no convenient population makes every count whole, the products can remain fractional: they represent probability-scaled counts, and the ratio still works.
For multiple groups, the same logic applies. Each group contributes a count equal to its group size multiplied by the conditional probability of the evidence in that group. Add all contributions to find the denominator, then take the share from the group of interest. This approach depends on the groups covering all possibilities and on the total evidence probability being positive.
Common Mistakes
- Using the wrong denominator. Dividing by the number in the target group answers a forward conditional question. For \(P(A\mid E)\), divide by all cases matching \(E\).
- Keeping only the target group's evidence count. The denominator must include evidence from every group in the partition.
- Reversing a conditional without recalculating. \(P(E\mid A)\) and \(P(A\mid E)\) generally differ because they use different reference groups.
- Conditioning on impossible evidence. If the evidence count is zero, then \(P(E)=0\), and \(P(A\mid E)\) is undefined. Do not report a ratio with a zero denominator.
- Treating hypothetical counts as observed data. Counts created from stated probabilities illustrate the calculation; they are not evidence that an actual group of that size was sampled.
Check Your Understanding
For each question, identify the evidence denominator before calculating or explaining the answer.
- In a hypothetical group of 2,000 items, 160 are both from Group A and match evidence \(E\), while 240 are from other groups and match \(E\). What is the natural-frequency estimate of \(P(A\mid E)\)?
- Why is the number of cases matching \(E\), rather than the number in group \(A\), the denominator for \(P(A\mid E)\)?
- Suppose \(P(A)=0.4\), \(P(A^c)=0.6\), \(P(E\mid A)=0\), and \(P(E\mid A^c)=0\). Is \(P(A\mid E)\) defined? Explain.
- A problem gives \(P(E\mid A)=0.7\). Explain why this alone does not determine \(P(A\mid E)\).
- When a hypothetical population produces fractional counts, what do those counts represent, and can their ratio still be used?