Introduction: Choose the Method From the Model
A mixed probability problem can contain both point masses and continuous outcomes, a latent condition that changes a distribution, or a sum whose exact distribution is inconvenient. The first step is not to select a formula by keyword. It is to identify the support of each variable, locate any atoms, and determine which quantities are independent or conditionally independent. Those features indicate whether conditioning, a cdf transformation, direct integration, an mgf, or an approximation is appropriate.
The methods are connected. Conditioning breaks a model into simpler cases; integration averages over a continuous set of cases; a cdf transformation handles events involving a function of a random variable; and an mgf can summarize sums of independent variables. Approximation is useful only after the model and its moments are understood.
A Practical Decision Process
Record the distribution, parameters, support, atoms, and any conditional assumptions. In particular, do not silently replace a mixed distribution by its continuous component.
For a transformed variable \(Y=g(X)\), solve \(g(X)\le y\) as a set of possible values of \(X\). For a mixture, identify how the event behaves in each component.
Apply total probability for discrete cases, integrate conditional probabilities over a continuous mixing distribution, or use a cdf transformation when the preimage of the event is clear.
Products of mgfs require independence. A normal approximation requires a suitable sum and a defensible approximation; having a mean and variance alone does not make an exact normal model.
Transformations: Keep the Atoms and the Continuous Part
For a transformed variable \(Y=g(X)\), the general cdf identity is \(F_Y(y)=P(g(X)\le y)\). This identity applies whether \(X\) is discrete, continuous, or mixed. Its advantage is that it does not require \(g\) to be one-to-one. The calculation must, however, include every part of the support that satisfies the event, including any atoms.
When \(X\) has a density \(f_X\), a probability over a region \(B\) of the continuous support is calculated by integrating \(f_X(x)\) over \(B\). If \(X\) also has atoms, their probabilities must be added separately whenever the atoms lie in \(B\). This is often safer than trying to apply a density transformation formula to a model that is not purely continuous.
Worked Example: An Absolute-Value Transformation of a Mixed Variable
A hypothetical loss index \(X\) equals 0 with probability 0.30. With probability 0.70, \(X\) is uniform on \((0,4)\). Let \(Y=|X-2|\). Find \(F_Y(y)\), \(P(Y\le1)\), and \(P(Y>0.8)\). The method is to translate each event into a region for \(X\) and account separately for the atom and the continuous component.
For \(0\le y<2\), the continuous-component event \(|X-2|\le y\) is \(2-y\le X\le2+y\). This interval has length \(2y\), so its conditional probability given the uniform component is \(2y/4=y/2\). The atom at \(X=0\) maps to \(Y=2\), so it does not contribute for \(y<2\). Therefore \(F_Y(y)=0.70(y/2)=0.35y\) on this interval. For \(y<0\), the cdf is zero. At \(y=2\), both the continuous component and the atom are included, so the cdf is one:
Thus \(P(Y\le1)=F_Y(1)=0.35\), or 0.3500 to four decimal places. For the second probability, the atom contributes 0.30 because it maps to \(Y=2>0.8\). In the uniform component, \(|X-2|>0.8\) means \(X<1.2\) or \(X>2.8\). The two intervals have total length \(1.2+1.2=2.4\) out of 4, giving conditional probability \(2.4/4=0.60\). Hence
The answer is 0.7200 to four decimal places. As a check, \(P(Y>0.8)=1-F_Y(0.8)=1-0.28=0.72\). The common mistake is to ignore the atom or to treat it as if it were spread across the uniform interval; the atom maps to the single value \(Y=2\).
Conditioning and Direct Integration
If a conditional model is simple but its parameter varies randomly, condition on that parameter first. For a discrete mixing variable \(Z\), the law of total probability gives \(P(A)=\sum_z P(A\mid Z=z)P(Z=z)\). For a continuous mixing variable \(\Theta\) with density \(f_\Theta\), the corresponding formula is
This is not an approximation: it averages the conditional probability over all possible values of the mixing variable. The integration limits must match the support of \(\Theta\).
Worked Example: A Poisson Count With a Continuous Random Rate
A hypothetical monitoring system has a random activity rate \(\Theta\), uniformly distributed on \((0,2)\). Given \(\Theta=\theta\), the count \(N\) during a fixed observation period is Poisson with mean \(\theta\). Find \(P(N=0)\), \(P(N\ge2)\), and \(\operatorname{Var}(N)\). The method is to condition on \(\Theta\) and integrate over its uniform density \(1/2\) on \((0,2)\).
For a Poisson variable with mean \(\theta\), \(P(N=0\mid\Theta=\theta)=e^{-\theta}\). Therefore
For \(N\ge2\), the conditional probability is \(1-e^{-\theta}(1+\theta)\). Thus
An antiderivative of \(e^{-\theta}(1+\theta)\) is \(-(\theta+2)e^{-\theta}\). Its integral from 0 to 2 is \(2-4e^{-2}\), so the required probability is \(1-(2-4e^{-2})/2=2e^{-2}\approx0.2707\). The probabilities are reported to four decimal places.
For the variance, conditional on \(\Theta\), \(E[N\mid\Theta]=\Theta\) and \(\operatorname{Var}(N\mid\Theta)=\Theta\). The law of total expectation gives \(E[N]=E[\Theta]=1\). The law of total variance gives
As a check, \(P(N=0)+P(N=1)+P(N\ge2)=1\): here \(P(N=1)=\frac12\int_0^2\theta e^{-\theta}d\theta=(1-3e^{-2})/2\), and the three terms sum to one. The trap is to substitute the average rate \(E[\Theta]=1\) into a single Poisson model. That would give \(P(N=0)=e^{-1}\), which is not the mixture probability \((1-e^{-2})/2\).
Discrete Mixtures, Moments, and mgfs
A mixture selects a component first and then draws from that component’s distribution. Its probability is found by weighting the component probabilities, and its mgf is the weighted average of the component mgfs. In general, this mixture is not itself a member of the same distribution family with the average parameter. In particular, a mixture of Poisson distributions is generally not Poisson.
Worked Example: Why an Average Poisson Rate Is Not Enough
A hypothetical environment is calm with probability 0.80 and busy with probability 0.20. Conditional on calm, \(N\) is Poisson with mean 1; conditional on busy, \(N\) is Poisson with mean 3. Find \(P(N=0)\), \(P(N\ge2)\), and \(\operatorname{Var}(N)\). The method is to condition on the environment for probabilities and moments.
The zero-count probability is
For one count, \(P(N=1\mid\lambda)=\lambda e^{-\lambda}\), so
Consequently, \(P(N\ge2)=1-P(N=0)-P(N=1)=0.3715632394\ldots\), or 0.3716 to four decimal places. For the moments, \(E[N]=0.80(1)+0.20(3)=1.4\). A Poisson variable with mean \(\lambda\) has second moment \(\lambda+\lambda^2\), so \(E[N^2]=0.80(1+1^2)+0.20(3+3^2)=0.80(2)+0.20(12)=4\). Therefore
The mgf also shows the mixture structure. The conditional Poisson mgf is \(\exp(\lambda(e^t-1))\), so here
This is not the mgf of a Poisson variable with mean 1.4. Indeed, the mixture variance 2.04 exceeds its mean 1.4, whereas a Poisson variable has variance equal to its mean. As a check, the three probabilities for \(N=0\), \(N=1\), and \(N\ge2\) sum to one. Replacing both conditional rates by their average would give \(P(N=0)=e^{-1.4}\approx0.2466\), not the correct 0.3043.
mgfs and Approximation for Independent Sums
For a random variable \(X\), its mgf is \(M_X(t)=E[e^{tX}]\) for values of \(t\) where the expectation exists. If \(X_1,\ldots,X_n\) are independent, the mgf of their sum \(S=\sum_{j=1}^nX_j\) is \(M_S(t)=\prod_{j=1}^nM_{X_j}(t)\). Independence is essential to this product rule. For identically distributed independent variables, \(M_S(t)=[M_X(t)]^n\). Derivatives at zero give moments when the mgf exists near zero.
An mgf does not automatically make a tail probability easy to calculate. It can instead provide the mean and variance needed for a normal approximation. For a sufficiently large sum of independent, identically distributed variables with finite variance, the central limit theorem motivates approximating the standardized sum by a standard normal distribution. The approximation is not an exact identity; skewness, sample size, and the event’s location all affect its accuracy.
Worked Example: A Sum of Zero-or-Exponential Payments
For each of 80 independent hypothetical devices, a payment \(Y_i\) is zero with probability 0.75. With probability 0.25, the payment has an exponential distribution with rate 0.5, so its conditional mean is 2. Estimate \(P(S>50)\), where \(S=\sum_{i=1}^{80}Y_i\). The method is to obtain the moments from the mixture and then apply a normal approximation to the independent sum.
For the exponential component, \(E[Y_i\mid Y_i>0]=2\) and \(E[Y_i^2\mid Y_i>0]=2(2^2)=8\). Including the zero-payment component gives
Thus \(\operatorname{Var}(Y_i)=2-(0.5)^2=1.75\). Independence gives \(E[S]=80(0.5)=40\) and \(\operatorname{Var}(S)=80(1.75)=140\), so the standard deviation is \(\sqrt{140}\approx11.8322\). The standardized threshold is
The summands are continuous except for a point mass at zero, but \(S>50\) excludes that atom. No continuity correction is needed for this continuous tail threshold. Using the central limit theorem approximation, \(P(S>50)\approx1-\Phi(0.8452)\approx0.1990\), to four decimal places. The mgf of one payment, for \(t<0.5\), is \(M_{Y_i}(t)=0.75+0.25(0.5/(0.5-t))\); independence gives \(M_S(t)=[M_{Y_i}(t)]^{80}\), consistent with the moment calculation. The check is that the threshold 50 is \(10\) above the mean, or about \(0.8452\) standard deviations. Treating the mixture payment as always exponential would discard its zero payments and produce incorrect moments.
Common Method-Selection Traps
- Dropping an atom during a transformation. Solve the event for all of the support, then check whether each point mass maps inside the event.
- Using an average parameter as if it removed mixing. Average component probabilities or conditional moments; do not assume a mixture retains the component family.
- Integrating over the wrong support. A continuous mixing variable must be averaged using its density and its actual range.
- Multiplying mgfs without independence. For dependent variables, the mgf of a sum is \(E[e^{t(X+Y)}]\); it cannot generally be factored into \(M_X(t)M_Y(t)\).
- Using a normal approximation before checking the model. Establish independence, moments, and the relevant scale first. A continuity correction applies to a discrete target, not automatically to every problem involving a sum.
These method choices support integrated problem solving aligned with the SOA Exam P (Probability) syllabus. A clear solution makes its model and assumptions visible, explains why its calculation applies, and checks the result against support, total probability, or moments.
Check Your Understanding
For each problem, decide which method fits the model before calculating.
- A variable \(X\) equals 1 with probability 0.20 and otherwise is uniform on \((0,3)\). Set \(Y=|X-1.5|\). Find \(P(Y\le0.5)\), including the atom. (Answer: 0.4667)
- \(\Theta\) is uniform on \((0,1)\), and conditional on \(\Theta=\theta\), \(N\) is Poisson with mean \(2\theta\). Find \(P(N=0)\). (Answer: 0.4323)
- A count is Poisson with mean 2 with probability 0.25 and Poisson with mean 0.5 with probability 0.75. Find its unconditional probability of zero. (Answer: 0.4887)
- Each of 25 independent payments is zero with probability 0.60 and otherwise exponential with rate 1. Find the mean and variance of the total payment. (Answer: mean 10, variance 16)
- Let \(B\sim\operatorname{Binomial}(100,0.30)\). Use a normal approximation with continuity correction to estimate \(P(B\ge38)\). (Answer: approximately 0.0509)