Why the Distinction Matters
A statistical method begins with a sample, but not every quantity involving sample data is a statistic. The distinction depends on how an expression is defined, not on whether it can be calculated after data have been recorded. In particular, a statistic must be a function of the sample that does not involve unknown population parameters.
This distinction keeps three ideas separate: the sample observations before they are collected, the fixed but possibly unknown features of the population, and the numerical results obtained when the observations are recorded. It is aligned with the SOA VEE Mathematical Statistics syllabus.
Random Sample, Parameter, Statistic, and Observed Value
A random sample is written \(X_1,\ldots,X_n\), with capital letters indicating random variables. Under the usual random-sample model, the observations are independent and have the same distribution. The distribution is described by one or more population parameters, represented here by \(\theta\). A parameter is fixed for a particular population or model, even when its value is unknown.
A statistic is a rule applied to the random sample. For example, \(X_1+X_2\) is a statistic: the rule is to add the first two observations, and no unknown parameter appears. If the observed values are \(x_1=4\) and \(x_2=7\), the observed value is \(4+7=11\). The number 11 is not the rule and is not a population parameter; it is the result of applying the statistic to this particular sample.
| Object | Typical notation | What it represents |
|---|---|---|
| Random sample | \(X_1,\ldots,X_n\) | The random observations before their values are known |
| Observed data | \(x_1,\ldots,x_n\) | The realized numerical values recorded after sampling |
| Parameter | \(\theta\), or a component such as \(\mu\) | A fixed feature of the model, often unknown |
| Statistic | \(T=g(X_1,\ldots,X_n)\) | A parameter-free rule applied to the random sample |
| Observed statistic value | \(g(x_1,\ldots,x_n)\) | The numerical result from applying the rule to the observed data |
The requirement that a statistic contain no unknown parameters concerns the formula defining the rule. A statistic’s probability distribution may depend on an unknown parameter; that does not make the statistic itself parameter-dependent. For instance, the sample sum is a statistic whether or not its distribution changes with the population mean.
Worked Example: Calculating a Statistic and Its Observed Value
A laboratory takes a random sample of four sealed packets from a production run and records the number of loose granules in each packet. Assume the packet measurements are independent and identically distributed, with an unknown population mean \(\mu\). The recorded values are \(2, 5, 4,\) and \(9\). Define the sample-average statistic \(T=(X_1+X_2+X_3+X_4)/4\). Identify the parameter, the statistic, the observed data, and the observed statistic value.
Step 1: Identify the parameter. The population mean \(\mu\) is fixed for the production run but is unknown. It is a parameter, not an observation from these four packets.
Step 2: Identify the statistic. The rule is to add the four sample observations and divide by four: \(T=(X_1+X_2+X_3+X_4)/4\). The rule contains no \(\mu\), so \(T\) is a statistic.
Step 3: Substitute the observed data. The observed values are \(x_1=2\), \(x_2=5\), \(x_3=4\), and \(x_4=9\). Their sum is \(2+5+4+9=20\), so the observed statistic value is \(20/4=5\).
Check. The result is an observed sample average of 5 granules per packet. It is not a claim that the unknown population mean \(\mu\) equals 5. The method avoids treating an observed statistic as if it were automatically the population parameter.
How to Test Whether an Expression Is a Statistic
For a proposed expression, first write it as a rule in the random variables \(X_1,\ldots,X_n\). Then identify every symbol that represents an unknown population quantity. If the rule requires such a quantity, it is not a statistic under the definition above. If the rule uses only sample observations, the sample size, and known constants, it is a statistic.
For example, if \(\mu\) is unknown, \(X_1-\mu\) is not a statistic because its value cannot be determined from the sample alone. By contrast, \(X_1-X_2\) is a statistic. A known constant may appear: if a fixed threshold \(c=10\) has been specified in advance, the count of observations exceeding 10 is a statistic. The threshold is known, not an unknown feature of the population.
Keep the distinction between \(X_i\), a random observation, and \(x_i\), its recorded value.
Check whether a symbol such as \(\mu\), \(\sigma^2\), or \(\theta\) is an unknown model parameter.
If the rule requires an unknown parameter, it is not a statistic. If it uses only the sample and known quantities, it is.
Substitute the observed values to find the observed statistic value; do not confuse that number with either the rule or a population parameter.
Worked Example: Statistics in a Bernoulli Sample
A hypothetical screening procedure records whether each of five randomly selected test units triggers a signal. Let \(X_i=1\) if unit \(i\) triggers a signal and \(X_i=0\) otherwise. Assume the observations are independent Bernoulli random variables with unknown trigger probability \(p\). The observed sequence is \(1,0,1,1,0\). Consider the expressions \(T=(X_1+\cdots+X_5)/5\), \(U=p\), and \(V=(T-p)^2\). Classify each and calculate the observed value of \(T\).
Step 1: Classify \(T\). The expression adds the five random observations and divides by the known sample size 5. It contains no unknown parameter, so \(T\) is a statistic.
Step 2: Calculate the observed value of \(T\). Substituting the data gives \(t=(1+0+1+1+0)/5=3/5=0.6\). Thus \(0.6\) is the observed value of the statistic \(T\).
Step 3: Classify \(U\). The quantity \(p\) is the fixed probability that a randomly selected unit triggers a signal. It describes the model and is unknown; it is a parameter, not a statistic.
Step 4: Classify \(V\). Although \(T\) is a statistic, \(V=(T-p)^2\) explicitly includes the unknown parameter \(p\). Therefore \(V\) is not a statistic as written. It cannot be evaluated from the sample alone.
Check. The sample contains three signals out of five, giving \(t=0.6\). That observed proportion need not equal \(p\). The method avoids treating a statistic’s observed value as though it removes the unknown parameter from a different expression.
Why a Statistic Is Random Before Sampling
The word “statistic” refers to a random quantity before the observations are available. This follows directly from its definition as a function of random observations. If \(T=g(X_1,\ldots,X_n)\), then different possible sample outcomes can produce different values of \(T\). Once one sample has been observed, the function is evaluated at that one outcome and gives a fixed numerical result.
This reasoning also explains why the probability distribution of a statistic can depend on an unknown parameter without violating the definition. The function \(g\) must be parameter-free; the probabilities of the sample outcomes to which \(g\) is applied may still depend on \(\theta\).
The probability of this event is determined by the distribution of the random sample. That sample distribution may involve \(\theta\), so the distribution of \(T\) may involve \(\theta\) as well. However, if \(g\) itself contains no unknown parameter, \(T\) still meets the definition of a statistic. The function and its distribution are different objects.
Worked Example: A Threshold Count and an Unknown-Parameter Expression
Suppose four independent measurements \(X_1,\ldots,X_4\) follow an exponential model with unknown rate parameter \(\lambda\). The observed measurements, in hours, are \(0.4, 1.0, 1.6,\) and \(2.2\). Define \(C\) as the number of observations strictly greater than the known threshold of 1 hour, and define \(Q=\lambda(X_1+X_2+X_3+X_4)\). Determine which expression is a statistic and evaluate the statistic from the data.
Step 1: Classify \(C\). The rule counts observations above a fixed, known threshold. It uses only the sample and the known number 1, so \(C\) is a statistic.
Step 2: Apply the strict comparison. The observations \(0.4\), \(1.0\), \(1.6\), and \(2.2\) are compared with 1. The values \(1.6\) and \(2.2\) are strictly greater than 1; the value \(1.0\) is equal to 1 and is not counted. Therefore the observed value is \(c=2\).
Step 3: Classify \(Q\). The expression includes the unknown rate parameter \(\lambda\). Even though the sum of the observations can be calculated, \(Q\) cannot be evaluated from the sample alone. It is not a statistic as written.
Step 4: Check the sample sum. The observed sum is \(0.4+1.0+1.6+2.2=5.2\) hours. This does not make \(Q\) a statistic: its value would still require \(\lambda\), which is unknown.
Check. The threshold count is 2, and the sum is 5.2 hours. Counting equality as exceeding the threshold would give the wrong count; multiplying the sum by an unknown rate does not produce a statistic.
Common Mistakes and Exam Traps
- Calling every sample-based expression a statistic. An expression such as \(X_1-\mu\) uses a sample observation but also requires the unknown parameter \(\mu\). It is not a statistic.
- Confusing the rule with its observed value. \(T=X_1+X_2\) is a statistic; if the data are 3 and 8, the observed value is 11. The statistic is not itself the number 11.
- Assuming a statistic cannot have a parameter-dependent distribution. The rule must not contain an unknown parameter. Its distribution can nevertheless depend on parameters that govern the sample.
- Replacing an unknown parameter without stating a new rule. If an expression uses unknown \(\mu\), substituting a sample-based quantity instead defines a different expression. It does not change the classification of the original expression.
- Forgetting that known constants are allowed. A rule using a threshold specified in advance is still parameter-free, even if the threshold appears numerically in the formula.
A reliable answer states the roles precisely: identify the sample observations, name any unknown population parameter, write the proposed function, and then say whether that function depends on the unknown parameter. If it is a statistic and data are supplied, evaluate the function and label the result as its observed value.
Check Your Understanding
For each problem, identify the statistic or parameter as requested and distinguish the rule from its observed value.
- A random sample of four measurements has observed values \(4,6,5,9\). Define \(T=(X_1+X_2+X_3+X_4)/4\) and \(C=\#\{i:X_i>6\}\). Which are statistics, and what are their observed values? (Answer: Both are statistics; \(t=6\) and \(c=1\).)
- Six independent Bernoulli observations have values \(1,1,0,1,0,0\), with unknown success probability \(p\). Classify \(R=X_1+\cdots+X_6\) and \(p\), and evaluate \(R\). (Answer: \(R\) is a statistic with observed value \(3\); \(p\) is an unknown parameter.)
- A random sample has observed values \(3.2,4.1,2.8\), and the population mean \(\mu\) is unknown. Is \(M=\max(X_1,X_2,X_3)\) a statistic? Is \(D=\bar X-\mu\)? Evaluate \(M\) and the observed sample mean. (Answer: \(M\) is a statistic with observed value \(4.1\); \(D\) is not a statistic; the observed sample mean is \(10.1/3\approx3.3667\).)
- Four measurements from an exponential model with unknown rate \(\lambda\) are \(0.3,1.0,1.4,2.2\). Define \(N=\#\{i:X_i>1\}\) and \(W=\lambda\sum_{i=1}^4X_i\). Classify both expressions and evaluate any statistic. (Answer: \(N\) is a statistic with observed value \(2\); \(W\) is not a statistic because it contains unknown \(\lambda\).)
- Suppose \(X_1,X_2,X_3\) are a random sample from a model with unknown parameter \(\theta\), and \(a=5\) is a known constant. Classify \(A=5(X_1-X_2)\) and \(B=\theta(X_1-X_2)\). Explain whether the distribution of \(A\) must be free of \(\theta\). (Answer: \(A\) is a statistic; \(B\) is not. The distribution of \(A\) may depend on \(\theta\), even though its defining rule does not.)