Why the Distinction Matters
A parameter describes a population or probability model; it is not directly observed in a finite sample. An estimator is a rule for using sample data to estimate that parameter. Once the observations are available, applying the rule produces a number: the estimate. These terms are closely related, but they do not name the same object.
The distinction becomes especially important when the target is a function of a parameter. For example, an actuary might want to estimate a model-based survival probability, not just the model’s mean lifetime. A plug-in rule first estimates the unknown parameter and then substitutes that estimate into the formula for the target. The result is still an estimator before data are observed, and becomes an estimate after they are observed.
From a Random Sample to a Number
Let \(X_1,\ldots,X_n\) denote the random observations in a sample. Before the sample is observed, an estimator is written as a function of those random variables, such as \(\hat\theta(X_1,\ldots,X_n)\). After observing the sample values \(x_1,\ldots,x_n\), the estimate is \(\hat\theta(x_1,\ldots,x_n)\). Uppercase letters emphasize random quantities; lowercase letters denote the realized observations.
The estimator is itself random before the data are collected: a different possible sample could produce a different value. The estimate is the particular value calculated from the sample at hand. Neither should be confused with \(\theta\), the fixed but unknown population parameter. For instance, writing \(\hat\theta=3.4\) without explaining whether \(3.4\) is a formula or a calculated value can obscure what is known and what remains uncertain.
Worked Example: Estimating an Event Probability
A hypothetical quality-control review records whether each of eight independently selected components passes a specified inspection. Let \(X_i=1\) if component \(i\) passes and \(X_i=0\) otherwise. Assume \(X_i\sim\operatorname{Bernoulli}(p)\), independently, where the unknown parameter \(p\) is the probability that a component passes. Use the sample proportion as the estimator of \(p\). The observed outcomes are \(1,0,1,1,0,1,0,1\). Identify the estimator and calculate the estimate.
Method. Write the estimator as a function of the random sample first, then evaluate it at the observed values.
There are \(n=8\) observations, so the sample proportion is
This formula is the estimator. It is random before observing the outcomes because the indicators \(X_i\) are random. In the observed sample, five entries equal 1, so their sum is \(5\). Substitution gives
Thus the estimate of \(p\) is \(0.625\), or \(62.5\%\). The unknown parameter remains \(p\); the data have not made \(p\) itself equal to \(0.625\).
Check and common mistake. The observed values contain five passes out of eight, so the calculated proportion is between 0 and 1 and agrees with the count. A common error is to call \(p\) the estimator or to report \(5/8\) as the estimator without stating the rule. Here, \(p\) is the unknown parameter, \(\frac{1}{8}\sum X_i\) is the estimator, and \(5/8\) is the estimate.
Plug-In Rules for Functions of Parameters
Often the quantity of interest is not the parameter itself but a function of it. Suppose the target is \(q=h(\theta)\), where \(h\) is a specified function and \(\hat\theta\) is an estimator of \(\theta\). The plug-in rule replaces the unknown \(\theta\) in the target formula by \(\hat\theta\). The resulting rule estimates \(q\).
The order matters: first identify the target function \(h\), then replace its unknown input by the estimator. In general, a plug-in estimate of \(h(\theta)\) is not the same as applying \(h\) separately to the observations and averaging. It is also not automatically a reliable or accurate estimate merely because the substitution is easy to perform.
Why the Plug-In Expression Is an Estimator
The justification follows from the definition of a statistic. Suppose \(\hat\theta=g(X_1,\ldots,X_n)\) is an estimator, where \(g\) has no unknown parameters as inputs. Applying a specified function \(h\) gives a new sample function \(h(g(X_1,\ldots,X_n))\). It depends on the random sample and on the specified functions, not on the unknown value of \(\theta\). Therefore, when defined, it is a statistic; if the target is \(h(\theta)\), it is the plug-in estimator of that target.
After the observations are recorded, the same function is evaluated at their realized values:
This argument establishes the form of the estimator, not its statistical quality. For example, the substitution alone does not show how close the estimator tends to be to the target, or how much it varies between samples. Those are separate questions about estimator properties.
Worked Example: Plugging an Estimated Lifetime Scale into a Survival Probability
In a hypothetical lifetime model, assume independent observations \(X_1,\ldots,X_n\) from an exponential distribution with scale parameter \(\theta>0\), so the density is \(f(x)=\frac{1}{\theta}e^{-x/\theta}\) for \(x>0\), and the mean lifetime is \(\theta\). Use the specified estimator \(\hat\theta=\bar X\). The model survival probability to be estimated is \(q=P(X>4)=e^{-4/\theta}\). For observed lifetimes \(2,3,5\), identify the plug-in estimator and calculate its estimate.
Method. Substitute the estimator \(\bar X\) for the unknown scale \(\theta\) in the survival-probability formula.
Before data are observed, the estimator of the target probability is
The sample sum is \(2+3+5=10\), and the sample size is \(3\). Therefore,
The observed estimate is obtained by substituting this value:
The estimate of the probability of surviving beyond time 4 is \(0.3012\), to four decimal places. The estimator \(e^{-4/\bar X}\) is the random rule; \(0.3012\) is the calculated estimate.
Check and common mistake. Since \(4/(10/3)=4(3/10)=1.2\), the exponent is \(-1.2\); \(e^{-1.2}=0.301194\ldots\), which rounds to \(0.3012\). A common error is to multiply \(4\) by \(\bar x\), producing \(e^{-4(10/3)}\). That is not the plug-in substitution: the original model has \(\theta\) in the denominator, so the correct expression is \(e^{-4/\bar x}\).
Worked Example: Estimating a Normal-Model Probability
Suppose a hypothetical measurement \(Y\) has distribution \(N(\mu,4)\), where the second argument is the variance, so the standard deviation is \(2\). The unknown mean is \(\mu\), and the variance is known. The target is the probability \(q=P(Y\leq 5)\) for a new measurement from this model. Use the specified estimator \(\hat\mu=\bar X\), based on four independent observations from the same model. The observed values are \(2,4,6,4\). Identify the plug-in estimator and calculate its estimate.
Method. Standardize the normal probability using the known standard deviation, then replace \(\mu\) by its estimator.
For a normal variable with mean \(\mu\) and standard deviation \(2\), the target is
where \(\Phi\) is the standard normal cumulative distribution function. Thus the plug-in estimator is
The observed sample sum is \(2+4+6+4=16\), so
Substitute this estimate of the mean into the target formula:
The estimated probability is \(0.6915\), to four decimal places. The estimator is the function \(\Phi((5-\bar X)/2)\), while the estimate is its observed value \(0.6915\).
Check and common mistake. The variance is \(4\), but the standard deviation used in standardization is \(\sqrt{4}=2\). If the variance were incorrectly used as the denominator, the calculation would give \(\Phi((5-4)/4)=\Phi(0.25)=0.5987\), which is not the stated normal probability. The model specifies variance \(4\), so standardization requires standard deviation \(2\).
Keep the Three Objects Separate
A clear solution identifies the parameter, writes the estimator as a function of the random sample, and then evaluates it using the observed sample. This sequence makes it easier to spot errors in notation, substitution, or interpretation. It is especially useful when a problem asks for a probability or another function of a model parameter, because the target and the parameter being estimated are not always the same quantity.
| Object | Meaning | Example from the exponential model |
|---|---|---|
| Parameter | Fixed, unknown quantity in the population model | \(\theta\), the exponential scale |
| Estimator | Sample-based rule, random before observations are known | \(\hat q=e^{-4/\bar X}\) |
| Estimate | Value of the rule after observed data are substituted | \(0.3012\) for the sample \(2,3,5\) |
For a plug-in calculation, check that the target has been written as a function of the parameter, that the estimator is substituted into the correct position, and that the resulting expression is defined. For probabilities, also verify that the final estimate lies between 0 and 1. These checks do not establish that an estimator has desirable properties, but they help ensure that the stated estimate actually follows from the stated rule.
This distinction is aligned with the SOA VEE Mathematical Statistics syllabus: point-estimation work depends on reading a rule correctly before evaluating it. A response that gives only a number may fail to distinguish the general sample rule from its one realized value. In a complete solution, show both when the question asks for them.
Check Your Understanding
For each problem, distinguish the estimator from its observed estimate. Assume any stated estimator is the rule to use.
- For eight independent Bernoulli observations with unknown success probability \(p\), use \(\hat p=\frac{1}{8}\sum_{i=1}^8X_i\). The observed outcomes are \(1,0,1,1,0,0,1,0\). State the estimator and calculate its estimate. (Answer: estimator \(\frac{1}{8}\sum_{i=1}^8X_i\); estimate \(4/8=0.5000\).)
- Independent lifetimes follow an exponential distribution with scale \(\theta\), and the specified estimator of \(\theta\) is \(\bar X\). For observations \(6,8,10,12\), state the estimator and calculate the estimate of \(\theta\). (Answer: estimator \(\bar X\); estimate \(36/4=9\).)
- Let \(Y\sim N(\mu,9)\), with known variance \(9\), and suppose \(\hat\mu=\bar X\). For a new observation, the target is \(P(Y>7)=1-\Phi((7-\mu)/3)\). If the sample is \(4,7,10,7\), give the plug-in estimator and its estimate. (Answer: estimator \(1-\Phi((7-\bar X)/3)\); \(\bar x=7\), estimate \(1-\Phi(0)=0.5000\).)
- Under an exponential model with scale \(\theta\), the target is \(q=e^{-3/\theta}\), and the specified estimator of \(\theta\) is \(\bar X\). For observations \(1,2,3,6\), write the plug-in estimator and calculate its estimate to four decimal places. (Answer: estimator \(e^{-3/\bar X}\); \(\bar x=12/4=3\), estimate \(e^{-1}=0.3679\).)
- A parameter of interest is \(q=\theta^2+1\), and the specified estimator of \(\theta\) is \(\bar X\). For observations \(1,2,4,5\), identify the plug-in estimator and calculate its estimate. Explain why averaging the transformed observations is not the specified plug-in rule. (Answer: estimator \(\bar X^2+1\); \(\bar x=12/4=3\), estimate \(3^2+1=10\); the rule transforms the sample mean, rather than averaging the individual squared values plus one.)