Parameters, Estimands, and Estimates

Why These Distinctions Matter

An empirical question may ask for the average earnings of a population, the share of households with a particular characteristic, or the difference in average outcomes between two groups. These questions concern quantities we want to know about a defined population. In practice, we often calculate a number from observed data instead. A clear analysis distinguishes the target quantity from the sample result.

The terms parameter, estimand, and estimate name related but different things. The distinctions are useful before any statistical procedure is chosen: without a well-defined target, there is no clear standard for judging what a calculation means.

From an Empirical Question to an Estimand

An empirical question needs more precision than a general topic. “What are earnings like?” does not specify which people, which measure of earnings, or which summary is wanted. “What is the mean annual earnings, in dollars, of full-time workers in Region R during year T?” specifies a target population, an outcome, a time period, and a summary. Together, those choices define the estimand.

For example, let \(X\) denote annual earnings in a specified population. The population mean is the parameter \(\mu_X\). If \(x_1,\ldots,x_n\) are the observed earnings for a sample, the sample mean \(\bar{x}\) is an estimate of that parameter:

The first expression refers to the population distribution; the second is calculated from the observed values. The parameter is fixed once the population and outcome are specified, even if its value is unknown. The estimate is a known number after the sample has been observed, and it can differ across samples.

Before the data are observed, the rule that will produce an estimate is called an estimator. For a sample mean, that rule is to add the observed values and divide by \(n\). Once the data are in hand, the resulting value \(\bar{x}\) is the estimate. This distinction is a matter of whether we mean a data-dependent rule or its realized value.

Proportions as Parameters and Estimates

A population share can be represented as the mean of a binary variable. Let \(Z\) equal 1 when a specified characteristic is present and 0 otherwise. The population proportion with that characteristic is \(p=P(Z=1)\). In a sample, if \(z_i\) records the same 0-or-1 outcome for observation \(i\), the sample proportion is \(\hat{p}=\sum_{i=1}^{n}z_i/n\). It is also the number of sampled observations with the characteristic divided by the sample size.

Group Contrasts Need a Defined Target

An empirical question may compare two groups. A descriptive contrast in mean outcomes is defined by specifying the population mean in each group and subtracting one from the other. For an outcome \(X\) and group label \(G\), one possible estimand is \(\delta=\mu_A-\mu_B\), where \(\mu_A\) is the mean of \(X\) among members of group A and \(\mu_B\) is the mean among members of group B. The order matters: this definition makes a positive contrast mean that group A has the higher mean.

The corresponding sample estimate subtracts the two sample means, \(\bar{x}_A-\bar{x}_B\). This calculation estimates a difference in observed group means. By itself, it does not define or establish a causal effect. A causal question requires a different target and additional assumptions; a numerical difference between groups should not be given a causal interpretation just because it is called a contrast.

How the Estimate Relates to the Target

For these examples, the estimate uses a sample analogue of the target quantity: an observed sample mean for a population mean, an observed sample fraction for a population proportion, and a difference between sample means for a difference between population means. This substitution gives a clear calculation rule, but it does not by itself prove that the estimate equals the target or is close to it.

The relationship can be seen directly for the group contrast. Its target is the difference between two population means. Replacing each population mean with its corresponding sample mean gives the estimate:

The arrow indicates the calculation used to form the estimate, not a guarantee of accuracy. The target is fixed by the question and population definitions. The estimate is determined by the particular observed data. A different sample can produce different sample means and therefore a different estimate.

This distinction also clarifies what can and cannot be concluded from a calculation. Reporting \(\hat{\delta}=5.0\) hours per week describes the sample contrast. To make a claim about the corresponding population contrast, an analyst needs a reason to connect the sample data to that population. Defining the estimand is necessary for that reasoning, but it is not, on its own, a justification that the estimate recovers the target.

A Practical Checklist

When translating an empirical question into a quantity, write down the pieces explicitly. This prevents the number produced by a spreadsheet or statistical program from silently replacing the question the analysis was meant to answer.

Common Mistake: Treating the Sample Number as the Population Quantity

A common mistake is to calculate a sample mean and write, “The population mean is 50 thousand dollars.” That statement confuses an estimate with its target. In the earnings example, the calculation establishes that the five sampled workers have a mean of 50 thousand dollars. It does not establish that all full-time workers in the region have that mean.

The correction is to state both quantities and their roles: “The target is the mean annual earnings of full-time workers in the specified region and year. The sample estimate is 50 thousand dollars per worker per year.” This wording preserves the distinction without suggesting that the estimate is necessarily inaccurate; it simply avoids claiming more than the calculation shows.

A related mistake is to call any difference between group estimates a causal effect. In the hours example, the estimate compares urban and rural sample means. It is not automatically an estimate of what would happen to the same people if their residential setting changed. First identify exactly which estimand the question defines; do not infer a causal interpretation from the subtraction alone.