Before giving a formal definition of consistent estimator, let us briefly highlight the main elements of a parameter estimation problem:
a sample , which is a collection of data drawn from an unknown probability distribution (the subscript is the sample size, i.e., the number of observations in the sample);
a function, called an estimator, that associates an estimate to each sample that could possibly be observed.
Before being observed, the sample is regarded as a random variable. Therefore, , which depends on , is also a random variable.
When needed, we can writeto highlight the fact that the estimator is a function of the sample .
Now, imagine we are able to collect new data and increase our sample size indefinitely, so as to obtain a sequence of samples and a sequence of estimators . If this "imaginary" sequence of estimators converges in probability to the true parameter value, then it is said to be consistent.
Definition A sequence of estimators is said to be consistent if and only ifwhere denotes convergence in probability.
Note that we have defined "consistent sequences of estimators". But what do we mean by "consistent estimator"? The latter locution is often informally used to mean that 1) the same predefined rule is used to generate all the estimators in the sequence and that 2) the sequence is consistent. Thus, the concept of consistency extends from the sequence of estimators to the rule used to generate it. For instance, suppose that the rule is to "compute the sample mean", so that is a sequence of sample means over samples of increasing size. If converges in probability to the mean of the distribution that generated the samples, then we say that is consistent. By a slight abuse of language, we also say that the sample mean is a consistent estimator.
The following table contains examples of consistent estimators (with links to lectures where consistency is proved).
|Estimator||Estimated parameter||Lecture where proof can be found|
|Sample mean||Expected value||Estimation of the mean|
|Sample variance||Variance||Estimation of the variance|
|OLS estimator||Coefficients of a linear regression||Properties of the OLS estimator|
|Maximum likelihood estimator||Any parameter of a distribution||Maximum likelihood|
An estimator which is not consistent is said to be inconsistent.
You will often read that a given estimator is not only consistent but also asymptotically normal, that is, its distribution converges to a normal distribution as the sample size increases. You might think that convergence to a normal distribution is at odds with the fact that consistency implies convergence in probability to a constant (the true parameter value). In other words, you might ask yourself: "Is convergence to a constant or to a distribution?". To answer this question, we should give a more precise definition of asymptotic normality.
Consider the ratio When is consistent, both the difference and the standard deviation converge to zero as tends to infinity. However, their ratio can converge to a distribution. When it converges to a standard normal distribution, then the sequence is said to be asymptotically normal.
The practical consequence of asymptotic normality is that, when is large, we can approximate the above ratio with a standard normal distribution. It follows that can be approximated by a normal distribution with mean and standard deviation . But the latter converges to zero, so that the distribution becomes more and more concentrated around the mean, ultimately converging to a constant.
Consistency is discussed in more detail in the lecture on Point estimation.
Previous entry: Conditional probability mass function
Next entry: Convergence criterion
Most of the learning materials found on this website are now available in a traditional textbook format.