Repeated Measurements and Serial Dependence

Why Repeated Measurements Need Special Attention

A data set can contain many rows without containing that many independent sources of information. If each person is measured at several visits, one row may describe that person's outcome today and another row the same person's outcome next month. A person's health, habits, and circumstances tend to persist, so outcomes close together in time can be more alike than outcomes from different people.

This dependence matters for regression inference. A regression mean model can describe how an outcome relates to predictors while its usual standard-error calculation still treats all rows as independent. If repeated outcomes from one person are correlated, that calculation can understate uncertainty. The key is to distinguish the mean relationship being estimated from the dependence among observations used to quantify its uncertainty.

A basic repeated-measurement regression can be written as \(Y_{it} = \beta_0 + \beta_1 x_{it} + \epsilon_{it}\), where \(Y_{it}\) is the outcome for person \(i\) at time \(t\), \(x_{it}\) is a predictor, and \(\epsilon_{it}\) is the error around the specified mean. This equation does not itself require errors from the same person to be independent. The usual independent-observation standard-error calculation does require that kind of independence, however, and therefore may not suit repeated measurements.

For the discussion below, suppose the mean relationship is correctly specified, predictor information is adequate for estimating its coefficients, and different people are independent. Within a person, errors may be correlated across occasions. These are substantive assumptions: clustering standard errors by person does not make an incorrect mean model correct, nor does it remove dependence between different people.

Recognizing the Time Pattern

The index \(t\) carries more information than a generic row number. It can indicate visit order, calendar time, or time since enrollment. If one person is measured on consecutive days, dependence may be strongest between measurements a day apart and weaker between those several weeks apart. That pattern is called serial because it follows an ordered sequence.

A simple way to describe this idea is to imagine that errors have the same variance \(\sigma^2\) at each occasion and correlation \(\rho\) between adjacent occasions. Under a first-order autoregressive pattern, the correlation \(k\) occasions apart is \(\rho^k\), with \(|\rho|<1\). This is one possible pattern, not a default truth: real data may have a different time structure, and a person-clustered standard error does not require choosing this particular pattern.

How Serial Correlation Changes a Simple Variance

A short derivation shows why repeated observations need not contribute as much information as the same number of independent observations. Consider the average of \(T\) errors from one person. Assume each error has mean zero and variance \(\sigma^2\), and assume the first-order autoregressive correlation pattern \(\operatorname{Corr}(\epsilon_{it},\epsilon_{is})=\rho^{|t-s|}\). The variance of a sum is the sum of all variances and covariances. Each pair of occasions \(k\) steps apart contributes covariance \(\sigma^2\rho^k\), and there are \(T-k\) such pairs.

The factor of 2 appears because each covariance between two different occasions appears twice in the variance expansion: once for each ordering of the pair. If the errors are independent across occasions, then \(\rho=0\), all covariance terms vanish, and the variance is \(\sigma^2/T\). With positive serial correlation, the covariance terms are positive, so the variance of the average is larger than \(\sigma^2/T\).

Person-Clustered Uncertainty and Longitudinal Modeling

One approach is to keep the chosen regression mean model and calculate uncertainty with a person-clustered sandwich variance. In the usual large-cluster-count setting, this allows the errors to be correlated in an unspecified way among records from the same person, while relying on independence across people. Its key advantage is that it need not specify whether same-person correlation follows an autoregressive pattern, an equal-correlation pattern, or another pattern.

In matrix notation, let \(X_i\) contain the design rows for person \(i\), and let \(\hat e_i\) be that person's vector of fitted residuals. A basic form of the clustered sandwich variance estimator is:

Here \(G\) is the number of people. The middle term aggregates each person's contribution as a whole, rather than treating each row as a separate independent contribution. Inference based on this estimator relies on having a sufficiently large number of independent people; a large row count created by repeatedly measuring only a few people does not substitute for many independent clusters. Small-sample adjustments and few-cluster concerns require particular care.

A longitudinal model, by contrast, specifies more directly how measurements from the same person relate over time. For example, it might use an autoregressive error structure or another explicit covariance model. Such a model can use the assumed pattern to estimate the mean relationship and its uncertainty, and a well-chosen structure may improve efficiency. Its conclusions then depend on whether the specified longitudinal structure is reasonable.

These approaches answer related but different questions. Person-clustered standard errors ask, in effect, how uncertain the regression coefficients are when within-person dependence is allowed without specifying its detailed form. A dedicated longitudinal model makes the time-related dependence itself part of the model. Neither approach automatically repairs a misspecified mean relationship, addresses confounding, or establishes a causal effect.

Common Mistakes and Practical Meaning

• Counting rows as independent people. Four visits for each of 50 people are 200 records, but only 50 independent people under the stated independence-across-people assumption.
• Assuming clustered standard errors model time. Clustering by person permits within-person correlation for uncertainty estimation; by itself, it does not specify how correlation changes with the time gap.
• Thinking clustering changes the estimated coefficient. With the same mean model and data, changing only the variance estimator changes standard errors and related inference, not the fitted coefficient estimates.
• Assuming a larger record count solves few-cluster problems. Adding more time points to the same small set of people does not provide the same independent information as adding people.
• Interpreting a corrected standard error as a cure-all. Neither a clustered variance nor a longitudinal covariance assumption fixes a poor mean model or justifies a causal interpretation on its own.

The practical choice depends on the goal and the data. If the primary goal is inference for a regression mean and there are many independent people, person-clustered uncertainty can be a useful way to avoid asserting a detailed within-person correlation pattern. If the timing pattern is central to the scientific question, or if the analysis needs to use an explicit model of that pattern, a longitudinal approach is more directly suited to the task. The assumptions should be stated plainly either way.