Checking Independence With Residual Order Plots

Why Collection Order Matters

An ANOVA model can describe differences among treatment or group means, but its conclusions also rely on how observations vary around those means. In particular, the model ordinarily assumes that the errors for distinct observations are independent. If measurements collected close together tend to be similar, or if a process changes during collection, that assumption may be questionable even when a plot of residuals against fitted values looks unremarkable.

A residual order plot displays residuals in the order the observations were collected, measured, or otherwise placed in a meaningful sequence. It can reveal drift, cycles, long runs of positive or negative residuals, clustering, or an abrupt change. Such patterns are clues to investigate, not proof of a particular cause or a formal test of independence.

For the usual one-way ANOVA model, errors have mean zero and a common variance \(\sigma^2\), in addition to being independent. Normality is a separate condition used for the exact small-sample ANOVA \(F\) inference; it is not what a residual order plot assesses. The order plot focuses on whether residual behavior appears related to sequence. It does not check every model assumption.

Constructing and Reading an Order Plot

Keep the order that has meaning for how the data were generated: for example, run number, measurement time, or sequence of collection. Plot each residual against that order and include a horizontal reference line at zero. If treatments or groups are interleaved, retain their actual order and mark group membership with symbols, colors, or labels. Sorting observations by residual value or grouping all observations together by treatment destroys the sequence information the plot is meant to show.

Look for patterns rather than isolated departures. A gradual rise or fall can indicate drift. A repeating wave or alternating pattern can suggest a cycle. A long sequence of residuals on the same side of zero is a run; repeated clusters or a sudden level change can suggest dependence or a process change. A roughly patternless cloud around zero is consistent with independence, but does not establish it: some kinds of dependence are difficult to see, especially with few observations.

The collection plan matters when interpreting the plot. If all observations from one treatment were collected first and all observations from another treatment later, treatment and time are confounded in the sequence. A visible change could reflect a treatment difference, a time change, or both. Interleaving or randomizing treatment order, where appropriate to the study, helps separate these possibilities. Also consider whether observations from the same person, animal, batch, or machine are naturally connected; a plot alone cannot make dependent observations independent.

Residuals Can Be Correlated Even When Errors Are Independent

Residuals are calculated after estimating the group means, so they are not simply the original errors. This estimation imposes a constraint: residuals within each group sum to zero. As a result, residuals within a group are correlated even under the model in which the errors themselves are independent. This fact is important when reading apparent patterns in an order plot.

Here is the derivation. The group mean error is \(\bar{\epsilon}_{i\cdot}=(1/n_i)\sum_{r=1}^{n_i}\epsilon_{ir}\), so \(e_{ij}=\epsilon_{ij}-\bar{\epsilon}_{i\cdot}\). For distinct \(j\) and \(\ell\), independence gives \(\operatorname{Cov}(\epsilon_{ij},\epsilon_{i\ell})=0\). Each of \(\operatorname{Cov}(\epsilon_{ij},\bar{\epsilon}_{i\cdot})\) and \(\operatorname{Cov}(\epsilon_{i\ell},\bar{\epsilon}_{i\cdot})\) equals \(\sigma^2/n_i\), while \(\operatorname{Var}(\bar{\epsilon}_{i\cdot})=\sigma^2/n_i\). Therefore:

For the correlation, each residual has variance \(\operatorname{Var}(e_{ij})=\sigma^2(1-1/n_i)\). Dividing the covariance by this variance gives \(\operatorname{Corr}(e_{ij},e_{i\ell})=[-\sigma^2/n_i]/[\sigma^2(1-1/n_i)]=-1/(n_i-1)\). For \(n_i=2\), the zero-sum constraint gives \(e_{i2}=-e_{i1}\), matching correlation \(-1\). If \(n_i=1\), the residual is always zero and its correlation is undefined. Residuals from different groups have covariance zero under this model because they are functions of disjoint sets of independent errors.

This induced correlation is not evidence that the errors are dependent; it comes from estimating a mean. It also does not make an order plot useless. Rather, it is a reason to interpret patterns with the group structure in view, particularly when groups are small or when observations from one group occur in a block.

Worked Examples

How to Use the Diagnostic Responsibly

A common mistake is to say, “The residuals are scattered around zero, so the observations are independent.” The correction is that a patternless-looking plot offers only limited evidence against visible sequence structure; it cannot prove independence, and dependence may not be apparent in a small sample. The opposite mistake is to treat every run of signs as proof of dependence. Residuals are constrained by group-mean estimation, and chance patterns occur. Describe what the plot shows, account for the design and residual constraints, and investigate plausible causes before drawing a conclusion.

Order plots matter because independence is about the way observations relate to one another, not simply about the shape of a single group’s residual distribution. A plot can expose a missing time or process feature that is invisible when observations are summarized only by treatment. It should be read alongside the study design and the other residual diagnostics, not as a stand-alone pass-or-fail rule.