Why Sparse Cells Need Careful Interpretation
A rating cell may contain little exposure, few policies, or only a handful of claims. In such a cell, one additional claim can materially change the observed rate and produce a large residual, even when the fitted model is not broadly miscalibrated. Conversely, a striking residual in a thin class can be diluted when results are summarized over a much larger segment.
The key distinction is between an unusual observation under the fitted model and evidence that the model is systematically wrong. Residual diagnostics help identify where to investigate; they do not make that distinction automatically. This tutorial focuses on interpreting extreme residuals when the underlying cells are sparse and assessing the consequences for segment-level conclusions.
Residuals in a Thin Poisson Cell
Suppose \(Y_i\) is the claim count for observation or aggregated cell \(i\), with exposure \(e_i\). Under a Poisson frequency GLM, conditional on the fitted rating variables, assume independent counts with mean \(\mu_i=e_i\lambda_i\), where \(\lambda_i\) is the fitted claim frequency per unit of exposure. The exposure enters as an offset in the log-mean model. The Poisson variance is also \(\mu_i\).
The Pearson residual compares the observed count with its fitted mean in units of the model standard deviation. The deviance residual measures the signed contribution of the observation to Poisson deviance. For a Poisson count, the formulas are:
When \(y_i=0\), the deviance term \(y_i\log(y_i/\mu_i)\) is defined by its limit as zero, so \(r_i^D=-\sqrt{2\mu_i}\). These formulas apply to the Poisson model as stated; a different response distribution requires its own variance function and deviance. For a GLM with estimated parameters, a commonly used leverage-adjusted Pearson residual is \(r_i^P/\sqrt{1-h_{ii}}\), where \(h_{ii}\) is the GLM leverage. This adjustment is a diagnostic approximation, not an exact standard normal transformation.
Worked Example: A Large Residual With a Low Expected Count
A hypothetical small rating cell has a fitted Poisson claim count with mean \(\mu=0.25\). It records \(y=2\) claims. Its GLM leverage is \(h_{ii}=0.20\). Calculate its Pearson residual, deviance residual, leverage-adjusted Pearson residual, and the model probability of at least two claims. Explain what the results do and do not establish.
Method. Use the Poisson residual formulas and the fitted Poisson distribution. The leverage adjustment uses the stated diagnostic approximation.
Pearson residual. The observed-minus-fitted count is \(2-0.25=1.75\), and the model standard deviation is \(\sqrt{0.25}=0.5\). Therefore,
Deviance residual. The log ratio is \(\log(2/0.25)=\log(8)=2.0794\). Hence the quantity inside the square root is \(2[2(2.0794)-1.75]=4.8178\), giving \(r_i^D=\sqrt{4.8178}=2.1949\). The sign is positive because \(y>\mu\).
Leverage adjustment. Since \(\sqrt{1-h_{ii}}=\sqrt{0.80}=0.8944\), the adjusted residual is \(3.5/0.8944=3.9131\).
Exact Poisson tail probability. The probability of at least two claims under a Poisson mean of \(0.25\) is
The four calculations show that this cell warrants investigation. They do not, by themselves, prove a systematic rating relativitiy is wrong: a thin cell can yield an unstable rate, and the residual is one diagnostic observation. A common mistake is to treat the adjusted Pearson residual of \(3.9131\) as an exact normal \(z\)-score and interpret it without checking the count scale or data. The Poisson tail probability is a separate, distribution-based calculation, not a residual or a complete model-validation test.
Cell-Level Extremes Versus Segment Calibration
A residual describes a difference between an observed response and a fitted response on a particular scale. A segment-level observed-to-predicted ratio instead compares sums of observed and fitted amounts. For counts, define the ratio as total observed claims divided by total fitted claims. The ratio is exposure-weighted implicitly through the fitted means; it is not generally the unweighted average of cell-level frequency ratios.
A sparse cell can have a very high observed-to-predicted ratio because its denominator is small. That ratio may be materially reduced when combined with more exposure. The broader ratio is useful for summarizing calibration, but it can also conceal opposing deviations between cells. Therefore, examine both the detail and the aggregation relevant to the business decision.
Worked Example: One Claim in a Thin Class
A hypothetical rating segment has two classes, both fitted at a claim frequency of \(0.20\) claims per exposure unit. The thin class has \(0.5\) exposure units and one observed claim. The other class has \(99.5\) exposure units and 20 observed claims. Calculate each class's fitted count and observed-to-predicted ratio, the thin-class Pearson residual, and the combined segment ratio.
Method. Under the stated Poisson frequency model, fitted count equals exposure times fitted frequency.
Thin class. Its fitted count is \(0.5(0.20)=0.10\). Its observed-to-predicted ratio is \(1/0.10=10.00\). The Pearson residual is
Other class. Its fitted count is \(99.5(0.20)=19.90\), so its observed-to-predicted ratio is \(20/19.90=1.0050\).
Combined segment. Total fitted claims are \(0.10+19.90=20.00\), and total observed claims are \(1+20=21\). Thus the segment ratio is \(21/20=1.05\).
The thin class is at ten times its prediction, while the segment as a whole is only five percent above prediction. That contrast reflects the small denominator in the thin class, not a mathematical contradiction. A common mistake is to average the class ratios, \((10.00+1.0050)/2=5.5025\), and call that the segment ratio. The correct segment ratio is the ratio of sums, \(21/20=1.05\).
How Sparse Cells Can Distort or Conceal Conclusions
A single extreme cell can change the narrative about a segment, especially if the segment has few cells or a decision-maker focuses on its largest residual. The reverse can also happen: a satisfactory aggregate ratio can mask a thin class with a large deviation, or opposing cell deviations can cancel. Neither the cell residual nor the aggregate ratio should be read in isolation.
Worked Example: Opposing Results Within a Segment
A hypothetical segment contains a thin class with fitted mean \(0.25\) and observed count \(2\), and a broader class with fitted mean \(19.75\) and observed count \(18\). Assume the fitted count means come from the same Poisson frequency model. Calculate the thin-class Pearson residual, the segment observed-to-predicted ratio, and the ratio after removing the thin class. Explain the effect on interpretation.
Method. Calculate residuals from cell-level counts and means, then calculate each observed-to-predicted ratio from sums.
Thin class. The residual is \((2-0.25)/\sqrt{0.25}=1.75/0.5=3.5\).
Whole segment. Observed claims total \(2+18=20\), and fitted claims total \(0.25+19.75=20\). The segment ratio is \(20/20=1.0000\).
Without the thin class. The ratio for the broader class alone is \(18/19.75=0.9114\), to four decimal places. The aggregate result of exactly \(1.0000\) arises because the thin class's two claims offset the broader class's shortfall.
The segment ratio alone would miss both the high thin-class residual and the lower result in the broader class. Conversely, presenting only the thin-class residual would overstate how far the whole segment is from its prediction. The common mistake is to infer that a segment is uniformly calibrated because its combined ratio equals one. The correction is to show the aggregate and inspect meaningful subgroups, while recognizing that the thin subgroup estimate is less stable.
A Practical Diagnostic Sequence
A disciplined review separates data reliability, statistical instability, and persistent model structure. The following sequence helps keep a dramatic residual from becoming an unsupported conclusion.
Check claim counts, exposure units, dates, rating-variable coding, and whether the record was assigned to the intended cell. A unit or coding error can create an extreme residual without indicating a model relationship problem.
Record the cell's exposure and expected count, not just its residual. For frequency, one claim against an expected count of \(0.10\) has a different stability context from one claim against an expected count of \(10\).
A high-leverage observation can affect the fitted model itself. Use leverage and influence diagnostics as context, and compare the observation with the rest of its rating class. A residual alone does not measure its effect on all fitted coefficients.
Review observed and fitted totals at the thin-cell level and at defensible broader aggregations. Look for offsetting deviations that a total ratio can conceal, and avoid an unweighted average of cell ratios.
Compare related classes and the relevant rating variables. A one-off extreme in a low-exposure cell is weaker evidence of structural misfit than a consistent pattern across adjacent or otherwise comparable cells.
Interpretation and Exam Traps
A residual is a diagnostic signal, not a verdict. Sparse counts make the distinction especially important because low expected counts make large standardized residuals more plausible than a simple normal approximation might suggest. The exact count distribution can provide useful context, but it does not replace checking exposure, data quality, leverage, and broader patterns.
- Do not equate a large residual with a failed model. It may reflect a low expected count, a data issue, a high-leverage point, or genuine model misfit. Investigate which explanation is plausible.
- Do not equate a good aggregate ratio with good calibration everywhere. Opposing deviations can cancel when counts are summed.
- Do not compare unweighted class ratios as if they were a segment ratio. Use total observed divided by total predicted for the aggregate.
- Do not apply a Poisson formula to a different response distribution. The variance and deviance residual depend on the assumed GLM distribution.
- Do not treat the leverage-adjusted Pearson residual as an exact probability statement. It is a diagnostic adjustment; sparse count distributions are discrete and may be far from normal.
Check Your Understanding
For each problem, calculate the requested diagnostic before reading the answer.
- A Poisson cell has fitted mean \(0.16\) and one observed claim. Calculate the Pearson residual and the probability of at least one claim under the fitted model. (Answer: \(r^P=2.1000\); \(P(Y\ge1)=1-e^{-0.16}=0.1479\).)
- A Poisson cell has fitted mean \(0.40\) and zero observed claims. Calculate its Pearson and deviance residuals and the probability of zero claims. (Answer: \(r^P=-0.6325\), \(r^D=-0.8944\), and \(P(Y=0)=e^{-0.40}=0.6703\).)
- Two classes have fitted frequency \(0.50\). The first has exposure \(0.4\) and one claim; the second has exposure \(39.6\) and 19 claims. Calculate their fitted counts, the first-class observed-to-predicted ratio and Pearson residual, and the combined segment ratio. (Answer: fitted counts \(0.20\) and \(19.80\); first-class ratio \(5.00\); first-class Pearson residual \(1.7889\); combined ratio \(20/20=1.0000\).)
- A Poisson observation has \(y=3\), \(\mu=1\), and leverage \(h_{ii}=0.36\). Calculate its Pearson residual and the leverage-adjusted Pearson residual. (Answer: \(r^P=2.0000\); adjusted residual \(2/\sqrt{0.64}=2.5000\).)
- A fitted Poisson count has mean \(0.6000\). Calculate the probability of at least two claims, rounding the probability to four decimals, and state why a normal approximation should be treated cautiously. (Answer: \(1-e^{-0.6}(1+0.6)=0.1219\); the low expected count produces a discrete distribution for which a normal approximation may be poor.)