Why Sparse Interaction Cells Matter
A categorical interaction allows the effect of one rating factor to vary across levels of another. That flexibility can reveal genuine differences, but it divides the data among more combinations. If only a few claims occur in one combination, its estimated rate can change substantially when a claim is added or removed. A large fitted relativity from such a cell may therefore reflect limited information rather than a reliable difference in underlying risk.
The first diagnostic is to examine both exposure and claim count in each combination. Exposure describes how much insured risk contributes to a frequency estimate; the number of observed claims indicates how much event information the cell contains. A cell with substantial exposure but few claims can still have an imprecise frequency estimate. A cell with little exposure may also be important to flag, even if its observed rate happens to resemble neighboring rates.
The information available depends on the response being modeled. In a claim-frequency model, the number of claims is central to the precision of a cell's frequency estimate. In a claim-severity model, the number of observed claims is more directly relevant than policy exposure alone. A cell table should therefore show counts and exposure in terms appropriate to the model, rather than treating exposure as a substitute for claim information.
How Sparse Counts Produce Unstable Relativities
Consider independent Poisson claim counts \(N_j\) in cells with known exposures \(E_j\), where the expected count in cell \(j\) is \(E_j\lambda_j\) and \(\lambda_j\) is its claim frequency. The cell frequency estimate is \(N_j/E_j\). Under this model, the count's variance equals its mean, so a small count has high relative sampling variability. In particular, the approximate relative standard error of \(N_j/E_j\) is \(1/\sqrt{N_j}\), when \(N_j>0\).
An interaction can be summarized by comparing a ratio of cell frequencies. For two binary factors, label their four combinations \(00, 01, 10,\) and \(11\). The multiplicative interaction contrast is the ratio of the frequency ratio across one factor at level 1 of the other factor to that same ratio at level 0:
With no interaction on the log-frequency scale, \(I=1\). An estimate far from 1 can indicate interaction, but if one or more counts are small, the estimate may be highly uncertain. When the four counts are positive, a common large-sample approximation under independent Poisson counts is:
This approximation makes the sparse-cell problem visible: a small count contributes a large reciprocal term. It also provides a useful screening calculation, not a guarantee that a normal approximation is reliable. With very small counts, an interval based on this approximation can be a rough indicator of uncertainty rather than dependable formal inference.
Deriving the Approximate Uncertainty
The formula follows from the Poisson model and a first-order approximation. For a Poisson count \(N_j\) with mean \(\mu_j=E_j\lambda_j\), the variance is also \(\mu_j\). When \(\mu_j\) is sufficiently large, the delta method applied to \(\log N_j\) gives:
The exposure is treated as known, so \(\log(N_j/E_j)=\log N_j-\log E_j\) has the same approximate variance. Also, \(\log \widehat I\) is a sum of four log-frequency estimates, with signs \(+,+,-,-\). If the cell counts are independent, the variance of that sum is the sum of their variances: signs disappear when terms are squared. Replacing each unknown mean by its observed count gives the stated reciprocal-count approximation. It is least convincing precisely when counts are so small that the large-count approximation is weak.
Worked Example: Screening Cells Before Interpreting an Interaction
A hypothetical insurer fits a frequency model with vehicle group and territory as categorical factors. The table shows earned exposure in policy-years and observed claims.
| Vehicle group | Territory | Exposure | Claims | Observed frequency |
|---|---|---|---|---|
| Standard | Central | 400 | 40 | 0.100 |
| Standard | Coastal | 300 | 30 | 0.100 |
| Utility | Central | 180 | 18 | 0.100 |
| Utility | Coastal | 10 | 1 | 0.100 |
Method. Calculate each observed frequency as claims divided by exposure, then assess the cell information separately from the point estimate.
For Standard/Central, the frequency is \(40/400=0.100\) claims per policy-year. For Standard/Coastal, it is \(30/300=0.100\). For Utility/Central, it is \(18/180=0.100\). For Utility/Coastal, it is \(1/10=0.100\). All four observed frequencies agree, but the last cell has only one claim and 10 policy-years.
Under the Poisson approximation, the relative standard error of the last cell frequency is \(1/\sqrt{1}=1\), or 100%. For the Standard/Central cell it is \(1/\sqrt{40}\approx0.158\), or about 15.8%. Equal point estimates do not mean equally precise estimates: the Utility/Coastal observation is much less informative.
Check. Each count divided by its exposure returns 0.100. The screening conclusion is not that the sparse cell's true frequency differs; it is that its observed rate supplies weak evidence about that frequency. A common mistake is to see the matching rate and treat the cells as equally credible.
Quantifying a Sparse Interaction Contrast
A saturated two-factor model makes the four cell frequencies explicit. The following calculation illustrates how a contrast can look substantial while remaining uncertain. Assume independent Poisson counts with known exposures; the estimate of each cell frequency is its claim count divided by exposure.
Worked Example: A Large Contrast With a Wide Uncertainty Range
For a hypothetical product, four factor combinations each have 100 policy-years of exposure. The counts are \(N_{11}=4\), \(N_{00}=20\), \(N_{10}=25\), and \(N_{01}=1\). Estimate the interaction contrast \(I=\lambda_{11}\lambda_{00}/(\lambda_{10}\lambda_{01})\), then calculate an approximate 95% Wald interval.
Method. Estimate the four frequencies from the counts and exposures, form the contrast, then use the reciprocal-count standard error on the log scale. The common mistake is to pair counts with the wrong cell labels or to multiply the frequencies incorrectly.
The estimated frequencies are:
Substitute using the defined contrast:
Thus the estimated interaction contrast is 3.2, not 32. Its logarithm is \(\log(3.2)\approx1.1632\). The approximate standard error is:
The log-scale 95% interval is \(1.1632\pm1.96(1.1576)\). The margin is approximately \(2.2689\), so the endpoints are \(-1.1057\) and \(3.4320\). Exponentiating gives an interval of approximately \([0.331,30.94]\) for the interaction contrast.
The estimate is above 1, but the interval is very wide and includes 1. The count of one in cell \(01\) is the main source of the large approximate standard error. Because some counts are small, this Wald interval should not be treated as precise evidence; it is a transparent warning about instability. The check is that exponentiating the log-scale endpoints preserves their order and gives a positive interval containing 1.
Choosing a Defensible Simplification
Two common responses to sparse interaction cells are combining factor levels and removing an interaction. They impose different assumptions and should not be treated as interchangeable.
| Choice | What changes | Key assumption and trade-off |
|---|---|---|
| Combine levels of a factor | Two or more categories become one group, reducing the number of distinct combinations. | The combined categories are sufficiently similar for the intended rating purpose. This changes the factor's classification and may remove meaningful differences. |
| Remove an interaction | The factor levels remain distinct, but their effects no longer vary jointly in the model. | The interaction is unnecessary or too weakly supported. Main effects can still vary across each factor's levels. |
Combining levels is most defensible when the categories are substantively similar and their separation is not needed for rating or regulatory reasons. Removing an interaction is more appropriate when the factors remain meaningful separately but there is not enough evidence to estimate a distinct effect for every combination. Neither choice should be made merely to force a particular estimate toward a preferred value.
Worked Example: Comparing Level Combination With Interaction Removal
A hypothetical frequency model uses vehicle group (car or truck) and territory (metro or rural). The cell data are: car/metro, 36 claims on 400 policy-years; car/rural, 18 on 200; truck/metro, 18 on 200; and truck/rural, 1 on 10. Compare the implications of combining territories with removing the vehicle-by-territory interaction.
Method. First calculate the cell rates and simple pooled summaries. Then distinguish what each model change assumes; the pooled summaries are diagnostic, not replacements for fitted GLM estimates.
The observed rates are \(36/400=0.090\) for car/metro, \(18/200=0.090\) for car/rural, \(18/200=0.090\) for truck/metro, and \(1/10=0.100\) for truck/rural. The last rate is based on just one claim. Pooling the truck observations across territory gives \(19/210\approx0.09048\); pooling the car observations gives \(54/600=0.090\). These crude summaries are close, but do not prove that all territories or vehicle groups have identical underlying rates.
If the two territories are combined, territory distinctions are removed for every vehicle group. This is reasonable only if the territory categories are suitably similar for the model's purpose. If instead the interaction is removed, metro and rural remain separate levels, as do car and truck; the model simply requires the territory effect to be the same for both vehicle groups on the log-frequency scale. That is a narrower restriction than eliminating territory altogether.
The sparse truck/rural cell alone is not enough to select either change. An actuary should consider whether the categories are conceptually comparable, whether the interaction has credible support across the table, and whether a simpler specification remains suitable for the intended rating use. A common mistake is to describe interaction removal as if it merged factor levels; it does not.
Practical Checks and Common Traps
A disciplined review separates data support from model structure. Before interpreting a fitted interaction, tabulate the exposure and claim count for every combination, calculate observed rates where appropriate, and identify cells that have little information. Then assess the interaction's uncertainty and decide whether the modeling assumption is supportable.
- Do not use exposure alone as a precision measure. A frequency cell with high exposure but very few claims can still have an imprecise estimate.
- Do not rank cells by observed rate alone. The same rate based on one claim and on dozens of claims does not carry the same information.
- Do not ignore zero counts. In a saturated Poisson cell model, a zero count gives an estimated frequency of zero, whose logarithm is undefined. A finite interaction estimate may depend on other model structure or pooling assumptions; it is not directly supplied by the zero cell.
- Do not treat a large estimated relativity as proof of a real interaction. Sparse counts can produce large estimates and wide uncertainty at the same time.
- Do not combine categories without a rationale. Fewer cells can improve numerical stability, but inappropriate grouping can conceal meaningful risk differences or conflict with rating constraints.
- Do not overinterpret a simple Poisson standard error. The reciprocal-count calculation assumes independent Poisson counts and is an approximation. Extra variation or dependence can make it too optimistic.
Check Your Understanding
For each problem, state the calculation or model implication that supports your answer.
- A hypothetical frequency table has two cells: 6 claims on 120 policy-years and 12 claims on 240 policy-years. Calculate both observed frequencies and say whether their equal rates imply equal precision. (Answer: Both are 0.050 claims per policy-year; the 12-claim cell has more claim information, so the estimates do not have equal precision.)
- Independent Poisson counts in four cells are \(9,16,4,\) and \(1\). Using the reciprocal-count approximation, calculate the standard error of the log interaction contrast. (Answer: \(\sqrt{1/9+1/16+1/4+1}\approx1.193\).)
- In a 2-by-2 table, the frequencies for cells \(11,00,10,01\) are respectively \(0.05,0.05,0.10,0.05\). Calculate the multiplicative interaction contrast. (Answer: \((0.05)(0.05)/((0.10)(0.05))=0.5\).)
- A hypothetical category has 2 claims on 20 policy-years and a substantively similar category has 8 claims on 80 policy-years. If they are combined, calculate the pooled frequency and state the assumption this entails. (Answer: \((2+8)/(20+80)=0.100\) claims per policy-year; combining assumes the categories can reasonably share one level for this use.)
- A model has distinct vehicle and territory levels, but only one claim in the truck/rural cell. Explain the difference between removing the vehicle-by-territory interaction and combining rural with metro. (Answer: Removing the interaction retains both factors and their main effects but constrains their effects not to vary jointly; combining territories removes that territory distinction.)