Use a Shallow Tree for Claim Triage

Why Use a Shallow Tree for Claim Triage?

A claim triage model can help staff decide which claims should receive an additional review. A classification tree is useful when the operational team needs rules that can be followed step by step: for example, “if the report is late, check for a documentation inconsistency; otherwise, route the claim to review.” A shallow tree limits the number of such decisions and the number of terminal groups, making its logic easier to inspect.

The tree predicts a defined categorical outcome, such as whether a claim was later found to have a material documentation deficiency. It does not establish that a feature caused the outcome, and an observed outcome rate in a terminal group is not a guarantee about the next claim in that group. The target, observation unit, and predictors must be defined consistently, and every predictor must be available when the triage decision is made.

Definition: A classification tree repeatedly divides the observations in a node into child nodes using rules based on predictor values. A terminal node (or leaf) is a node that is not split further. Each terminal node can be assigned a predicted class, often the class observed most frequently among its training observations, and can be translated into an operational rule by tracing the splits from the root to that leaf.

Impurity and Split Selection

For a binary outcome, call the target class “positive” when a claim meets the stated review criterion and “negative” otherwise. If a node contains a proportion \(p\) of positive observations and \(1-p\) negative observations, its Gini impurity is

$$ G=1-p^2-(1-p)^2=2p(1-p). $$

Impurity is zero when all observations in a node have the same class. It is largest at \(p=0.5\), where the classes are evenly mixed. For a proposed split into child nodes \(j=1,\ldots,J\), calculate each child’s impurity and weight it by that child’s share of the parent observations. The split’s weighted impurity is

G_{\text{split}}=\sum_{j=1}^{J}\frac{n_j}{n}G_j,

where \(n\) is the parent node size and \(n_j\) and \(G_j\) are the size and Gini impurity of child \(j\). The reduction in impurity is \(G_{\text{parent}}-G_{\text{split}}\). A larger positive reduction means that the split separates the observed classes more effectively according to this criterion. It does not, by itself, show that the split will perform well on future claims.

Why the Gini formula works: Imagine drawing two observations independently from a node, with replacement. The probability that both are positive is \(p^2\), and the probability that both are negative is \((1-p)^2\). Thus the probability that their classes differ is \(1-p^2-(1-p)^2\), exactly the Gini impurity. A more mixed node has a greater chance of producing different labels in this pair.

Worked Example: Choose Between Two Candidate Splits

A hypothetical training node contains 20 claims: 8 are positive and 12 are negative. Candidate A divides them into two groups of 10, with 6 positives in the first group and 2 in the second. Candidate B divides them into a group of 5 with 4 positives and a group of 15 with 4 positives. Calculate the weighted Gini impurity for each split and select the better split by this criterion.

Method. Calculate the parent impurity, then calculate each child impurity and weight it by the child’s share of the 20 observations. This comparison avoids choosing a split merely because one child looks especially pure while ignoring the other child.

The parent positive proportion is \(8/20=0.4\), so its impurity is \(2(0.4)(0.6)=0.48\).

For Candidate A, the child positive proportions are \(6/10=0.6\) and \(2/10=0.2\). Their impurities are \(2(0.6)(0.4)=0.48\) and \(2(0.2)(0.8)=0.32\). The weighted impurity is

\frac{10}{20}(0.48)+\frac{10}{20}(0.32)=0.24+0.16=0.40.

For Candidate B, the positive proportions are \(4/5=0.8\) and \(4/15\). The first child’s impurity is \(2(0.8)(0.2)=0.32\). The second child’s impurity is \(2(4/15)(11/15)=88/225\approx0.391111\). Therefore,

\frac{5}{20}(0.32)+\frac{15}{20}\left(\frac{88}{225}\right) =0.08+0.293333\ldots =0.373333\ldots .

Candidate A reduces impurity by \(0.48-0.40=0.08\). Candidate B reduces it by \(0.48-0.373333\ldots=0.106666\ldots\). Candidate B has the lower weighted impurity and the greater reduction, so it is preferred by Gini impurity. The exact weighted impurity for B is \(28/75\), and the reduction is \(8/75\). These calculations check against the rounded values.

Common mistake avoided: comparing only the purest child would favor the first child of Candidate B, but a split must account for every child and its size. The weighted calculation does that.

Worked Example: Check the Whole Shallow Tree

A hypothetical insurer studies 60 historical claims. The positive outcome is a material documentation deficiency identified through a completed review. At the root, 9 claims are positive. A first split separates 15 claims reported within two days from 45 claims reported later. The early-report group contains 6 positive claims; the later-report group contains 3. The later-report group is then split by whether a documentation inconsistency was recorded: 15 claims have an inconsistency, including 2 positives, and 30 do not, including 1 positive. The early-report group is left terminal. Calculate the impurity of the root, the first split, and the final three-leaf tree.

Method. Use the positive proportion in each node to calculate its Gini impurity, then use observation counts to weight child impurities. Keeping the branch sizes visible prevents a common root-weighting error.

At the root, the positive proportion is \(9/60=0.15\). The root impurity is

G_{\text{root}}=2(0.15)(0.85)=0.255.

The early-report child has \(6/15=0.4\) positives, so its impurity is \(2(0.4)(0.6)=0.48\). The later-report child has \(3/45=1/15\) positives, giving

G_{\text{late}}=2\left(\frac{3}{45}\right)\left(\frac{42}{45}\right) =\frac{28}{225} \approx0.124444.

The first split’s weighted impurity is therefore

\frac{15}{60}(0.48)+\frac{45}{60}\left(\frac{28}{225}\right) =0.12+0.093333\ldots =\frac{16}{75} \approx0.213333.

Its reduction from the root is \(0.255-16/75=1/24\approx0.041667\).

Now calculate the two children formed by splitting the 45 later claims. The inconsistency child has \(2/15\) positives, so

G_{\text{inconsistency}} =2\left(\frac{2}{15}\right)\left(\frac{13}{15}\right) =\frac{52}{225} \approx0.231111.

The no-inconsistency child has \(1/30\) positive, so

G_{\text{no inconsistency}} =2\left(\frac{1}{30}\right)\left(\frac{29}{30}\right) =\frac{29}{450} \approx0.064444.

The final tree has three leaves: 15 early claims with impurity \(0.48\), 15 later claims with an inconsistency with impurity \(52/225\), and 30 later claims without an inconsistency with impurity \(29/450\). Weighting each leaf by its share of all 60 root observations gives

\begin{aligned} G_{\text{tree}} &=\frac{15}{60}(0.48) +\frac{15}{60}\left(\frac{52}{225}\right) +\frac{30}{60}\left(\frac{29}{450}\right)\\ &=0.12+0.057778\ldots+0.032222\ldots\\ &=0.21. \end{aligned}

The late-report parent contributed \((45/60)(28/225)=0.093333\ldots\) before its split; its two children contribute \(0.057778\ldots+0.032222\ldots=0.09\). Thus the additional reduction from that split is \(0.003333\ldots=1/300\). The final tree’s impurity, \(0.21\), is lower than the first-split impurity by exactly that amount. This confirms the whole-tree arithmetic.

Common mistake avoided: the second leaf has root weight \(15/60\), not \(10/60\) or another adjusted weight. Each terminal node’s impurity is weighted by its count divided by the original root count. The three weights must sum to one.

Translate Leaves into Operational Rules

The tree above gives three explicit routes. The positive proportions in the leaves are \(6/15=0.40\), \(2/15\approx0.1333\), and \(1/30\approx0.0333\). Those are observed training rates, not certain predictions for individual claims. A team may define a review rule using a specified operational criterion, but it should state that criterion separately from the tree structure.

Check report timing.
If the claim was reported within two days, route it to the early-report terminal group.

Check later reports for inconsistency.
For claims reported later, route claims with a recorded documentation inconsistency to that terminal group.

Apply the remaining route.
Later-reported claims without a recorded inconsistency belong to the final terminal group.

For example, if an operational policy says to send a claim to additional review whenever its terminal group’s training positive rate is at least 0.10, then the first two groups meet the stated rule and the third does not. This is a policy application to the displayed rates, not proof that the cutoff is optimal. The rule should also be checked against review capacity, the consequences of missed deficiencies, and any fairness requirements.

Worked Example: Assess Review Workload and Captured Positives

A separate hypothetical tree is fitted to 80 claims and has three terminal groups. Group A contains 20 claims, of which 8 are positive. Group B contains 30 claims, of which 3 are positive. Group C contains 30 claims, of which 1 is positive. An operational instruction says to review a group if its observed positive rate is at least 0.25. Find the review workload, the number of positive claims captured, and the share of all positive claims captured.

Method. Compute each leaf rate, apply the given group-level rule, and then count reviewed observations and positives. This keeps the decision rule distinct from the impurity criterion used to grow a tree.

Group A’s rate is \(8/20=0.40\), so it is reviewed. Group B’s rate is \(3/30=0.10\), and Group C’s rate is \(1/30\approx0.0333\); neither meets 0.25. Thus only Group A is reviewed. The workload is \(20\) of \(80\), or \(20/80=0.25\), which is 25% of claims.

There are \(8+3+1=12\) positive claims in total. The reviewed group contains 8, so the captured share is \(8/12=2/3\approx0.6667\), or 66.67% when expressed as a percentage to two decimal places. The calculation checks: 4 positive claims remain outside review, and \(8+4=12\).

Common mistake avoided: the workload is not the number of terminal nodes selected; it is the number of claims in those nodes. Here one selected node accounts for 20 claims.

Is the Tree Suitable for the Task?

A useful triage tree must be more than mathematically coherent. Its splits should be understandable, use information available at the point of decision, and result in actions that staff can carry out consistently. Before adopting a tree, ask:

Is the target appropriate? Define what counts as a positive outcome and ensure it represents the review objective rather than an imperfect substitute.
Are the split variables available in time? A feature recorded only after a claim has been reviewed cannot support initial triage.
Are the leaves large enough to interpret? A very small leaf can have an unstable observed rate and may encode an accidental feature of the training data.
Does each rule make operational sense? Check whether the data definition, cut point, and handling of missing or unusual values can be implemented as written.
Is simplicity worth the loss of detail? A shallow tree is easier to explain, but may leave substantial variation within its leaves or fail to distinguish useful patterns.

A shallow tree is not automatically a fair or stable decision process. Review its rules for unjustified proxies, uneven consequences across groups, and data-quality problems. Likewise, a lower training impurity is not evidence on its own of better performance on new claims. Treat leaf rates as descriptive summaries of the training data, and use an appropriate evaluation design before making claims about future results. The tree should support a documented decision process, not replace professional judgment or review of limitations.

Key takeaway: Build and inspect a shallow classification tree by comparing weighted child impurity with parent impurity, then translate each terminal path into a precise rule. Check every node weight against the root count, and assess whether the resulting rules are usable and appropriate for the triage task.

Check Your Understanding

Work through each problem before checking its final answer. Each scenario is hypothetical.

A node has 10 claims, 4 positive and 6 negative. A split creates a child of 4 claims with 3 positives and a child of 6 claims with 1 positive. Calculate the parent impurity, weighted child impurity, and impurity reduction. (Answer: Parent \(0.48\); weighted child impurity \(19/60\approx0.3167\); reduction \(49/300\approx0.1633\).)
A node has 12 claims, 3 positive. A proposed split creates two groups, each of size 6, with 2 positives in the first and 1 in the second. Calculate the parent and weighted child impurities. (Answer: Parent \(2(3/12)(9/12)=0.375\); child impurities \(4/9\) and \(5/18\); weighted child impurity \(13/36\approx0.3611\).)
A terminal group contains 25 claims, 5 positive. A separate terminal group contains 15 claims, 3 positive. Under a rule to review groups whose observed positive rate is at least 0.18, which groups are reviewed, and how many claims are sent to review? (Answer: Both rates are \(0.20\), so both groups are reviewed; 40 claims are sent to review.)
A tree has leaves of sizes 10, 20, and 30, with Gini impurities \(0.32\), \(0.18\), and \(0.10\), respectively. Calculate its root-weighted impurity. (Answer: \((10/60)(0.32)+(20/60)(0.18)+(30/60)(0.10)=0.053333\ldots+0.06+0.05=0.163333\ldots\), or \(49/300\).)
A hypothetical tree sends a group of 12 claims to review; 5 of those are positive. Across all terminal groups there are 20 positive claims among 100 claims. Find the reviewed workload share and the share of positive claims captured. (Answer: Workload share \(12/100=0.12\), or 12%; captured positive share \(5/20=0.25\), or 25%.)

What You'll Learn