Who Got the Payment: The Attribution Problem

A letter, a text and a call precede a payment, and the call gets the credit because it came last. On 1,000 accounts, 708 pay, 620 of them would have paid anyway, and last-touch attribution hands those 620 to whichever channel happened to be used last. The call appears to cost $6.36 per cure. It costs $150 per cure it caused. The budget then follows the wrong number.

Three touches, one payment

An account goes thirty days past due. A letter goes out on day one. A text follows on day five. A call connects on day six. On day eight, the consumer pays.

Which of the three gets the credit? In most operations, the call does, because it was the last thing that happened before the money arrived. That rule is called last-touch attribution, and it is the default in collections, in marketing and in most of underwriting, because it needs no model and no experiment. It only needs a timestamp.

Run the arithmetic on 1,000 such accounts with figures that are stylized but not unusual. Per our early-stage analysis, about 62% of them would have paid with no contact at all. The letter reaches everyone and adds 4 points of cures. The text reaches the 85% who consented and adds 4 points among them. The call connects with 45% and adds 3 points among those.

In total, 708 accounts cure, of which 620 would have cured anyway. The three channels together caused 88 payments. The other 88% of the "successes" belonged to nobody.

What last-touch reports instead

Now assign each of the 708 cures to the last channel that touched the account. Calls touched 45% of accounts last. Texts were the last touch on the 47% that got a text but no connected call. Letters were last on the 8% that got nothing else.

The error is not random. It has a shape. Last-touch hands every self-cure to whichever channel happened to be used last, and the channel used last is usually the most intrusive and the most expensive, because operations escalate. So the method systematically flatters calls, systematically underweights letters, and reports a cost per cure for every channel that is mostly the cost of claiming credit for payments that were coming regardless.

Then the budget follows the report. Per our dialer analysis, an operation that believes calls cost $6.36 per cure will buy more dialer capacity, and the marginal $1,000 spent on calls produces about 7 additional payments while the same $1,000 in texts, up to the limit of consent, produces over a thousand. Per our cost to collect analysis, that is how a cost-per-account metric can improve while cost per dollar recovered gets worse.

Why this is a different problem from the counterfactual

Per our counterfactual analysis, the deepest measurement problem in lending is that you cannot see what would have happened to the accounts you did not treat. Attribution is the next problem along: even when you know the treatments worked in aggregate, you cannot see which of several treatments did the work on any one account.

The two problems have different fixes. The counterfactual needs a holdout, a group that gets nothing. Attribution needs the holdout and a design that varies channels independently, so each channel's contribution can be separated. An operation that runs a no-contact holdout but sends every treated account the same sequence has solved the first problem and left the second untouched: it knows the sequence works and still cannot say which part of it to cut.

Why the wrong method persists

• It is free. Last-touch needs no experiment and no statistics, only a log.
• It is flattering. Every channel owner can show a cost per cure that looks good, because the self-cures are shared out among them.
• The alternative requires withholding treatment. Per our small numbers analysis, a holdout large enough to measure a 3-point effect runs to thousands of accounts, and nobody wants to be the person who did not call them.
• The error compounds quietly. Per our denominator analysis, a rate built on the wrong base can look stable for years while being wrong the whole time.
• Incentives align with it. Per our incentive analysis, a collector paid on payments received after their call has every reason to defend the rule that credits them.

The third point is the honest obstacle. A holdout costs real recoveries in the short run. But per our prioritization analysis, an operation that cannot tell which channel works is already spending badly every month, and the cost of a holdout is small beside the cost of permanently routing budget to the wrong place.

What can actually be measured

1. A no-contact holdout, randomly assigned, large enough for the effect size you care about. This gives the baseline that every other number depends on.
2. Channel-level holdouts. Randomly withhold one channel at a time from a slice of accounts. The difference against the full sequence is that channel's incremental effect.
3. Order experiments. Vary the sequence. If text-first performs as well as letter-first at a fraction of the cost, the letter was doing less than its position implied.
4. Report incremental cures and cost per incremental cure, and retire cost per credited cure from every dashboard.
5. Treat self-cure as its own line. Per our early-stage analysis, the 62% is the largest number in the whole operation, and a report that does not show it cannot be read correctly.
6. Check the measurement by segment. Per our language access analysis, a channel that works on average can be worthless for a group that cannot use it, and the average hides that.

Item four is the one that changes behaviour. A dashboard that shows calls at $150 per incremental cure, beside texts at $0.75 and letters at $16, produces a different budget from one that shows $6.36, $0.08 and $11.14. Both are arithmetic on the same accounts. Only one of them describes what the money did.

Testable implications

• Operations using last-touch should allocate a rising share of budget to the most intrusive channel over time, even where channel-level holdouts show its incremental effect is smallest.
• Introducing a no-contact holdout should cut every channel's reported effectiveness by roughly the self-cure share, here 88%, and the cut should be largest for the channel used last.
• Firms that run channel-level holdouts should shift spend toward low-cost channels relative to otherwise similar firms that do not, with no loss in recovery.
• Cutting an early, cheap channel should reduce the measured performance of later channels with a lag, as the later channels lose the accounts the early one was warming.

The fourth implication is the one that surprises people. Per our alert analysis and our monitoring analysis, an operation rarely monitors the interaction between its channels, so the decline shows up as a mystery in the channel that was left rather than as the consequence of the channel that was cut.

Frequently asked questions

This analysis is educational and does not constitute financial, legal, or operational advice. All figures are stylized illustrations using assumed self-cure rates, channel reach, incremental effects, and channel costs, and do not describe any particular operation. Real channel effects interact and vary by portfolio, and experimental designs must comply with debt collection and communication consent rules that vary by jurisdiction. Consult qualified advisers before changing contact strategies.