Common Errors In Empirical Loss And Survival Summaries

Why Small Data-Handling Errors Matter

Empirical summaries are built from counts, but a count is only meaningful when its rule is clear. A threshold written as “at most 2,000” includes observations equal to 2,000; “greater than 2,000” does not. Likewise, a recorded follow-up time is not necessarily an observed event time. Confusing these distinctions can change a reported proportion, event count, or risk set.

This tutorial gives a practical audit method for empirical loss and survival summaries. It focuses on tied values, strict versus inclusive thresholds, event indicators, risk-set membership, and incomplete observations. The aim is to identify what the records actually say before calculating a summary.

Audit Thresholds Before Counting

Ties make threshold language especially important. The expressions “below \(x\),” “at or below \(x\),” “above \(x\),” and “at least \(x\)” correspond to different sets. For exact observations, the boundary value belongs in the count for “at or below” and “at least,” but not in the count for “below” or “above.”

A useful check is to partition the sample into three groups: values less than \(x\), values equal to \(x\), and values greater than \(x\). If their counts are \(L\), \(E\), and \(G\), respectively, then \(L+E+G=n\). The relevant summaries follow directly: below \(x\) is \(L/n\), at or below \(x\) is \((L+E)/n\), above \(x\) is \(G/n\), and at least \(x\) is \((E+G)/n\).

Do Not Confuse an Observed Time with an Event

In a right-censored survival record, the observed time alone does not say whether the event occurred. Let \(T\) denote the event time and \(C\) the censoring time. The recorded time is \(Z=\min(T,C)\), and the event indicator is \(\delta=1\) if \(T\le C\) is observed, and \(\delta=0\) if observation ends before the event is observed. Thus \((Z,\delta)=(5,1)\) records an event at time 5, whereas \((Z,\delta)=(5,0)\) records follow-up through time 5 without an observed event.

The indicator answers a different question from the time: \(Z\) says when observation ended, and \(\delta\) says whether an event was observed then. A count of events must use \(\delta\), not simply count every row or every distinct recorded time. Constructing the record counts does not require an assumption that censoring is independent of event time. Interpreting incomplete records to estimate a population distribution does require appropriate assumptions; the raw counts alone do not solve that inferential problem.

What a Censored Record Does—and Does Not—Establish

For a right-censored record \((Z,\delta)=(c,0)\), observation ended at \(c\) without an observed event. It establishes that the event was not observed by the end of that follow-up; it does not establish that the event occurred at \(c\). In particular, replacing \(T\) by \(c\) treats a lower bound on the event time as if it were the exact event time.

This distinction also matters for empirical loss summaries. A loss reported as right-censored at \(c\) tells us that the loss exceeded the observed threshold under the stated censoring convention. It does not give the exact loss amount. Adding \(c\) as though it were the loss can distort a sample mean or empirical distribution. Excluding all such records and averaging only exact losses can also be misleading: that is a complete-case summary, not automatically a summary of the full group.

A Short Derivation for Risk-Set Updates

A risk-set audit can be justified directly from its definition. Suppose \(Y(t)\) counts all records whose observed time has not ended before \(t\), so \(Z_i\ge t\). At time \(t\), let \(d_t\) be the number of event records and \(c_t\) the number of censored records with \(Z_i=t\). Every record in these two groups belongs to the risk set immediately before \(t\), and each leaves it immediately afterward. No record with \(Z_i>t\) leaves at \(t\). Therefore the update follows by subtracting exactly the records that end observation at \(t\):

where \(Y(t+)\) denotes the risk-set size immediately after time \(t\). This is a bookkeeping identity, not a population estimate. It remains valid with tied events and censorings because the tied records are all counted before the update and all removed in the same update. For the preceding example, \(Y(4+)=3-1-1=1\).

Common Errors to Catch Before Reporting

• Changing the inequality while counting. “At most” includes equality; “below” excludes it. Write the set being counted before calculating.
• Counting all observed times as events. A record with \(\delta=0\) is censored, even though it has a numerical time.
• Dropping tied records from the pre-time risk set. Under the stated convention, every record with \(Z_i=t\) is included immediately before \(t\), whether its indicator is 0 or 1.
• Keeping earlier records in the risk set. If \(Z_i<t\), observation ended before \(t\), so that record is not at risk immediately before \(t\).
• Treating a censoring time as an exact loss or event time. The record gives incomplete information, not an exact value. Any summary must respect that distinction.
• Reporting a complete-case average as if it described everyone. State which records were included and avoid claiming that an exact-only subset represents the full group without justification.