Choosing Between Bounds And A Point Estimate

Why the Choice Matters

When assumptions and observed data identify a causal effect only within an interval, the analyst faces a reporting decision: retain that interval or add assumptions that yield a point estimate. The choice is not simply between a “less precise” and a “more precise” number. It is between conclusions supported by different amounts of information and, often, by different assumptions. A point estimate may be useful, but its apparent precision is not evidence that the underlying uncertainty has been resolved.

The practical question is: what can be said without claiming more than the analysis identifies? Sometimes every value in the interval supports the same substantive conclusion. In other cases the interval includes values with meaningfully different implications. Reporting that distinction is more informative than selecting a convenient number from the interval.

Definition: Let \(\theta\) be a causal estimand and let the maintained assumptions and available information imply the identified set \(\Theta_I=[L,U]\), where \(L\leq U\). The effect is partially identified if this set contains more than one value. It is point identified if the set contains exactly one value, \(L=U\). A point estimate is a single value computed from data or a model; it is not, by itself, proof that the population effect is point identified.

The interval \([L,U]\) is an identified set under stated assumptions, not automatically a confidence interval. In a sample, the endpoints may themselves be estimated and subject to sampling uncertainty. A confidence set for the identified set addresses that additional uncertainty; it does not turn a population identified interval into a point-identified effect. Keep these two sources of uncertainty separate when deciding what to report.

A Decision Rule for Choosing the Form of the Result

Retain the bounds as the primary result when the interval is a valid description of what the maintained assumptions identify and no additional assumptions needed for point identification are adequately justified. A point estimate can be added as a secondary, explicitly assumption-dependent analysis, provided the extra assumptions are stated and assessed. Choosing a point merely because a decision-maker asks for one does not supply those assumptions.

A useful decision is whether the substantive conclusion is invariant across the identified set. For example, if every possible effect in the set is beneficial, the direction of effect is robust to the unresolved identification, even if its magnitude is not. If the set crosses zero, the sign is not determined by the bounds alone. For a threshold \(c\), check the location of the entire interval rather than comparing \(c\) with a selected point such as the midpoint.

Threshold interpretation: For \(\Theta_I=[L,U]\), if \(U<c\), every admissible value is below \(c\); if \(L>c\), every admissible value is above \(c\). If \(L<c<U\), the identified set contains values on both sides, so the bounds alone do not determine which side contains \(\theta\). Endpoint cases are distinct: if \(L=c<U\), then \(\theta\geq c\); if \(L<c=U\), then \(\theta\leq c\); and if \(L=U=c\), then \(\theta=c\). These statements follow from \(\theta\in[L,U]\), not from a probability distribution over that interval.

The exact language also depends on the decision criterion. For a requirement that the effect be strictly greater than \(c\), the requirement is guaranteed when \(L>c\), ruled out when \(U\leq c\), and unresolved when \(L\leq c<U\). For a requirement that the effect be at least \(c\), it is guaranteed when \(L\geq c\), ruled out when \(U<c\), and unresolved when \(L<c\leq U\). State whether the threshold is strict; equality can change the conclusion.

Worked Example: A Clinical Threshold Inside the Bounds

Suppose a hypothetical analysis bounds the average reduction in symptom days over a specified follow-up period at 0.8 to 2.4 days per patient. A reduction of 1.5 days is the prespecified threshold for a clinically important average reduction. Let \(\theta\) denote the reduction in days, so larger positive values mean fewer symptom days.

The identified interval is \([L,U]=[0.8,2.4]\) days, and the threshold is \(c=1.5\) days. Since \(0.8<1.5<2.4\), the interval contains values below and above the threshold. Thus the bounds do not establish that the effect reaches the clinical threshold, nor do they establish that it falls short.

However, because \(L=0.8>0\), every value in the interval indicates a reduction in symptom days. The direction is robustly favorable, while clinical importance is not determined by the bounds. The midpoint is \((0.8+2.4)/2=1.6\) days, but that arithmetic does not make 1.6 the identified effect or establish that the clinical threshold is met. Reporting the interval preserves the distinction between the conclusion supported throughout the set and the conclusion that depends on the unknown value.

Justifying a Bounds-Based Conclusion

The logic behind robust conclusions is simple but important. The analysis asserts that the true estimand belongs to the identified set. If every value in the set satisfies a claim, then the true value satisfies that claim under the maintained assumptions. If the set contains values that satisfy the claim and values that do not, membership alone cannot settle it.

\theta\in[L,U],\quad [L,U]\subseteq A \quad\Longrightarrow\quad \theta\in A.

For instance, for the claim \(\theta>c\), the set is wholly inside the claim region when \(L>c\), because then every \(\theta\in[L,U]\) is at least \(L\) and therefore greater than \(c\). If \(L\leq c<U\), there are admissible values at or below \(c\) and values above \(c\), so the set does not settle the claim. This is a logical guarantee about the identified set, not a statement about the probability that a particular value is true.

Worked Example: A Risk Difference That Does Not Fix the Direction

In a hypothetical evaluation, the risk difference is defined as the outcome risk under the intervention minus the outcome risk under comparison. Suppose the identified set is \([-0.03,0.07]\), measured in probability units. Negative values mean lower risk under the intervention; positive values mean higher risk.

The interval crosses zero because \(-0.03<0<0.07\). Therefore, some values consistent with the bounds indicate a 3-percentage-point reduction, while others indicate as much as a 7-percentage-point increase. The bounds do not determine the direction of the effect. They do establish that the effect is no smaller than \(-0.03\) and no greater than \(0.07\), under the assumptions used to derive them.

The midpoint is \((-0.03+0.07)/2=0.02\), or a 2-percentage-point increase. That value is one point in the set, not a conclusion implied by the bounds. Unless an additional assumption, a defensible model, or a separately stated decision rule supports using it, describing the intervention as increasing risk by 2 percentage points overstates what this identified set establishes.

When a Point Estimate May Be Useful

A point estimate can be worth reporting when a decision requires a single input, when a credible scientific assumption narrows the identified set to one value, or when a model-based estimate answers a clearly specified secondary question. In each case, distinguish the point estimate’s role from the bounds’ role. If the point depends on a parametric restriction, extrapolation, or another added assumption, report that dependence rather than presenting the number as though it followed from the weaker assumptions used for the bounds.

The interval midpoint, \((L+U)/2\), is especially tempting. It is a descriptive summary of the endpoints, but it is not generally a preferred estimate of \(\theta\). The bounds do not say that values near the midpoint are more plausible than values near an endpoint, nor that the set is a probability distribution. A midpoint may be chosen as a decision convention under a specified loss function, but that is a decision rule—not additional identification.

Worked Example: When the Whole Interval Clears a Decision Threshold

Suppose a hypothetical program’s effect is measured as additional completed training hours per eligible participant, with positive values indicating an increase. The analysis yields \([L,U]=[3,8]\) hours. A policy criterion requires an increase of at least 2 hours.

Here \(L=3\geq2=c\), so every value in the identified set meets the criterion. The lower endpoint is already 1 hour above the threshold, and the upper endpoint is 6 hours above it. The supported conclusion is that the effect is at least 3 hours, under the assumptions defining the bounds; consequently it is at least 2 hours. The precise effect remains unknown within the interval, but that uncertainty does not affect whether the stated criterion is met.

A point estimate is unnecessary for this threshold decision: all admissible values lead to the same classification. A point estimate might still be useful for planning, but it would need a separate justification and should not replace the interval’s information about possible magnitudes.

Diagnostics, Comparisons, and Common Pitfalls

Before choosing how to present a result, check that the bounds actually support the intended comparison. Confirm that \(L\leq U\), that both endpoints use the same estimand, population, outcome scale, and units, and that the interval was derived under assumptions relevant to the question. Review the assumption audit and diagnostics for informative bounds: a narrow interval can still be misleading if it depends on an implausible restriction, while a wide interval can be a valid statement of limited information. If endpoints are estimated, also report the relevant sampling uncertainty rather than treating them as known constants.

The main alternatives answer different questions. Bounds report the range compatible with the information and maintained assumptions. A point-identified analysis adds enough restrictions to single out one population value; its estimate may be more decisive, but the conclusion then relies on those restrictions. A sensitivity analysis can show how conclusions change under selected assumptions, but it does not make a particular assumption true. These approaches can be presented together when clearly labeled; they should not be blended into one apparently assumption-free number.

Do not call the midpoint “the effect.” It is a summary of two endpoints unless a separate estimation or decision rationale is supplied.
Do not treat a bounds interval as a confidence interval. The former represents identification uncertainty under assumptions; sampling uncertainty in estimated bounds is an additional issue.
Do not infer a threshold conclusion from a point inside the interval. Compare the threshold with both endpoints and handle equality explicitly.
Do not overstate what a favorable interval establishes. An interval wholly above zero supports a positive direction, but may still include effects too small to matter in practice.

Key takeaway: Choose bounds when they are the defensible implication of the information and assumptions available. State a conclusion as robust only when it holds for every value in the identified set. If a point estimate relies on added assumptions or a decision convention, label that basis and keep it distinct from what the bounds alone establish.

Check Your Understanding

For each question, first reason from the whole identified interval. Solutions follow.

An effect is bounded by \([1,4]\) units. Is its direction determined? Is an effect above a threshold of 2 units determined?
An interval is \([0,5]\), and the decision criterion is strictly more than 0 units. Does the interval guarantee that criterion?
An interval is \([-2,3]\), and the threshold is \(c=3\). What can be concluded about whether \(\theta\leq c\)?
For bounds \([2,6]\), a report calls 4 the “identified effect.” What is wrong with that wording?

Solutions. (1) The direction is positive throughout because \(L=1>0\). The threshold is not established: \(1<2<4\), so the interval contains values below and above 2. (2) No. Here \(L=c=0<U\), so the interval implies \(\theta\geq0\), not \(\theta>0\); equality is allowed. (3) Since \(L=-2<3=U\), the endpoint case gives \(\theta\leq3\). In fact every value in the interval satisfies the weak criterion, although the interval does not guarantee \(\theta<3\). (4) The bounds identify a set, not the value 4. Four is the midpoint, \((2+6)/2=4\), but no assumption or decision rule has been given that identifies or estimates the effect as that midpoint.