Tipping-Point Interpretation on Different Outcome Scales

Why the Scale of a Tipping Threshold Matters

A tipping point describes where a specified sensitivity analysis changes its conclusion. Its numerical value is expressed on some scale: for example, a shift of \(-6\) points, a shift of \(-0.5\) standard deviations, or a log-scale shift of \(\log(0.8)\). Those numbers can refer to the same missing-outcome scenario, or they can encode different scenarios. The distinction matters because a threshold without its scale and parameterization is difficult to interpret.

The basic setting is a pattern-mixture sensitivity analysis in which missing outcomes are imputed under a reference assumption and then adjusted by a sensitivity parameter. In a simple continuous-outcome model, let \(Y_{\mathrm{mis}}(0)\) denote an imputed missing outcome under the reference assumption. An additive raw-scale adjustment is \(Y_{\mathrm{mis}}(\delta)=Y_{\mathrm{mis}}(0)+\delta\). A tipping threshold is a value of \(\delta\) at which a stated decision rule changes. For example, the rule might concern whether an estimated mean difference is positive or whether a confidence interval excludes zero. The threshold depends on that rule as well as on the data and model.

Earlier material on delta adjustment establishes that the direction and meaning of \(\delta\) depend on how it is defined. Here the question is narrower: how should a threshold be interpreted when outcomes are expressed in different units or transformed? A useful principle is to first identify the outcomes that each scenario assigns to missing cases. Only then translate the threshold.

Raw Units and Standardized Units

For a raw-scale additive adjustment, \(\delta\) has the same units as the outcome. If \(Y\) is a symptom score measured in points, a threshold of \(-6\) means that the assumed missing outcomes are shifted downward by 6 points relative to their reference imputations. This is often the most direct clinical or operational interpretation.

A standardized shift divides the raw shift by a chosen reference standard deviation. Let \(s_{\mathrm{ref}}\) be a fixed, positive standard deviation, such as the standard deviation of the outcome in the observed reference group. Define the standardized outcome as \(Z=(Y-c)/s_{\mathrm{ref}}\), where \(c\) is a fixed centering constant. A raw shift \(\delta_Y\) then corresponds to a standardized shift \(\delta_Z=\delta_Y/s_{\mathrm{ref}}\). The centering constant cancels because the same \(c\) is used before and after adjustment.

This conversion is exact for a fixed linear rescaling. It is not enough to say “standardized units,” however: the reference standard deviation must be named or otherwise defined. If the standard deviation is recalculated separately for each sensitivity scenario, the denominator can also change, so the converted values need not represent a simple translation of one fixed raw-scale threshold.

Transformed Outcomes: Translate the Scenario, Not Just the Number

A nonlinear transformation requires more care. Let \(g\) be a one-to-one transformation over the outcome values under consideration. If a reference-imputed value \(y_0\) is changed to \(y_1\), the corresponding change on the transformed scale is \(g(y_1)-g(y_0)\). In general, that difference depends on \(y_0\). Therefore, one raw additive shift need not correspond to a constant additive shift on the transformed scale for every missing case.

For a positive outcome, the logarithm is a common example. An additive shift \(\delta_{\log}\) to the log outcome gives \(\log(Y_{\mathrm{mis}})=\log(Y_{\mathrm{mis}}(0))+\delta_{\log}\). Exponentiating both sides shows that the original-scale outcome is multiplied by \(\exp(\delta_{\log})\). Thus the log-scale shift has a multiplicative interpretation. A negative log shift is a reduction; it does not generally mean a fixed reduction in the outcome’s original units.

Why a Nonlinear Scale Can Define a Different Analysis

A nonlinear transformation can do more than change the units used to display a threshold. Imposing a constant additive adjustment on the transformed scale generally creates outcome assignments different from those made by a constant raw-scale adjustment. This is not an error if the transformed-scale assumption is intended, but it should not be described as a mere conversion of the raw-scale threshold.

Decision Rules, Comparisons, and Diagnostics

A threshold is interpretable only alongside the decision rule that defines it. For an effect-direction rule, it may mark where an estimated effect changes sign. For an inferential rule, it may mark where a confidence-interval conclusion changes. Those are different tipping definitions, even if the sensitivity parameter and outcome scale are identical. Changing an outcome scale can also change the estimand or the fitted model, so a conclusion about the mean on one scale cannot automatically be carried over to a mean on another.

For a fixed scenario and a decision rule defined on the same underlying quantity, translating units does not change the scenario itself. For example, converting a score from points to a linearly rescaled score changes the numerical threshold but not the adjusted outcomes represented by it. With a nonlinear transformation, it is safer to reconstruct the missing outcomes and state what is being summarized than to assume that a test or mean-based conclusion is invariant.

Several diagnostics help reveal a scale mismatch before a threshold is reported:

• Check the parameter units. A raw-scale delta has outcome units; a standardized delta is dimensionless; a log-scale delta is also dimensionless but has a multiplicative interpretation after exponentiation.
• Reconstruct adjusted outcomes. Apply the proposed shift to representative reference imputations. If the implied changes are not the intended ones, the proposed “conversion” changed the scenario.
• Check the transformation’s domain. For example, a logarithm requires positive values. A proposed transformed-scale adjustment that produces invalid values cannot be used as stated.
• Hold the reference scale fixed when standardizing. Re-estimating the denominator across scenarios can make apparent changes in standardized thresholds reflect both a changing shift and a changing scale.
• Recheck the estimand and decision rule. A transformation can change what is being averaged or tested. Confirm that a reported change in conclusion is the one the analysis set out to assess.

A practical reporting comparison is between the raw-unit additive model, which is easy to interpret as a fixed absolute change, and a transformed-scale model, which can express changes such as a fixed proportion. Neither is universally preferable. The appropriate choice depends on the outcome, the intended assumption about departures from reference imputations, and the target estimand. A standardized summary can help compare the magnitude of shifts across contexts, but it does not replace the clinically meaningful raw-unit description.

Deriving a Valid Scale Translation

The conversion rule can be justified directly. For standardization, both the reference and adjusted outcome are transformed with the same fixed center and scale. Subtracting them eliminates the center, leaving the raw difference divided by the reference standard deviation. For a general one-to-one transformation, the same adjusted outcome \(y_1\) must be obtained after transforming back. The corresponding transformed shift is therefore determined by the transformed endpoints, not by a universal conversion factor.

For standardization, take \(g(y)=(y-c)/s_{\mathrm{ref}}\). Then \(\delta_g=(y_1-c)/s_{\mathrm{ref}}-(y_0-c)/s_{\mathrm{ref}}=(y_1-y_0)/s_{\mathrm{ref}}\), which is constant for a fixed raw difference. For the logarithm, \(g(y)=\log(y)\), so \(\delta_g=\log(y_1/y_0)\). This depends on the ratio rather than the raw difference, explaining why a fixed raw shift usually produces different log shifts at different reference values.

Common Mistake and Practical Limits

Wrong reasoning: “The tipping threshold is \(-0.5\) on both scales, so the missing outcomes have the same adjustment.” This ignores what one unit means on each scale. A half-unit standardized shift, a half-unit log shift, and a half-unit raw shift are not interchangeable. The correction is to specify the transformation and reference scale, then translate the implied missing outcomes. For a log shift, exponentiate the parameter to obtain the multiplier; for a standardized shift, multiply by the stated reference standard deviation.

Scale translation also has limits. Standardization depends on a defensible reference standard deviation, and a nonlinear transformation may be difficult to interpret or may be undefined for some outcomes. More fundamentally, expressing a threshold in a different scale does not make the sensitivity assumption empirically verified. It makes the assumption easier to communicate or compare; conclusions still depend on the chosen missing-data model, the observed data, and the decision rule.