Interpreting the Direction of a Delta Shift

What the Direction of a Delta Shift Says

A delta value is meaningful only relative to a stated reference. In a sensitivity analysis, a baseline imputation model supplies a reference distribution for missing values; a delta shift then represents an assumption about how the distribution for those missing values differs from that reference. The sign describes the direction of the assumed shift on the specified scale. It does not establish what the unobserved outcomes truly are.

For interpretation, keep three things in view: which group is being shifted, which group or model provides the reference, and how the outcome is coded. For example, a positive shift in a symptom score coded so that larger values mean worse symptoms assumes worse outcomes among the missing cases. If the score is coded so that larger values mean better health, that same positive shift assumes better outcomes.

The interpretation is conditional on the modeling choices that define the reference. In particular, the baseline model, the chosen reference group, the outcome coding, and the scale must be clear. A delta shift is a sensitivity assumption, not a quantity identified from the observed outcomes alone when the missingness mechanism is not identified by those data.

Assumptions Behind the Interpretation

Before interpreting a shift, verify that the shift has a coherent meaning in the model being used. For the examples below, assume the baseline imputation model adequately represents the reference conditional distribution given the modeled predictors; the relevant reference group and covariate values are specified; and the delta is applied to the missing outcomes as described. A constant shift also assumes the chosen delta applies uniformly across the cases to which it is assigned. That may be a useful sensitivity scenario, but it is not guaranteed by the data.

For an additive shift, assume the outcome’s numerical scale is meaningful for differences and that adding delta does not produce impossible values. If every imputed value is translated by the same amount, the distribution’s shape and spread are retained while its location moves. For a binary outcome, assume the model represents the event probability and the delta is added to its logit. The event must be defined consistently: changing which category is called the event reverses the direction of the log-odds contrast.

In the binary expression, \(p_0\) is the reference probability of the specified event and \(p_{\delta}\) is the probability after the log-odds shift. The logit is strictly increasing in probability, so a positive delta increases the event probability and a negative delta decreases it, provided the event coding is held fixed.

Reading an Additive Delta in Outcome Units

On an additive scale, the arithmetic interpretation is direct: the delta is added to the reference value. If delta is \(-0.4\) points on a symptom score, the adjusted missing outcomes are 0.4 points lower than the corresponding reference draws. This statement concerns the shifted distribution under the scenario. It does not claim that every missing person’s unobserved outcome would be exactly 0.4 points below an observed person’s outcome.

The units are essential: a delta of \(-5\) points is not interchangeable with \(-5\) percent or with a standardized difference. Similarly, if an outcome is a count or is bounded, an unrestricted additive shift can yield invalid values. A positive sign alone does not guarantee a plausible scenario; its consequences must be checked against the outcome’s support and the study context.

Reading a Log-Odds Delta

A log-odds delta is not a change in probability by the same numerical amount. Exponentiating the delta gives the multiplier on the odds of the defined event. The corresponding probability change depends on the reference probability: the same odds multiplier generally produces different probability changes at different starting probabilities.

This follows by exponentiating both sides of the logit equation: \(\exp(\operatorname{logit}(p_{\delta}))=\exp(\operatorname{logit}(p_0)+\delta)\). Since \(\exp(a+b)=\exp(a)\exp(b)\), the adjusted odds equal the reference odds multiplied by \(\exp(\delta)\). Because \(\exp(\delta)>1\) for positive delta and is between 0 and 1 for negative delta, the direction of the odds change follows the sign of delta.

The calculation uses the unrounded reference odds when multiplying by 2; rounding \(0.428571\ldots\) first can create a small arithmetic discrepancy. In general, carry extra precision through intermediate steps and round only the reported result.

Direction on the Outcome Scale Is Not the Direction of an Effect

A delta’s sign describes an outcome shift relative to its reference, not automatically the direction of a treatment contrast, a regression coefficient, or a causal effect. Those quantities depend on how shifts are assigned across groups and on the scale used to compare groups. Even a common log-odds shift can change a risk difference because the probability transformation is nonlinear.

If a delta is applied to one group but not another, or different deltas are used across groups, the resulting contrast depends further on that assignment. State the shifted group, the reference group, the event definition, and the effect scale before describing the sensitivity result.

Common Mistakes, Diagnostics, and Alternatives

• Reading the sign without checking coding. A positive shift means a higher value on the modeled scale. Check whether higher outcome values mean improvement, worsening, or a different event category.
• Calling a log-odds delta a probability difference. A delta of 0.5 is not a 50-percentage-point increase. Its odds multiplier is \(\exp(0.5)\approx1.649\), and the probability change depends on \(p_0\).
• Equating missing with observed individuals one by one. The shift describes a modeled distributional contrast, conditional on the stated reference and predictors. It is not a claim about individual unobserved values.
• Assuming a common shift leaves every effect measure unchanged. A common log-odds shift changes probabilities nonlinearly and can change risk differences.
• Ignoring the outcome’s support. For additive shifts, inspect whether adjusted draws remain allowed. For binary outcomes, verify that the event and non-event categories have not been reversed.

Useful checks include calculating several adjusted probabilities at representative reference probabilities, reviewing the shifted values against the outcome’s possible range, and confirming that the delta is attached to the intended missing cases and reference. If signs or magnitudes appear implausible, first check the coding and reference definition, then inspect the implied adjusted distribution. A correct formula cannot rescue an incorrectly specified reference.

The main alternatives answer different modeling questions. An additive shift is natural when differences in outcome units are interpretable. A log-odds shift encodes a multiplicative change in event odds and keeps probabilities between 0 and 1 through the logistic transformation. A probability-scale adjustment instead specifies a direct probability difference, but it must respect the 0-to-1 bounds. A no-shift reference scenario leaves the baseline imputation distribution unchanged. These are alternative sensitivity assumptions, not competing estimates of a known missing value; each has limitations, and none identifies the unobserved outcomes from observed data alone.