Why the Scale of a Selection Coefficient Matters
A selection coefficient can look small or large simply because of the units used to record an outcome. A coefficient per one symptom-score point, per one standard deviation, and per a clinically meaningful change may describe the same fitted relationship. Comparing the numbers without converting their units can therefore be misleading. The useful question is not just “What is the coefficient?” but “What change in the outcome does it represent, and what change in dropout odds does the model imply?”
Consider a discrete-time dropout model at visit \(t\). Among participants still under observation just before that visit, let \(D_t=1\) indicate dropout at \(t\), and let \(Y_t\) be the current outcome used in the selection equation. For fixed time and any other variables held fixed, write:
Here \(\beta_t\) is the change in conditional log odds of dropout for a one-unit increase in \(Y_t\). The model assumes that, over the range being interpreted, this log-odds relationship is linear in the coded outcome. Interpreting a coefficient also requires stating what a larger outcome value means. The equation describes a model association; it does not, by itself, establish that changing the outcome would cause dropout.
How the Coefficient Changes Under Rescaling
Let \(Z_t=(Y_t-a)/c\), where \(a\) is a fixed centering value and \(c>0\) is a fixed scale. Then \(Y_t=a+cZ_t\). Substituting this identity into the original model gives the rescaled model:
Thus the coefficient for \(Z_t\) is \(c\beta_t\), and the intercept becomes \(\alpha_t+\beta_ta\). Centering alone (\(c=1\)) changes the intercept but not the slope. A scale change alters the slope’s numerical value because one unit of \(Z_t\) represents \(c\) units of \(Y_t\). The fitted linear predictor, and therefore the fitted model for the same outcome values, is unchanged.
If \(c\) is an outcome standard deviation, the rescaled coefficient is the change in log odds per standard deviation. If \(c\) is a prespecified clinical increment, it is the change in log odds per such increment. In either case, the corresponding odds multiplier is the exponential of the rescaled coefficient. A positive rescaling preserves the sign; reversing the outcome’s direction instead reverses the sign and requires careful interpretation.
Worked Example: Raw and Standardized Symptom Scores
Suppose a hypothetical model uses a symptom score measured in points, with \(\alpha_t=-3.00\) and \(\beta_t=0.080\) per point. Assume the reference outcome distribution has standard deviation \(8.00\) points. Compare model-implied dropout odds for scores of 15 and 20, first in raw units and then in standard-deviation units.
A five-point increase changes the linear predictor by \(0.080(20-15)=0.400\). The odds multiplier is \(\exp(0.400)=1.492\). Thus, at the same visit and with other conditioning variables fixed, the model implies 1.492 times the dropout odds at score 20 as at score 15.
The scores also allow a direct check using the model’s odds. At 15 points, the log odds are \(-3.00+0.080(15)=-1.80\), giving odds \(\exp(-1.80)=0.1653\). At 20 points, the log odds are \(-3.00+0.080(20)=-1.40\), giving odds \(\exp(-1.40)=0.2466\). Their ratio is \(0.2466/0.1653=1.492\), up to rounding.
Set \(Z_t=(Y_t-20)/8\). Then \(Y_t=20+8Z_t\), so the standardized coefficient is \(8(0.080)=0.640\) per standard deviation and the intercept is \(-3.00+0.080(20)=-1.40\). The score of 15 corresponds to \(Z_t=-0.625\), and 20 corresponds to \(Z_t=0\). Their linear predictors are \(-1.40+0.640(-0.625)=-1.80\) and \(-1.40\). The same 0.400 log-odds difference and 1.492 odds multiplier follow. Standardization has changed the coefficient’s units, not the fitted comparison.
Choosing a Scale for Interpretation
Raw units are useful when they are familiar and substantively interpretable. Standard-deviation units can make coefficients from different outcome scales easier to compare, provided the reference standard deviations and outcome definitions are appropriate. A clinical increment can be clearest when stakeholders recognize a particular change as meaningful. These scales answer different reporting needs; none makes a selection model more identified or less dependent on its assumptions.
| Scale reported | Coefficient represents | Odds multiplier for the stated change | Main caution |
|---|---|---|---|
| Raw outcome | Log-odds change per recorded unit | \(\exp(\beta_t\Delta)\) for \(\Delta\) raw units | Units may be unfamiliar or too small to be useful |
| Standardized outcome | Log-odds change per reference standard deviation | \(\exp(\beta_t\sigma)\) for one SD when \(\beta_t\) is per raw unit | The reference SD must be stated; it depends on the population or sample used |
| Clinical increment | Log-odds change per specified clinical amount | \(\exp(\beta_t c)\) for an increment of \(c\) raw units | The increment and direction must be explicit |
For a sample-based standardization, the sample standard deviation \(s\) can be used in place of a population standard deviation. Report which value was used and where it came from. If \(s\) is used, the standardized coefficient is \(s\hat{\beta}_t\); the choice of reference group or time point can therefore affect comparisons across analyses.
Worked Example: A Clinically Meaningful Change
Suppose a hypothetical fatigue scale is coded so that higher scores indicate worse fatigue, and the estimated dropout coefficient is \(\hat{\beta}_t=-0.025\) per point. A decrease of 10 points is the clinical change of interest. Calculate the model-implied odds multiplier for that decrease.
The change is \(\Delta=-10\) points, so the log-odds change is \((-0.025)(-10)=0.250\). The odds multiplier is \(\exp(0.250)=1.284\). Therefore, the model implies 1.284 times the dropout odds at a score 10 points lower than at the comparison score, conditional on the same visit and other modeled variables.
The coefficient for a 10-point higher score is \((-0.025)(10)=-0.250\), with odds multiplier \(\exp(-0.250)=0.7788\). This is the reciprocal of 1.284, apart from rounding. The sign is not a contradiction: a 10-point decrease and a 10-point increase describe opposite comparisons. Since higher values mean worse fatigue, the model’s negative coefficient associates worse fatigue with lower conditional dropout odds in this specification.
Uncertainty and a Complete Conversion
A coefficient’s uncertainty must be converted along with the coefficient. For a fixed positive scaling constant \(c\), if the raw-scale estimate is \(\hat{\beta}_t\) with standard error \(SE(\hat{\beta}_t)\), then the rescaled estimate and its standard error are \(c\hat{\beta}_t\) and \(cSE(\hat{\beta}_t)\). A confidence interval for the rescaled log-odds coefficient can be exponentiated at both endpoints to obtain an interval for the corresponding odds multiplier. The interval is then expressed for the specified outcome increment, not for a one-unit raw change.
This follows from multiplying the coefficient by a constant: its sampling standard deviation is multiplied by the absolute value of that constant. With \(c>0\), both the estimate and standard error are multiplied by \(c\). Consequently, the Wald statistic \(\hat{\beta}_t/SE(\hat{\beta}_t)\) is unchanged. The same fitted relationship has not gained or lost statistical evidence merely because the outcome was recorded on another positive linear scale.
Worked Example: Rescaling an Estimate and Its Interval
In a hypothetical analysis, the estimated coefficient is \(0.060\) per outcome point, with standard error \(0.020\) per point. A four-point increase is the chosen clinical increment. Calculate its estimated odds multiplier and a 95% Wald confidence interval for that multiplier.
For four points, the coefficient is \(4(0.060)=0.240\), and its standard error is \(4(0.020)=0.080\). The estimated odds multiplier is \(\exp(0.240)=1.271\). Under the usual large-sample Wald approximation, the 95% interval on the log-odds scale is \(0.240\pm1.96(0.080)\), or \([0.0832,0.3968]\). Exponentiating gives \([\exp(0.0832),\exp(0.3968)]=[1.087,1.487]\).
The model-based 95% confidence interval for the conditional dropout-odds multiplier associated with a four-point increase is approximately 1.087 to 1.487, at the same visit and with other conditioning variables fixed. It is not a 95% probability statement about a fixed parameter. The interval describes uncertainty under the fitted model and its large-sample approximation.
Diagnostics, Alternatives, and Limitations
Before interpreting a scaled coefficient, verify that the outcome was transformed as intended, the scale constant is positive and correctly reported, and the direction of higher scores is clear. Recompute at least one odds multiplier from the original coefficient and the stated outcome increment. It should agree with the multiplier from the converted coefficient. Disagreement often signals a unit conversion error, a reversed contrast, or an intercept mistakenly treated as a slope.
- Check linearity on the log-odds scale. The conversion is exact for the specified linear-logit model, but it does not establish that the model fits. A relationship that bends over the relevant outcome range may not be summarized well by one constant slope.
- Check the range of interpretation. A meaningful increment should be plausible in the data and in the scientific setting. Applying a per-unit slope to a very large change can imply a multiplier based on extrapolation.
- Check the outcome coding. If higher scores mean improvement rather than worsening, the verbal interpretation must reflect that. Reversing the coding reverses the coefficient sign.
- Keep the model’s conditioning fixed. The multiplier compares specified outcome values at the same visit and with other modeled variables held fixed. It is not, by itself, a marginal population comparison.
The main alternative is not a different fitted model but a different reporting scale. Raw-unit reporting preserves direct connection to data collection; standardization aids scale-free comparison; clinical-unit reporting prioritizes interpretability. In contrast, replacing a linear outcome term with a nonlinear term or changing the link function changes the model, not merely its units. A transformed coefficient from that different model cannot be obtained by simple multiplication alone.
Scaling does not resolve the central limitations of selection models. The interpretation still depends on the chosen outcome model, selection equation, and assumptions about outcomes not observed after dropout. A standardized coefficient is not inherently more clinically meaningful, and a clinical-scale odds multiplier is not evidence of a causal effect. These are unit choices for expressing a model parameter, not remedies for identification or model misspecification.
Check Your Understanding
Work each problem first, then compare with the concise solutions below.
- A model has a coefficient of \(0.04\) per point. What coefficient corresponds to a 6-point scale, and what is the implied odds multiplier for a 6-point increase?
- An outcome is standardized as \(Z=(Y-30)/5\), and the raw-scale coefficient is \(-0.03\) per point. Find the standardized coefficient and explain the intercept’s role in the transformation.
- A coefficient is \(0.05\) per unit, and a 10-unit decrease is being compared with the starting value. Calculate the odds multiplier and state why the direction matters.
- Why can a coefficient be numerically larger after standardization even though the fitted model has not changed?
Solutions. (1) The coefficient per 6 points is \(6(0.04)=0.24\), and the odds multiplier is \(\exp(0.24)=1.271\). (2) The standardized coefficient is \(5(-0.03)=-0.15\) per standard deviation. Centering at 30 also changes the intercept to \(\alpha+30\beta\); it does not change fitted linear predictors. (3) The change is \(\Delta=-10\), so the log-odds change is \(0.05(-10)=-0.50\), and the odds multiplier is \(\exp(-0.50)=0.6065\). The decrease has the opposite direction from a 10-unit increase, whose multiplier is \(\exp(0.50)=1.649\). (4) One standardized unit represents several raw units when the standard deviation exceeds one. The coefficient is multiplied by that standard deviation, while the outcome values are expressed in corresponding units, leaving each fitted linear predictor unchanged.