Why Identification Requires Normalizations
An ordered response model describes outcomes whose categories have a natural order, such as low, medium, and high. Its latent-index foundation represents those observed categories as ranges of an unobserved continuous variable. The boundaries between ranges are thresholds. The observed data can reveal how covariates shift outcomes across categories, but they do not provide a natural unit or origin for the unobserved variable. As a result, some latent parameters can be changed together without changing any observable implication of the model.
This is an identification issue, not a computational inconvenience. An identified parameter is uniquely determined by the model's observable implications, subject to its assumptions. If different parameter values imply exactly the same distribution of observed outcomes, the data cannot distinguish them. Ordered logit and ordered probit therefore impose conventions for latent location and scale before interpreting coefficients or thresholds.
Write \(F\) for the cumulative distribution function of the error after its location and scale have been separated. The model's cumulative category boundary is determined by the standardized difference between a threshold and the latent index. In particular, the relevant argument is \((\tau_j-\alpha-x_i'\beta)/s\). This expression makes the two indeterminacies visible: only differences in location matter, and those differences are measured relative to the error scale.
Location Indeterminacy
Suppose the latent variable, its intercept, and every threshold are shifted by the same constant \(a\). The latent variable becomes \(Y_i^{*\,new}=Y_i^*+a\), its intercept becomes \(\alpha^{new}=\alpha+a\), and its thresholds become \(\tau_j^{new}=\tau_j+a\). The error and slope coefficients are unchanged. The category assignment is unchanged because \(\tau_j^{new}-Y_i^{*\,new}=\tau_j-Y_i^*\) for each threshold.
The same conclusion follows from the standardized boundary argument: \((\tau_j^{new}-\alpha^{new}-x_i'\beta)/s=(\tau_j-\alpha-x_i'\beta)/s\). Thus the observed outcomes cannot determine the overall origin of the latent scale. In a specification with an intercept, it cannot separately identify that intercept and the common location of all thresholds.
Worked Example: Shifting the Latent Origin
Consider a hypothetical three-category model with \(\alpha=2\), thresholds \(\tau_1=1\) and \(\tau_2=5\), and any fixed slope vector \(\beta\). Shift the latent origin by \(a=7\).
The transformed parameters are \(\alpha^{new}=2+7=9\), \(\tau_1^{new}=1+7=8\), and \(\tau_2^{new}=5+7=12\). For any covariate value \(x_i\), the first boundary contrast is unchanged: \[ \tau_1^{new}-\alpha^{new}-x_i'\beta=8-9-x_i'\beta=1-2-x_i'\beta. \] The second is unchanged as well: \[ \tau_2^{new}-\alpha^{new}-x_i'\beta=12-9-x_i'\beta=5-2-x_i'\beta. \] Every outcome is assigned to the same category under the original and shifted parameterizations. The shift changes the reported latent origin, not the model's implications for observed categories.
A conventional solution is to omit the intercept and use thresholds relative to it. Define \(c_j=\tau_j-\alpha\); then the boundary contrast is \(c_j-x_i'\beta\). Alternatively, one can set a location parameter to a fixed value. The important point is that the model must not try to estimate an unrestricted intercept and an unrestricted common threshold location as if they were separately determined by the data.
Scale Indeterminacy and Its Normalization
There is also no observed unit for \(Y_i^*\). For any constant \(b>0\), multiply the latent variable, its intercept and slopes, and all thresholds by \(b\); multiply the error by \(b\) as well. Since \(b\) is positive, threshold ordering and category assignments are preserved. In the boundary expression, both numerator and error scale are multiplied by \(b\), leaving their ratio unchanged.
For a model with intercept, the standardized boundary argument can be written as \((\tau_j-\alpha-x_i'\beta)/s\). Under the transformation, it becomes \((b\tau_j-b\alpha-x_i'b\beta)/(bs)\), which equals the original argument. Therefore, without a scale restriction, the data identify ratios such as \(\beta/s\), not the absolute size of \(\beta\) and \(s\) separately. With the intercept absorbed into the thresholds, the same logic applies to \(c_j/s\) and \(\beta/s\).
Worked Example: Equivalent Scale Choices
Suppose a hypothetical ordered logit parameterization uses thresholds \((-1,\;0.5,\;2)\), slope \(\beta=0.8\), and logistic error scale \(s=1\). To display the scale indeterminacy, consider instead a model that allows \(s=2\), and multiply each threshold and the slope by \(2\). The transformed thresholds are \((-2,\;1,\;4)\), and the transformed slope is \(1.6\).
At \(x=1\), the first standardized boundary argument in the original parameterization is \[ \frac{-1-(0.8)(1)}{1}=-1.8. \] In the rescaled parameterization it is \[ \frac{-2-(1.6)(1)}{2}=\frac{-3.6}{2}=-1.8. \] For the second threshold, the corresponding arguments are \((0.5-0.8)/1=-0.3\) and \((1-1.6)/2=-0.3\). All threshold arguments agree, not just those at \(x=1\), because both thresholds and the slope have been multiplied by the same factor as the scale.
The standard ordered logit convention chooses the first parameterization, with logistic scale fixed at one. Under that convention, the coefficient is expressed in standard-logistic latent units. If the scale were left free, the data would identify the ratio \(\beta/s\), which is \(0.8/1=1.6/2=0.8\), rather than \(\beta\) and \(s\) separately.
Threshold Ordering
The thresholds must respect the observed category order. For \(J\) categories, the required constraint is \(\tau_1<\tau_2<\cdots<\tau_{J-1}\), or equivalently \(c_1<c_2<\cdots<c_{J-1}\) after absorbing the intercept. The first threshold separates categories 1 and 2, the next separates categories 2 and 3, and so on. Reversing two thresholds would contradict the meaning of the category labels.
Strict inequalities have a further implication when the error distribution has a strictly increasing cumulative distribution function. At any finite covariate value, the probability of an interior category \(j\) is the difference between the cumulative probabilities at its upper and lower boundaries. Since its upper boundary is strictly higher, that difference is positive. If two adjacent thresholds coincide, the interval assigned to the intervening category has zero width, and that category has zero probability under the model. Such a boundary case does not describe a model in which every category can occur.
Worked Example: Checking an Ordered Set of Thresholds
A hypothetical five-category response model has candidate thresholds \((-1.2,\;0.3,\;0.3,\;2.1)\). Check whether these values satisfy the strict ordering constraint.
The first comparison, \(-1.2<0.3\), holds. The next comparison would require \(0.3<0.3\), which is false, so the thresholds do not satisfy strict ordering. The interval assigned to category 3 has endpoints \(0.3\) and \(0.3\); it has zero width. Under a continuous error distribution, that category consequently has zero probability at every covariate value.
A valid illustrative set is \((-1.2,\;0.3,\;0.8,\;2.1)\), since \(-1.2<0.3<0.8<2.1\). Its thresholds preserve the stated category sequence and leave a nonempty latent interval for each interior category. Estimation procedures must respect this ordering rather than treating the cutpoints as unrelated unrestricted numbers.
What Covariate Shifts Are Identified?
Once location and scale have been normalized, the covariate coefficients describe shifts in the latent index relative to the chosen error scale. If a covariate vector changes from \(x\) to \(x+\Delta x\), the latent-index location changes by \(\Delta x'\beta\). Before normalization, the corresponding scale-free shift is \(\Delta x'\beta/s\). After fixing \(s=1\), the coefficient vector is reported on that normalized latent scale.
These are not changes in the units of the observed ordinal outcome. For example, a coefficient of \(0.6\) does not mean that a one-unit covariate increase raises the observed category by \(0.6\) categories. It means that the covariate shifts the latent index by \(0.6\) normalized units, holding other covariates fixed. The consequences for category probabilities depend on where the thresholds lie relative to the index. The coefficient's direction is identified under the chosen orientation: a positive coefficient shifts the index toward higher-numbered categories.
Worked Example: Interpreting a Normalized Covariate Shift
Consider a hypothetical ordered probit model with no intercept, thresholds \((-1,\;0.5,\;2)\), and coefficient \(\beta=0.6\) on a covariate \(x\), with standard-normal error scale fixed at one. Compare \(x=0\) with \(x=1\).
The latent-index location changes from \(0.6(0)=0\) to \(0.6(1)=0.6\), a shift of \(0.6\) normalized latent units. The threshold-minus-index contrasts at \(x=0\) are \((-1,\;0.5,\;2)\). At \(x=1\), they are \[ (-1-0.6,\;0.5-0.6,\;2-0.6)=(-1.6,\;-0.1,\;1.4). \] Each contrast falls by \(0.6\), because the covariate raises the latent index while the thresholds stay fixed. Thus the model places the latent index closer to higher categories; it does not assert a \(0.6\)-category increase in the observed response.
If the error scale had instead been left as \(s=2\), the same observable shift would be described by a coefficient of \(1.2\): its scale-free value would still be \(1.2/2=0.6\). The normalized probit coefficient \(0.6\) is therefore a unit-dependent representation of the identified shift relative to error variation.
Derivation: The Identified Boundary Contrast
The identification argument can be summarized directly from the latent threshold rule. For a given threshold, category membership depends on whether \(Y_i^*\) lies above or below that threshold. Substituting the latent index into that comparison gives a contrast between the threshold and the covariate-driven location:
If \(\varepsilon_i\) has location zero and scale \(s\), standardizing the error divides the right-hand boundary by \(s\). Therefore the observable boundary depends on \((\tau_j-\alpha-x_i'\beta)/s\), not on the individual unscaled quantities. A common translation leaves the differences unchanged. A common positive rescaling multiplies both the differences and \(s\), so the standardized boundary remains unchanged. Fixing the error location and scale, and absorbing or fixing the latent intercept, removes these redundant degrees of freedom. Strictly ordered thresholds then preserve the ordered category structure.
Use thresholds relative to the intercept, or impose a location normalization, so a common shift cannot be separately assigned to the intercept and every threshold.
Use the standard-normal error for probit or the standard-logistic error for logit. This makes coefficients and thresholds interpretable relative to a stated scale.
Require each threshold to be strictly greater than the preceding threshold when every category is to have positive probability.
Common Pitfall and Practical Significance
A common mistake is to interpret a larger ordered-logit coefficient as evidence of a larger substantive effect without checking the latent scale. The coefficient is tied to the error-scale convention: if the scale were allowed to vary, multiplying all latent coefficients and thresholds by a positive constant while multiplying the error scale by the same constant would leave the model's observable implications unchanged. Comparisons across models with different error-scale conventions therefore require care.
A second mistake is to estimate an intercept and all thresholds freely, then treat each estimate as independently meaningful. Their common location is not separately identified. A third is to allow thresholds to cross or coincide and then interpret the resulting parameters as boundaries for the stated ordered categories. The correction in each case is structural: choose an explicit location and scale normalization and enforce the threshold order. Those choices make the model's parameterization identifiable; they do not add information about a natural origin or unit that the ordinal observations do not contain.
Check Your Understanding
Use the latent-index model and its identification constraints to answer the following questions.
- Why does adding the same constant to the intercept and every threshold leave category assignments unchanged?
- In an ordered probit model, what scale restriction is imposed on the error, and what does it imply about the units of the coefficients?
- For thresholds \((-2,\;0,\;0,\;1)\), which category has a zero-width latent interval?
- A covariate coefficient is positive. What does this say about the latent index, and why does it not mean the observed category rises by that many units?
- Explain why the standard logistic error has variance \(\pi^2/3\) even though its scale is normalized to one.