Does multivalued always mean ordered dose levels?

No. Usage can include unordered alternatives, though ordered levels are a common application.

Can the effect between adjacent levels differ?

Yes. A flexible analysis allows different increments rather than imposing a straight line by default.

Is one propensity score probability enough for every contrast?

The required balancing or weighting construction depends on the estimand and assumptions; the full treatment structure must be respected.

AI Glossary / Statistics and probability / Advanced

Multivalued treatment

Definition

A multivalued treatment is a treatment variable with more than two possible values, such as no participation, low-intensity participation and high-intensity participation. The values may represent ordered levels or distinct alternatives, depending on the application. Causal analysis defines a potential outcome under each treatment value and compares specified values for a stated population. A multivalued treatment is not the same as a treatment effect that happens to differ across individuals. Terminology overlaps with multiple treatments, so the actual intervention values and their interpretation should be made explicit.

What it means

A binary treatment asks what changes between two conditions. A multivalued treatment permits several contrasts and, for ordered intensities, a dose-response question. Moving from level zero to one may have a different effect from moving from one to two. Coding levels as 0, 1 and 2 does not establish equal spacing in a meaningful dose or equal outcome increments. Only one potential outcome is normally observed for each unit, so comparisons still face the fundamental problem of unobserved counterfactuals. Observational selection into treatment levels can confound crude comparisons.

How it works

Define the treatment levels, outcome, timing and target population. Specify whether the estimand is a contrast between levels, a set of potential-outcome means or a dose-response function. Use random assignment or an identification strategy appropriate to the observational setting. For discrete levels, a generalized propensity score includes the probability of receiving each level conditional on measured pretreatment covariates. Weighting, matching or outcome models can support estimation under suitable assumptions, including consistency, exchangeability and overlap. Examine balance and extreme weights. Continuous treatment requires density-based methods and additional care rather than simply treating every observed dose as a separate category.

Three practical examples

Program evaluation: compare no course, a short course and an intensive course with clearly defined content and follow-up.
Product experimentation: assign users to several specified levels of a feature's exposure or intensity.
Policy analysis: estimate outcomes under alternative service levels for a common eligible population. These examples require attention to whether treatment versions are comparable and whether each level could realistically be assigned to the relevant units.

Think of it this way

Think of a dial with several settings rather than an on-off switch. The outcome may respond differently to each change in setting. Observing people who chose different settings does not reveal the dial's effect unless differences between those people are addressed by the design or defensible assumptions.

Worked example

Suppose a hypothetical target population has potential-outcome means of 50, 60 and 64 under treatment levels 0, 1 and 2. The average contrasts are 10 for level 1 versus 0, 4 for level 2 versus 1, and 14 for level 2 versus 0. A straight line connecting levels 0 and 2 has slope 7 and predicts 57 at level 1, missing the stated mean of 60. The dose-response pattern is therefore not linear in the numerical level code. Separately, if a unit's probabilities of receiving levels 0, 1 and 2 are 0.2, 0.3 and 0.5, its inverse-probability weight when observed at level 2 is 1/0.5 = 2. A probability of 0.02 would instead produce weight 50. These original calculations illustrate contrast definition and overlap sensitivity; they are not an estimated causal result from real observations.

Illustrative case study

In a fictional case, Oak Learning evaluates three training intensities. Analysts initially enter the treatment code as a single linear predictor, implicitly assuming equal outcome increments. A more flexible specification allows the intermediate level to differ from that pattern. The team also finds that highly experienced employees almost never select the lowest intensity, creating weak overlap for some comparisons. It reports the limitation and revisits the target population rather than relying on a complex model to invent unsupported counterfactual outcomes.

Common misconceptions and mistakes

More than two treatment values do not imply a linear dose-response. An ordinal label need not represent equal numerical distances, and different treatment types may have no natural order at all. A multinomial model for treatment assignment does not by itself identify causal effects. Balancing measured covariates cannot guarantee that unmeasured confounding is absent. Very small assignment probabilities can create unstable weighting, while trimming observations changes the target population. Contrasts between levels must refer to compatible populations if they are to form a coherent comparison.

Business use

Businesses should define intervention levels precisely enough that the estimated contrast corresponds to an implementable decision. Plan allocation and sample sizes for the comparisons that matter, rather than spreading resources across many poorly supported levels. Report potential-outcome means or contrasts with uncertainty and clarify the population to which they apply. Examine overlap and sensitivity to model choices before ranking levels. If greater intensity costs more, evaluate incremental benefit alongside incremental cost. The highest observed average outcome need not identify the best causal or economic choice.

Check your understanding

Treatment levels are coded 0, 1 and 2. Does that coding justify assuming that the effect from 0 to 1 equals the effect from 1 to 2?

Answer: No. Equal increments are an additional modeling assumption, not a consequence of the labels.

Related terms

Multiple treatments
potential outcome
dose-response
generalized propensity score
inverse-probability weighting
overlap
treatment contrast
exchangeability

Learn next

Define potential-outcome means for several levels and compare adjacent with nonadjacent contrasts. Then study generalized propensity scores and how limited overlap changes what can be estimated credibly.

Propensity scores

Sources

Guido Imbens, The Role of the Propensity Score in Estimating Dose-Response Functions: https://www.nber.org/papers/t0237 StataCorp, An ordered-probit inverse probability weighted estimator, multivalued treatment weighting: https://blog.stata.com/2016/09/13/an-ordered-probit-inverse-probability-weighted-ipw-estimator/

Previous: Multivalued Fact Next: Multivariable model