Combined Likelihood Profile (CLIP)

• Combined Likelihood Profile (CLIP) is a statistical method that builds a profile likelihood from disjoint datasets sharing common parameters while independently handling nuisance parameters.
• It enhances inference by combining per-dataset likelihoods without pooling raw data, allowing for tailored treatment of backgrounds and systematics.
• CLIP is vital in complex analyses, including astrophysics and dark matter searches, where preserving dataset-specific characteristics improves uncertainty quantification.

The phrase “Combined Likelihood Profile” (“CLIP”, Editor’s term) denotes a profile-likelihood construction built from a joint likelihood over multiple disjoint datasets that share one or more parameters of interest while retaining dataset-specific nuisance parameters. In the terminology of “Using Likelihood for Combined Data Set Analysis” (Anderson et al., 2015), the underlying object is the joint likelihood together with likelihood profiling: common parameters are estimated by multiplying per-dataset likelihoods, and nuisance parameters are handled independently for each dataset rather than by pooling raw data. In this sense, CLIP is not a separate statistical formalism but a compact name for the profile of a combined likelihood. The same acronym also appears in a distinct sense in “Whitened CLIP as a Likelihood Surrogate of Images and Captions” (Betser et al., 11 May 2025), where CLIP refers to Contrastive Language-Image Pre-training rather than combined profile-likelihood analysis.

1. Statistical definition

For a single dataset $\mathcal{D}$, the standard likelihood is

$$L(\alpha \mid \mathcal{D}) = P(\mathcal{D} \mid \alpha),$$

where $\alpha$ contains both parameters of interest and nuisance parameters (Anderson et al., 2015). For binned count data, the construction used in the joint-likelihood paper is a binned Poisson likelihood,

$$\mathcal{L}_d(\boldsymbol{\mu}, \boldsymbol{\theta}_d \mid \mathcal{D}_d) = \prod_k \frac{\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)^{n_{d,k}} e^{-\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)}}{n_{d,k}!},$$

with common parameters $\boldsymbol{\mu}$ and dataset-specific nuisance parameters $\boldsymbol{\theta}_d$ (Anderson et al., 2015).

For $N$ disjoint datasets $\mathcal{D}_d$, the joint likelihood is

$$\mathcal{L}_{\rm joint}(\boldsymbol{\mu}, \{\boldsymbol{\theta}_d\}\mid \{\mathcal{D}_d\}) = \prod_{d=1}^{N} \mathcal{L}_d(\boldsymbol{\mu}, \boldsymbol{\theta}_d \mid \mathcal{D}_d),$$

and the corresponding log-likelihood is additive,

$$\ln \mathcal{L}_{\rm joint}(\boldsymbol{\mu}, \{\boldsymbol{\theta}_d\}) = \sum_{d=1}^{N} \ln \mathcal{L}_d(\boldsymbol{\mu}, \boldsymbol{\theta}_d).$$

The CLIP object is the profile over nuisance parameters,

$$L(\alpha \mid \mathcal{D}) = P(\mathcal{D} \mid \alpha),$$0

or, equivalently in log form,

$$L(\alpha \mid \mathcal{D}) = P(\mathcal{D} \mid \alpha),$$1

This construction is explicitly motivated by the fact that joining likelihoods, rather than the data itself, allows nuisance parameters to be dealt with independently (Anderson et al., 2015).

2. Profile-likelihood inference

Inference in a combined likelihood profile is based on the profile likelihood ratio. For a null value $$L(\alpha \mid \mathcal{D}) = P(\mathcal{D} \mid \alpha),$$2, the test statistic is

$$L(\alpha \mid \mathcal{D}) = P(\mathcal{D} \mid \alpha),$$3

where the numerator uses the constrained maximum over nuisance parameters and the denominator uses the global maximum-likelihood estimate (Anderson et al., 2015). For confidence intervals and upper limits, one uses the profile log-likelihood difference,

$$L(\alpha \mid \mathcal{D}) = P(\mathcal{D} \mid \alpha),$$4

For a single parameter of interest in the asymptotic regime, a one-sided $$L(\alpha \mid \mathcal{D}) = P(\mathcal{D} \mid \alpha),$$5 confidence upper limit corresponds to $$L(\alpha \mid \mathcal{D}) = P(\mathcal{D} \mid \alpha),$$6, or equivalently $$L(\alpha \mid \mathcal{D}) = P(\mathcal{D} \mid \alpha),$$7 (Anderson et al., 2015). The same paper notes that, within the asymptotic regime, confidence limits are set at levels corresponding to the $$L(\alpha \mid \mathcal{D}) = P(\mathcal{D} \mid \alpha),$$8 probability density function.

When the signal parameter is constrained to be non-negative, the null TS distribution follows the Chernoff form described in the paper as a $$L(\alpha \mid \mathcal{D}) = P(\mathcal{D} \mid \alpha),$$9: half of the realizations give $\alpha$0, and the remainder follow a $\alpha$1 distribution with one degree of freedom (Anderson et al., 2015). This boundary effect is central whenever CLIP is used for positive-definite signal strengths.

Profile likelihood should also be distinguished from Bayesian marginalisation. In “Evidence for extra radiation? Profile likelihood versus Bayesian posterior” (Hamann, 2011), the profile likelihood for a parameter $\alpha$2 is defined by maximising over nuisance directions, $\alpha$3, whereas the marginalised posterior integrates over them. The paper further warns that intervals derived from $\alpha$4 may not have the desired frequentist coverage properties if the profile likelihood is not Gaussian (Hamann, 2011).

3. Relation to stacking methods

The central methodological contrast in (Anderson et al., 2015) is between joint likelihood and stacking. In data stacking, one merges the data and evaluates a single likelihood on the union of datasets; in joint likelihood, each dataset preserves its own model and nuisance structure. This difference is consequential whenever datasets have different backgrounds, exposures, or systematic treatments.

Approach	Construction	Characteristics
Joint likelihood	$\alpha$5	Dataset-specific nuisance parameters retained
Simple data stacking	Likelihood on $\alpha$6	Different S/B ratios and uncertainties are smeared
Residual stacking	$\alpha$7	Usually a more viable alternative than simple stacking

The paper’s discussion is explicit about the drawbacks of naive stacking: it mixes datasets with different backgrounds, exposures, and systematics into a single effective dataset; dataset-specific nuisance parameters become harder or impossible to model precisely in a single pooled likelihood; and poor-quality or low signal-to-noise data can get “washed out” by large but less constraining data (Anderson et al., 2015). By contrast, in the joint likelihood each dataset retains its own $\alpha$8, and nuisance parameters can differ in functional form, probability distribution, or prior.

For simple models, residual stacking can match joint likelihood in statistical performance, but the paper emphasizes that joint likelihood is easier to generalize to more complex nuisance models, including non-Gaussian or constrained nuisance treatments (Anderson et al., 2015). A plausible implication is that CLIP is most useful not merely because of asymptotic efficiency, but because it preserves model heterogeneity across constituent datasets.

4. Posterior-based constraints and volume effects

The comparison between profile-based and posterior-based inference becomes especially important in high-dimensional problems. In the cosmological analysis of $\alpha$9 in (Hamann, 2011), the marginalised posterior preferred higher values than the profile likelihood when using WMAP7+ACT alone: the posterior mode was approximately $$\mathcal{L}_d(\boldsymbol{\mu}, \boldsymbol{\theta}_d \mid \mathcal{D}_d) = \prod_k \frac{\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)^{n_{d,k}} e^{-\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)}}{n_{d,k}!},$$0, while the profile mode was $$\mathcal{L}_d(\boldsymbol{\mu}, \boldsymbol{\theta}_d \mid \mathcal{D}_d) = \prod_k \frac{\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)^{n_{d,k}} e^{-\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)}}{n_{d,k}!},$$1. When HST information was added, the discrepancy became small compared to the uncertainty, with posterior modes around $$\mathcal{L}_d(\boldsymbol{\mu}, \boldsymbol{\theta}_d \mid \mathcal{D}_d) = \prod_k \frac{\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)^{n_{d,k}} e^{-\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)}}{n_{d,k}!},$$2–$$\mathcal{L}_d(\boldsymbol{\mu}, \boldsymbol{\theta}_d \mid \mathcal{D}_d) = \prod_k \frac{\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)^{n_{d,k}} e^{-\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)}}{n_{d,k}!},$$3 and a profile mode of $$\mathcal{L}_d(\boldsymbol{\mu}, \boldsymbol{\theta}_d \mid \mathcal{D}_d) = \prod_k \frac{\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)^{n_{d,k}} e^{-\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)}}{n_{d,k}!},$$4 (Hamann, 2011).

The paper traces this discrepancy to a volume effect. The CMB constrains the redshift of matter-radiation equality,

$$\mathcal{L}_d(\boldsymbol{\mu}, \boldsymbol{\theta}_d \mid \mathcal{D}_d) = \prod_k \frac{\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)^{n_{d,k}} e^{-\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)}}{n_{d,k}!},$$5

more directly than it constrains $$\mathcal{L}_d(\boldsymbol{\mu}, \boldsymbol{\theta}_d \mid \mathcal{D}_d) = \prod_k \frac{\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)^{n_{d,k}} e^{-\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)}}{n_{d,k}!},$$6 itself (Hamann, 2011). Because the width of slices in other parameters grows with $$\mathcal{L}_d(\boldsymbol{\mu}, \boldsymbol{\theta}_d \mid \mathcal{D}_d) = \prod_k \frac{\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)^{n_{d,k}} e^{-\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)}}{n_{d,k}!},$$7, marginalisation accumulates more posterior volume at larger $$\mathcal{L}_d(\boldsymbol{\mu}, \boldsymbol{\theta}_d \mid \mathcal{D}_d) = \prod_k \frac{\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)^{n_{d,k}} e^{-\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)}}{n_{d,k}!},$$8, whereas the profile likelihood follows only the ridge of maximum likelihood. In the idealised Gaussian case, profiling and marginalisation would yield the same one-dimensional Gaussian shape; their disagreement is therefore a diagnostic of non-Gaussianity and $$\mathcal{L}_d(\boldsymbol{\mu}, \boldsymbol{\theta}_d \mid \mathcal{D}_d) = \prod_k \frac{\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)^{n_{d,k}} e^{-\lambda_{d,k}(\boldsymbol{\mu}, \boldsymbol{\theta}_d)}}{n_{d,k}!},$$9-dependent volume.

For CLIP-style analyses, the direct lesson is methodological rather than domain-specific: combine at the likelihood level, then inspect both profile and posterior summaries when degeneracies are strong. This suggests that a discrepancy between a combined profile and a marginalised posterior is not necessarily a contradiction; it can instead reveal the geometry of the likelihood in nuisance directions (Hamann, 2011).

5. Toy-model behavior and astrophysical use

The toy Monte Carlo study in (Anderson et al., 2015) provides the most direct empirical characterization of the joint-likelihood profile. In the one-bin model,

$\boldsymbol{\mu}$0

$\boldsymbol{\mu}$1 is a shared signal scale factor, $\boldsymbol{\mu}$2 is the known signal expectation per unit $\boldsymbol{\mu}$3, and $\boldsymbol{\mu}$4 is a background nuisance parameter for dataset $\boldsymbol{\mu}$5. The corresponding joint likelihood is

$\boldsymbol{\mu}$6

Using two datasets with signal-to-background ratios $\boldsymbol{\mu}$7 and $\boldsymbol{\mu}$8, the study found that all methods showed coverage at the nominal level, that limits improved approximately as $\boldsymbol{\mu}$9, that $\boldsymbol{\theta}_d$0 matched the joint likelihood in performance, and that simple $\boldsymbol{\theta}_d$1 was less constraining because it threw away information (Anderson et al., 2015).

The same study examined the scaling of upper limits with the number of datasets $\boldsymbol{\theta}_d$2. In a moderate-background case with $\boldsymbol{\theta}_d$3 and $\boldsymbol{\theta}_d$4 total events per set, $\boldsymbol{\theta}_d$5 CL upper limits on $\boldsymbol{\theta}_d$6 improved as $\boldsymbol{\theta}_d$7. In a very low-background regime with $\boldsymbol{\theta}_d$8, $\boldsymbol{\theta}_d$9, and the constraint $N$0, the upper limit scaled approximately as $N$1, because the log-likelihood becomes approximately linear near $N$2 and the slope adds across datasets (Anderson et al., 2015). The analytic scaling given in the paper is

$N$3

Applications cited in the paper include galaxy cluster emission searches, large extra dimensions, and dark matter searches in dwarf spheroidal galaxies with Fermi-LAT (Anderson et al., 2015). For the dwarf-spheroidal case, the structure remains the same but with many bins and more complicated per-target backgrounds. The paper’s account identifies dwarf spheroidals as independent observations of the same dark-matter signal parameter, making the joint likelihood with profiling the canonical way to combine them (Anderson et al., 2015).

6. Assumptions, caveats, and uncertainty propagation

The most important assumption in CLIP is that the datasets are disjoint and independent. The product over datasets assumes no overlapping events and no shared noise realization. The overlap study in (Anderson et al., 2015) shows that if datasets overlap, the joint likelihood double counts information, inflates TS values, and distorts null distributions; as overlap between two equivalent datasets increases, TS increases roughly in proportion to the overlap, and at $N$4 overlap the null TS distribution becomes skewed upward (Anderson et al., 2015). This leads to over-estimation of significance and, under certain criteria, increased type II errors.

A second caveat concerns asymptotics. The use of $N$5 or Chernoff limits assumes large enough counts and regular models; when those conditions fail, TS distributions must be obtained numerically, for example via Monte Carlo (Anderson et al., 2015). A third caveat concerns nuisance modeling: joint likelihood is flexible precisely because it can incorporate non-Gaussian, bounded, or otherwise complicated nuisance models, but that flexibility does not remove the need for correct specification.

For combined analyses that propagate fitted results into further combinations, the decomposition of uncertainty is itself part of the CLIP problem. “Uncertainty components in profile likelihood fits” (Pinto et al., 2023) argues that the routinely quoted “impacts” of nuisance parameters are not adequate for such applications: they are not variance components, they do not add in quadrature to the total uncertainty, and using them as building blocks in combinations leads to inconsistent results. The paper instead defines uncertainty components that satisfy

$N$6

and, in the Gaussian linear regime, derives the corresponding covariance decomposition,

$N$7

$N$8

In the nuisance-parameter representation, the contribution of systematic source $N$9 to the uncertainty of parameter of interest $\mathcal{D}_d$0 is, in the Gaussian regime, the absolute value of the POI–NP covariance, $\mathcal{D}_d$1 (Pinto et al., 2023). For non-Gaussian profile-likelihood fits, the same paper proposes shifted-observable or Monte Carlo procedures that preserve the interpretation of each component as variance induced by a specific fluctuating input.

Taken together, these results define the scope of CLIP with some precision. It is a likelihood-level combination method for common parameters across disjoint datasets; it derives inference from profile likelihood ratios rather than marginalisation; it is usually preferable to naive data stacking when nuisance structures differ; and, when results are propagated into higher-level combinations, it requires true uncertainty components rather than impact-based surrogates [(Anderson et al., 2015); (Hamann, 2011); (Pinto et al., 2023)].