Why Parameter Counts Matter
AIC and BIC balance fit against model complexity. Their penalty depends on \(k\), the number of free parameters estimated for the candidate model. Counting incorrectly can change the comparison: omitting an estimated variance makes a model appear too simple, while counting a derived quantity as an extra parameter makes it appear too complex.
The relevant count is not the number of symbols in a model description, nor the number of values reported by software. It is the dimension of the parameter space that is independently fitted from the data. A fixed input contributes no fitted parameter. A quantity determined by other fitted parameters contributes no additional parameter.
For the usual forms \( \mathrm{AIC}=-2\ell(\hat{\theta})+2k \) and \( \mathrm{BIC}=-2\ell(\hat{\theta})+k\ln n \), \( \ell(\hat{\theta}) \) is the maximized log-likelihood and \(n\) is the sample size under the convention being used. This tutorial focuses on determining \(k\); the same count is used in either criterion. For AICc, \(k\) is also the number of estimated parameters in the correction.
A Consistent Counting Procedure
First list the model's unknown parameters. Then identify which are fixed by the problem, which are estimated independently, and which are constrained or derived. This is especially important when the model includes regression coefficients, dispersion or scale parameters, component-specific quantities, or restrictions tying parameters together.
List the unknown quantities that enter the likelihood, such as regression coefficients, a variance, a mixture weight, or a severity parameter.
A specified deductible, known exposure offset, or externally fixed variance is not estimated from the likelihood and does not add to \(k\).
An equality constraint can reduce the number of free parameters. A constraint that only defines a derived quantity does not add another one.
Include all fitted parameters in that likelihood, including nuisance parameters such as an error variance, even if they are not the main subject of the model.
A regression model with \(p\) freely estimated regression coefficients and an estimated residual variance has \(k=p+1\). If the variance is known and fixed, it has \(k=p\). In a model with several groups, a shared parameter is counted once; distinct group-specific parameters are counted separately.
Regression Coefficients and Scale Parameters
In normal linear regression with fixed predictors, suppose the response has conditional mean \( \beta_0+\beta_1x_1+\cdots+\beta_rx_r \) and independent errors distributed as \(N(0,\sigma^2)\). If all \(r+1\) coefficients and \(\sigma^2\) are estimated, then \(k=r+2\): the intercept, the \(r\) slopes, and the variance. The variance is a fitted likelihood parameter, even though it describes scatter rather than the conditional mean.
Worked Example: Counting Parameters in Two Regression Fits
A hypothetical insurer models log claim severity using 40 independent observations and fixed predictor values. Both candidate models assume independent normal errors with an estimated common variance. Model A has an intercept and two slopes; its maximized log-likelihood is \(-81.2\). Model B has an intercept and one slope; its maximized log-likelihood is \(-83.0\). Find each model's \(k\), AIC, and BIC.
Method and count. Count the regression coefficients and the error variance. Model A has \(1+2=3\) regression coefficients and one variance, so \(k_A=4\). Model B has \(1+1=2\) coefficients and one variance, so \(k_B=3\). The 40 predictor rows are fixed design information, not additional parameters.
For Model A, \(-2\ell_A=-2(-81.2)=162.4\). Therefore, \( \mathrm{AIC}_A=162.4+2(4)=170.4 \). Since \(\ln(40)\approx3.688879\), \( \mathrm{BIC}_A=162.4+4(3.688879)=177.155516 \), or 177.156 to three decimals.
For Model B, \(-2\ell_B=-2(-83.0)=166.0\). Therefore, \( \mathrm{AIC}_B=166.0+2(3)=172.0 \), and \( \mathrm{BIC}_B=166.0+3(3.688879)=177.066637 \), or 177.067 to three decimals.
Check and trap avoided. Model A has one additional parameter and a higher maximized likelihood, but its AIC is lower by \(172.0-170.4=1.6\). Its BIC is slightly higher. Omitting the fitted variance would incorrectly use \(k_A=3\) and \(k_B=2\), reducing both penalties by 2 and misrepresenting the models' fitted parameter counts.
Mixture Models and Probability Constraints
A mixture model combines component distributions using weights. For a \(K\)-component mixture, the \(K\) weights sum to 1. If they are all estimated, there are \(K-1\) independent weights: once \(K-1\) are chosen, the last is determined. If every component has \(q\) independently estimated distribution parameters, the usual count is \(Kq+(K-1)\). This assumes no parameter sharing or further constraints.
For example, a two-component normal mixture with separately estimated component means, a shared variance, and an estimated mixing probability has \(2+1+1=4\) parameters. The two means are fitted separately, the common variance is counted once, and the two probabilities contribute one free weight because they sum to 1. Merely labeling the components differently does not reduce the count. An ordering convention such as \(\mu_1<\mu_2\) can distinguish labels but does not ordinarily remove a continuous degree of freedom.
Worked Example: A Two-Component Severity Mixture
A hypothetical sample contains 60 positive claim amounts. Compare a single normal model for a transformed severity with a two-component normal mixture. The mixture has separate component means, one shared variance, and an estimated mixing probability \(p\), with component weights \(p\) and \(1-p\). Both models estimate all their distribution parameters. Suppose their maximized log-likelihoods are \(-51.2\) for the single normal model and \(-46.7\) for the mixture. Calculate \(k\) and AIC for each.
Single normal model. Its fitted parameters are one mean and one variance, so \(k_1=2\). The fit term is \(-2(-51.2)=102.4\), giving \( \mathrm{AIC}_1=102.4+2(2)=106.4 \).
Mixture model. It estimates two means and one shared variance. Its two weights add to 1, so only one weight is free. Thus \(k_2=2+1+1=4\). The fit term is \(-2(-46.7)=93.4\), giving \( \mathrm{AIC}_2=93.4+2(4)=101.4 \).
Check and trap avoided. The mixture's AIC is \(106.4-101.4=5.0\) lower. Counting both weights as independent would give \(k=5\) and AIC \(103.4\), an incorrect penalty: after choosing \(p\), the other weight is exactly \(1-p\). Conversely, counting only one mean would also be wrong because the component means are separately estimated.
Constraints, Fixed Inputs, and Derived Quantities
An independent equality constraint reduces the number of free parameters because it restricts the parameter space to a lower-dimensional set. If a model begins with \(m\) freely estimated quantities and imposes \(r\) independent equality constraints, the usual count is \(m-r\). This rule assumes the constraints are genuinely independent; redundant equations do not each reduce the count again.
By contrast, a fixed quantity is not estimated at all, and a derived quantity is not an extra degree of freedom. A predicted mean at a given predictor value is calculated from fitted coefficients; it is not a new parameter for every observation. A likelihood's normalization may depend on parameters already in the model, but that dependence does not itself create another parameter.
Inequality constraints need more care. A restriction such as \(p>0\) and \(p<1\) sets the allowed parameter range but ordinarily does not reduce its dimension: \(p\) remains one free parameter. If an estimate is forced onto a boundary, standard information-criterion and likelihood approximations may require additional caution; the boundary does not justify casually subtracting a parameter from the count.
Worked Example: Shared Exponential Scale Under an Equality Constraint
A hypothetical portfolio has two groups of positive losses. Group A records 3 and 3 losses; Group B records 3 and 5 losses. Assume losses are independent and exponentially distributed with mean parameter \(\theta\). Compare a model with a separate mean \(\theta_A,\theta_B\) against a constrained model with the common mean \(\theta_A=\theta_B=\theta\). Calculate the maximum log-likelihood, \(k\), and AIC for each. Use the exponential density \(f(x;\theta)=\theta^{-1}e^{-x/\theta}\), for \(x>0\).
Separate means. For \(n\) exponential observations with sum \(s\), the log-likelihood is \(\ell(\theta)=-n\ln\theta-s/\theta\). Differentiating gives \(-n/\theta+s/\theta^2=0\), so the maximum occurs at \(\hat{\theta}=s/n\).
Group A has \(n_A=2\) and \(s_A=3+3=6\), so \(\hat{\theta}_A=6/2=3\). Its maximized log-likelihood is \(-2\ln3-6/3=-2\ln3-2\approx-4.197225\). Group B has \(n_B=2\) and \(s_B=3+5=8\), so \(\hat{\theta}_B=8/2=4\). Its maximized log-likelihood is \(-2\ln4-8/4=-2\ln4-2\approx-4.772589\). The total is \(\ell_{\mathrm{sep}}\approx-8.969813\). There are two independently estimated means, so \(k_{\mathrm{sep}}=2\), and \( \mathrm{AIC}_{\mathrm{sep}}=-2(-8.969813)+2(2)=21.939627 \), or 21.940 to three decimals.
Common mean. Pooling the four observations gives \(n=4\) and \(s=3+3+3+5=14\). Therefore \(\hat{\theta}=14/4=3.5\), and \( \ell_{\mathrm{shared}}=-4\ln3.5-14/3.5=-4\ln3.5-4\approx-9.011052 \). The equality constraint leaves one free mean, so \(k_{\mathrm{shared}}=1\). Its AIC is \( -2(-9.011052)+2(1)=20.022104 \), or 20.022 to three decimals.
Check and trap avoided. The shared model has a slightly lower maximized likelihood, but its AIC is lower by about \(21.940-20.022=1.918\). Counting two means in the shared model would ignore the equality constraint. Counting zero parameters because the common mean is derived from the pooled data would also be wrong: it is still estimated from the observations.
Parameter-Count Justification
The dimension rule for mixture weights and equality constraints follows from how many coordinates can vary independently. For weights \(w_1,\ldots,w_K\), the condition \(w_1+\cdots+w_K=1\) determines \(w_K=1-\sum_{j=1}^{K-1}w_j\). Thus the weights have \(K-1\) free coordinates, provided the weights are not fixed or subject to additional independent restrictions.
Similarly, if three regression coefficients satisfy \(\beta_1+\beta_2+\beta_3=0\), then \(\beta_3=-\beta_1-\beta_2\). The likelihood can be parameterized by \(\beta_1,\beta_2\) rather than all three coefficients. It therefore has two free coefficient coordinates, not three. More generally, independent equality constraints remove one coordinate each because one parameter can be solved for in terms of the others. This is why the count follows the dimension of the allowed parameter space rather than the number of symbols originally written down.
Common Counting Traps
- Leaving out the variance. If an error variance or severity scale is estimated, it is a parameter even if the model is described mainly by its mean structure.
- Counting constrained quantities independently. Mixture weights summing to 1 contribute \(K-1\), not \(K\), when all weights are estimated.
- Counting fixed values. A known threshold, deductible, or offset does not add to \(k\). If that quantity is estimated instead, it generally does.
- Counting every reported output. Fitted values, residuals, and transformations calculated from the fitted parameters are not additional free parameters.
- Assuming every restriction reduces the count. An inequality or label-ordering convention usually restricts the parameter space without reducing its continuous dimension.
A reliable final check is to ask: if all but one of the quantities were held fixed, could this quantity still vary independently over an interval while remaining within the model? If yes, it is usually a free parameter, subject to the model's actual equality constraints. This counting discipline makes information-criterion comparisons consistent and is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.
Check Your Understanding
For each hypothetical model, determine the number of free parameters. Use the stated assumptions and count only quantities estimated from the data.
- A normal linear regression has an intercept, four slopes, and one estimated common error variance. How many parameters are counted? (Answer: 6.)
- A five-component mixture estimates two parameters per component and all five weights, which sum to 1. There is no parameter sharing. What is \(k\)? (Answer: \(5(2)+4=14\).)
- A model has four regression coefficients and an estimated variance, with one independent equality constraint tying two coefficients together. How many free parameters remain? (Answer: \(4+1-1=4\).)
- A two-group exponential model estimates a common mean, and uses a known fixed exposure multiplier for each group. How many parameters are counted? (Answer: 1; the shared mean is estimated, while the multipliers are fixed.)
- A three-component normal mixture estimates three means, three component-specific variances, and three weights summing to 1. The weights are otherwise unrestricted. What is \(k\)? (Answer: \(3+3+2=8\).)