如何对随时间变化的离散年龄分布建模？含贝叶斯方法需求

Great question—you’re right that modeling each age in isolation misses the big picture of how the entire distribution shifts over time. Bayesian approaches are ideal here because they let you capture smooth, structured changes in the distribution while quantifying uncertainty. Here are practical, actionable methods tailored to your discrete age data:

1. Hierarchical Bayesian Dirichlet Models #

Since age groups are categorical (discrete), a Dirichlet-multinomial framework is a natural fit. This models the proportion of individuals in each age group at time t as a Multinomial distribution, with the underlying probabilities governed by a Dirichlet prior that evolves over time.

• How to implement it:

  • For each time point t, define the probability vector θ_t (one value per age group) such that θ_t ~ Dirichlet(α_t), where α_t are concentration parameters that control the spread of the distribution.
  • To model time variation, make α_t a function of time. You can use:
    • A linear trend: α_t = α₀ + β*t (simple, works if changes are steady)
    • A random walk: α_t = α_{t-1} + ε_t (captures smooth, unstructured changes)
    • Splines: α_t = spline(t) (flexible for non-linear shifts, like the move from 80 to 50-70 you observed)
  • The observed age counts at time t then follow counts_t ~ Multinomial(N_t, θ_t), where N_t is the total number of individuals at time t.

• Example snippet (conceptual Stan code):

  data {
    int<lower=1> T; // number of time points
    int<lower=1> K; // number of age groups
    int counts[T, K]; // age counts per time point
    int N[T]; // total individuals per time point
  }
  parameters {
    vector[K] alpha0; // initial concentration parameters
    vector[K] beta; // time trend for concentrations
  }
  model {
    alpha0 ~ normal(1, 1);
    beta ~ normal(0, 0.5);
    for (t in 1:T) {
      vector[K] alpha_t = alpha0 + beta * t;
      counts[t] ~ multinomial( dirichlet(alpha_t) );
    }
  }
  

2. Bayesian Nonparametric Time-Varying Mixtures #

If you don’t want to constrain the distribution to a specific form (like Dirichlet), time-varying Dirichlet Process Mixtures (DPMs) let the data dictate the shape of the evolving distribution.

• The idea: Each time point’s age distribution is a mixture of components, where the number of components and their parameters change smoothly over time. You can model the mixture weights and component parameters with Gaussian Processes (GPs) to enforce temporal smoothness.
• This is great for capturing complex shifts—like the sudden drop in concentration at 80 and spread to 50-70—without pre-defining the structure.

3. Bayesian Generalized Linear Mixed Models (GLMMs) for Age Counts #

Treat each age group’s count as a response variable, and model how it changes with time while accounting for shared variability across age groups.

• Model structure (log-link for count data):

  log(count_ij) = β₀ + β₁*age_i + β₂*time_j + β₃*(age_i * time_j) + u_i
  

  • count_ij: Number of individuals in age group i at time j
  • β₃*(age_i * time_j): The critical interaction term—this captures how the effect of time varies across age groups (exactly the shift you’re observing)
  • u_i: Random effect for age groups, to account for unmeasured similarities between adjacent ages

• Bayesian twist: Assign priors to all coefficients (e.g., β ~ Normal(0, 1), u_i ~ Normal(0, σ_u)), which lets you quantify uncertainty in how each age group’s count changes over time. You can then derive the full distribution at each time point by normalizing the predicted counts.

Bonus: Validate with Visualization #

After fitting any of these models, pair the results with your ggridges plots to validate:

• Plot predicted distribution ridges alongside observed ones to check if the model captures the shift from 80 to 50-70.
• Create a heatmap of predicted probabilities (age groups on y-axis, time on x-axis) to visualize how each age group’s share changes over time.

内容的提问来源于stack exchange，提问作者bdetweiler