基于KFAS R包的含T与R矩阵的自定义状态空间模型估计问题

Hey there, let's work through this state space model estimation problem together! It sounds like you're hitting a tricky numerical error in the KFAS package while trying to implement the model from Modeling data revisions: Measurement error and dynamics of “...”, and the existing AR(1) transition equation guide didn't quite get you to your expected results. Let's break this down step by step:

1. First, Fix the "NA/Infinite System Matrices" Error #

That error points to issues in matrices other than Z (so your transition matrix T, disturbance covariance Q, or initial state covariance P0). Here's how to diagnose and fix it:

• Inspect your matrices directly: After building your SSModel object, use R's base functions to check for problematic values:

  # Replace with your actual model object
  my_ssm <- SSModel(...)
  # Check key matrices for NA/inf values
  str(my_ssm$T)
  summary(my_ssm$Q)
  str(my_ssm$P0)
  

• Watch for parameter feasibility: If you're estimating an AR(1) transition, make sure your initial guess for the AR coefficient stays within the stationary range (-1 < phi < 1). Setting it to 1 or outside this range can create singular matrices or infinite values.
• Use log-transforms for variances: When estimating covariance parameters, use log-transformed initial values to ensure they stay positive during optimization. This prevents negative variance estimates that would break your covariance matrices:

  # Example initial values for optimization
  init_params <- c(phi = 0.6, log_sigma_eta = 0, log_sigma_e = 0)
  

2. Align Your Model with the "Modeling Data Revisions" Framework #

That paper's core structure typically combines measurement error with a dynamic latent state (often an AR process). Let's make sure your SSModel matches this structure:

• Measurement Equation: Ensure you're explicitly modeling measurement error in H. For example, if you have observed data that's a noisy version of a latent true value:

  # y = observed data, alpha = latent true state
  ssm_rev_model <- SSModel(
    y ~ -1 + SSMcustom(
      T = matrix(phi), # AR(1) coefficient for latent state
      R = matrix(1),
      Q = matrix(exp(log_sigma_eta)), # Convert log variance back to positive
      P0 = matrix(exp(log_sigma_eta) / (1 - phi^2)), # Stationary AR(1) initial covariance
      a0 = 0
    ),
    H = matrix(exp(log_sigma_e)) # Measurement error covariance
  )
  

• Multivariate case (if using revised data): If you're working with both initial and revised observations, adjust your Z and H matrices to map each observation to the latent state correctly. For example, two observations sharing the same latent state but with different measurement errors would have a Z matrix like matrix(c(1,1), nrow=2) and a diagonal H matrix with two distinct variances.

3. Troubleshoot Output vs. Expected Results #

Once the matrix error is fixed, if your results still don't match expectations:

• Check for local optima: Try multiple sets of initial parameters—optimization algorithms like BFGS can get stuck in local minima if starting values are off.
• Validate the Kalman Filter/Smoother: Run the filter manually with KFAS::KFAS() and inspect the filtered/smoothed states. Compare these to your expected values to see if the discrepancy comes from the model structure or the estimation process.
• Verify model identification: Ensure your model isn't over- or under-identified. For example, if you're trying to estimate too many unconstrained parameters relative to your data, the results might be unstable or nonsensical.

If you can share a minimal reproducible example of your code and data, we could dig even deeper—but these steps should help you resolve the matrix error and align your model with the paper's framework.

内容的提问来源于stack exchange，提问作者m.ziembinski