基于spatstat构建Gibbs点过程模型时parres偏残差图拟合异常的技术咨询

Hey there, let's break down your problem and walk through possible fixes. First off: seeing the fitted curve fall outside the confidence interval in your partial residual plot is not ideal, but it's definitely a common issue that crops up when point process models don't fully capture the relationship between your covariates and the point pattern. Let's unpack why this might be happening and how to address it.

Is this result normal/common? #

Short answer: It's common enough to signal a model fit issue, but not "normal" in the sense that it indicates your current model isn't perfectly capturing the data's structure. Partial residual plots are meant to show how well your model's specified covariate relationship matches the observed residual pattern—when the fitted line strays outside the CI, it means the model's assumed functional form (or other components) aren't aligning with the data.

Possible Causes & Fixes #

1. Incorrect functional form for the FB0lin covariate #

You're using a linear term for FB0lin, but the true relationship between habitat suitability and bird occurrence might be nonlinear. Log-transforming the covariate doesn't always guarantee a linear response, especially if the underlying relationship is curved (e.g., saturating at high suitability values).

Fixes to try:

• Test a nonlinear functional form using splines. spatstat works with mgcv's spline terms, so you can adjust your model like this:

  library(mgcv)
  # Fit model with a smooth spline for FB0lin
  mod.ois.echlin_spline <- ppm(ois.ppp, ~ s(FB0lin) + Nbdate + pAccess, 
                               interaction = Geyer(r=401, sat=9), eps=100)
  # Generate and plot updated partial residuals
  res.FB0.echlin_spline <- parres(mod.ois.echlin_spline, covariate="FB0lin")
  plot(res.FB0.echlin_spline, main="FB0 LinCost (Spline Term)", legend=FALSE)
  

• Double-check your normalization/log-transformation steps. Did you handle zero values properly before log-transforming? If FB0lin has zeros, adding a small offset (like log(FB0lin + 1)) might be necessary to avoid bias.

2. Suboptimal Geyer interaction parameters #

You used profilepl to select r=401 and sat=9, but it's worth verifying if these parameters are truly the best fit. If the profile likelihood surface is flat or the chosen parameters are at the edge of your search range, the model might not be capturing spatial interactions correctly—which can distort covariate effects in partial residuals.

Fixes to try:

• Plot the profile likelihood results to confirm your parameter choice:

  plot(term.interlin)
  

  Look for clear peaks in the likelihood surface. If the optimal r or sat is near the bounds of your seq() or 1:40 range, expand the search space (e.g., seq(1, 1501, by=50) for r, 1:60 for sat) and re-run profilepl.
• Consider alternative spatial interaction models. Geyer is great for saturated clustering, but maybe a Strauss process or a pairwise interaction model fits your data better. Test this by swapping the interaction argument:

  # Try a Strauss model as a comparison
  mod.ois.echlin_strauss <- ppm(ois.ppp, ~ FB0lin + Nbdate + pAccess, 
                                interaction = Strauss(r=401), eps=100)
  

3. Inappropriate resolution for virtual points (eps) #

The eps=100 creates a 100x100m grid of dummy points, but this resolution might not match the scale of your data or covariates. If the grid is too coarse, the model might miss fine-scale variation; if too fine, it could introduce unnecessary noise.

Fixes to try:

• Test different eps values (e.g., 50m or 150m) and see if the partial residual plot improves:

  mod.ois.echlin_eps50 <- ppm(ois.ppp, ~ FB0lin + Nbdate + pAccess, 
                               interaction = Geyer(r=401, sat=9), eps=50)
  res.FB0.echlin_eps50 <- parres(mod.ois.echlin_eps50, covariate="FB0lin")
  plot(res.FB0.echlin_eps50, main="FB0 LinCost (eps=50m)", legend=FALSE)
  

4. Unaccounted sampling bias or missing covariates #

You included Nbdate (sampling effort) and pAccess (可达性), but these might not fully control for sampling bias. For example, if sampling effort varies in a way that's correlated with FB0lin but isn't captured by Nbdate, this could skew the partial residuals.

Fixes to try:

• Check if Nbdate or pAccess need transformation (e.g., log-transformation for effort, which often has a right-skewed distribution).
• Run additional model diagnostics to rule out residual spatial patterns:

  # Generate Pearson residuals and plot them
  res_pearson <- residuals(mod.ois.echlin, type="pearson")
  plot(res_pearson, main="Pearson Residuals of the Model")
  

  If you see clustered or systematic residual patterns, it means the model is missing key spatial or covariate effects.

Final Notes #

Start with the simplest fixes first (like testing a spline for FB0lin or verifying Geyer parameters) before moving to more complex adjustments. Partial residual plots are a great diagnostic tool, but they work best when paired with other checks like residual spatial plots and likelihood profile analyses.

内容的提问来源于stack exchange，提问作者Elie Morin