Diagnose Precision Index Variation for Scale Estimation

The scale submodel log(sigma) ~ log_cv requires meaningful spread in log_cv across observations. If all observations have nearly identical pcov (e.g., all in the well-determined midrange), beta1 is not identifiable and the model falls back to intercept-only sigma (uniform weights).

Usage

diagnose_cv(df)

Arguments

df

Data frame with cv_i and log_cv columns, typically from prepare_cv().

Value

Named list:

n_finite

Number of observations with finite cv_i > 0.

cv_min, cv_median, cv_max

Summary statistics of cv_i.

log_cv_sd

Standard deviation of log_cv.

log_cv_range

Range (max - min) of log_cv.

use_cv_slope

Logical: TRUE if sd(log_cv) >= 0.05.

message

Human-readable summary of the diagnosis.

Details

The threshold is sd(log_cv) >= 0.05. Below this, there is not enough contrast in measurement precision across the concentration range to distinguish differential weighting from uniform weighting.

Examples

data(example_assay)
dat_sub <- example_assay[example_assay$antigen == "prn" &
                         example_assay$feature == "IgG1", ]
d <- prepare_cv(dat_sub, pcov_col = "pcov")
cv_diag <- diagnose_cv(d)
cat(cv_diag$message, "\n")
#> OK: sd(log_cv) = 0.487; beta1 identifiable from 506 observations 
cat("sd(log_cv) =", cv_diag$log_cv_sd, "\n")
#> sd(log_cv) = 0.4872 
cat("use slope?", cv_diag$use_cv_slope, "\n")
#> use slope? TRUE