Skip to contents

Diagnostic thresholds

Source: vignettes/articles/diagnostic-thresholds.Rmd
diagnostic-thresholds.Rmd

A cutoff turns a continuous test into a yes or a no. Where it sits decides how many people with the condition are missed and how many without it are sent on for more tests, and what a positive result means depends on how common the condition is where the test is used. ggdiagnostic() shows all of it at once, for any cutoff and any prevalence.

A first explorer

The Pima Indians diabetes data in MASS record plasma glucose two hours after a glucose load. A cutoff of 126 mg/dL is taken as prespecified here, and the test is imagined in a population where one person in ten has diabetes:

pima <- rbind(MASS::Pima.tr, MASS::Pima.te)
dx <- ggdiagnostic(type ~ glu, pima, cutoff = 126, prevalence = 0.1,
                   labels = c("No diabetes", "Diabetes"),
                   marker_label = "Plasma glucose (mg/dL)",
                   title = "Plasma glucose for diabetes")
dx

Drag the cutoff across the distributions, or use the slider above the plot. The shaded tails are the people who test positive, the point on the ROC curve and its crosshair of intervals move with it, and the grid of 1,000 people recolors: found, missed, false alarms and correctly cleared. The second slider sets the prevalence, which changes the predictive values and the grid but not sensitivity or specificity. The sentence above the plot says what the current setting does.

dx$accuracy
#>       measure  estimate     lower     upper
#> 1 sensitivity 0.6666667 0.5943294 0.7319232
#> 2 specificity 0.7690141 0.7224322 0.8098363
#> 3         ppv 0.2428181 0.2052492 0.2847992
#> 4         npv 0.9540513 0.9435977 0.9626441
#> 5 lr_positive 2.8861789 2.3243049 3.5838794
#> 6 lr_negative 0.4334554 0.3492499 0.5379633
dx$auc
#> $auc
#> [1] 0.7939763
#> 
#> $lower
#> [1] 0.753043
#> 
#> $upper
#> [1] 0.8349096

What to keep in mind

  • A cutoff chosen by looking at the same data, such as the one that maximizes Youden’s index, which is the default when cutoff is not given, looks better here than it will in new patients. The plot says so.
  • The prevalence in a case control sample is not the prevalence in practice. Set prevalence to the one that applies.
  • A test cutoff is not a treatment threshold: deciding who to treat also weighs the harms of treatment.

Intervals

Sensitivity and specificity have Wilson score intervals, the predictive values the logit intervals of Mercaldo and colleagues, the likelihood ratios the log method and the area under the curve DeLong’s. An interval that needs a count of zero is not shown.

Options

argument effect
direction whether "higher" or "lower" values point to the condition; "auto" by default
labels names for those without and with the condition
marker_label the marker’s axis label, with its unit
level the confidence level