Epidemiology

Screening

Applied at the population level among asymptomatic individuals.
Screening should only be done if there are effective interventions among presymptomatic individuals. Else it may be worth waiting for symptoms.
An individual who does not actually have the disease but is predicted to have it will be enrolled into the healthcare system unnecessarily.
Disease prevalence should be expected to be much lower than in diagnostic settings - remember issues with PPV from Demonstration 1.
If screening has any side effects, than many individuals may be exposed to them.

Screening leads to early detection - early detection does not have any benefit.
Two patients die at the same time, but one was screened and cancer detected earlier.
The first individual is followed for longer.
Lead time bias: when length of follow-up makes you conclude that screening is useful: "Individuals lived longer due to screening".
Lead time bias: when you attribute benefit to simply observing people.

Same number of fast and slow growing tumors over a period of time.
By chance, screening will find more slow growing tumors.
People with fast growing cancers will be diagnosed once symptoms are detected (or postmortem).
Length time bias: concluding that screening has an effect on cancer progress and risk of death when in fact slower growing cancers are oversampled by screening.

Predicitive assessments requires us to know the true value (Y+ or Y-).
If no gold standard exists which is efficient in classifying individuals then this adds further uncertainty to our estimates.

Predictions based on data for which there is systematic missingness - findings are biased.
Cannot generalise as easily, since datapoints collected are not from a random sample.
If we continuously screen, but the screening oversamples individuals with high risk, then we create a selection bias and PPV may be inflated.

Even with a perfectly random sample from the correct population we may still be measuring the wrong thing.
Different measurement methods results in different Y+ and Y- classifications, biasing the estimate of prevalence.
Not always a problem (!) - The link between cancer and smoking is not dependent on the person reporting smoking 20 or 25 cigarettes per day.
Smoking cessation interventions targeting "smokers" may be helpful regardless of the exact number of cigarettes smoked.