medicalRecords.RdAn entirely artificially constructed data set and context designed for classroom discussion and analysis.
A medical data mining context is given in detail below. In light of the context, interesting scientific questions will arise as to the data collection, and how the results should, or should not, be interpreted. It should also raise questions on what might be done in any follow up studies.
Instructors might choose to invent their own context.
A data frame with 16 rows and 5 variables (providing the counts for a 2x2x2x2 contingency table).
A two level factor recording one of two age groups: "20-39" or "40-59".
A two level factor recording sex: "Male" or "Female".
A two level factor recording the treatment received: "A" or "B".
A two level factor recording patient outcome after treatment: "Recovered" or "Died".
The frequency count of patients having that combination of factors.
One fictional context (constructed in March 2020) for this data set is given below (in the PPDAC style of Mackay and Oldford (2000)).
Problem:
A disease epidemic has broken out in the population of some country. It is thought that adults under the age of 60 appear to be particularly vulnerable. Both men and women contract the disease and need to be treated. Those who go untreated die within 5 days of contracting the disease.
The medical community has tried two quite different approaches to treat patients having the disease -- call these `Treatment A` and `Treatment B`. For the health of the country, it is important to determine which of these two treatments is more effective.
Plan:
To investigate which is the better treatment, it was decided to mine the medical records from another country of those who had contracted the disease and had been treated with one of the two treatments. Patients treated with either A or B survive the disease and recover fully; some however still die.
Electronic medical records available from several of the more populous districts are accessible. These can be searched to provide records from patients that have received treatment. It was decided that there should be the same number of records drawn for each treatment.
Moreover, concern was raised that the investigation have gender balance (i.e. equal numbers of males and females). So, to make sure that both sexes were equally represented, it was also decided that the number of female patients would be the same as the number of male patients.
Finally, it was desirable to detect even small differences in success rates of the two treatments since small differences could mean many more lives being saved. A sample size of about n = 3,000 was decided on.
Records would be collected until 3,000 were found, 1500 of which were treated with `A`, 1500 with `B`, and there were equal numbers of males and females in the study.
Data:
In this stage, the plan is executed. Instead of 1500 records of treatment `A` and `B`, 1600 of each were found. The number of males and females was kept equal (now 1600 of each sex).
The process was to search the records in order, selecting those first encountered to get 1600 for each treatment and 1600 of each sex. Many records might be discarded whenever one quota was met and the search continued to meet the other quotas. It was also noticed that the patient's age was available for each record, so that the effect of treatment on younger and older adults might also be considered.
The counts which fell into the various categories were assembled into the data presented here.
R.J. MacKay and R.W. Oldford 2000, `Scientific Method, Statistical Method, and the Speed of Light`, Statistical Science, Volume 15, No. 3, pp. 254-278. <doi:10.1214/ss/1009212817>
R.W. Oldford