From datasets to resultssets in Stata

In general, a Stata dataset, or a dataset in any other format, should contain one observation per thing, and data on attributes of things. For instance, in the medical sector, a dataset might have one observation per patient, and data on the patient’s baseline characteristics. Alternatively, a dataset might have one observation per visit to a health centre, and data on the state of the patient who made the visit. Usually, the variables in a dataset are either primary key variables, which identify the things corresponding to the observations, or non-key variables, which identify interesting attributes of those things. For instance, if there is one observation per patient, then one of the variables is usually a patient ID number, which identifies the observation uniquely. Or, if there is one observation per visit, and a patient may only make one visit per day, then the primary key variables are usually patient ID and visit date. Statisticians (and other data analysts) are typically provided with (or collect for themselves) a dataset with one observation per “experimental or observational unit”, where a unit may be a patient, a patientday, a country-year, a car model, or some other thing. However, they are typically paid to produce plots for presentation, or tables for publication. To do this, they really need datasets with one observation per plotted data point, or per Y -axis label, or per X-axis label, or per table row. Axis labels and table rows do not often correspond to the original units in the original dataset. For instance, in the auto data, shipped with official Stata, there is one observation per car model, and the primary key is the single variable make. Figure 1 gives confidence intervals from a regression model measuring differences in fuel efficiency (in miles per gallon) in cars from inside and outside the US, with various 1978 repair records, compared with a “reference car”, made by a US company and with a repair record of 3. It was created using the eclplot package, which creates confidence interval plots, and requires, as input, a dataset with one observation per confidence interval to be plotted and data on estimates and confidence limits. Table 1 gives the same data (plus P -values) as a table. It was created using the listtex package, which creates tables which can be cut and pasted into a TEX, LTEX, HTML or word processor document, and requires, as input, a dataset with one observation per table row. Both eclplot and listtex are downloadable from SSC, but neither can use directly as input the original auto data, with one observation per car model. Instead, resultsset-generating and resultsset-processing packages are used, taking as input the original auto data, and creating, as output, datasets that can be input to eclplot and listtex. This survey explains how to do this, using example do-files distributed with this document on the conference website at http://www.stata.com/support/meeting/10uk/ . These do-files will work under Stata 8 if the user has installed the required unofficial Stata packages, listed at the top of each do-file. These packages are downloadable from the SSC archive at http://ideas.repec.org/s/boc/bocode.html using the Stata ssc command. The full list of required packages comprises the resultsset-generating packages descsave, parmest, xcollapse and xcontract, together with the resultsset-processing packages eclplot, listtex, dsconcat, sencode, sdecode, factext, factmerg and ingap, and the estimation package somersd.