Do what I mean

04 August 13. [link] PDF version

This interstitial post gives some discussion of the design goals of Apophenia. I'll refer to it a few times when I get to things that some systems do automatically that Apophenia insists that users do manually.

The key point is the difference between interactive and script-oriented design. It's increasingly the case that people expect to be able to use one language both interactively and as an unsupervised and reliable moving part in a bigger machine, but the two goals are in some sense at odds. As an interactive user, I want the machine to do what I mean (DWIM), not force me to explicitly enumerate tædium and minutiæ; as a reliable component, I don't want the system to make guesses at what should happen, because sometimes those guesses will be wrong.

Why your R script broke

At work, people take me as an authority on R, even though there are many people in the building far better with it than I. So I get to hear a lot of complaints about R scripts that don't work. Passing on the package management, version mismatches, and other sort of environment issues, here are the top three reasons a script worked yesterday but didn't today:

All three of these are situations where the system did something for the user without the user's knowledge, allegedly to help the user out, and what it did was too situation-specific to generalize. To give you a clearer image of the first two cases, here is a typical story:

My point here is not to kvetch, and please do not send me defensive emails explaining to me the secret trick that my coworkers missed to making automatic and implicit factor conversions reliable. My point is to show that R was designed for interactive use first, and uploading to the server for batch processing second. The auto-casting DWIM features are a clear indication of those priorities, and really exactly what you want if you have a single data set in front of you and can check and correct any missteps as they happen. And if you think I'm reading intent into details of syntax, get a copy of Becker, Chambers, and Wilks's The New S Language, and read about how much the designers of S (of which R is an implementation) pushed interactivity and reacting to output. The thought of running S headless only gets a half-page discussion on p 57: “There are occasions when interaction with S is unnecessary, e.g., when a large computation is to be carried out from a source file.''

Apophenia's intent is to allow kickass narrative model development (if you've been reading along, you know what I mean by that now) and to provide a statistics and data processing facility for the open source stack. You should be able to write a statistic-calculating function, post it on GitHub, and be reasonably confident that people with data nothing like yours don't have to bend over backward to use the function. Replicable processing is the priority, but we develop those reliable programs interactively, so the UI has to not hurt.

Thanks for reading. Tomorrow I'll post the logit regression I'd promised, and now you know why it will have a line explicitly casting a text variable to numeric values.

[Previous entry: "A test for independence"]
[Next entry: "A Logit and some elasticities"]