5 Specification Curves
The third and fourth verbs summarise and visualise the distribution of estimates. A specification curve is the natural first picture.
5.1 Reading the curve
spec_curve() sorts the estimates from smallest to largest, draws each with its confidence interval, and colours points by whether the interval excludes zero.
spec_curve(results)Three features carry most of the message: where the mass of estimates sits, how wide it spreads, and what share crosses the null. A curve that stays on one side of zero signals a finding robust to the choices in the grid. A curve that straddles zero, with many intervals covering it, signals an association that depends on specification.
5.2 Ordering and annotation
Sorting by the estimate makes central tendency and spread legible at a glance. An analyst can also sort by a specific choice to see whether one decision drives the pattern, for example whether the largest estimates share a particular adjustment set. The companion volcano in the next chapter makes that dependence explicit by colouring on model size.
5.3 Inference across specifications
The curve describes a distribution, and it can also support a test. Simonsohn, Simmons, and Nelson propose resampling under the null to ask whether the observed curve is more extreme than chance would produce (Simonsohn et al. 2020). That inferential layer sits on top of the descriptive curve and is a reasonable extension once the descriptive picture is in hand.