Visual search of regression results

This article is about ideas on ways to search the MetaMetriks-database. It derives design-ideas from WhereDoesMyMoneyGo.

Top layer: Choose Y
The mosaic of coloured squares shows the dependent variables contained in the database, the size of each square represent the quantity of regressions stored for the respective Y see Figure 1.

Second layer: Choose X
Having chosen the dependent variable the user is presented with a similar mosaic for the regressors that this dependent variable has been explained by.

Third layer: Choose basic functional form
Several mosaics next to each other; one for level regressions, changes, logs, … Within each mosaic (within each basic functional form), squares show the breakdown of results: significant positive, insignificant positive, significant negative, insignificant negative

Fourth layer: Equations view
Clicking on a result (e.g. significant positive) a list of the relevant regression results is shown. To not overflow the reader with the standard regression tables, visualisations of regression results are used instead: The left-hand side of Figure 2 shows a medium-sized square with the full labelling of the y-variable (no codes to separately look up) followed by a “=” and a second square labelled with the x-variable chosen. The size of the second square is scaled in proportion to the first by the coefficient estimate. So a standard regression result “y= ...+2.41X+...” would be visualised by a sequence of squares where the square representing the X-variable would be 2.41-times larger than that of the dependent variable. Numbers could be provided when holding the mouse over a square but otherwise the picture would not be clogged up by coefficients. After the square for the x-variable those of coefficients would follow, separated by plus and minus signs (depending on coefficient estimates) and lastly by a square for the intercept.

The visualisation should could also depict some other aspects of the equation estimated. Two-stage regressions could be represented by a vertical chain of squares, starting from the 	item for which the first-stage regression. Also multi-collinearity could be visualised. Let's assume that in the coming years it becomes 	more standard to openly provide one's dataset. Then our application could automatically calculate covariances and use this information when displaying the equation. The squares 	that we use for representing regressors could be presented adjacent and close to each other where covariances are high.

Direct mode
2. The frontpage should feature also a way to directly get to the regression view; as in just selecting variables from a drop-down list.