Paired meta-analyses visualization / calculation

[christinabergmann]

Some MAs are pairs of two conditions that ideally are displayed in two ways:

1. separated by condition, but in the same visualization
2. as difference between conditions

I'll try and work through an example: https://docs.google.com/spreadsheets/d/1rZl-Ub5qOE_WJwHg23g8ijgANdRC82PQUXRbh6TJeKQ/edit#gid=0 Is the old version of what is now the AGL meta-analysis: https://docs.google.com/spreadsheets/d/1yCKAsu-vvQGemPga22F1cdtn90gs7Vp8UCto7HH10cE/edit#gid=818895612

There is one "control" condition and a paired "treatment" condition (btw: we should add columns and common labels to the codebook). Which condition it is is coded in the column expt_condition (unimodal = control, bimodal = treatment, I'd say). An additional column pair notes for each control experiment which entry is the matching treatment condition based on their label in unique_exp.

(A simpler solution might be to label treatment and control and have matching unique_exp labels.)

Effect sizes in the paired spreadsheet are then derived from the pairs, but I am not sure how @alecristia derived them. There should be a simple decision tree, similar to effect size calculation.

Note: Thinking it through I now believe that we need a separate visualization for showing data separated by condition, because we want to also show effect sizes etc for both conditions. Unless there is a better solution, of course... What would be ideal would be flipping back and forth between the difference scores and the two conditions. Adding them to MetaLab separately doesn't seem sensible btw.

[amsan7]

from our meeting (w/ @mcfrank), we are thinking:

• in compute_es pipeline, look for paired MAs, and compute ES for paired conditions (associating these only with the row in the data that has the treatment condition)
• in the visualization, add a selector for when there is a paired MA, to either "show all effect sizes" or "show effect sizes on comparison to control conditions."
• when "show effect sizes on comparison..." is checked, then data are filtered to only treatment rows and paired ES measures are used for all visualizations.

how does that sound, @alecristia?

[christinabergmann]

Hi all, Alex and I in consultancy with Sho and Molly arrived at this solution:

We specify which meta-analyses have a control condition and those should then have the column `condition_type' like in this spreadsheet: https://docs.google.com/spreadsheets/d/163QOxPcb_bNbnMrnoZee3peIKvLlo8EYJUHVd06W81I/edit#gid=0

Visualization is then either split into control' versuscritical' or their interaction, i.e. condition_type is a moderator in the model and the estimates for that moderator effect are used for effect size display and visualization. The interaction would be the default.

Does that make sense? It is both the more feasible and more transparent solution to this issue compared to calculating effect size differences between paired conditions. Based on our standard rma model, this solution also takes into account when papers test the same infant in both conditions (correct me if I am wrong here).

Alex will revert her data in AGL to add the column, but for now the above-mentioned dataset should be a good test case.

Let me know if anything is unclear.

[christinabergmann]

Ok, I think we need to implement this now for the mispronuciation meta-analysis and possibly also other dataset, is anyone on it?

[shotsuji]

Hi Alessandro, would this be possible to solve before you leave? Happy to discuss! @christinabergmann