Timeline for A more definitive discussion of variable selection
Current License: CC BY-SA 3.0
14 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jun 1, 2019 at 23:58 | comment | added | Alexis | 2/2 Of course, not including $L$ still leaves $E[Y|A]$ biased (from $A-L-U_{2}-Y$). Resolving the DAG I describe to a causally valid estimate necessitates measuring either $U_{1}$ or $U_{2}$ or a strong set of proxies for either or both, so in that sense, some confounders must be included in the model. But there's nuance in this. | |
| Jun 1, 2019 at 23:53 | comment | added | Alexis | 1/2 +1 Lovely answer. A niggle: "Confounders must be included in models." Depending on the specific causal relationships involved doing so may bias the associations away from a causal estimate. For example, in the DAG described by $L \rightarrow A$, $U_{1} \rightarrow L$, $U_{1} \rightarrow A$, $U_{2} \rightarrow Y$, and $U_{2} \rightarrow L$, $L$ is a confounder, both in the backdoor criterion sense, and in the classic (i) associated with $A$ & $Y$, (ii) a cause of $A$ or $Y$, and (iii) prior to $A$ & $Y$ sense. However, $E[Y|A,L]$ is biased (from $A-U_{1}-L-U_{2}-Y$). | |
| Jul 23, 2016 at 20:57 | history | bounty ended | gung - Reinstate Monica | ||
| Jul 21, 2016 at 17:51 | comment | added | AdamO | @DavidC.Norris "Educational and Psychological Measurement" is the journal in question and alas, the Babyak article only discusses automated stepwise regression. I have seen in some articles authors who discuss their "hands on" approach to including and refitting models as being "stepwise" (albeit not automated). I would argue they've correctly described the procedure they used, but I still have grave contentions with this approach. | |
| S Jul 20, 2016 at 3:26 | history | suggested | David C. Norris | CC BY-SA 3.0 |
Corrected spelling of 'Munchausen'
|
| Jul 20, 2016 at 3:17 | comment | added | David C. Norris | (+1) Outstanding response, AdamO! Appreciating that you've already done a huge amount of work, I wonder if you would tell anything more you know about the current adoption status of the journal policy you mention. I still fly into fits of rage on seeing the word 'stepwise' in the JAMA journals at least. Can you cite an editorial on this policy? | |
| Jul 20, 2016 at 2:59 | review | Suggested edits | |||
| S Jul 20, 2016 at 3:26 | |||||
| Jul 16, 2016 at 21:12 | vote | accept | sharper_image | ||
| Jul 15, 2016 at 11:06 | comment | added | Scortchi♦ | (+1) On point #1: G&H do say these are "general principles for building regression models for prediction" [my italics] so your caveat needn't apply. (Perhaps those rare smokers with good lung function do indeed tend to be especially fit.) When they get on to discussing causal models they make the same point (Ch. 9.7). | |
| Jul 14, 2016 at 20:27 | history | edited | AdamO | CC BY-SA 3.0 |
added 381 characters in body
|
| Jul 14, 2016 at 18:20 | comment | added | sharper_image | Really appreciate your detailed commentary - I've never had anyone answer my questions in such detail before. I'm going through your links now! | |
| Jul 14, 2016 at 17:41 | history | edited | AdamO | CC BY-SA 3.0 |
added 616 characters in body
|
| Jul 14, 2016 at 17:35 | history | edited | AdamO | CC BY-SA 3.0 |
added 616 characters in body
|
| Jul 14, 2016 at 17:27 | history | answered | AdamO | CC BY-SA 3.0 |