Table 1.

Assumptions and general guidelines for making partly conditional inference

Method	Assumption under which method is valid	Guidelines for use
IPW	u-MAR (1) u-MAR is equivalent to mortal-cohort dDTIC and missingness-independent death (2) u-MAR is the discrete death-time version of MAR in Kurland and Heagerty (2005)	Model probability of dropout at visit for those who survived up to visit Do not include in dropout models Appropriate when longitudinal and survival processes depend on one another, and survival process does not depend on dropout process u-MAR assumption can be tested if survival is known up to the end of study
IPW	p-MAR Note: p-MAR is a weaker assumption than u-MAR	Model probability of dropout at visit for those who survived up to and including visit ,
IPW	f-MAR (1) f-MAR is a weaker assumption than u-MAR (2) f-MAR is the discrete death-time version of MAR-S in Kurland and Heagerty (2005)	For visit each , model the probability of dropout at visit for those who die between visits and (i.e. whose ) Include in dropout models
MI/LI	mortal-cohort dDTIC and independent death	Do not include in the imputation models Delete imputed post-death outcomes before analysis Appropriate when survival and dropout processes depend on one another, and survival process does not depend on missing outcome process
MI/LI	f-MAR	Include in imputation models Delete imputed post-death outcomes before analysis
AIPW	either i) u-MAR or ii) p-MAR, f-MAR and independent death	Do not include in dropout and imputation models Delete imputed post-death outcomes At least as efficient as IPW if dropout and imputation models are correctly specified and survival process does not depend on longitudinal process (death still truncates longitudinal and dropout processes)
AIPW	f-MAR	Valid as long as either dropout or imputation models are correctly specified Include in dropout and imputation models Delete imputed post-death outcomes At least as efficient as IPW if dropout and imputation models are correctly specified and data are f-MAR