How to analyze work productivity loss due to health problems in randomized controlled trials? A simulation study

Table 2 Number of simulated databases with quasi-complete separation

Distribution of productivity loss outcome	Number of Observations in each arm	Truncated negative binomial distributions of productivity loss outcomes in the two arms	Number of databases with quasi-complete separation^a
80:15:5/60:30:10	100	Equal Scale	3
80:15:5/60:30:10	100	Unequal Scale	7
80:15:5/60:30:10	200	Equal Scale	0
80:15:5/60:30:10	200	Unequal Scale	0
60:35:5/40:50:10	100	Equal Scale	0
60:35:5/40:50:10	100	Unequal Scale	0
60:35:5/40:50:10	200	Equal Scale	0
60:35:5/40:50:10	200	Unequal Scale	0
50:40:10/30:55:15	50	Equal Scale	0
50:40:10/30:55:15	50	Unequal Scale	2
50:40:10/30:55:15	100	Equal Scale	0
50:40:10/30:55:15	100	Unequal Scale	0
50:40:10/30:55:15	200	Equal Scale	0
50:40:10/30:55:15	200	Unequal Scale	0

^aThe maximum likelihood estimate may not exist while running multinomial logistic regression for the three-part models. For number of observations = 50 and 5% max loss, a large number of databases with quasi-separation and thus the three-part model was not considered for these scenarios

ISSN: 1471-2288