The use of complete-case and multiple imputation-based analyses in molecular epidemiology studies that assess interaction effects

Desai, Manisha; Esserman, Denise A; Gammon, Marilie D; Terry, Mary B

doi:10.1186/1742-5573-8-5

Epidemiologic Perspectives & Innovations

Table 1 Description of Scenarios Used in Two Sets of Simulation Studies.

From: The use of complete-case and multiple imputation-based analyses in molecular epidemiology studies that assess interaction effects

Table	Set	Scenario	Median % Missing X1	Nature of Missing	Auxiliary Relationship	Type of Missing Variable
3a. Impact of Auxiliary Relationship Under Condition 1	1	A	22%	Condition 1 ^a	None ³	Binary
	1	B	22%	Condition 1 ^a	Moderate ⁴	Binary
	1	C	22%	Condition 1 ^a	Strong ⁵	Binary
3b. Impact of Auxiliary Relationship Under Condition 2	1	D	20%	Condition 2 ^b	None ³	Continuous
	1	E	20%	Condition 2 ^b	Moderate ⁶	Continuous
	1	F	20%	Condition 2 ^b	Strong ²	Continuous
3c. Impact of Auxiliary Relationship Under Condition 3	1	G	20%	Condition 3 ^c	None ³	Continuous
	1	H	20%	Condition 3 ^c	Moderate ⁶	Continuous
	1	I	20%	Condition 3 ^c	Strong ²	Continuous
4a. Impact of Conditions Under Set of Auxiliary Variables with Varying Strength When Missing Genotype Data	2	J	21%	Condition 1 ^d	Realistic ⁷	Binary
	2	K	24%	Condition 3 ^e	Realistic ⁷	Binary
4b. Impact of Conditions Under Set of Auxiliary Variables with Varying Strength When Missing Exposure Data	2	L	21%	Condition 1 ^d	Realistic ⁷	Binary
	2	M	21%	Condition 2 ^f	Realistic ⁸	Ordinal

^aCondition 1: X1 is 12.2 times more likely to be missing if X1 = 1
^bCondition 2: Extreme values of X1 are more likely to be missing (probability of missing is a quadratic function of X1 or the log odds of missing X1 = γ₀+γ₁ X1 + γ₂ X1², where γ₁ = -1 and γ₂ = 2.)
^cCondition 3: A 1-unit increase in X1 corresponds to a 7.4 times decrease in the probability of missing for controls, but a 7.4 times increase for cases
^dCondition 1: Those with fast metabolizing genotype are 12.2 times more likely to be missing data on genotype. Missingness is also related to other observed covariates (mammogram, education, race, breastfeeding oral contraceptive use, hormone therapy use, and smoking status) as informed by the real data set for all subjects.
^eCondition 3: Exposed controls with fast genotype are 7.4 times more likely to be missing genotype than those without, while exposed cases with fast genotype are 7.4 times less likely to be missing genotype. Unexposed subjects with fast genotype are 2.7 times more likely to be missing genotype. Missingness is also related to other observed covariates (mammogram, education, race, breastfeeding oral contraceptive use, hormone therapy use, and smoking status) as informed by the real data set for all subjects.
^fCondition 2:Extreme values of exposure are more likely to be missing (probability of missing is a quadratic function of X1 or the log odds of missing X1 = γ₀+γ₁ X1 + γ₂ X1², where γ₁ = -3 and γ₂ = 1.) Missingness is also related to other observed covariates (mammogram, education, race, breastfeeding oral contraceptive use, hormone therapy use, and smoking status) as informed by the real data set for all subjects.
¹Strong: Those with X1 = 1 have Z values (SD = 1), that are 3 units higher on average than those with X1 = 0
²Strong: Average correlation between X1 and Z is 0.97
³None: X1 and Z are independent variables
⁴Moderate: Those with X1 = 1 have Z values (SD = 1) that are 1 unit higher on average than those with X1 = 0
⁵Strong: Those with X1 = 1 have Z values (SD = 1) that are 4 units higher on average than those with X1 = 0
⁶Moderate: Average correlation between X1 and Z is 0.57
⁷Realistic: Auxiliary variables are informed by real data set and include all variables used to generate missing data mechanism of genotype (case-control status, exposure, mammogram, education, race, breastfeeding behavior, oral contraceptive use, hormone therapy use, and smoking status) as well as those that relate to genotype (exposure, race, breastfeeding behavior), a subset of the former.
⁸Realistic: Auxiliary variables are informed by real data set and include all variables used to generate missing data mechanism of exposure (case-control status, genotype, mammogram, education, race, breastfeeding behavior, oral contraceptive use, hormone therapy use, and smoking status) as well as those that relate to exposure (genotype, education, race, oral contraceptive use, hormone therapy use, and smoking status), a subset of the former.

Back to article page