You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardexpand all lines: README.md
+7-1
Original file line number
Diff line number
Diff line change
@@ -7,7 +7,7 @@ Simulation 2: Fixed batch effects in two steps or one step.
7
7
8
8
Simulation 3: More extreme selection and fixed plate effects in two step or one step.
9
9
10
-
Reference: Odintsova VV, Willemsen G, Dolan CV, Hottenga JJ, Martin NG, Slagboom PE, Ordoñana JR, Boomsma DI. Establishing a Twin Register: An Invaluable Resource for (Behavior) Genetic, Epidemiological, Biomarker, and 'Omics' Studies. Twin Res Hum Genet. 2018 Jun;21(3):239-252.
10
+
Simulation 4: estimating genetic and non-genetic variance components in the twin design
11
11
12
12
13
13
How to allocate twins to batch in assay of metabolites in an extremely discordant and concordant (EDAC) twin design?
@@ -22,18 +22,21 @@ We make the following assumptions concerning the analysis. We assume that the tw
22
22
The association between metabolite and aggression is accommodated by means of a fixed effect, i.e., regression of metabolite on binary (0/1) or continuous aggression score.
23
23
24
24
Simulation 1. Random effects model.
25
+
25
26
The metabolite explains 5% (r2=.05) of the variance in aggression. The heritability of metabolite is .6 (h2m=.5), the heritability of the residual of aggression is .5 (h2ag=.5). The number of batches is 70, the number of twin pairs is 600 (NMZtot=600). The true phenotypic variances of metabolite and aggression are both set to equal 1 and the variance is .25 (platevar=.25). All variables have zero mean. So the metabolite variance is 1.25. The selection is based on a mean split (high aggr>0; low aggr<0). The number of replications is 50. All 600 pairs are in the analysis. Estimation method is by restricted ML (REML).
26
27
Conclusion:
27
28
Allocation regime (pairs vs. individuals) has no effect on the estimate of the parameter of interest. We note that, as expected, regression on continuous predictors confers more power than regression on binary predictor (0/1). The variance components (Additive genetic, Environment and Batch) appear to be slightly downwards biased in the allocation by pair, but accurate in the allocation by individual condition.
28
29
29
30
Simulation 2. Fixed batch effects in two steps or one step.
31
+
30
32
It may be expedient to carry out analyses in two steps, i.e., first correct for batch effects, and second carry out the analysis of actual interest. We compared one and two step analyses in simulation 2. We used linear mixed modeling in simulation 1 (estimating the twin covariance conditional on predictor and batch). Here we use GEE (generalized estimating equations), i.e., we correct the standard errors after the analyses using a sandwich correction. That is:
31
33
One step: using GEE regress metabolite on predictor and batch simultaneously.
32
34
Two step: first correct metabolite for batch effect and then use GEE to analyze the residuals.
33
35
Conclusions:
34
36
Conclusions are the same as those based on Simulation 1. Allocation regime (pairs vs. individuals) has little effect on the test of the parameter of interest. Again, as expected, regression of continuous predictors confers more power that regression on binary predictor (0/1). We see little differences between one and two step procedure.
35
37
36
38
Simulation 3: More extreme selection and fixed plate effects in two step or one step.
39
+
37
40
This simulation is the same as simulation 2. However here we employ a more extreme selection criterion rather than a mean split (Simulations 1 and 2), the selection of high and low scoring twins is on the basis of the criteria > .5 std unit or < -.5 std units. As in simulation 2, we carry out one and two step analyses using GEE. Given the selection, we set the total sample size to 5000 (random sample) and select from this sample based on the criteria mentioned.
38
41
Conclusions:
39
42
The conclusions are consistent with those of simulations 1 and 2. The allocation regime, i.e., pairs vs. individuals, has little effect on the test of the parameter of interest. The regression of continuous predictors confers more power that regression on binary predictor (0/1), as expected.. There is little difference between the results of the one and two step procedure.
@@ -42,12 +45,15 @@ The conclusions are consistent with those of simulations 1 and 2. The allocatio
42
45
How to allocate twins to batches in assay of metabolites in the classical twin design
43
46
44
47
Simulation 4: estimating genetic and non-genetic variance components in the twin design.
48
+
45
49
In simulation 1, we noted that batch allocation by twin pair appeared to result in a slight bias in the estimates of the variance components. In simulation 4, we examined the effect on variance components by fitting an ACE model to twin data. Here we treat batch as a random and as a fixed effect, and we carry out both one step and two step analyses. We consider relatively small sample sizes. We use linear mixed modeling with REML (restricted maximum likelihood) estimation.
46
50
Conclusion:
47
51
The results demonstrate that allocation regime has little effect in the one step analyses, regardless of whether this is based on random effects or fixed effects modeling of batch. In the two step procedures, we note a downward bias in the estimates of the A (additive genetic) and C (common environment) variance components. This bias is greater in given the allocation by twin pairs, and greater as the number of batches increases (compare 15, 25, and 40 batch conditions).
48
52
49
53
For details of the simulations and Settings in the R code see:
0 commit comments