In the below we fit mixed effects models of various structures to the albumin secretion data.
First we fit the full model with fixed effects for condition, day, and their interaction, as well as nested random effects for lab, study, and circuit. From this we find that labs contribute a large proportion to the total variance, whereas studies within labs contribute very little. The variability seen between studies may stem from lab-to-lab variability rather than study-to-study variability.
We then fit the model where we ignore the factor of study, i.e., the above model but with circuits nested directly within labs. We find that the variance components are very similar to before.
Next we fit a model where we ignore the layer of labs, and treat the three studies as completely independent. In this case the proportion of variability assigned to circuits within studies is slightly larger than in the previous models, although not much. Also the proportion of variability assigned to residuals is slightly larger, but overall the allocations are similar.
Finally we also investigate the case where we look at only studies from the first lab (studies 1 and 2) and fit the complete model with random effects for labs, studies, and circuits. Notably the results again indicate that indeed there is not much study-to-study variability within labs but more variability between labs.
We also investigate the BIC of each of the models using all data, and find that the mixed model with random effects for lab and circuit (and not study) has the lowest BIC. Based on this and the variance decomposition we claim that this model is the best.
| npar | AIC | BIC | logLik | deviance | Chisq | Df | Pr(>Chisq) | |
|---|---|---|---|---|---|---|---|---|
| alb_circuit_mixed | 30 | -90.86427 | 38.86650 | 75.43213 | -150.8643 | NA | NA | NA |
| alb_lab_circuit_mixed | 31 | -106.80508 | 27.25005 | 84.40254 | -168.8051 | 17.940815 | 1 | 0.0000228 |
| alb_study_circuit_mixed | 31 | -105.41903 | 28.63609 | 83.70952 | -167.4190 | 0.000000 | 0 | NA |
| alb_lab_study_circuit_mixed | 32 | -104.81858 | 33.56091 | 84.40929 | -168.8186 | 1.399546 | 1 | 0.2367996 |
[1] “Optimal model (based on BIC): alb_lab_circuit_mixed” [1] “Optimal model (based on AIC): alb_lab_circuit_mixed”
Next we look at each study at its own and investigate the amount of circuit-to-circuit variability. Clearly study 2 has larger circuit-to-circuit variability than studies 1 and 3.
In the below we investigate the estimated disease effect for each study separately at day 13, using a mixed model with fixed effects for condition, day and the interaction, and a random effect for circuit (that is, the same models as generated the previous pie-charts).
In the below we fit mixed effects models of various structures to the ketone bodies data.
First we fit the full model with fixed effects for condition, day, and their interaction, as well as nested random effects for lab, study, and circuit. From this we find that labs contribute a large proportion to the total variance, whereas studies within labs contribute very little (singular fit). The variability seen between studies may stem from lab-to-lab variability rather than study-to-study variability.
We then fit the model where we ignore the factor of study, i.e., the above model but with circuits nested directly within labs. We find that the variance components are very similar to before.
Next we fit a model where we ignore the layer of labs, and treat the three studies as completely independent. In this case the proportion of variability assigned to circuits within studies is slightly larger than in the previous models, although not much. Also the proportion of variability assigned to residuals is slightly larger, but overall the allocations are similar.
Finally we also investigate the case where we look at only studies from the first lab (studies 1 and 2) and fit the complete model with random effects for labs, studies, and circuits. Notably the results again indicate that indeed there is not much study-to-study variability (singular fit) within labs.
We also investigate the BIC of each of the models using all data, and find that the mixed model with random effects for lab and circuit (and not study) has the lowest BIC. Based on this and the variance decomposition we claim that this model is the best.
## boundary (singular) fit: see help('isSingular')
## boundary (singular) fit: see help('isSingular')
| npar | AIC | BIC | logLik | deviance | Chisq | Df | Pr(>Chisq) | |
|---|---|---|---|---|---|---|---|---|
| ket_circuit_mixed | 10 | 91.08592 | 120.9902 | -35.54296 | 71.08592 | NA | NA | NA |
| ket_lab_circuit_mixed | 11 | 89.96745 | 122.8622 | -33.98373 | 67.96745 | 3.118465 | 1 | 0.0774097 |
| ket_study_circuit_mixed | 11 | 91.04527 | 123.9400 | -34.52264 | 69.04527 | 0.000000 | 0 | NA |
| ket_lab_study_circuit_mixed | 12 | 91.96745 | 127.8526 | -33.98373 | 67.96745 | 1.077820 | 1 | 0.2991857 |
[1] “Optimal model (based on BIC): ket_circuit_mixed” [1] “Optimal model (based on AIC): ket_lab_circuit_mixed”
Next we look at each study at its own and investigate the amount of circuit-to-circuit variability. Clearly study 2 has larger circuit-to-circuit variability than studies 1 and 3.
In the below we investigate the estimated disease effect for each study separately at day 13, using a mixed model with fixed effects for condition, day and the interaction, and a random effect for circuit (that is, the same models as generated the previous pie-charts).
No data from study 1, meaning that lab and study are completely confounded. In the following we can therefore not separate between lab-to-lab variability and study-to-study variability.
In the below we fit mixed effects models of various structures to the IL-1R2 data.
First we fit the full model with fixed effects for condition, day, and their interaction, as well as nested random effects for lab (or equivalently, study) and circuit.
## Warning in checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, :
## Model failed to converge with max|grad| = 0.455226 (tol = 0.002, component 1)
| npar | AIC | BIC | logLik | deviance | Chisq | Df | Pr(>Chisq) | |
|---|---|---|---|---|---|---|---|---|
| il1_circuit_mixed | 14 | 106.8026 | 142.7034 | -39.40128 | 78.80256 | NA | NA | NA |
| il1_lab_circuit_mixed | 15 | 104.7682 | 143.2335 | -37.38412 | 74.76824 | 4.034329 | 1 | 0.0445834 |
[1] “Optimal model (based on BIC): il1_circuit_mixed” [1] “Optimal model (based on AIC): il1_lab_circuit_mixed”
Next we look at each study at its own and investigate the amount of circuit-to-circuit variability. Clearly study 2 has larger circuit-to-circuit variability than studies 1 and 3.
In the below we investigate the estimated disease effect for each study separately at day 13, using a mixed model with fixed effects for condition, day and the interaction, and a random effect for circuit (that is, the same models as generated the previous pie-charts).
For GTT endpoints we only include data from the TissUse lab, and hence we cannot look at any lab-to-lab effects or variability. We also look specifically at measurements at day 13, as that is the only day where we have sampled both diseased and healthy conditions.
In the below we fit mixed effects models of various structures to the GTT glucose data.
First we fit the full model with fixed effects for condition, day, and their interaction, as well as nested random effects for study and circuit.
| npar | AIC | BIC | logLik | deviance | Chisq | Df | Pr(>Chisq) | |
|---|---|---|---|---|---|---|---|---|
| glu_day13_circuit_mixed | 18 | -380.6086 | -329.2720 | 208.3043 | -416.6086 | NA | NA | NA |
| glu_day13_study_circuit_mixed | 19 | -380.3064 | -326.1178 | 209.1532 | -418.3064 | 1.697785 | 1 | 0.192578 |
[1] “Optimal model (based on BIC): glu_day13_circuit_mixed” [1] “Optimal model (based on AIC): glu_day13_circuit_mixed”
Next we look at each study at its own and investigate the amount of circuit-to-circuit variability.
In the below we investigate the estimated disease effect for each study separately at day 13, using a mixed model with fixed effects for condition, hour and the interaction, and a random effect for circuit (that is, the same models as generated the previous pie-charts).
For GTT endpoints we only include data from the TissUse lab, and hence we cannot look at any lab-to-lab effects or variability. We also look specifically at measurements at day 13, as that is the only day where we have sampled both diseased and healthy conditions.
In the below we fit mixed effects models of various structures to the GTT insulin data.
First we fit the full model with fixed effects for condition, day, and their interaction, as well as nested random effects for study and circuit. Notably, there is a very large study-to-study variability for this endpoint.
| npar | AIC | BIC | logLik | deviance | Chisq | Df | Pr(>Chisq) | |
|---|---|---|---|---|---|---|---|---|
| ins_day13_circuit_mixed | 18 | 250.4113 | 300.1304 | -107.20565 | 214.4113 | NA | NA | NA |
| ins_day13_study_circuit_mixed | 19 | 201.1124 | 253.5937 | -81.55618 | 163.1124 | 51.29894 | 1 | 0 |
[1] “Optimal model (based on BIC): ins_day13_study_circuit_mixed” [1] “Optimal model (based on AIC): ins_day13_study_circuit_mixed”
Next we look at each study at its own and investigate the amount of circuit-to-circuit variability.
## fixed-effect model matrix is rank deficient so dropping 2 columns / coefficients
## fixed-effect model matrix is rank deficient so dropping 2 columns / coefficients
In the below we investigate the estimated disease effect for each study separately at day 13, using a mixed model with fixed effects for condition, hour and the interaction, and a random effect for circuit (that is, the same models as generated the previous pie-charts).
Analyses were conducted using the R Statistical language (version 4.2.1; R Core Team, 2022) on macOS Ventura 13.5.2, using the packages magrittr (version 2.0.3; Bache S, Wickham H, 2022), lme4 (version 1.1.34; Bates D et al., 2015), Matrix (version 1.5.1; Bates D et al., 2022), lmerTest (version 3.1.3; Kuznetsova A et al., 2017), emmeans (version 1.8.2; Lenth R, 2022), report (version 0.5.5; Makowski D et al., 2021), tibble (version 3.2.1; Müller K, Wickham H, 2023), ggplot2 (version 3.4.0; Wickham H, 2016), stringr (version 1.5.0; Wickham H, 2022), forcats (version 1.0.0; Wickham H, 2023), tidyverse (version 1.3.2; Wickham H et al., 2019), readxl (version 1.4.1; Wickham H, Bryan J, 2022), dplyr (version 1.1.2; Wickham H et al., 2023), purrr (version 1.0.2; Wickham H, Henry L, 2023), readr (version 2.1.3; Wickham H et al., 2022) and tidyr (version 1.3.0; Wickham H et al., 2023).