-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmismatchmodelpresentation20240919.qmd
123 lines (79 loc) · 4.22 KB
/
mismatchmodelpresentation20240919.qmd
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
---
title: "Multilevel trait-microbiome mismatch model"
author: Quentin D. Read
format:
revealjs:
embed-resources: true
transition: slide
engine: knitr
---
## The problem
- We have maternal plants from different populations
- We want to measure the association between seedling traits and seed microbiome in their offspring
+ And whether this association is different by population but that's just an extra wrinkle
:::: {.columns}
::: {.column}
:::
::: {.column}
:::
::::
## The problem
- But it is not possible to observe seed microbiome and seedling traits for the same seed
- The seed can either be germinated to measure traits, or destroyed to measure microbiome, but not both
:::: {.columns}
::: {.column}
:::
::: {.column}
:::
::::
## How to resolve the problem
- We can consider this to be a pathological missing data situation
- Any individuals missing $X$ (microbiome) are missing $y$ (trait), and vice versa
- But we can still estimate random intercepts for each mother plant, for each one
- Then model the relationship between $X$ and $y$ at the offspring level
- Bayesian approach imputes the missing data as parameters in the model
## Simulate some data
- The real dataset turns out to be a poor one to actually demonstrate the model!
- Simulate maternal microbiome with a multivariate normal distribution
+ $n_m$: number of mothers (20)
+ $n_o$: number of offspring per mother (10)
+ $n_t$: number of taxa (200)
+ $\mathbf{X}_m \sim \text{MVNormal}(0, \mathbf{I}_{n_t})$
## Simulate some data, part 2
- Simulate offspring microbiome, also with a multivariate normal distribution
- Use the maternal means $\mathbf{X}_m$ we simulated in the previous step, and the covariance matrix $\boldsymbol\Sigma_m$ of the simulated maternal means
- For each mother $i$ her offspring microbiome is:
+ $\mathbf{X}_{oi} \sim \text{MVNormal}(\mathbf{X}_{mi}, \boldsymbol\Sigma_m)$
## Simulate some data, part 3
- Assign only a few nonzero effects of taxa on trait. The rest are 0.
+ $\boldsymbol\beta = [50, 10, -50, 0, 0, 0, ...]$
- Use these, with some normally distributed random noise added, to get the offspring trait means
- For each mother $i$ her offspring trait values are:
+ $\mathbf{y}_i \sim \text{Normal}(\mathbf{X}_{oi}\boldsymbol\beta, 1)$
## Description of the model {.smaller}
- Random maternal intercepts for microbiome, random maternal intercepts for trait
- Covariance between random intercepts for each taxon set to 0
- Trait has fixed effects for all microbiome taxa (no interaction effects or other covariates)
- I will write the formal statistical notation for this soon
## Priors
- Use "regularized horseshoe priors" on the fixed effects, which can handle very wide data
+ Piironen & Vehtari 2017, Electronic Journal of Statistics
+ https://avehtari.github.io/modelselection/regularizedhorseshoe_slides.pdf
- Gamma priors (weakly informative) on variance of random effects
- Missing values follow the same prior distributions as non-missing values
## Fit the model to the data
- Fit the model to the full dataset
- Then set half the trait data to be missing, and the other half of the microbiome data to be missing, and refit
- Compare the fixed effect estimates
## Comparison of fixed effects
- Only taxa 1-3 were simulated to have nonzero fixed effect (most of the 200 taxa aren't shown here)
## Future directions
- The model could be extended to multivariate trait outcomes and interactions between taxa, *at least in theory*
- A major issue is scaling up the model
+ With 200 taxa, it took several hours to sample the posterior but this will blow up as community size increases
+ Full Bayesian computation on anything much bigger than that is going to be very tough