Bayesian Inference of deterministic population growth models -- Brazilian Meeting on Bayesian Statistics
1 like772 views
The document presents a Bayesian inference approach for deterministic population growth models, emphasizing the importance of temperature-dependent parameters in populations. It discusses model specification, data collection, and the use of Hamiltonian Monte Carlo methods for posterior approximation using Stan. Results include population estimates for the Chagas disease vector Rhodnius prolixus, and the authors highlight the advantages of Bayesian methods for parameter interpretability and efficient sampling.
![Bayesian inference of deterministic population growth models
Luiz Max F. de Carvalho∗
[lmax.procc@gmail.com]
Claudio J. Struchiner [stru@fiocruz.br]
Leonardo S. Bastos [lsbastos@fiocruz.br]
Scientific Computing Programme (PROCC), Oswaldo Cruz Foundation (Fiocruz), Rio de Janeiro,
Brazil
March, 2014
12th EBEB - Atibaia – SP](https://image.slidesharecdn.com/carvalhoebeb2014-140317102315-phpapp02/85/Bayesian-Inference-of-deterministic-population-growth-models-Brazilian-Meeting-on-Bayesian-Statistics-1-320.jpg)
Bayesian Inference of deterministic population growth models -- Brazilian Meeting on Bayesian Statistics
- 1. Bayesian inference of deterministic population growth models Luiz Max F. de Carvalho∗ [lmax.procc@gmail.com] Claudio J. Struchiner [stru@fiocruz.br] Leonardo S. Bastos [lsbastos@fiocruz.br] Scientific Computing Programme (PROCC), Oswaldo Cruz Foundation (Fiocruz), Rio de Janeiro, Brazil March, 2014 12th EBEB - Atibaia – SP
- 2. Nice to meet you! • BSc. in Microbiology, UFRJ (2013); • Statistics Assistant, Pan American Health Organization, 2010-2013; • Currently at PROCC and DME/IM-UFRJ (MSc); • Soon to be moving to the University of Edinburgh for a PhD in Evolutionary Biology. 2 of 17
- 3. Motivation • Deterministic models are widely used in Science, let alone Biology; ◦ Population Growth; ◦ Disease Spreading; ◦ Cell and molecular interactions. • They provide a crude but easily interpretable representation of reality; • Temperature is a key factor to the growth of several organisms. ◦ Disease-carrying arthropds; ◦ Pathogenic bacteria; ◦ Economically important plants. • With a deterministic model and some time series data at hand, how to learn about model parameters? 3 of 17
- 4. Background • Consider a deterministic model M(·); ◦ Let x ∈ X ⊂ Rp be the set of model inputs and y ∈ Y ⊂ Rn be the model outputs. The deterministic model M(x; θ) = y, where θ ∈ Θ ⊂ Rq is a q-dimensional parameter vector, completely specifies the relationship between x and y (Poole & Raftery, 2000); ◦ In our particular case, we have laid our dirty hands on some data y and inputs x that we think can be modelled as y = M(x; θ) • We are now interested in learning about θ 4 of 17
- 5. Temperature-dependent Population Growth • Consider the ordinary non-linear differential equation (Verhulst, 1838): dP dt = r 1 − P K P ∴ P(t) = K 1 + K−N0 N0 e−rt (1) • We formulate a modified version of (1), with temperature-dependent parameters P(t, T) = K(T) 1 + K(T)−N0 N0 e−r(T)t (2) 5 of 17
- 6. Temperature-dependent Population Growth (cont.) • To complete model specification, we propose two smooth functions on temperature T: K(T) = cK exp − (T − aK )2 bK (3) r(T) = cr exp − (T − ar )2 br (4) We want to learn about θ = {aK , bK , cK , ar , br , cr } 6 of 17
- 7. Likelihood • Assume P(t, T) to be a Gaussian process with fixed variance τ2 ; • Let y = {y1, y2, . . . , yN } be an output vector with N measurements, which we observe directly; • Moreover, let t = {t1, t2, ..., tN } and T = {T1, T2, ..., TN } be the vectors with observed times and temperatures. Then yi |ti , Ti , N0, θ ∼ N(µ(ti , Ti , N0; θ), τ2 ) (5) µ(ti , Ti , θ) = K(Ti ; θK ) 1 + K(Ti ;θK )−N0 N0 e−r(Ti ;θr )ti , ∀i = 1, 2, . . . , N (6) which is equivalent to writing yi = M(ti , Ti , N0; θ) + , ∼ N(0, τ2 ). 7 of 17
- 8. Priors • Biologically motivated, proper priors, elicited to maintain functional form while remaining diffuse. aK , ar ∼ Normal(20, 10) bK , br ∼ Gamma(4, 1/5) cK ∼ Gamma(1, 1/1000) cr ∼ Normal(1/2, 2) τ2 ∼ Gamma(1/10, 1/10) 8 of 17
- 9. Posterior • From the Bayes theorem p(θ|y, t, T) ∝ p(y|θ, t, T)π(θ|t, T) (7) • The model for P(t, T) is thus hierarchical and depends on two latent quantities, r(T) and K(T). 9 of 17
- 10. Posterior Approximation – Stan • Hamiltonian Monte Carlo (HMC): ◦ Avoids Random Walk behaviour; ◦ Allows big moves, with high acceptance probability; ◦ Does not suffer with highly correlated posteriors; • We used the stan package of the R Statistical Computing Environment to approximate the posterior through HMC. ◦ Fast C++ implementation; ◦ No-U-Turn Sampler (NUTS); ◦ Neat BUGS-like syntax for model specification; ◦ Smooth interface with R. • MCMC was run for 50, 000 iterations with 25, 000 burn-in and convergence was assessed by inspecting the trace- and autocorrelation plots and potential scale reduction factor. 10 of 17
- 11. Results I – Simulation study • From a set of parameters θ∗ and a grid {t, T} we generate Q data sets of size N = nt × nT by sampling from y|θ∗ , t, T; • We then obtain Q posterior estimates and calculate MSE, normalized bias and coverage probability of the 95% credibility intervals; Parameter Value Posterior Mean Bias MSE Coverage aK 30.00 29.44 0.01 7.99 0.93 ar 23.00 22.71 0.00 3.31 0.86 bK 10.00 13.08 0.95 450.26 0.94 br 15.00 16.82 0.22 16.77 0.88 cK 700.00 692.17 0.09 7203.06 0.96 cr 0.40 0.43 0.00 0.04 0.85 τ 3.16 4.89 0.94 67.02 0.88 ◦ Consistent results for N = 350, 630 and 1600. 11 of 17
- 12. Results II – Rhodnius prolixus data • Important Chagas disease vector; ◦ We fear climatic change may increase suitability in previously uncolonized areas; • Population sizes measured in 10 temperatures in the range 16 − 38 ◦ C for 35 days (N0 = 30 for all T); (a) K(T) (b) r(T) 12 of 17
- 13. Results II – Rhodnius prolixus data (cont.) Posterior Mean (95% C.I.) Prior Mean (95% C.I.) aK 19.23 (17.56 – 21.09) 25.00 (5.40 – 44.60) ar 25.73 (25.44 – 26.10) 25.00 (5.40 – 44.60) bK 106.17 (75.25 – 137.31) 20.00 (5.44 – 43.84) br 26.77 (22.59 – 32.19) 20.00 (5.44 – 43.84) cK 1023.32 (898.28 – 1165.40) 1000.00 (25.31 – 3688.87) cr 0.66 (0.58 – 0.76) 0.50 (-3.41 – 4.41) τ 177.33 (166.10 – 191.78) 1.00 (0.00 – 9.78) 13 of 17
- 14. Results II – Rhodnius prolixus data (cont.) (c) Data (d) Posterior 14 of 17
- 15. Conclusions and Perspectives • We stress the importance of using Bayesian Inference to learn about model parameters ◦ Parameters retain direct interpretability • Stan ◦ Efficient sampling through HMC; ◦ NUTS drops the need for hand-tuning; ◦ Consirable speed-up and quicker convergence. • Perspectives ◦ Dynamic variance, τ2 (t) ; ◦ Other data sets, e.g, bacterial growth; ◦ Complete treatment of uncertainty: Bayesian melding. 15 of 17
- 16. Thank you! • References D. Poole and A. E. Raftery, “Inference for deterministic simulation models: the bayesian melding approach,” Journal of the American Statistical Association, vol. 95, no. 452, pp. 1244–1255, 2000. P.-F. Verhulst, “Notice sur la loi que la population suit dans son accroissement. correspondance math´ematique et physique publi´ee par a,” Quetelet, vol. 10, pp. 113–121, 1838. • Acknowledgements ◦ My advisors, Claudio and Leo; ◦ Leonardo B. Santos, INPE; ◦ PROCC, who provided an office and a Gourmet coffee machine! 16 of 17