Some tricks for estimating space-time point process models. Rick Paik Schoenberg, UCLA

Transcription

1 Some tricks for estimating space-time point process models. Rick Paik Schoenberg, UCLA 1. Nonparametric estimation. 2. Parametric estimation of clustering. 3. Separable estimation for covariates. 1

2 1. Nonparametric estimation of clustering. Estimate the rate λ(t,x,y,m) = µ(x,y,m) + g(t-t i, x-x i, y-y i, m i ). i:ti<t from Marsan and Lengliné (2008). 2

3 1. Nonparametric estimation. Marsan, D. and Lengliné, O. (2008). Extending earthquakes' reach through cascading. Science 319,

4 λ(t,x,y,m) = µ(x,y,m) + g(t-t i, x-x i, y-y i, m i ). The idea in Marsan and Lengliné (2008) is to use a kind of E-M algorithm: 0. Start with some naive estimate of µ and g. 1. Using this, compute w ij = P(eq i triggered earthquake j), and i:ti<t w 0j = P(earthquake j was a mainshock). 2. Using w ij and w 0j, estimate µ and g by MLE. Repeat steps 1 and 2 until convergence. With µ a constant, the estimate of g in step 2 is simply g(δt, Δr, m) = A w ij / [N m S t], where A is the set of pairs of points whose distance ~ Δt, Δr, m i ~ m, i.e. t i t j in (Δt - t, Δt + t), etc. N m is the number of eqs with m i ~ m, and S is the total surface covered by the disk with radius Δr +/- r. 4

5 λ(t,x,y,m) = µ(x,y,m) + g(t-ti, x-xi, y-yi, mi). i:ti<t This nonparametric estimate can then be used as a guide to come up with a parametric form for the triggering function, g. Fox et al., 2015.

6 2. Parametric estimation of clustering. The Hawkes process (Hawkes 1971) is a useful form for modeling clusteted processes, where λ(t,x,y,m) = µ(x,y,m) + g(t-t i, x-x i, y-y i, m i ). i:ti<t An example is the Epidemic-Type Aftershock Sequence (ETAS) model of Ogata (1988, 1998), which has been used for earthquakes as well as invasive species (Balderama et al. 2012) and crime (Mohler et al. 2011). With ETAS, µ(x,y,m) = µ(x,y) f(m), and e.g. 6

7 parametric point process estimation is easy?! a. Point processes can generally be very well estimated by maximum likelihood estimation (MLE).!!MLEs are generally consistent, asymptotically normal, and efficient!!!fisher (1925), Cramer (1946), Ogata (1978), Le Cam (1990).! b. Maximizing a function is fast and easy.! 400 yrs of work on this since Newton (1669). Every computer package has a strong quasi-newton minimization routine, such as optim() in R.!!For instance, suppose the function you want to minimize is!!!!f(a,b) = (a-4.19) 4 + (b-7.23) 4.!!f = function(p) (p[1]-4.19)^4 + (p[2]-7.23)^4!!p = c(1,1) ## initial guess at (a,b)!!est = optim(p,f)$par! c. The likelihood function is fairly simple.! In practice, it s convenient to compute log(likelihood), and minimize that.!!log(likelihood) = log{λ(t i,x i,y i )} - λ(t,x,y) dt dx dy.! ETAS (Ogata 98), λ(t,x,y) = µ ρ(x,y) + g(t-t j, x-x j, y-y j; M j ),! where g(t,x,y,m) = K (t+c) -p e a(m-m0) (x 2 + y 2 + d) -q! or K (t+c) -p {(x 2 + y 2 )e -a(m-m0) + d} -q.! 7

8 Parametric point process estimation is easy?! ETAS (Ogata 98), λ(t,x,y) = µ ρ(x,y) + g(t-t j, x-x j, y-y j; M j )! where g(t,x,y,m) = K (t+c) -p e a(m-m0) (x 2 + y 2 + d) -q! or K (t+c) -p {(x 2 + y 2 )e -a(m-m0) + d} -q.! To make these spatial and temporal parts of g densities, I suggest writing g as!!g(t,x,y,m) = {K (p-1) c p-1 (q-1) d q-1 / π} (t+c) -p e a(m-m0) (x 2 + y 2 + d) -q! or!!g(t,x,y,m) = {K (p-1) c p-1 (q-1) d q-1 / π} (t+c) -p {(x 2 + y 2 )e -a(m-m0) + d} -q.! Two reasons for this:!!1) It is easy to see how to replace g by another density.!!2) For each earthquake (t j,x j,y j,m j ),!!! g(t-t j, x-x j, y-y j; M j ) dt dx dy = K e a(m-m0).! 8

9 Actually, parametric point process estimation is hard!!!!!!!!! Main obstacles:!!a. Small problems with optimization routines.!!!extreme flatness, local maxima, choosing a starting value.!!b. The integral term in the log likelihood.!!!log(likelihood) = log{λ(t i,x i,y i )} - λ(t,x,y) dt dx dy.!!!!!the sum is easy, but the integral term is extremely difficult to compute.!!by far the hardest part of estimation (Ogata 1998, Harte 2010).! Ogata 98: divide space around each pt into quadrants and integrate over them.! PtProcess, Harte 10: uses optim() or nlm(). User must calculate the somehow.! Numerical approximation is slow and is a poor approx. for some values of Θ.! 1,000 computations x 1,000 events x 1,000 function calls = 1 billion computations.!!problem B. contributes to reluctance to repeat estimation and check on A.! 9

10 Suggestions (Schoenberg 2013). 1. Write the triggering function as a density in time and space. 2. Approximate the integral term as K e a(mj- M0). 3. Given data from time 0 to time T, estimate ETAS progressively, using data til time T/100, then 2T/100,..., and assess convergence. 10

11 What is behind this integral approximation?! Recall that log(likelihood) = log{λ(t i,x i,y i )} - λ(t,x,y) dt dx dy,!!! where λ(t,x,y) = µ ρ(x,y) + g(t-t j, x-x j, y-y j; M j ).! After writing g as a density,!!! g(t-t j, x-x j, y-y j; M j ) dt dx dy = K e a(m-m0),! if the integral were over infinite time and space.! Technically, in evaluating λ(t,x,y) dt dx dy, the integral is just over time [0,T] and in some spatial region S.! I suggest ignoring this and approximating g(t-t j, x-x j, y-y j; M j ) dt dx dy ~ K e a(m-m0).! Thus, λ(t,x,y) dt dx dy = µt + g(t-t j, x-x j, y-y j; M j ) dt dx dy!!!!!! = µt + g(t-t j, x-x j, y-y j; M j ) dt dx dy!!!!! ~ µt + K e a(mj-m0),! which doesn t depend on c,p,q,d, and is extremely easy to compute.! 11

12 Simulation and example.! 100 ETAS processes, estimated my way.! 12

13 Simulation and example.!!% error in µ, from 100 ETAS simulations.! 13

14 Simulation and example.! CA earthquakes from 14 months after Hector Mine, M 3, from SCSN / SCEDC, as discussed in Ogata, Y., Jones, L. M. and Toda, S. (2003). 14

15 Simulation and example.! Hector Mine data. Convergence of ETAS parameters is evident after ~ 200 days.! Note that this repeated estimation over different time windows is easy if one uses the integral trick. This seems to lead to more stability in the estimates, since a little local maximum is less likely to appear repeatedly in all these estimations.! 15

16 Simulation and example.! Hector Mine data. ETAS parameter estimates as ratios of final estimates.! 16

17 Parametric point process estimation. Suggestions (Schoenberg 2013). 1. Write the triggering function as a density in time and space. 2. Approximate the integral term as K e a(mj- M0). 3. Given data from time 0 to time T, estimate ETAS progressively, using data til time T/100, then 2T/100,..., and assess convergence. It s fast, easy, and works great. Huge reduction in run time. Enormous reduction in programming. 17

18 3. Separable Point Process estimation for covariates. for example, when modeling wildfire incidence as a function of fuel moisture, precipitation, temperature, windspeed,...

19 Burning Index (BI) NFDRS: Spread Component (SC) and Energy Release Component (ERC), each based on dozens of equations. BI = [10.96 x SC x ERC] 0.46 Uses daily weather variables, drought index, and vegetation info. Human interactions excluded. Predicts: flame length, area/fire? # of fires? # of fires? Total burn area?

20 Some BI equations: (From Pyne et al., 1996:) Rate of spread: R = I R ξ (1 + φ w + φ s ) / (ρ b ε Q ig ). Oven-dry bulk density: ρ b = w 0 /δ. Reaction Intensity: I R = Γ w n h η M η s. Effective heating number: ε = exp(-138/σ). Optimum reaction velocity: Γ = Γ max (β / β op ) A exp[a(1- β / β op )]. Maximum reaction velocity: Γ max = σ 1.5 ( σ 1.5 ) -1. Optimum packing ratios: β op = σ A = 133 σ Moisture damping coef.: η M = M f /M x (M f /M x ) (M f /M x ) 3. Mineral damping coef.: η s = S e (max = 1.0). Propagating flux ratio: ξ = ( σ) -1 exp[( σ 0.5 )(β + 0.1)]. Wind factors: σ w = CU B (β/β op ) -E. C = 7.47 exp( σ 0.55 ). B = σ E = exp(-3.59 x 10-4 σ). Net fuel loading: w n = w 0 (1 - S T ). Heat of preignition: Q ig = M f. Slope factor: φ s = β -0.3 (tan φ) 2. Packing ratio: β = ρ b / ρ p.

21 Good news for BI: BI is positively associated with wildfire occurrence. Positive correlations with number of fires, daily area burned, and area per fire. Properly emphasizes windspeed, relative humidity.

22 Some problems with BI Correlations are low.! Corr(BI, area burned) = 0.09! Corr(BI, # of fires) = 0.13! Corr(BI, area per fire) = 0.076! Corr(date, area burned) = 0.06! Corr(windspeed, area burned) = Too high in Winter (esp Dec and Jan) Too low in Fall (esp Sept and Oct)

23

24

25

26 Separable Estimation for Point Processes Consider λ(t, x 1,, x k ; θ). [For fires, x 1 =location, x 2 = area.]

27

28

29

30

31 Separable Estimation for Point Processes Consider λ(t, x 1,, x k ; θ). [For fires, x 1 =location, x 2 = area.] Say λ is multiplicative in mark x j if λ(t, x 1,, x k ; θ) = θ 0 λ j (t, x j ; θ j ) λ -j (t, x -j ; θ -j ), where x -j = (x 1,,x j-1, x j+1,,x k ), same for θ -j and λ -j If λ is multiplicative in x ~ j and if one of these holds, then ^ θ j, the partial MLE, is consistent (Schoenberg 2015): S λ -j (t, x -j ; θ -j ) dµ -j = γ, for all θ -j. S λ j (t, x j ; θ j ) dµ j = γ, for all θ j. ^ ~ S λ j (t, x; θ) dµ = S λ j (t, x j ; θ j ) dµ j = γ, for all θ.

32 Individual Covariates: Suppose λ is multiplicative, λ j (t,x j ; θ j ) = f 1 [X(t,x j ); β 1 ] f 2 [Y(t,x j ); β 2 ], X and Y are independent, and the log-likelihood is differentiable w.r.t. β 1, then the partial MLE of β 1 is consistent. Suppose λ is multiplicative and the jth component is additive: λ j (t,x j ; θ j ) = f 1 [X(t,x j ); β 1 ] + f 2 [Y(t,x j ); β 2 ]. If f 2 is small S f 2 (Y; β 2 ) 2 / f 1 (X; ~ β 1 ) dµ / Τ -> p 0], then the partial MLE β 1 is consistent.

33 Impact Model building. Model evaluation / dimension reduction. Excluded variables.

34 (sq m) r = 0.16

35

36 (sq m) (F)

37

38 Model Construction Wildfire incidence seems roughly multiplicative. (only marginally significant in separability test) Windspeed. RH, Temp, Precip. Tapered Pareto size distribution f, smooth spatial background µ. [*] λ(t,x,a) = f(a) µ(x) β 1 exp(β 2 RH + β 3 WS) (β 4 + β 5 Temp)(max{β 6 - β 7 Prec,β 8 }) Relative AICs (Poisson - Model, so higher is better): Poisson RH BI Model [*]

39 Comparison of Predictive Efficacy False alarms per year % of fires correctly alarmed BI 150: Model [*]: BI 200: Model [*]:

40 Earthquake weather example, Schoenberg model with temp model without temp

41 Conclusions: Start with nonparametric estimates of triggering and bg rate. For a parametric model with clustering, write the triggering function g as a density times K, so you can estimate by MLE, essentially approximating the integral of g as K. Build initial attempts at point process models by examining each covariate individually.