The BayGMST statistical model

This vignette describes the statistical model implemented by fit_baygmst(), connecting the code to the underlying statistics and to the literature it descends from. For a hands-on walkthrough of the software, see vignette("baygmst-intro", package = "BayGMST").

Background

Global mean surface temperature (GMST) is a fundamental state variable of Earth’s climate system and a central target of paleoclimate reconstruction efforts. Because direct instrumental measurements extend back only about 150 years, estimates of GMST over longer periods must be inferred from indirect observations preserved in paleoclimate archives: tree rings, glacier ice, corals and sclerosponges, lake and marine sediments, and historical documents (Mann et al., 1999; Moberg et al., 2005). A variety of statistical strategies have been developed to extract the climate signal contained in these “proxies” (Tingley et al., 2012; Smerdon and Pollack, 2016), whether the target is a spatial climate field (Smerdon et al., 2023) or an aggregated index such as GMST.

Intellectual ancestry

BayGMST’s inference framework builds directly on the work of Barboza et al. (2014, 2019) and Wang (2020), who utilized a reduced proxy Bayesian approach that leverages not only the climate proxy network, but also estimates of past forcings (e.g., greenhouse gas concentrations, solar, and volcanic activity), which are known to influence GMST. The model below is their hierarchical formulation; BayGMST packages it, with the implementation choices noted in this vignette.

The hierarchical model

The approach is based on computing a “reduced proxy” \boldsymbol{P} = (P_1, \ldots, P_t, \ldots, P_n)^T, an aggregate of the available paleoclimate proxy observations. Several methods exist to do so, like principal component regression, partial least squares, sliced inverse regression, or a simple composite (Barboza et al., 2019); see reduce_proxies() for the methods implemented here, and as_baygmst_proxy() for supplying a reduced or composited series built elsewhere. This quantity is then modeled as autoregressive and linearly dependent on the GMST, denoted \boldsymbol{T} = (T_1, \ldots, T_t, \ldots, T_n)^T (subscript t refers to time in years CE).

The model’s data level:

P_t = \phi_P P_{t-1} + \alpha_0 + \alpha_1 T_t + \sigma_P \epsilon_t, \qquad \epsilon_t \overset{\text{iid}}{\sim} N(0, 1)

where \phi_P denotes the lag-1 autocorrelation of P, the \alpha’s are coefficients, and \sigma_P controls the amplitude of the unit-variance noise term \epsilon.

The process level models T_t as autoregressive and linearly dependent on the climate forcings G_t (greenhouse gas concentrations), V_t (related to volcanic aerosol optical depth), and S_t (solar irradiance), already transformed as described below (see transform_forcings()):

T_t = \phi_T T_{t-1} + \beta_0 + \beta_G G_t + \beta_V V_t + \beta_S S_t + \sigma_T \eta_t, \qquad \eta_t \overset{\text{iid}}{\sim} N(0, 1)

where \phi_T denotes the lag-1 autocorrelation of \boldsymbol{T}, the \beta’s are constant coefficients, and \sigma_T governs the amplitude of \eta_t. Gaussian white noise for both \boldsymbol{\eta} and \boldsymbol{\epsilon} provides good results (Wang, 2020). Here |\phi_P|, |\phi_T| < 1 and \sigma_P, \sigma_T > 0.

Missing (pre-instrumental) temperatures are treated as parameters and estimated jointly with the model parameters, so the posterior yields both the reconstruction and its uncertainty.

Mapping to the code

Notation Stan parameter R argument / role
T_t y (transformed parameter) instrumental_T (observed part), y_mis (reconstructed part)
P_t z (data) proxy, see reduce_proxies() / as_baygmst_proxy()
\phi_P phi_R
\sigma_P sigma_z
\alpha_0, \alpha_1 alpha0, alpha1
\phi_T phi_T
\sigma_T sigma_y
\beta_0, \beta_G, \beta_V, \beta_S beta0, betaG, betaV, betaS
G_t, V_t, S_t G, V, S (data) forcing_G, forcing_V, forcing_S

Note on naming: the Stan code calls the proxy-equation parameters phi_R and sigma_z (“R” for “reduced proxy”, matching this implementation’s single-series design); this vignette writes them \phi_P and \sigma_P for readability. They are the same quantities.

Priors

As specified in the Stan model’s model block (weakly informative):

\begin{aligned} \alpha_0, \alpha_1, \beta_0, \beta_G, \beta_V, \beta_S &\sim N(0, 0.5) \\ \phi_P, \phi_T &\sim N(0, 0.3) \\ \sigma_P, \sigma_T &\sim \text{Exponential}(1) \end{aligned}

Because the data level estimates its own intercept \alpha_0 and slope \alpha_1, the proxy series does not need to be pre-calibrated to temperature units; these priors do, however, assume P_t is on roughly unit scale (a standardized composite or a temperature-anomaly-scale series). See as_baygmst_proxy().

Forcing data

Per Barboza et al. (2019, section 2.3), the three external forcings used in this family of models are:

All these series measure perturbation to the top of the atmosphere energy budget, and are thus expressed in \mathrm{W\,m^{-2}}, following the climate science convention. To effect the conversion from the original units (aerosol optical depth for V_t; mole fraction for G_t), transform_forcings() implements

G'_t = \texttt{co2\_coef} \times \log(G_t / \texttt{co2\_c0}), \qquad V'_t = -\lvert\texttt{vol\_coef}\rvert \,(1 - e^{-V_t}),

where co2_coef = 5.35 is the simplified CO2 radiative forcing coefficient of Myhre et al. (1998), adopted throughout the IPCC assessment reports; co2_c0 = 280 ppm is the standard pre-industrial atmospheric CO2 reference concentration; and vol_coef = 25 is the stratospheric aerosol optical depth-to-forcing scaling factor (in \mathrm{W\,m^{-2}} per unit optical depth) used in IPCC AR5, traced to the GISS ModelE lineage of Hansen et al. (2005). Solar irradiance is simply centered on its own mean, S'_t = S_t - \bar{S}. These physically motivated formulations are a slight departure from the transformations used by Barboza et al. (2014, 2019).

Modeling assumptions worth knowing

References