SIMULATION-BASED INFERENCE IN NONLINEAR STATE-SPACE MODELS: APPLICATION TO TESTING THE PERMANENT INCOME HYPOTHESIS? Roberto S. Mariano University of Pennsylvania, USA and Hisashi Tanizaki Kobe University, Japan (Revised November 1996) *Presented at the Conference on Simulation-Based Inference in Econometrics - Minneapolis Federal Reserve Bank, Minnesota, November 17-18, 1995 1SIMULATION-BASED INFERENCE IN NONLINEAR. STATE-SPACE MODELS: APPLICATION TO TESTING THE PERMANENT INCOME HYPOTHESIS Roberto S. Mariano and Hisashi Tanizaki November 1996 Abstract This paper deals with the wpplication of nonlinear /nonnormal filtering tech- niques to testing the permanent income hypothesis. We set up a general non- inear state-space model for unobserved permanent consumption, extending the hi rature on this celebrated topic vo a simultaneous consideration of transitory consumption, variable interest rates, and nonlinearity of the Euler equation de- rived from utility maximization by the representative agent. Because of the nonlinear complexities in the model, stochastic simulations are used to obtain numerical maximum likelihood estimates of unknown model parameters. Our modeling approach and use of stochastic simulations also allow us to estimate unobserved permanent consumption over the sample period. The permanent income hypothesis is stated in terms of unknown parame- ters in the formulated state-space model. Various nonlinear filters are used to estimate these parameters and calculate likelihood ratio tests of the permanent income hypothesis. Annual data for the US and Japan are used in these calcula- tions. The alternative filters used are: the extended Kalman filter, Kitagawa’s numerical integration filter, importance sampling filter, density-based Monte Carlo filter, and rejection sampling filter. Alll the filtering techniques used here produced very similar results ~ perhaps because our model is only mildly nonlinear. Our numerical results clearly reject the permanent income hypothesis. Our estimates of permanent consumption én the US and Japan point to uifferences in degree of rationality in consumer behavior in the US and Japan. Japan behaves somewhat more rationally than the US ~ the estimated ratio of permanent consumption to total is substantially higher in Japan than in the US — and, as the calculations for the oil-crisis period indicate, even more so in a recessionary period.1 INTRODUCTION Nonlinear filtering has been developed in various directions beyond the classic lin- ear/normal filtering theory of Kalman (1960) and Kalman and Bucy (1961). One approach is to apply a Taylor Series expansion to linearize the nonlinear mea- surement and transition equations. The linear recursive algorithm is then applied to this modified system. Procedures in this gendre ~ such as the extended Kalman filter and the second-order nonlinear filter are discussed in Wishner, Tabaczynski and Athans (1969), Sorenson and Alspach (1971), Alspach and Sorenson (1972), Gelb (1974), Anderson and Moore (1979) and Tanizaki and Mariano (1992). Proceeding in another direction, others have sought to avoid normality assump- tions by working on recursive algorithms for upd2ting probability density functions. Kitagawa (1987) and Kramer and Sorenson (1983) evaluate densities through nu- merical integration. Alternatively, simulation techniques are a natural tool for eval- uating these densities. Monte-Carlo integration with Gibbs sampling is explored in Carlin, Polson and Stoffer (1992). Monte-Carlo integration with importance sam- pling in this context is discussed in Tanizaki (1993), Tanizaki and Mariano (1994) and Mariano and Tanizaki (1995). Further modifications of these procedures — using rejection samplii g - are discussed also in Tanizaki and Mariano (1995). In this paper, we consider the application of these nonlinear/nonnormal filter- ing techniques to testing the permanent income hypothesis. Numerous papers have revisited this celebrated hypothesis. Flavin (1981), Hall and Mishkin (1982), and Campbell and Mankiw (1987) examined this issue taking transitory consumption into account. Hayashi (1985a, 1985b) introduced liquidity constrainsts and durable goods in testing this hypothesis while Mankiw (1981) and Muelbauer (1983) intro- duced variable interest rates into the model. Here, for our analysis of the permanent income hypthesis as proposed by Hall (1978), we set-up a nonlinear state space model for unobserved permanent consumption, which simultaneously takes into ac- count transitory ccasumption, variable interest rates, and nonlinearity of the Euler equation. With our approach - of formulating a nonlinear state-space model of per- manent consumption - the permanent income hypothesis can be stated in terms of unknown parameters in the model. Because of nonlinearities in the model, nonlinear filtering techniques based on stochastic simulations are used to perform two major steps in the analysis: © estimate unknown model parameters and calculate likelihood ration tests of the permanent income hypothesis, and * estimate unobserved permanent consumption over the sample period.Thus, our approach differs from earlier studies in two significant aspects. First, our analysis copes at the same time with these issues of of presence of transitory consumption, variability of interest rates, and nonlinearity of first order conditions. Secondly, stochastic simulation play a key role in the empirical implementation of the analysis because of nonlinear complexities in the model. Annual data for the US and Japan are used in these calculations. Filters are also used to estimate permanent consumption in the U.S. and Japan. Difference in permanent consumption as a percent of total are observed for these two coun- tries showing differences in degree of rationality in consumer behavior in these two countries Before going into this application, we provide a brief overview of filtering in nonlinear state space models. A more complete discussion is in Tanizaki and Mariano (1994, 1995). 2 NONLINEAR STATE-SPACE MODELS AND NONLINEAR FIL- TERING Consider the following nonlinear state-space model: Measurement equation: ys = hy (a, €+) (Q) Transition equation: a = fr (a+-1,m) (2) The vector y; is observed over the sample period t = 1,2,..., 7’ while ay is a vector of unobserved “state” variables. The vector functions hy and f, are specified known functions. The disturbances (1,7) are assumed to have zero means and to be jointly independent over time. Their probability distributions are known and are not necessarily normal. Given the sample {y: : where =1,2,...,7}, the main problem is to evaluate E (a,| Is), information at time s 3 {is ¥2s00Yo}- ®) Standard terminology refers to the problem as Prediction — if r>s, Filtering - if r=s, (4) Smoothing - if r P (ae | fea) /P (ye | a) The denominator in (10) is obtained by integrating the numerator with respect to a. P(y; | ar) is obtained directly from the measurement equation with the function hy and the distribution of ¢ both specified. Finally, P(a: | J:-1) comes from the prediction algorithm in (9). Formulas (9) and (10) provide the recursive algorithm (as summarized in the following flow chart) for prediction and filtering, starting from initial conditions, say ao, for the state variables.Recursive Algorithm Nonlinear Prediction and Filtering Observations Measurement Fite/ no > Plyslar) Pah) — Pll) = h — Plyler) Peal) P(elh) — we —+ Pipes) —+ Pimalb) ete. Transition Equation P (as\a0) P (@2|01) P (aslaz) Next, we move on to the calculation of P(az41 | J) and P (a, | J;). As we pointed out earlier, Kitagawa (1987) and Kramer and Sorenson (1988) used numerical inte- gration in their calculations. Monte Carlo simulation with importance sampling proceeds as follows. Denoi2 the importance density on a; by P, (ai). Then we can write Pau | h)= [Plo | at) P (ae | Le) /Pa (ax)] - Pa (a+) dow (11) And, for n random draws from P, (a), say {@ :i=1,2,....n}, we obtain the fol- lowing approximations to (9) and (10). Paugs | Le) = DP (aes | ie) - wee (ie) /n P (ay | Le) = P (yt | an) P (ae | Le-1) /De Bi = YP (ye | Ga) - ween Ge) /n where We) (Bie) = P (Bie |e) [Pp (Bit) We jena (Git) = P (Bie | Je-1) /P» (Git) 6 (12) (13) (14) (15) (16)Recursion formulas for wy), and w;|4-1 are given in Tanizaki and Mariano (1995). The above formulas also give us a recursive approximation to the likelihood of the observations. This is particularly useful for a numerical maximization of the likelihood funeticn, Pr) = BP ta) =TLLP (ue | a) P(ae | es} den a7) & (O/myF) EP (ve | Ge) ways Gi). As before, the Gj; are random draws from the importance density P, (a1). A modified simulation approach, using rejection sampling as proposed in Tanizaki and Mariano (1995) gets random draws Gj, , from P(a, | J,). For prediction, we then have, Poca | Le) ® DP (ae4s | Ginete) /r- (18) This is because of (9) where P (au4a | Le) = Bact? (241 | a2)- Given &,1}1, we can generate G,r41]1 by Bieaaye = fess ( Biveyes Aiea)» (19) For any integrable function g, we get E(9(ai41) | h) = Xs (Gie4r4e) + (20) For filtering, how do we generate &,1\ ? The formulas are Pa: | L-1) = J P (at, at-1| Lt-1) dara @ = J P(atlar-1):P (1-1 | f-1) der. This implies that, from (10), P(a|I)= MJ f Nedax ' where Nie [ Plaid ox)-P (os) ax-1)+ P (ara | een) dara. (22) Thus, P (a: | J) is proportional to N;, which, in turn, can be approximated by Ne= Yo P (uel ax) P (at | Geeajeea) /n- (23) To get draws recursively from this approximate filtering density in (23), apply the following “rejection” sampling scheme:1. Given (Gie-apens Gut. nt-1|t-1)s choose one of these randomly and get, say wv . Sample P (a; | G.4-1) to get a. calculated as Ont = fe (Gata Fie) (24) 3. Calculate w (as1,44), where we construct w; such that OS wae) $1, w (aut) & P (ys [ant 4, Include a, in the draw from P (a; | J;) with probability calculated in step 3 and denote this by a,¢4:- a . Continue doing this until you get n draws &;,4)1- ° . Finally, calculate E (g(a) |W) = D9 (Hee) /n (25) 3 TESTING THE PERMANENT INCOME HYPOTHESIS To test the permanent income hypothesis, we now consider the following state-space model Measurement Equation C= CP+ OP ten (26) Transition Equation @\7 ARs (#) =itn, 7) where (€,,7) are bivariate normal, independent over time, with mean zero and covariance matrix (am 0 ) 0. @ (28) The above model is driven by the behavior of the representative agent, with the following definitions:per capita total consumption per capita permanent consumption consumption component independent of the permanent income hypothesis e= C7 per capita transitory consumption, assumed random, with zero mean = gross rate of return on savings between t and t +1 per capita disposable income = population size sep = random disturbance term in the transition equation CP and CT are unobservable, while C;, ¥;, Lr, R, are observed. CP is treated as the state variable. This is estimated by nonlinear filtering; in the process, we can thus estimate permanent and transitory consumption separately. A more detailed discussion of this model follows. 3.1 Measurement Equation The measurement equation is simply the identity that total consumption is the sum of permanent consumption, transitory consumption and a part of consumption which is independent of the permanent income hypothesis. ‘The assumption on transitory consumption is based on earlier authors’ treatment of the subject. In a cross-sectional framework, Lor= is assumed by Friedman(1974), where CZ denotes transitory consumption of the it household at time t. Furthermore, Friedman(1957) assumed that C7 is indepen- dent of the permanent income, transitory income and permanent consumption-see Branson (1979). Aggregate transitory consumption is represented as cT= Ler/L. Assuming that OF are identically and independently distributed with mean zero and variance (o-Y;)? for all i and t, the transitory consumption of the representative agent (i.e., C7 ) is given by a random shock with mean zero and variance (0.¥;)?/L:.It might be plausible to assume that the transitory consumption increases (decreases) as the income level increases (decreases). Cj represents an exogenous part of consumption which does not depend on the permanent income hypothesis proposed by Hall (1978). It is well known that variables other than lagged consumption appear to play a significant role in the determination of current consumption (see Diebold and Nerlove (1989)). Accord- ingly, C7 is a part of consumption which depends on other variables such as income. Therefore, it is assumed in this paper that C7 is a function of income, i.e., Cf =nhi+ Xia, (29) where y and 72 are unknown parameters to be estimated. Under the permanent income hypothesis, C7 = 0, which is equivalent to 7 = Y2 = 0 in (29). ¥; denotes an instrument variable of ¥;, because ¥; is correlated with 1m. In this section, to test whether the permanent income hypothesis holds, we use the likelihood ratio test for the null hypothesis Ho: = 2 = 0. (30) 3.2 Transition Equation The transition equation corresponds to the Euler equation, which is derived as fol- lows. Consider the problem of choosing a consumption sequence {C?} by the repre- sentative agent which maximizes the following expected utility: A(S B'u(C2)), 7 subject to: Aur = Ri(Ar+Wi- Ci), G = CP+cr+cT, (31) where Ap is given. The representative utility function u(.) is twice continuously differentiable, bounded, increasing and concave. A;4; is the stock of assets at the beginning of time t+1, W; is noncapital or labor income at t, and R; is the gross rate of return on savings between t and t+1. E,(-) denotes the mathematical expectation, given information known at t. f is the discount rate. It might be plausible in empirical studies that the discount rate ( is less than one. However, Kocherlakota (1990) showed that well-behaved competitive equilibria with positive interest rates 2Since Y; is correlated with C;, we use a proxy variable of ¥; as Y; to exclude correlation between Y; and Cy. In this paper, the instrument variable ¥; is taken as the predicted value of ¥; regressed on a constant term, a time trend, Y-; and C,-1. 10may exist in infinite horizon growth economies even though individuals have discount factors larger than one, which implies that when per capita consumption is growing over time, it is possible for equilibria to exist in representative consumer endowment economies even though f > 1. Therefore, we do not have to pay much attention to the possibility of the discount rate being greater than one. Thus, under the above setup, maximizing the expected utility with respect to the permanent consumption sequence {C?},? we obtain the Euler equation, BReaw'(CP)\ _ (Bee) =1l+m, (32) where the error term 7; is assumed to be truncated normal. u’(-) represents the first derivative of the underlying utility function. Taking the utility function as u(CP) = log(C?), (33) The Euler equation corresponding to this particular form of the utility function reduces to the transition equation in (27). 4 ESTIMATION RESULTS The likelihood function of the innovation form is maximized by a simple grid search with respect to the unknown parameters (i.€., 71, 12) 0, B, and o»). We test the permanent income hypothesis (i.e., the null hypothesis Ho : 71 = 2 = 0) and estimate permanent consumption for both Japan and the U.S. *Usually, the utility function is maximized with respect to total consumption C;, not permanent consumption C?. We assume that the utility function depends on C, as well as wx, where us denotes a sum of the transitory consumption and the other part of consumption independent of the permanent income hypothesis, i.e., «= Cy +C7 in this paper. Taking the utility function as u(C; — :), we obtain the Euler equation (32). ‘Note that we have 1+ > 0. Therefore, the exact distribution of m is represented as: Palm) -1/oy is enough small. Accordingly, we approximate the density function of m as the normal density, ie., m ~ N(0,02). ll4.1 Data Annual data from 1955 to 1993 are used for Japan and the U.S., and the estimation period is from 1957 to 1993. For Japanese data, C;, Y;, P; and L; denote per capita final consumption cx- penditure of households (Japanese yen at 1985 price, annual data, per capita data divided by number of population), per capita national disposable income of house- holds (Japanese yen at 1985 price, annual data, per capita data divided by number of population), implicit price deflator of final consumption expenditure of households (1985=1.00), and number of population, taken from Annual Report on National Accounts (Economic Planning Agency, Government of Japan). The gross rate of re- turn on savings (Rx) is defined as Ry = (1 +1,/100)P;/Pi41, where r; is installment savings of banks (annual rate, percent). For the U.S. data, C;, ¥%; P; and L; denote per capita personal consumption expenditure (U.S. dollars at 1987 price, annual data, per capita data divided by number of population), per capita national disposable income (U.S. dollars at 1987 price, annual data, per capita data divided by number of population) and implicit price deflator of personal consumption expenditure (1987=1.00), and number of population, from Economic Report of the President. The gross rate of return on savings (Rj) is defined as R; = (1+1;/100)P;/Pia1, where r; is Aaa (annual rate, percent). In Tables 1 ~ 5, each parameter is estimated by a simple grid search method, i.e., the log-likelihood function is maximized by changing the parameter value by 0.001 for B, by 0.0001 for o,, 0.001 for 1 and 49, and 0.1 for o., respectively. log L in Tables 1 - 5 denotes the maximized log-likelihood function. 4.2 Nonlinear Filtering Techniques ‘We estimate the state-space model (26) and (27) by five nonlinear filtering methods: Extended Kalman Filter in Table 1 and Figure 1, Numerical Integration Filter in Table 2 und Figure 2, Importance Sampling Filter in Table 3 and Figure 3, Density- Based Monte-Carlo Filter in Table 4 and Figure 4, and Rejection Sampling Filter in Table 5 and Figure 5. The parameter estimates are similar for Tables 1 - 5. In the extended Kalman filter, the transition equation (27) is approximated by the first-order Taylor series expansion and applied to the standard Kalman filter algorithm, i., the linear recursive algorithm. In the numerical integration filter, we take n = 200. The first half of n nodes (ie., m nodes) are obtained from [aj,_y—4,/Diy_y»@%j-1+44/Dyq] and the second 12half of n nodes (i.e., n ~m nodes) are from [aj ~ 44/Sj,,a3, + 4,/S5,)- For both intervals, the distance between two nodes are equal. a is, m nodes are from oj, +42 a and n — m nodes are from . 2i-1-m Qt a ae [ria where i= 1,-++,m and m = 100 is taken. afy_y, Bij. 1» @jj, and D4, are obtained from the extended Kalman filter algorithm. In the importance sampling filter, n = 200 random draws are generated from the following importance density: 2 fos = lvvat, 45% Pr(ai) = SN (ahp1 AEG ea) + 5M (Oi ABH) where aj), 4, Dia» @jj, and Dj, are obtained from the extended Kalman filter algorithm. For density approximation, the importance density should have broader tails than the prediction density and the filtering density (see Tanizaki (1993), Tanizaki and Mariano (1994) and Mariano and Tanizaki (1995)). Therefore, for the im- portance density, we choose a larger variance than the variance obtained from the extended Kalman filter. In the density-based Monte-Carlo filter, n=10,000 random draws are generated from the transition equation (27). In the rejection sampling filter, the random draws are generated from the tran- sition equation (27), where the number of random draws is taken as n = 1000. The acceptance probability is given by the exponential part of the normal density obtained from the measurement equation (26), i.e., ah) = -——_ _(¢,_cr_o: ’) wlan) = em (—seyeg (Ge OF - C7?) ‘Under the above setup of each nonlinear filter, the unknown parameters (i.e., 8, on; 71, 12 and g;) are estimated in Tables 1 - 5. The parameter estimates are similar for all the tables. For each country in each table, we have three estimation results, ie., (i) 11 =0 and 2 = 0, (i) 71 # 0 and = 0, and (iii) 1, # 0 and y #0. For all the 13estimation methods and both the countries, the hypothesis Ho : 1 = 12 = 0 is rejected according to the likelihood ratio test.* The likelihood ratio statistics are given in Table 6. This indicates that the per- manent income hypothesis does not hold for both Japan and the U.S. This test takes into account transitory consumption, nonlinearity of the Euler equation and the variable interest rate. Numerous earlier papers testing the permanent income hypothesis consider some of these three issues, but none of them deals with the three issues simultaneously. Our result shows that even if all three issues are in- cluded in the model, we still find significant evidence against the permanent income hypothesis. Next. in Figures 1 - 5, we plo* the ratio of estimated permanent consumption relative to total consumption, i.e., 10007,/C;, where C}, denotes the filtering esti- mate of per capita permanent consuriiption at time ¢, for Japan and the U.S. A high ratio of permanent consumption implies that a large amount of people behave under the permanent income hypothesis. In the U.S., about 15 % - 20 % of consumption is based on the permanent income hypothesis during the estimation period. The ratio is almost constant over time although it is slightly small in 1974 and 1981, the years of the two oil crises. The ratio in Japan is larger than that in the U.S. for all the periods. Around 25 % of total consumption is permanent consumption over the . whol? sample period except for the first-oil crisis. During the first oil shock, the ratio of permanent consumption to total consumption fell to about 20% in Japan. For the U.S., however, there was hardly any downward adjustment in the ratio. These aumerical results indicate that Japan behaves somewhat more rationally than the U.S. and the experience in the first oil crisis indicates that this occurs even more so in a recessionary period. In Japan, the first oil shock brought on a serious recession - with inflation and wage increases ranging between 23% and 25% in 1974. From this experience, the Japanese economy learned quickly and managed to steer a more stable course in the second oil shock. In 1981, wage and price increases in Japan were kept at a normal rate of 5%. Consequently, as Figures 1-5 show, there was hardly any adjustment in Japanese permanent consumption as the second oil shock took place. "Note that the likelihood ratio test statistic of the hypothesis Ho : 11 = 12 = 0 is asymptotically distributed as a Chi-squared random variable with two degrees of freedom. The critical values of a Chi-squared distribution with two degrees of freedom are given by 5.99 for 0.05 % and 9.21 for 0.01 %, respectively. 145 SUMMARY In this paper, for both Japan and the U.S., we have formulated a nonlinear state space model to test the permanent income hypothesis. The state-space model is based on the Euler equation derived from a utility maximization by a representative agent. A contribution of the paper to the wide literature on this celebrated topic comes from the fact that our approach simultaneously takes into account the pres- ence of transitory consumption, nonlinearity of the Euler equation and the effects of variable interest rates. A second contribution of the paper shows how stochastic sim- ulations, applied to nonlinear filtering, are utilized not only to calculate numerical maximum likelihood estimates of model parameters but also to estimate unobserved permanent consumption over the sample period. Annual consumption data from 1955 to 1993 are used for both Japan and the USS. All the filtering techniques used here produced similar results. Taking into account transitory consumption, nonlinearity of the Euler equation and the variable interest rate simultaneously, our results reject the permanent income hypothesis and point to differences in degree of rationality in consumer behavior in the U.S. and Japan. References Alspach, D.L. and H.W. Sorenson, 1972, “ Nonlinear Bayesian Estimation Using Gaussian Sum Approximations, ” IEEE Transactions on Automatic Control, Vol.AC-17, No.4, pp.439 - 448. Anderson, B.D.O. and J.B. Moore, 1979, Optimal Filtering, Prentice-Hall, New York. Branson, W.H., 1979, Macroeconomic Theory and Policy (second edition), Haper & Row, Publishers, Inc. Campbell, J-Y. and N.G. Mankiw, 1987, “ Are Output Fluctuations Transitory?, ” Quarterly Journal of Economies, Vol.102, pp.857 ~ 880. Diebold, F.X. and M. Nerlove, 1989, “ Unit Roots in Economic Time Series: A Selective Survey, ” in Advances in Econometrics, Vol.8, pp.3 - 69, JAI Press. Flavin, M.A., 1981, “ The Adjustment of Consumption to Changing Expectations about Future Income, ” Journal of Political Economy, Vol.89, No.5, pp.974 - 1009. Friedman, M., 1957, A Theory of the Consumption Function, Princeton University Press. 15Gelb, A., 1974, Applied Optimal Estimation, MIT Press. Hall, R-E., 1978, “ Stochastic Implications of the Life Cycle-Permanent Income Hypothesis: Theory and Evidence, ” Journal of Political Economy, Vol.86, No.6, pp.971 - 987. Hall, R.E., 1990, The Rational Consumer, The MIT Press. Hall, RE. and F.S. Mishkin, 1982, “ The Sensitivity of Consumption to Transitory Income: Estimates from Panel Data on Households. ” Econometrica, Vol.50, No.2, pp.461 ~ 481. Harvey, A.C., 1989, Forecasting, Structural Time Series Models and the Kalman Filter, Cambridge University Press. Hayashi, F., 1985a, “ The Effects of Liquidity Constraints on Consumption: A Cross-sectional Study, ” Quarterly Journal of Economics, Vol.100, pp.183 — 206. Hayashi, F., 1985b, “ The Permanent Income Hypothesis and Consumption Dura- bility: Analysis Based on Japanese Panel Data,” Quarterly Journal of Eco- nomics, Vol.100, pp.1083 ~ 1113. Kocherlakota, N.R., 1990, “ On the ‘Discount’ Factor in Growth Economies,” Jour- nal of Monetary Economics, Vol.25, pp.43 ~ 48. Kalman, R.E., 1960, “ A New Approach to Linear Filtering and Prediction Prob- lems,” Journal of Basic Engineering, Transactions ASME, SerD, Vol.82, pp.35 - 45. Kalman, RE. and R.S. Bucy, 1961, “ New Results in Linear Filtering and Prediction Theory, ” Journal of Basic Engineering, Transactions ASME, Ser.D, Vol.83, pp.95 - 108. Karlin, B-P., N.G. Polson and D.S. Stoffer, 1992, “ A Monte Carlo Approach to Nonnormal and Nonlinear State Space Modeling, ” Journal of the American Statistical Association, Vol.87, pp.493 ~ 500. Kitagawa, G., 1987, “ Non-Gaussian State-Space Modeling of Nonstationary Time Series,” Journal of the American Statistical Association, Vol.82, pp.1032 — 1063 (with discussion). Kramer, S.C. and Sorenson H.W., 1988, “ Recursive Bayesian Estimation Using Piece-wise Constant Approximations, ” Automatica, Vol.24, No.6, pp.789 - 801. 16Mankiw, N.G., 1981, “ The Permanent Income Hypothesis and The Real Interest Rate, ” Economics Letters, Vol.7, pp.307 - 311. Mariano, R.S. and H. Tanizaki, 1995, “ Prediction of Final Data with Use of Pre- liminary and/or Revised Data, ” Journal of Forecasting, Vol.14, No.4, pp.351 — 380. Muellbauer, J., 1983, “ Surprises in the Consumption Function, ” Economic Journal (Supplement), pp.34 - 50. Sorenson, H.W. and D.L. Alspach, 1971, “ Recursive Bayesian Estimation Using Gaussian Sums, ” Automatica, Vol.7, pp.465 - 479. Tanizaki, H., 1993, Nonlinear Filters: Estimaticn and Applications, (Lectu.e Notes in Mathematical Economics and Systems, No.400), Springer-Verlag. Tanizaki, H. and R.S. Mariano, 1992, “ Nonlinear Filters Based on Taylor Series Expansions, ” Unpublished Manuscript. Tanizaki, H. and R.S. Mariano, 1994, “ Prediction, Filtering and Smoothing in Nonlinear and Nonnormal Cases Using Monte-Carlo Integration,” Journal of Applied Econometrics, Vol.9, No.2, pp.163 ~ 179. Tanizaki, H. and R.S. Mariano, 1995, “ Nonlinear and Nonnormal State-Space Mod- cling with Monte-Carlo Stochastic Simulation, ” Journal of Econometrics, forthcoming. Wishner, R.P., J.A. Tabaczynski and M. Athans, 1969, “ A Comparison of Three Non-Linear Filters, ” Automatica, Vol.5, pp.487 — 496. 17Table 1: Extended Kalman Filter B on (a ee eee 1.060 0.0415 — — 41.5 -435.322 Japan | 1.065 0.0714 0.596 — 33.2 —411.017 1.059 0.0685 0.317 0.317 68.3 —409.503 0.989 0.0269 — — 269 -251.419 US. 0.992 0.0734 0.720 — 13.6 -—231.589 0.994 0.0789 0.506 0.246 49.3 —230.316 18Figure 1: Ratio of Permanent Consumption: Case 71 # 0 and 2 #0 Extended Kalman Filter 40 au ee ° % cere ee s . . ee 20 : - « er 10 1960 1970 1980 1990 Year eeooe Japan sooo US. Table 2: Numerical Integration Filter B On n eee: logL 1051 0.0405 — = — 405 —435.149 Japan | 1.039 0.0687 0.586 — 30.3 —410.193 1.040 0.0674 0.311 0.298 66.2 -409.056 0.986 0.0270 — — 26.5 251.805 US. | 0.965 0.07388 0.717 — 7.5 230.458 0.974 0.0828 0.495 0.232 43.6 —229.781 19Figure 2: Ratio of Permanent Consumption: Case 1 #0 and y #0 Numerical Integration Filter 40 30} o. ee feat oote, % . ° 204 *,e, patentee, oe 10 O+ T T T : T T 1960 1970 1980 1990 cress US. Table 3: Importance Sampling Filter B on n we % logL 0.0407 ~— — 424 —436.070 Japan 0.0722 0.602 — 20.3 408.768 1.031 0.0684 0.323 0.323 65.4 —409.117 0.985 0.0272 — — 274 -251.313 US. | 0.965 0.0735 0.718 — 82 —231.813 0.958 0.0809 0.509 0.248 45.6 —231.234 20Figure 3: Ratio of Permanent Consumption: Case 1 #0 and 72 #0 Importance Sampling Filter 40 30 % eer ua ne wee te, oe s+,,° aes ewkt* ae 10 | | 1960 1970 1980 1990 ‘ear eoooe Japan see US. Table 4: Density-Based Monte-Carlo Filter B On n 2% log L 1061 0.04144 — — 414 —446.788 Japan | 1.064 0.0713 0.597 — 33.1 —421.532 1.058 0.0685 0.318 0.317 68.3 —420.637 0.989 0.0269 — — 26.9 ~258.077 US. | 0.992 0.0733 0.720 — 13.6 —237.904 0.993 0.0789 0.506 0.246 49.3 —237.388 21Figure 4: Ratio of Permanent Consumption: Case 1 # 0 and 7 #0 Density-Based Monte-Carlo Filter 40 30 ve % fete as ° sooo oe 20 et Cgtte vere te, 10 or — T T T T T 1960 1970 1980 1990 Year ceooe Japan see U.S. Table 5: Rejection Sampling Filter ai 12 fe 13 1.060 0.0414 — — 41.5 446.784 Japan | 1.065 0.0714 0.596 — 33.0 —421.430 1.058 0.0686 0.316 0.317 68.3 420.418 0.982 0.0269 — — 26.9 —258.233 US. 0.992 0.0734 0.719 — 13.6 -—237.432 0.994 0.0789 0.506 0.246 49.3 —236.925 22Figure 5: Ratio of Permanent Consumption: Case 1 # 0 and 2 #0 Rejection Sampling Filter 40 30 ed % eee aan° a 20 : se Fs 10 o+ 7 : 7 7 T 7 1960 1970 1980 1990 Year eeooe Japan seeee US. Table 6: Likelihood Ratio Test Statistics (Ho: 1 = 72 = 0) Japan US. Figure 435.322 — 409.503) = 51.638 | 2(251.419 — 230.316) = 42.206 | Figure 1 2(251.805 — 229.781) = 44.048 | Figure 2 2(251.313 — 231.234) = 40.158 | Figure 3 pie 2¢ (446.788 — 420.637) (446.784 — 420.418) = 52.732 (258.077 — 237.388) = 41.378 | Figure 4 2 2 2(436.070 — 409.117) = 53.906 2¢ 2¢ (258.233 — 236.925) = 42.616 | Figure 5 23