MIT LIBRARIES 3 9080 02617 6898 Bf. ,.;.: Digitized by the Internet Archive in 2011 with funding from Boston Library Consortium IVIember Libraries http://www.archive.org/details/threestrikesyourOOcaba ;l W^ OH" Massachusetts Institute of Technology Department of Economics Working Paper Series Three Strikes and You're Out: Reply to Cooper & Willis Ricardo J. Caballero Eduardo M.R.A. Engel Working Paper 04-10 March 2004 Room E52-251 50 Memorial Drive Cambridge, This MA 02142 paper can be downloaded without charge from the Paper Collection at Social Science Research Network http://ssm.com/abstract=5 2602 1 iMASSACHUSETTS INSTITUTE C'i^ TECHNOLOGY 1 MAR 1 1 200^ Three Strikes and You're Out: Reply By Ricardo J. to Cooper and Willis Caballero and Eduardo M.R.A. EngeP March, 2004 Abstract Cooper and Willis (2003) is the latest in a sequence of criticisms of our methodology for estimating aggregate nonlinearities when microeconomic adjustment based on "reproducing" our main findings using artificial microeconomic agents face quadratic adjustment costs. sults lumpy. Their case That Their mistakes range from misinterpreting their is data generated by a model where is, they supposedly find our re- where they should not be found. The three claims on which they base incorrect. to is own their case are simulation results to failing understand the context in which our procedures should be applied. They also claim that our approach assumes that employment decisions depend on the gap between the target and current level of unemployment. This is incorrect as well, since the 'gap approach' has been derived formally from at least as sophisticated microeconomic models as the one they present. On a more positive note, the correct interpretation of Cooper and Willis's results shows our procedures are surprisingly robust to significant departures from the assumptions that made in our original derivations. JEL Codes: Keywords: C22, C43, D2, E2, E5. Adjustment hazard, aggregate nonlinearities, lumpy adjustment, observed and unobserved gaps, quadratic adjustment. *Caballero: Department of Economics, Massachusetts Institute of Teclmology, 50 MA Memorial Drive, Cambridge, 02142, and National Bureau of Economic Research (e-mail: caball@MIT.EDU); Engel: Department of Eco- nomics, Yale University, P.O.Box 208268, mail: eduardo.engel@yale.edu). New Haven, CT 06520, and National Bureau of Economic Research (e- Summary 1 of the case Cooper and Willis (2003), henceforth CW, is Labor Adjustment: Mind the Gap." In comment this "The Economics of the third version of the authors' CW argue that our finding (in Caballero and Engel (1993), henceforth CE, and Caballero, Engel and Haltiwanger (1997), henceforth lumpy microeconomic adjustment matters They base their case for aggregate on "reproducing" our main findings using where microeconomic agents face quadratic adjustment we show In this reply that the three claims mistakes range from misinterpreting their own on which they base is, model they supposedly find our employment decisions depend on the gap between is the target incorrect as well, since the 'gap approach' has phisticated CW's results partures from the assumptions CE Throughout, microeconomic distribution of and CEH shows made and current On more a microeconomic gaps are CE and CEH = (first other studies that at the is: 'kMi^^+yMi^\ (1) Ad^'^ is and desired employment the equation above simplifies to a linear on aggregates at least as so- explaining aggregate employment fluctuations. the cross section distribution of gaps between actual variable depends only of unemployment. positive note, the correct many represents the rate of growth of aggregate employment, and 0, level and examine whether the implied features of the usefiil in A£', and y = approach assumes that in our original derivations. level adjustments are lumpy,* When A, > Their that our procedures are surprisingly robust to significant de- take as an assumption validated in Specifically, the basic regression in AE that our been derived formally from microeconomic models as the one they present. interpretation of their case are incorrect. simulation results to failing to understand the con- which our procedures should be applied. They also claim text in where That not warranted. data generated by a artificial costs. is where they should not be found. results This employment dynamics, CEH) that moments of the the i-th moment of at the firm level. model where the left hand side cross section distribution of gaps). This case can be obtained either from a microeconomic model where agents adjust infrequently but with a probability that is independent of their gap (the constant hazard model of Calvo, 1983) or from a model where agents face quadratic adjustment When Y > 'CW 0, in favor and adjust all the time (Sargent, 1978).^ on the other hand, higher moments of the cross-section (2003) seem to agree with evidence" costs this assumption, in particular, in their distribution of gaps matter conclusion they refer to "overwhelming of it. "See Rotemberg (1987) for a formal proof of the aggregate equivalence of Calvo 's lumpy adjustment model and the quadratic adjustment cost model. 1 from a scenario where microeconomic adjust- for aggregate dynamics.^ This case can be obtained ment is lumpy and the probabiUty of such adjustment is increasing model of CE). There it is ample microeconomic evidence matters for aggregate adjustment. very significant y > and a We find that it in the gap (the increasing hazard for this behavior, the question is whether does, since our aggregate regressions large contribution of yvW/ to aggregate CW's critique has changed over time, but as of today, it can be employment show a fluctuations. claims, all of them split into three based on applying our procedures to data generated with a model with smooth microeconomic adjustment: When our measure • Claim 1: of microeconomic gaps are computed from there exist parameter configurations for there the • is Claim which estimates of y are similar to no microeconomic lumpiness or common denominator in nonlinearities. This has microeconomic been ours, even their though main claim, and CW (2001, 2002, 2003). When the microeconomic gaps 2: their artificial data, are not directly observed but can be estimated with used in data, the procedures CEH give nonsensical results when applied to their data. When • Claim 3: only aggregate data are used, coupled with the Kolmogorov equations quired to keep track of the simulated cross section distribution of gaps, as in CE, our mates can be found even when their Not only are these claims incorrect, as we will argue below, but they also reflect a fundamental We developed a methodology to study whether lumpy microeconomic adjustment has aggregate implications, not In section incorrect. microeconomic adjustments II we show In section III that we should not be used if to infer from aggregate data whether are lumpy. due to a basic interpretation error of their argue that since the identification strategy gaps with microeconomic data is built on the observation that own we results, Claim 1 is adopt for estimating microeconomic data are lumpy, microeconomic data are not lumpy. Therefore Claim 2 Furthermore, the fact that esti- (linear) data are used. misunderstanding of the point of our papers. the underlying re- is it not surprising. CW find nonsensical results while we find meaningful and statistically significant results, indicates that our findings do not arise when microeconomic adjustments are smooth. ^The higher moment it is that matters in specification (1) simple and shows up often both in our work and in is the third CW's moment. critique. We focus on this specification because Yet there are other specifications our work that involve higher moments different from the third moment, which explains 'higher moments'. why we in their and generically refer to IV we show In section that Claim 3 has nothing to nomic adjustment. Their finding comes from relaxing do with lumpy to what we found with to Somewhat to departures paradoxically, the non-lumpy microeoco- an extreme the maintained assumption of our analysis that the driving forces are random walks."* This result comparable vs. is neither new nor quantitatively actual data. CW work of can be used to show that our approach from the random walk assumption. In is robust nothing can be found with the serial fact, CW (2002), and (almost) nothing with the low serial correlation of 0.47 CW (2003) dropped further to 0.28, and even then the gain in R^ from correlation of 0.81 used in assumed in CW (2001). adding higher moments it is substantially less than half of what we found. Section V concludes. Their main critique 2 In the main part of their critique, CW compute from their artificial data the cross-sectional moments of static gaps and esfimate an equation analogous to (1): A£P^ = ;iywJ')'^^ + YyW;^)-^^, where A£'^^ and M^')'*^^ stand for the rate (2) of growth of aggregate employment and the /th moment of the cross section distribution of static gaps respectively, when the underlying data are generated with CW's quadratic adjustment cost model. Their main finding ent from the one is that they estimate a positive and statistically significant y, we find using actual data. Cooper and Willis then argue that this is evidence that a researcher testing for aggregate nonlinearities on their data are important for aggregate dynamics. our results may not very differ- It would conclude, erroneously, follows, they argue, that our methodology is that these flawed and well be due to misspecification error. However, finding similar values of y does not mean that a researcher will conclude that the nonlinear term to also look at cases. It is equally important for aggregate dynamics in the two cases. For whether the regressor that turns out that much smaller it is one needs multiplied by y has similar variability in the two does not: The variability of M^^^ in than that of the corresponding this, moment when CW's quadratic adjustment model is micro-adjustments are lumpy. Thus, ''in our derivations, and as is standard in much of the (5,5)-literature, we assumed that the driving forces follow a random walk, an assumption that cannot be easily rejected in the data. In this case, one can show that the static gap (the difference between cuiTent employment and the optimal level of employment if there are no adjustment costs) is equal gap (the difference between current employment and the optimal level of employment if adjustment removed only today) plus a constant that depends on the drift. This is a very usefiil result since the static gap to the frictionless costs are is straightforward to calculate while its frictionless counterpart involves more complex dynamic calculations. the contribution of yM^^^ simulated data, while is it minuscule in explaining aggregate employment is large and economically reported values of neither R^ nor A, The first column in Table 1 when table that the R'^ reported is and R^ falls based on Table la in estimating (2) is the CW (2003). the second By column of Table here, the moment of y, even though statistically less than 0.013 over the CEH, reported in R^ increases by 0.15 when adding a non-linear parameter and the variation of the speed of adjustment over the relevant range as that in CW's comment, the third same with or without contrast, in the corresponding exercise in Table 3 of 1 put, the apparent from their It is economically irrelevant, as the adjustment speed varies by relevant range of gaps.^ in CW's rises). gaps: 0.90. Similarly, the estimated value of the non-hnear parameter significant,^ is Simply significant in our findings. change when adding the nonlinear term while they change substantially in our setting (k volatility in is more than ten times as large CW's model. The economic striking in the irrelevance of the non-linearities estimated by Cooper and 2002 version of their comment, where they used a more Willis is even more realistic value for the first order correlation of productivity shocks (0.81 at an annual level). ^ There they report an I^ of 0.97, both for the model with and without the non-linear parameter,^ and the adjustment speed implied by their non-linear ^The model statistical significance their simulations, ^Where varies only by 0.005 over the relevant range of values taken by the gap. they find possibly reflects the fact that they use time series with 1000 observations in while CEH's estimates are based on 35 observations. the 'relevant range' is defined as the cross-section of static gaps, respectively. ^q ± 2<3a, with //q and Oq denoting the mean and standard deviation of A tedious but straightforward calculation from first principles shows that _ StgOfi 2 2 with: Gk "-^o-^' where G= ( 1 A+ p- p^ — 5)/(l — 5pj, X denotes the speed of adjustment in the partial adjustment representation of the quadratic adjustment cost model, 5 denotes the discount rate that results target as a present value of future static targets, in this model when calculating the (con-ect) dynamic Og denotes the standard deviation of firm-specific productivity shocks CW a is defined on p. 23 in (2002). ^As we pointed out the en-ors in the first and second versions of CW's and critique, they reacted by looking for a new parameter configurations and new model specifications that might help their case. Their lack of success, despite two major revisions of their original comment, possibly ^This is for the reported for R^ are is benchmark with high adjustment 1 .00. the best evidence of the robustness of our findings. costs. For the benchmark with low adjustment costs, both values Estimating unobserved gaps with microeconomic data 3 The second and third points of CW's critique stem from the fact that in practice the gaps are not observed and hence neither are the cross-sectional distributions of these gaps. They argue that our procedures to estimate these gaps and moments introduce positives and nonsensical tion), since Cooper and claims 2 and In (in this section) distinction and between the procedure that in Caballero Willis' specific critique differs Engel in Caballero, and Engel (1993) (in the next sec- between these cases (corresponding to their observe the microeconomic data but have no direct observation of the gaps. In order to construct the microeconomic gaps, when hours exceed are certain normal below normal. we relies heavily context, the relationship The use information on hours. level, there is a shortage idea being that of labor while the opposite true is one needs to estimate the mapping from the hours-gap Still, the employment-gap, and the equation that does this suffers Our way out to false 3, respectively). CEH we when hours which can lead errors Again, Cooper and Willis are mistaken. results. To explain why, we begin by making a and Haltiwanger (1997) new on our observation that from to classic simultaneity problems. microeconomic adjustment is between hours and employment gaps can be estimated lumpy. In if this one only uses observations where large adjustments took place; the basic logic behind this procedure being that during these episodes the variability of the regressor regression. Yet if one and Willis' data, knows that swamps the variability microeconomic data are not lumpy, as of the error term in that is the case with no sensible researcher would use our procedure. Cooper and Willis make the mistake of not understanding that the microeconomic estimation procedure in on the observation reality, 4 Cooper that microeconomic behavior a fact explicitly acknowledged is CEH is lumpy. Fortunately for us, the conditional latter holds in by CW.^ Estimating unobserved gaps using only aggregate data In Caballero and Engel (1993) cross-sectional we do moments from an not observe microeconomic data and hence generate the internally consistent model. This tablished fact that microeconomic adjustment Kolmogorov/Markov functional equation sponding to a given ^"[There Conclusion. is] set is model starts from the well es- lumpy, and uses this information to construct the for the evolution of the cross-section distribution corre- of parameters. Cooper and Willis apply our procedure to data generated by ovei-whelming evidence that plant level adjustment is nonlinear", CW (2003), first paragraph in the adjustment cost model, and find evidence that y in equation (2) their quadratic is when positive it should be zero. Here CW fail assumed correlation we assumed that the driving forces follow a the Introduction, in this case the static gap longer and within the {S,s) literature that is now depends on gap They then drop the is if the equal to the frictionless gap plus a constant. random walk assumption The the state. serial correlation step in first serial with a quadratic adjustment cost model. It is But results qualitatively similar to ours. CW in well gap no between static to rediscover this result."^ is around 0.28 (we report to to generate microeconomic data only then that they find, under some circumstances, new (we this is neither random walk assumption go on It is relaxed, the static is of the driving forces from one correlation coefficients at annualized rates) and departures from the low random walk. As mentioned a sufficient statistic for the probability of adjustment, so that the difference frictionless all serial the very is in their driving processes. In our derivations known which to identify the real reason behind their finding, and static knew already frictionless that for very large gaps could not be exchanged) nor quantitatively comparable to our findings. Paradoxically, the findings in the serial correlation is CW are encouraging for the gap approach, since dropped to very low the values of serial correlation used in levels that things start breaking CW (2001), which we found CE it is only down. In when fact, for would be no are already low," there significant false positive finding. Table 2 reports the gains in R^ that the (absolute) gap, versus those that in from adding a hazard term increasing would be obtained from doing the degrees of serial correlation in the driving processes.'^ Clearly, there of finding an increasing hazard when there (i.e., assumed random walk. then the gain in "'Although they the difference fit was fail to none) CW exercise with different is no risk if serial correlation is of false positives not too far from the CW had to stretch things a lot to find parameters similar to ours, and even less than half of the gain we found. highlight the connection between their sharp departure from the between both gap measures. Also, beginning claiming that our approach assumes CW, 2003) is in tliat in their abstract, the optimal policy depends random walk assumption and by repeatedly they mislead their readers on the gap. In the final sentence of Section they finally acknowledge that the "gap approach" can be derived from optimizing behavior follow a random walk, yet credit a previous version of their comment for this well known II (in when shocks result (see, for example, Nickell, 1985). "The aimualized serial correlation they use in driving force we RBC (see, e.g., models '"We used replicate in Cooley andPrescott, 1995) CW's CW Caballero and Engel (1993) is is (2001) is close to 0.50. above 0.80. Also note CE for this model procedure and use 1000 observations as they (0.75). serial correlation in the actual used to calibrate 0.81. aggregate shock in order to calibrate the R^ of the constant-hazard in The that the standard value / did. We also add i.i.d. quadratic adjustment cost nonnal noise model to match to the the R'- Final 5 The first Remarks paragraph in the conclusion of CW illustrates the flawed logic of their approach. concludes that "despite the overwhelming evidence that plant-level adjustment is It nonlinear, the question of whether this matters for aggregate employment dynamics remains an open issue." But if the goal is to show whether clearly established microeconomic lumpiness matters at the aggregate level, then the natural approach to start fi"om a is and determine whether aggregation removes what our methodology is all when that satisfy the little at all, we have made an wrong, or irrelevant, or driven test is is precisely with simulated data that whether a procedure designed to test false positives twisted logic at best. effort to take their claims seriously. than can be rescued from the sequel of either and start microeconomic lumpiness condition provides applied to their counterfactual data. This In our reply, however, of micro nonlinearities, which designed to do. Instead, Cooper and Willis does not resemble actual microeconomic data competing hypotheses traces model with microeconomic lumpiness CW's attempts. by an extraneous The But there is very results they claim to find are ingredient. Let us recap what they did and the conclusions they should have drawn: 1 CW relax both of our maintained assumptions — that microeconomic adjustments are lumpy and that driving forces follow a random walk lumpy microeconomic adjustment it in first an extreme fashion. The evidence on overwhelming, even Cooper and Willis acknowledge and our assumption of a random walk at times, assumption of an annual 2. is — is definitely closer to reality than their order correlation as low as 0.28. Correctly interpreted, their main result implies the exact opposite of their Claim microeconomic gaps are observed, our methodology does not detect ities when When the nomic data, microeconomic gaps are not observed but need one should not use our identification strategy is know to be estimated from microeco- (which sharp contrast with those when only we found with actual relies to be smooth. In rameter estimates they find with their counterfactual data are not Finally, significant nonlinear- CW. lumpiness) with their data, where adjustment 4. When the applied to data generated even with the major departures from our assumptions considered by 3. 1. on microeconomic any event, the pa- statistically significant, in data. aggregate data are available and the path of the cross-sectional distri- bution needs to be simulated, the assumptions about the serial correlation of the driving processes become more important. This is not new. The surprising feature of CW's results is that even after sion of their is less dropping the serial correlation as as they do in the most recent ver- comment, the explanatory power of the nonlinear terms than half of what we found in the data. If one adopts the persistence of the driving processes, essentially much no realistic and uses the values assumed no gain from adding higher moments their claim, there is more false positive finding in their to their regressions. in experiments assumptions on CW (2002), there is Again, and contrary to even when applying our methodology to highly unrealistic data. The other two paragraphs proach" First, is voodoo-economics and what they us from in their conclusion carry the implicit call "the that they are messages ready to deliver a superior gap-free alternative. microeconomic models as the one they present extensive literature on the optimality of {S,s) models).'^ Second, and perhaps in published from dynamic optimization work. In gap ap- gap approach" has been derived formally by us and many others before at least as sophisticated the methods derived that "the fact, the difficulties in that (for this, see the more importantly, do not "rely on gap measures" already exist measuring gaps was our own motivation for Caballero and Engel (1994, 1999).'^ To end on a more positive note, CW's approach contrasts with more constructive and interesting recent developments in the literature on the adjustments. For example, RBC model, macroeconomic implications of lumpy microeconomic Kahn and Thomas (2003) conclude matter for actual investment. But this gate data generated is fixed costs if the interest rate is be modified for first invest- as a demonstration that fixed costs do not And it among they also show that the aggre- features of actual aggregate data, such as the such spikes can be generated by microeconomic and fruitful area of research: to capture the nonlinearites that are this proof of optimality of (S,s) policies discussed in this reply see, that fact, not endogenized (confirming the results in Caballero and Engel, Let us hope that energy will be spent on '^The on aggregate not what they did. In 999). This finding points to an interesting to many by such a model misses important skewness caused by investment spikes. need an otherwise standard fixed costs of adjusting caphal do not have a significant impact ment.'^ This finding has been misinterpreted by 1 that within is How does the RBC model observed in aggregate investment? type of question. in Scarf (1960). For important extensions, relevant to the models others, Harrison, Sellke and Taylor (1983), Grossman and Laroque (1990), and the pedagogical survey in Dixit (1993). '"'in these papers we extended the (S,s) literature to incorporate stochastic adjustment costs model via maximum likelihood. '^See Veracierto (2002) for a similar conclusion in a model of iiTeversible investment. and estimate a structural REFERENCES Caballero, Ricardo J. and Engel, Eduardo M.R.A. "Microeconomic Adjustment Hazards and Aggregate Dynamics", Quarterly Journal ofEconomics, "Explaining Investment Dynamics in . proach," National Bureau of US May 1993, 108(2), pp. 359-383. Manufacturing: Economic Research (Cambridge, A Generalized {S,s) Ap- MA) Working Paper No. 4887, October 1994. "Explaining Investment Dynamics in . US Manufacturing: A Generalized {S,s) Ap- proach," Econometrica, July 1999, (57(4), pp. 741-782. and Haltiwanger, John C. "Aggregate Employment Dynamics: Building from Microe- conomic Evidence", American Economic Review, March 1997, 57(1), pp. 1 15-137. Calvo, Guillermo, "Staggered Prices in a Utility-Maximizing Framework," Journal of Monetary Economics, September 1983, 72(3), pp. 383-398. Cooley, Thomas Thomas F. F. and Prescott, Edward C. "Economic Growth and Business Cycles," in Cooley, ed.. Frontiers of Business Cycle Research. Princeton: Princeton Uni- versity Press, 1995, pp. 1-38. Cooper, Russell and WUlis, Johnathan. "The Economics of Labor Adjustment: Mind the Gap," National Bureau of Economic Research (Cambridge, MA) Working Paper No. 8527, October 2001. . "The Economics of Labor Adjustment: Mind the Gap." Federal Reserve Bank of Min- neapolis Research Department Staff Paper 310, August 2002. . "The Economics of Labor Adjustment: Mind the Gap." Federal Reserve Bank of Kansas Research Working Paper 03-05, July 2003. Dixit, Avinash. "A Simplified Treatment of the Theory of Optimal Control of Brownian Motion", Journal ofEconomic Dynamics and Control, October 1991, 75(4), pp. 657-673. Grossman, Sanford J. and Laroque, Guy. "Asset Pricing and Optimal Portfolio Choice in the Presence of Illiquid Durable Consumption Goods," Econometrica, January 1990, 55(1), pp. 25-51. Harrison, J. Michael; Sellke, Thomas L. and Taylor, Allison J. "Impulse Control of Brownian Motion," Mathematics of Operations Research, August 1983, 5(3), pp. 454-466. Khan, Aubin and Thomas, Do Cycle Models: Julia K. "Nonconvex Factor Adjustments in Equilibrium Business Nonlinearities Matter?", Journal of Monetary Economics, March 2003, 50(2), pp. 331-360. Nickell, Stephen. Oxford Bulletin of Economics and Statistics, Rotemberg, Julio An "Error Correction, Partial Adjustment and All That: J. "The May 1985, 47 {2), New Keynesian Microfoundations," in ley Fischer, eds.. National pp. 1 19-29. Olivier Bureau of Economics (Cambridge, Expository Note," MA) J. Blanchard and Stan- Macroeconomics An- nual, 1987, pp. 69-104. Scarf, Herbert E. "The Optimality of {S,s) Policies n the Dynamic Inventory Problem," in Ken- neth Arrow, Samuel Karlin, Patrick Suppes, eds.. Mathematical Methods in Social Sciences, Stanford University Press: Stanford, CA, 1960, 196-202. Veracierto, Marcelo. "Plant Level Irreversible Investment and Equilibrium Business Cycles", American Economic Review, 92{\), March 2002, 10 pp. 181-197. Table 1: ESTIMATION WITH STATIC Gap CW (quadr. adj .) CEH (lumpy R- without non-linear parameter 0.90 0.65 R^ with non-linear parameter: 0.90 0.79 Increase in R^ after adding non-linear parameter: 0.00 0.14 Minimum adjustment speed (non-linear model): Maximum adjustment speed (non-linear model): 0.19 0.31 0.20 0.46 Range of adjustment speeds (non-linear model); 0.01 0.15 adj .) CW column based on Table la in CW (2004). CEH based on Table 3 in CEH (1997). Maximum, minimum and range of adjustment speeds are calculated considering adjustment hazards in the range jjG ± 2aa, where /jq and Oa denote the mean and standard deviation of the cross-section of static gaps for the model under consideration. 11 Table 2: Estimation with Inferred Static Gap: Macroeconomic Data Data Driving force p (annual) BLS, as in CE Increase in R- OTS (1993) Sim. Quadr. Adj. 1.00 0.00 Sim. Quadr. Adj., 0.75" 0.01 0.47 0.03 028 0.05 CW (2001) Sim. Quadr. Adj., as in CW (2004) Sim. Quadr. Adj. as in 'Increase in Br' denotes the difference between the R- obtained constant hazard and the E} when imposing a constant hazard, in when estimating a model with a non- both cases using the methodology in CE (1993). 'Sim. Quadr. Adj.' stands for 'Simulated Quadratic Adjustment'. This value of p calibrating RBC somewhat below both the value in the driving force used models (see Cooley and Prescott, 1995). is 3837 023 12 in CE and the values used when MIT LIBRARIES 3 9080 02617 6898 i!;:iijii;i;ii!;;:ii;/i;iiiil;ii;iiil:l!i!i;iiiIU;miiiiijMl!i!iii;!:iiii