 Thank you very much. So I'm Emma Howard. I'm at Trinity College in Dublin and this is work that I've done with my colleague Carol Newman who's also in Trinity College and John Rand who's at the University of Copenhagen and Finn Tarp who's here in UNU wider. So just to say at the start this is preliminary work so any comments or suggestions are very welcome. So just to motivate the paper a little bit so there's lots of evidence, there's lots of literature that says that clustering facilitates growth of regions, of countries and of individual firms and there's a lot of empirical investigations into the evidence into the agglomeration economies that are present in a particular country. However where there's limited empirical evidence is linking clustering to firm performance and this is what we try to address particularly in a developing country context. So there's a number of reasons why we might think that firms that locate in clusters would be more productive. So there's a number of mechanisms through which these productivity gains can occur. So the first one is that if firms are located in a cluster with competitors this competition will incentivize them to cut slack and reduce costs so we've got a competition channel. Secondly if firms are located in close proximity to other firms this increases the potential for productivity spillovers. And then finally there's also a labor productivity channel whereby if firms are locating close to firms that hire similar types of workers a pool of labor emerges and you've got better matching of workers to jobs so workers that are better suited to their jobs are going to be more productive. So we've also got this labor productivity channel. The issue with this type of analysis in trying to identify the impact of locating in a particular cluster on a firm's productivity is that you've got this self selection issue. So it may be the case that more productive firms locate in productive clusters. So this is kind of the main issue that we address in this paper. So we look at a way around this to try and identify the actual impact on the firm of locating in the cluster. So our contribution then to the research we use a rich and unique data set of manufacturing firms in Vietnam. So we've got a developing country context. And what we do is we extend this Oli Pakes method of estimating productivity. So I'll talk a little bit about the Oli Pakes method in a minute. But basically what we do is we take their method and we extend it to also control for the productivity of the cluster to address the self selection issue. And this allows us to identify how locating in a cluster impacts on the firm productivity. And so as I said this is early work. So any comments or suggestions are welcome. But in future work we're also going to attempt to uncover the mechanisms behind these productivity spillovers. So just to talk briefly about the data as I said it's a very rich and unique data set. It's the Vietnamese Enterprise Survey. We've got a panel from 2002 to 2007. That's from the General Statistics Office in Vietnam. It's an unbalanced panel. So we have firms entering and firms exiting over this time period. And the data contains all registered firms with more than 30 employees plus a representative sample of those with fewer than 30 employees. And in that data then we have information on the commune in which the firm is located. So there's three main administrative areas in Vietnam. So province is the largest, district is smaller and then commune is the smallest geographical area. So there's about 12,000 communes in Vietnam just to give you an idea. And then we also have standard financial data on the firm so we know their employees, their assets and so on. Okay, so to talk a little bit about what we actually do then in the paper. So as I said we extend this Ollipakes methodology, their approach for estimating productivity. So we extend it to also control for the productivity of the cluster in which the firm is located when we're estimating the firm productivity. This is a slightly similar approach that Delocker uses in his 7. So he does something similar in that he extends this Ollipakes method to include a control for whether the firm is an exporter or not when estimating their productivity. So we do something similar here but we control for how productive the cluster is in which the firm is located. So basically there's two main parts to the analysis that I'm going to present today. Then the first part is to estimate the firm's productivity controlling for the productivity of the cluster in which they're located. And once we have that productivity measure then we can try to disentangle the impact that the cluster productivity has on the individual firm productivity. Okay, so just briefly, I'm not going to go through this in detail because of time constraints, but briefly to tell you what the Ollipakes estimation methodology does is if the traditional way of estimating a firm productivity was just to assume a Cobb Douglas production function and then take logs and run an OLS to measure your estimates. So there's a couple of issues with that and in that your estimates are biased. So there's two main biases that occur when you estimate productivity in that way. So the first one is the Simultaneity bias. Basically when the firm is choosing their inputs they know their productivity. But the productivity is unobserved to the econometrician so you have a bias that occurs there. Secondly you have a survival bias. So some people will correct for this by just using a balanced panel. The Ollipakes methodology controls for this survival bias in the estimation. So the survival bias occurs because you've got a negative relationship between the capital of a firm and the probability that they'll exit the market. Okay, so if you think that a firm's exit decision is going to be that they will exit the market if they experience a productivity shock that's greater than a particular threshold. That threshold is going to depend on the capital of the firm. So if the firm has more capital they've got a greater potential for future profitability so they're going to be able to take a bigger productivity shock than a firm with a lower level of capital. So that's the second type of bias that's controlled for in this OP procedure. And so OP controls for both of these biases in this three-step estimation procedure. So what we do is we extend this procedure slightly by also controlling for this self-selection problem. So by controlling for the productivity of the cluster that the firm's located in when we're estimating their productivity. And so there's a few assumptions in the OP procedure. So productivity follows a first-order Markov process. We assume a Cobb-Douglas production function and we have to assume that investment is monotonically increasing in productivity. Okay, what the OP procedure does then is proxies productivity by a function in investment and capital. So we extend this and we also include in this proxy cluster productivity. Okay, so then there's three stages to the estimation. We obtain a consistent coefficient for labor in the first stage. The second stage controls for this survival bias by predicting probability of survival and then using those predicted probabilities in the third stage to get the consistent estimation of the capital coefficient. Okay, so in our production function estimation then investment is simply given by the change in assets from one year to the other. Our output measure, we just use the total revenue of the firm, total number of employees as labor, capital is our total assets. We also need to compute the average productivity of the cluster to use in our estimation of the productivity of the firm. Okay, the way we do this is we just use an index number approach to measure the total factor productivity of every firm in the cluster. We define a cluster at three different levels. We conduct the analysis at three different levels. We define the cluster as either being the commune, the district or the province that the firm is located in. And to determine the average productivity of the cluster, for firm I, we take the TFP for all other firms in the cluster and take an average. So we exclude firm I when we're calculating the average productivity of the cluster. So our cluster measure then is firm specific and cluster specific. So first of all, just to show you, the difference then in the production function estimates when we use a standard OLS approach and when we use our extended OP approach for estimating the Cobb-Douglas production function. So as it's a Cobb-Douglas production function, our variables here are logged. So if we have a look first of all at the capital coefficients, the survival bias, so this negative relationship between the capital and the probability of exit causes the OLS estimates to have a downward bias. So as you can see, this is corrected for then when we use our extended OP approach. So the first column just shows you the OLS estimates. The second column is the extended OP when the cluster is defined at the commune. Third is when it's defined at the district and the last column is when we define the cluster at the province level. The second thing to look at is this labour coefficients, so the difference in the labour coefficients. We see an upward bias on the labour coefficient when we use OLS and that's as a result of this simultaneously bias. And this is corrected for then in our extended OP procedure in columns two, three and four. Okay, so we take these consistent estimates then for the coefficients on labour and productivity and we basically just back out our productivity of the individual firm. So we have this estimate for firm productivity which controls not only for the simultaneity and the survival bias, but also for this self-selection bias. So it controls for the productivity of the cluster that the firm is located in. So once we have that then we want to back out the impact to the firm of locating in a particular cluster. So we estimate this regression here which is just the productivity of the firm, regressed on past productivity, past investment and past average productivity of the cluster. Okay, so if we have a look at some results, again we conduct the results at three levels of analysis. So the first two columns are when the cluster is defined at the commune level, next two are at the district and the last two columns are when it's defined at the province level. So first thing to note is as you would expect, previous productivity is positive and significant in all specifications of the model. The more productive the firm was last period, the more productive there will be this period as we would expect. Second result to note then is that the investment again as expected is positive and significant across all specifications. So investment has a positive effect on productivity. What we're really interested in here is the impact of the average cluster productivity on the productivity of the firm. And as you can see, we have a positive and significant effect in the first specification of all three of these models. So the more productive the cluster is, the more productive the firm will be. To further try and disentangle this a little bit, the second specification of the model includes an interaction term between investment and cluster productivity. And when we include this interaction term, what we actually find is that the cluster productivity is no longer significant, but we have a positive and significant effect on this interaction term. So it seems that there's evidence of spillovers, but that for the firm to actually realize the benefit of these spillovers, it needs to also be investing. And this is something that comes out in the learning by exporting literature as well. So it's not an unexpected result. Okay, so to conclude, we have preliminary evidence of productivity spillovers, and that investment is necessary to benefit from these spillovers. So as I said, this is quite new work. So we have a number of issues that we want to address, some next steps. At the moment, we're estimating productivity for our entire dataset. We're going to separate out and estimate the productivity for each sector separately. We also want to do a number of robustness checks. So we're going to conduct the analysis using other cluster characteristics. So also average cluster labor productivity and the size of the cluster. And then finally, we're going to explore a little bit the mechanisms behind these productivity spillovers. So technology transfers we know from previous work are very important in Vietnam, so we're going to look at that channel. The presence of foreign firms also from previous work we know is important, and to investigate the competition channel as well, the competitors.