This paper presents evidence of the impact of a price information intervention on farmers’ price search and marketing behavior in India. The authors first develop an ex-ante bargaining model between traders and farmers and derive some key theoretical predictions. They then test these predictions using a randomized experiment that provided mandi-level potato price information to farmers in West Bengal through public and private treatments. Overall, the authors find that farmers were more likely to track price information, with similar effects for the two treatments. Nevertheless, the interventions did not appear to increase affect the average annual net prices that farmers received or the quantities of potatoes sold. The authors conclude that the absence of these effects are, in part, due to the relative monopsony power of traders on these markets, and are more consistent with an ex-post bargaining model between farmers and traders.
This paper addresses an important issue in economics: Namely, the role of asymmetric information between two contracting parties, and how this affects the optimal quantities traded and welfare. This speaks to an important policy issue as well: Given the rapid advancement of governmental and mobile phone-based price platforms, is the provision of this price information improving farmers’ access to information and their bargaining power vis-à-vis traders? And is this improving their welfare? Despite the importance of this work, and the innovation of the experimental design, there are several areas where the paper could be strengthened and clarified. The primary comments are below.
Theoretical models. The ex-ante bargaining model is a useful construct for thinking about trader-farmer bargaining and how access to an informative signal would affect the quantities sold under different bargaining weights. While the authors mention that some of the assumptions are derived from the local context (ie, limited intra-annual storage, asymmetric information), the other qualitative and quantitative information that the authors collected on the potato marketing system (which is substantial, and for which the authors should be commended), suggests that an ex-ante bargaining model is not a useful starting point in this context. (In addition, while the ex-ante bargaining model is useful for thinking through the theoretical predictions, it isn’t clear that these predictions are surprising – ie, that the quantities sold by farmers should increase with a reduction in asymmetric information, unless the traders have monopsony power.) With this in mind, it is unclear why the authors start with this ex-ante model, as opposed to the ex-post bargaining model. Is it because the ex-ante model drove the research design, and so the authors wanted to be transparent in their presentation of the initial hypothesis, tests and results?
In addition to the above: If the ex-ante bargaining model is a useful starting point, it would be helpful to incorporate ore of the actual findings on the potato market structure (Section 2) into the model. To clarify: A key result from the data provided in Section 2 is that there is high inter- and intra-annual price variation in the mandi markets, but little pass-through to farmers (which could in part explain the lower average farm-gate prices but also lower intra-annual farm-gate price variation, which was extremely low). Under the assumption that both the trader and the farmer are risk-neutral (and the trader doesn’t have a perfect monopsony), reducing the information costs would increase the average price that farmers receive. However, this framework doesn’t account for potential changes in this information asymmetry on farm-gate price variation. A risk-neutral farmer would only be concerned about average price levels, but a risk-averse farmer would care about both price levels and variation. While the authors state that the results also extend to the case where both parties are risk averse, it would be useful to think through a modification where the farmer is risk averse and the trader is risk neutral, and how this might (or might not) affect the theoretical predictions. This would seem to make the model more closely aligned with the potato market structure.
Suggestion: Rather than include the full formal ex-ante bargaining model, it seems as if authors could substantially shorten this model and present the theoretical predictions (which seem straightforward and intuitive). Then, they could state that the existing information on the potato marketing system in West Bengal suggests that an ex-post bargaining is more likely, and present this model. The empirical results could then follow from the presentation of these two models.
If, alternatively, the authors include the ex-ante bargaining models as is, then thinking through a model where the farmer is risk averse would affect the theoretical predictions.
Experimental design
The experiment is innovative and well thought out. At the same time, but there are several elements of the design that could be clarified. The program appeared to occur in several stages, with a baseline in 2006, some price provision (and data collection) in 2007, and some additional data collection in 2008 (with some price provision). Greater detail on the actual experiment, including a figure of the time frame of the data collection and the experiment itself, would be useful. In addition, while the phone recipients were randomly chosen, were they randomly chosen from among all farmers or a subset of farmers?
Suggestion: Provide a time frame of the experiment, including the timing of the surveys, the provision of information and how this overlays with the potato planting, harvesting and sales information (which is provided on p. 5 in the text).
The primary motivation for the experiment seems to be the assumption that farmers do not have perfect information, and that this asymmetric information is resulting in relatively high trader-level margins. As evidence of this, the authors use their marketing data to construct the trader marketing margins, suggesting that these represent 15-40 percent of the wholesale price and 34-89 percent of the farm-gate price. Despite the fact that these are upper bounds for the traders’ margins (since the costs are estimated from farmers’ data), it is a bit hard to think concrete about the magnitude of these margins, both for the context and in terms of economic theory. Are these margins are “large” or “small”? In other words, what is the benchmark that we would expect to see for more “reasonable” traders’ margins (other than the benchmark of zero profits)? In addition, is it reasonable to calculate trading margins for one year, or are traders maximizing their profits over a longer time horizon (so we would expect positive profits in some years and zero and negative profits in other years)?
The experiment provided average daily price information for two potato varieties for nearby mandis. Were these only the average prices on the mandi or maximum and minimum? Do prices change by quality and size? If so, what was the range of potato prices within a given mandi? Do the min and max prices on these markets reflect quality differences, size differences or both? Would farmers be able to generally infer the difference between the average price and their own quantity, quality and size produced? Were prices not provided during the planting season to avoid supply-side effects? Some more details on the experiment (in terms of the types of prices provided and how these accurately represent potato prices on these mandis – in terms of variation, quality and size – would be useful.
Empirical specification
What regression are the authors estimating in Table 5? While the table is relatively straightforward, it would be useful to present the empirical specification before presenting this table (as the estimation for this table seems to be the same as the key empirical equation. The only exception is for Columns 3-5, which uses a multinomial logit regression; for this latter specification, it might be helpful for the authors to briefly mention why the IIA should hold in this case, or why an alternative specification that collapses the choices into trader/friend or other would not be more informative).
Dependent variable: In the key empirical specification, the authors use the net price received, rather than the gross price. Is there a reason for this? Given measurement error associated with estimating costs, and the potential impact of classical measurement error on the s.e. (see below comment), it seem as if the authors would want to use the gross prices as well as the net prices. In addition, if there are no zero values in the price data, it might make more sense to use a logarithmic specification.
Dependent variable: While we are interested in the quantities sold and prices received, do the authors have any information on whether the farmers switched traders? This would give a better sense of the “outside option” of those farmers and whether the increased access to price information changed their marketing behavior. This might not be observed using the yearly averages, but might certainly be observed using the intra-annual data on sales (ie, whether they attempted to sell to certain traders at first, failed and then changed).
Should the subscripts on the “private information” and “public information” be “v” (for village) or “i”? My understanding was that the treatment was at the village level, with the only individual-level variation for the phone treatment. (In this case, then, it might also make sense to have an interaction between private information and phone recipient, which would then allow us to interpret the coefficient on private information as those without phones living in private information villages, and the coefficient on private information as the impact of receiving that phone in that village).
Empirical results.
First-stage results. The experiment was conducted in 72 villages, with 48 treatment villages and 24 control villages. Within the private information villages, only 4 farmers were treated (out of an average of 200 farmers within the village). Even with information-sharing among farmers within these villages, there was probably a high rate of imperfect compliance – ie, very few farmers actually received price information (either first-hand or second-hand). Do the authors have any data about the first-stage results of who received price information from one of the cited sources in the treatment villages? (Note: Columns 3-5 in Table 5 are somewhat analogous to a first stage, but do not include information on the specific “other” source). The authors mention that they didn’t want to prompt farmers to cite these specific sources in the treatment villages (hence the “other” category), but this information seems to be important to understand the impact of the intervention on the first-stage.
Statistical significance. The authors conclude that there are no statistically significant results in Table 7. While this is certainly true, it would be helpful to think about whether these results are economically and statistically insignificant, or whether the authors are simply unable to detect a statistically significant effect because they are underpowered (either in terms of our sample size, or because classical measurement error in the dependent variable). I was unable to calculate the economic significance of these coefficients (as data on the control group’s average price received or quantity sold is not provided in Table 7 (or the baseline means comparison in Table 4), although the coefficients on prices seem to represent about 4% of the average farm-gate prices (from Section 2.2), but it is important to differentiate between no effect and an underpowered effect for two reasons:
While the authors have 72 villages, with 48 treatment and 32 control, the minimum detectable effect would need to be relatively large for the authors to detect an effect with the given sample size (and in the face of imperfect compliance, as mentioned above). A very crude (back of the envelope) calculation shows that if we use the mean farm-gate prices farmers received before the program (2.55) and the treatment effect on net prices received (the coefficient on in Table 7), the authors only have power of 25% (assuming perfect compliance and a very low s.d.). If we take 80% power and their sample size as a given, then the minimum detectable effect would be 10 percent (about .20 Rupees/kg). This would be a relatively large effect in this context, as the traders’ margins are only .71 Rupees/kg – so this price intervention would represent 30% of traders’ margins. This seems like a potentially unrealistically assumption for a treatment effect.
Quantities and prices are notoriously difficult to measure. While the authors did a great job of collecting high frequency data, it is reasonable to assume that there is, at a minimum, classical measurement error – which would simply increase the s.e. on the treatment indicators and make it more difficult to detect an effect.
Suggestion: It would be helpful for the authors to provide the means and s.d. of the outcomes of interest in the regression specifications to better interpret the magnitude of the effect, as well as in the balance table. In addition, if at all possible, some discussion of compliance at the first stage (ie, use of the information treatment) would be useful to get a better sense of take-up. Finally, if the coefficient is relatively large, it would be helpful to include some discussion about whether there are no effects or the authors are underpowered to detect an effect (for the above reasons).
Alternative explanations.
The authors provide some evidence that the experiment is not affecting the market structure by assessing the impact of the intervention on traders’ market share and the Herfindhal index. As a first-order measure, would it be useful to also include the number of traders operating in each village as the dependent variable? The authors argue that the interventions made it easier for new phorias to enter the market; it seems as if this could be easily tested by looking at the number of traders.
One key concern that the authors mention – yet cannot address – is the fact that 2008 seems to have been a very “specific” year in the potato market. While the authors certainly cannot recreate the experiment, it would be useful to know why 2008 displayed such different pricing patterns (ie, whether it was a government intervention, or a supply or demand shock, etc) to get a better sense as to how this affected traders and farmers’ beliefs about prices, and the context of the experiment.
Other minor comments:
On p. 1, the authors mention that “improvements in information…to farmers would increase farm-gate prices, pass-through as well as quantities traded.” It seems as if the caveat should be added “in the absence of other market failures”, since this is a key caveat of the model (and the weakness of many governmental price provision programs).
The authors provide a thorough overview of the potato marketing system in India. Is there any caste distinction between traders (intermediaries and wholesalers) in India, or traders and farmers? If so, is this an element of the bargaining process (that cannot be captured but would affect the theoretical predictions and efficiency of bargaining outcomes)?
Table 6: The results in Panel A did not come through in my version of the paper.
1