The aim of this study is about the development of the capital stock on Finnish farms. Therefore the subject is closely related to investment behaviour of farmers. Due to structural change the average size of agricultural investment and capital stock have been increasing during previous decade. On the other hand, not all farms choose to enlarge, but part of them are quitting or continue as they were. Thus, large investments are concentrated to a subset of farms. The empirical goal of this study is to investigate what kind of farms do invest? (and hence are among those which shall continue production in the future). Related questions are how well investments can be forecasted by bookkeeping information and what kind of a process defines the growth of farms capital stock.
Before going into statistical modelling of farm investments, an extensive literature review is presented in the first section of the study. Topics of current economic theory that are related to investment behaviour of firms are brought up. The emphasis is on distinctive features of agriculture and in particular agriculture in Finnish operational environment. The standard derivation of demand of the capital is presented in the context of profit maximization models. These models consist of a simple static model and more complex dynamic models with adjustment costs. Furthermore, the effect of uncertainty in investment decision is analysed by using the real option model. It is pointed out that some features of agriculture may significantly restrict maximization behaviour. It is concluded that profit maximization models may describe actions of some farmers well, but may not be appropriate (in standard form) for others.
The empirical part of the study utilizes statistical methods in a panel data setting. The panel data contains observations from the same farms over different years, which makes it possible to assess the development in variable values. Data are from Finnish bookkeeping farms and the period covers years 1998-2011. There are observations from 1 580 farms in total. However, some farms drop out before year 2011 and are replaced by sampling new ones from the farm population. There are 800-900 farms annually in the panel. Variables in the dataset include statistics related to the farm size, the economic state of the farm, and to human resources of the farm family.
Data are first summarized by the means of exploratory analysis. Descriptive statistics show a large increase in the size of average investments and utilization of capital with respect to time. Furthermore, it is obvious that exogenous shocks have influenced on farm investments during certain years. These shocks reflect variation in profitability and the modification of investment aid scheme. Then attention is turned into analysing development paths of capital at the farm level. This reveals that investments in many cases are lumpy i.e. there are many periods of inaction followed by a large adjustment of the capital stock. This observation contradicts assumptions in a standard adjustment cost model, where smooth path of adjustment is assumed. Nevertheless, similar observations have been made with various real datasets (references are found in the text).
Next section concentrates on empirical challenges with this particular study. It is shown that the distributions are highly skewed and many of the variables include plenty of possible outliers. Still, these outliers are very likely to represent real values in measured quantity (as opposed to measurement errors). In most cases outliers reflect a whole new scale of farms in Finnish agriculture as a result of structural change. Therefore, outliers contain valuable information and cannot be dropped from the data. It is also assumed that various potential explanatory variables correlate heavily with each other. This is confirmed by the results of principal component analysis (PCA), which shows that data can be significantly condensed while preserving most of the original information. Moreover, results of PCA show that main variation in the dataset is within the economic size of the farms. That is to say, farms of various sizes are operating in Finnish agriculture. The final challenge relates to the drop out of farms from the panel. If drop out mechanism is not random, it may have an effect to generalization of results. Drop out mechanism is studied by forming a Random Survival Forest model. Results show that drop out is related on past investment behaviour and the farm size. Farms with very small past investments (and small size) are most likely to drop out. This indicates that drop out from panel may be related to quitting farming altogether.
The challenges above lead to choosing tree based models for analysis tool. The tree models may not offer as precise quantitative estimates as, for instance, linear regression models. However, tree models have various benefits in this particular setting. The distributional assumptions required from the data are somewhat less strict as compared to regression models. Furthermore, tree models are suitable in finding various forms of relations between dependent and explanatory variables and can handle a large number of possible covariates. Applied tree models were classification and regression trees (standard and ensemble CART trees, which forms Random Forest model). Comparison of the results was made with a newer model called GUIDE which utilizes the panel structure of the data, when the dependent variable can appear in a longitudinal form in the model.
The results from tree models show that enlargement investments are most probable in farms that have sufficient size to begin with. The best predictor of various measures of the farm size turns out to be the starting capital stock. Results depend somewhat on the length of the time series from particular farm. Logically, lengthening of observation period increases the probability of observing large investment. This is problematic because the panel is unbalanced. Surely, longer time series provides more reliable picture of the development of capital stock, but there exists significant trade-off between the required length of time series and the number of farms that have such number of observations.
Additionally, the results show that large past investments increase the probability of subsequent investments. That is to say, choosing expansion path results successive increases of farm size. This may be distinguishing characteristic for firms in the industry that are going through structural changes. Some authors (e.g. Nilsen and Schiantarelli (2003)) have had opposite results when studying industries in mature state.
Large investments may not lead to immediate improvement in profitability. In contrast they do in many cases result poor profitability during the year the investment is made and also short period after that. This form of adjustment costs should be taken into account when testing profit maximization models empirically. In general, it seems that incentives for investments are their long term prospects. However, test data is usually available for a rather short period, which may cause profit maximization models to fail empirically.
Number of issues remains to be clarified in continuation of research. One issue is whether different types of capital are substitutable or do they have to be utilized in almost constant proportions in farm production. It was shown that current theoretical models should be modified in order to obtain empirical support from the Finnish agricultural data. During the structural change the sole objective of profit maximization may not be adequate for all farms. Statistical tools in this area of research are developing rapidly and the data also allows more sophisticated analysis. Quantitative implications of differing investment aid schemes, or more precisely differing combination of investment and direct subsidies, can be studied by utilizing differences among support areas in a setting called natural experiment.