Analysis of Representativeness in the Swiss Farm Accountancy Network

Until 1998, the Farm Accountancy Data Network in Switzerland was based on a sample of so-called test farms. This sample was deliberately selected to include farms characterised by above-average economic results. However, one of the objectives of the redrafted legal mandate of 1999 was to illustrate the economic situation in agriculture with a sample of representative reference farms. Starting from existing data, the aim was to achieve this objective with the _Method 2000_ package of measures. The most important methodical changes comprise the delimitation of the population, the development of a new farm typolo-gy, the introduction of a stratified selection plan and weighting of the individual farm results. However, statistical conclusions from the sample to the population are problematic because there is no random selection of reference farms. Considerable bias must be expected for the averages estimated today. The quality of these estimates is assessed by several methodically different approaches. The results show that due both to the inadequate representation of specialist crops and pigs/poultry farms and to the exclusion of farm associations from the reference farms, today_s estimated family farm incomes show a negative bias in the order of 2% relative to the target population, i.e. the current estimate is too low. A calibration model for the adjustment of biased auxiliary variables (area, age, education, organic farms) does, however, indicate a positive bias of agricultural income of between 5% and 6%. The above-average education of the farm managers in the current sample constitutes a primary cause. Together with the impact of selective forces in sampling, due mainly to the high standards demanded of the accounting system, it can on the whole be concluded that the family farm incomes of the reference farms calculated and published today are between 3% and 8% too high relative to the figures of the population. The bias is greater or also smaller for other key figures. In the current general framework, a random selection necessary from the statistical point of view would lead to such low response rates that we could basically expect the bias observed today. However, a rough estimate shows that, with the measures under _Method 2000_, over 80% of the income difference between the previous test farms and the assumed average of the new population is adjusted. These measures can therefore be judged effective as well as efficient owing to the low implementation costs. However, there is further potential for improvement. The proposals for improving estimation quality, suitable for short-term implementation, relate to a dynamisation of thresholds for the delimitation of the population, adjustments in the weighting system and in the selection of under-represented farm groups, the examination of a new measurement of farm size, an adjustment of target population and frame population, and a simplified regular estimation of bias based on selected auxiliary variables. Associated in part with higher implementation costs are proposals for bias assessment in greater depth than is currently the case, extensions of the weighting system with additional calibration of auxiliary variables, and the introduction of random selection for sub-samples. The present paper gives priority to selection-related bias. However, total survey error needs to be taken into account if conceptual alternatives to the current survey system are to be assessed. Building on the selection errors analysed here, careful attention must at the same time be paid to interactions with sampling errors and measurement errors.