Next Generation Myanmar > 55 Weighting of data To improve accuracy, weights have been applied to the survey dataset on age, gender and education. Again, the 2014 census has been used as the sampling frame. Statistical reliability As in all surveys, the survey respondents are only a sample of the total survey universe or survey population, meaning that we cannot know with certainty whether the figures obtained are exactly those we would have, had all possible respondents been interviewed. It is nonetheless possible to predict the variation between the sample results and the true values that would be obtained under such circumstances. This can be predicted using two factors: knowledge of the size of the samples on which the results are based, and the number of times that particular answer is given. We can make the prediction with varying levels of confidence – the higher the level of confidence, the wider is the range of the variation. The conventionally accepted level of confidence used in reporting sampled survey results is 95 per cent, and will also be used here. The level of confidence of 95 per cent means that the chances are 95 in 100 that the true value will fall within a specified range. Table 3 illustrates the predicted ranges for different sample sizes and percentage results at the 95 per cent confidence interval. An indication of sampling error is also given. Please note that these intervals are indicative as they are calculated on the basis on unclustered simple random sample and do not take into account the effect of clustering and weighting. For example, with a sample of 2,471, for a question where 30 per cent of respondents gave a particular answer, the chances are 95 in 100 that the true value of this answer (which would have been obtained if the whole survey population, all people aged 18–30 in the project regions, had been asked this question) is between 28.2 per cent and 31.8 per cent. When results are compared between separate groups within a sample, the results may show a real difference in opinions between these sub-groups, or they may occur by chance and not be conclusive. Testing whether the difference is real or only occurred because of our limited sample depends on the same factors: size of the subsamples, the percentage of respondents selecting a particular answer, and the degree of confidence chosen. Table 4 shows conventional 95 per cent confidence whether the differences between the survey results of different sub-groups (such as regions or different age groups) are real or inconclusive. For example, if a particular question response in a region A with a sample of 100 is 50 per cent, and the objective is to compare it with the response in a region B (sample size 100 respondents), then any result below 36.1 per cent or above 63.9 per cent will show, with 95 per cent likelihood, a real observable difference in the opinions of people living in these two regions. Any result falling between 36.1 per cent and 63.9 per cent would not show a statistically significant difference between region A and B. On the other hand, when comparisons are made between gender groups, where there are larger numbers of respondents per group, the span within which differences are inconclusive is smaller. Similarly, answer distributions towards the extremes of 0 per cent and 100 per cent (50 per cent is ‘non-extreme’) yield a smaller span. If, for example, a particular question response among women is 80 or 20 per cent and the objective is to compare it with the response given by men, then any result below 76.8 per cent or above 83.2 per cent will show, with 95 per cent likelihood, a real observable difference in the opinions of women and men. Throughout the report, with a few exceptions, only statistically significant results have been reported. Data collection and language Data was collected through pen-andpaper interviewing. While this is associated with somewhat higher degrees of reliability error, it contains advantages in difficult-to-reach territory, insofar as it works entirely independently of web access and does not expose enumerators to potential risks from possessing visibly valuable items. Following data collection, data was coded by data clerks in Yangon. Data is most reliable and accurate when collected in the language most convenient to respondents. Recognising the linguistic diversity of Myanmar, surveys were conducted by enumerators proficient in local languages when the Myanmar language was not appropriate. These enumerators were able to facilitate respondents’ answering of the questionnaire. While there is a risk that certain variations may have appeared in the wording of questions due to this modality, it was not seen as possible/feasible to translate the tool to all relevant questions. As a way of mitigating impact, enumerators were carefully trained in the research objectives to ensure questions were understood as intended.