Non-sampling errors are not attributed to chance and are consequence of certain factors which are within human control. In other words, they are due to certain causes which can be traced and may arise at any stage of the enquiry, viz., planning and execution of the survey and collection, processing and analysis of the data.
Non-sampling errors are thus, present both in census surveys as well as sample surveys. Obviously, non-sampling errors will be of large magnitude in a census survey than in a sample survey because, they increase with the increase in the number of units to be examined and enumerated.
It is very difficult to prepare an exhaustive list of all the sources of non-sampling errors. We enumerate below some of the important factors responsible for non-sampling errors in any survey (census or sample).
1. Faulty planning, including vague and faulty definitions of the population or the statistical units to be used, incomplete list of population-members (i.e. incomplete frame in case of sample survey).
2. Vague and imperfect questionnaire which might result in incomplete or wrong information.
3. Defective methods of interviewing and asking questions.
4. Vagueness about the type of the data to be collected.
5. Exaggerated or wrong answers to the questions which appeal to the pride or prestige Or self- interest of the respondents. For example, a person may overstate his education or income or understate his age or he may give wrong statements to safeguard his self-interest.
6. Personal bias of the investigator.
7. Lack of trained and qualified investigators and lack of supervisory staff.
8. Failure of respondents’ memory to recall the events or happenings in the past.
9. Non-response and Inadequate or Incomplete Response, bias due to non-response results if in a house-to-house survey the respondent is not available in spite of repeated visits by the investigator or, if the respondent refuses to furnish the information.
Incomplete response error is introduced, if the respondent is unable to furnish information on certain questions or, if he is unwilling or even refuses to answer certain questions.
10. Improper Coverage, if the objectives of the survey are not precisely stated in clear cut terms, this may result in (i) the inclusion in the survey of certain units which are not to be included or (ii) the exclusion of certain units which were to be included in the survey under the objectives.
For example, in a census to determine the number of individuals in the age group, say, 15 years to 55 years, more or less serious errors may occur in deciding whom to enumerate unless particular community or area is not specified and also, the time at which the age is to be specified.
11. Compiling Errors, i.e. wrong calculations or entries made during the processing and analysis of the data. Various operations of data processing such as editing and coding of the responses, punching of cards, tabulation and summarising the original observations made in the survey are a potential source of error. Compilation errors are subject to control through verification, consistency checks, etc.
12. Publication Errors. Publication errors, i.e. the errors committed during presentation and printings of tabulated results are basically due to two sources. The first refers to the mechanics of publication- the proofing error and the like. The other, which is of a more serious nature, lies in the failure of the survey organisation to point out the limitations of the statistics.
Biased errors creep in because of:
(i) Bias on the part of the enumerator or investigator whose personal beliefs and prejudices are likely to affect the results of the enquiry.
(ii) Bias in the measuring instrument or the equipment used for recording the observations.
(iii) Bias due to faulty collection of the data and in the statistical techniques and the formulae used for the analysis of the data.
(iv) Respondents’ bias. An appeal to the pride or prestige of an individual introduces a bias called prestige bias by virtue of which he may upgrade his education, occupation, income etc. or understate his age thus, resulting in wrong answers. Moreover, respondents may furnish wrong information to safeguard their personal interests. For example, for income-tax purposes, a person may give an understatement of his salary or income or assets.
(v) Bias due to non-response.
(vi) Bias in the Technique of Approximations. If while, rounding off, each individual’s value is either approximated to next highest or lowest number so that all the errors move in the same direction, there is bias for overstatement or understatement respectively. For example, if the figures are to be rounded off to the next highest or lowest hundred, then each of the values 305 and 396 will be recorded as 400 and 300 respectively.
Owing to their nature, the biased errors have a tendency to grow in magnitude with an increase in the number of the observations and hence, are also known as Cumulative Errors. Thus, the magnitude of the biased errors is directly proportional to the number of observations.
The errors are termed as unbiased errors, if the estimated or approximated values are likely to err on either side, i.e., if the chances of making an over-estimate are almost same as the chance of making an under-estimate. Since, these errors move in both the directions, the errors in one direction are more or less neutralised by the errors in the opposite direction and consequently, the ultimate result is not much affected. For example, if the individual values, say, 385,415, 355,445 are rounded to the nearest complete unit, i.e., hundred, each one of them would be recorded as 400.
In this case, the values 385 and 355 give over estimating errors of magnitudes 15 and 45 respectively, while, the values 415 and 445 give under-estimating errors of magnitudes 15 and 45 respectively and in the ultimate result (approximation) they get neutralised.
Thus, if the number of observations is quite large, these unbiased errors will not affect the final result. Since, the errors in one direction compensate for the errors in the other direction, unbiased errors are also termed as compensatory errors.
Thus, we observe that the unbiased errors do not grow with the increase in the number of observations but, they have a tendency to get neutralised and are minimum in the ultimate analysis and the magnitude of the unbiased errors is inversely proportional to the number of items.