3 article(s) found.

Allocation of Dual-Frame Telephone Survey for Given Cost

**Keyword**cellphone survey; effective sample size; survey cost; dual-frame

telephone survey; unequal weighting effect

Show abstract

Full content

With the increase of cell phone usage in recent years, traditional landline surveys face a problem of incomplete coverage. It is now necessary to conduct dual-frame telephone surveys that includes cell phone samples and landline samples. Designing a dual-frame telephone survey requires a decision on the sample allocation. The allocation of the sample to the dual-frame associates with the unequal weighting effect and the survey cost. Therefore, this study aimed to illustrate an optimal allocation of respondents from landline and cellphone frames that result in the lowest unequal weight effect (i.e., the highest effective sample size) for a given cost by using the relative unit cost of obtaining a cell respondent compared to a landline respondent from a comparison study of survey cost, and an unequal weighting effect from “Public Value and Electronic Governance.” The results suggested that the optimal design will have 64.18% of the sample completes from the landline frame, and 35.82% of the sample completes from the cellphone frame in a cell-phone-only screened design. Additionally, this paper shows that the sample sizes of cell phone only could be a function of unequal weight effect and survey cost. Thus, the organizer of the cell-phone-only screened design could substitute parameters into the function depending on different situations.

Is Weighting a Routine or Something that Needs to be Justified? (in English)

Show abstract

Full content

Survey research as a method of collecting sample data is supposed to produce sample statistics which can estimate the corresponding population parameters if the sampling design is appropriate. However, for reasons such as unit non-response, survey data is usually weighted by the institutes that collect the data or by researchers who analyse the data in order to correct or diminish the discrepancies between sample and population. Sample statistics based on weighted data are more representative of the population parameters than unweighted data in terms of some demographic characteristics.Therefore, to some extent, it seems legitimate to weight data and this manipulation has become a routine when dealing with survey data.

It is true that to weight data could be helpful, but this manipulation needs justifications. This paper therefore tries to argue that to weight data is no panacea and should not be taken for granted when considering the examples in Taiwan’s Election and Democratization Studies (TEDS) surveys. The first section discusses why weighted data is not necessarily representative of the population. As the TEDS surveys show, the turnout, the vote shares of parties, and marital status become more deviant from the population parameters after weighting the data.

If the focus is the relationships between variables, the correlations may be changed by weighting the data in bivariate or multivariate analysis. However, it is not clear whether we manufacture relationships which do not exist or if weighting the data actually helps us approximate the relationships that already exist in the population. Besides, it should be noted that to weight data set as a whole only deals with the problem of unit non-response, but does not solve the problem of item non-response.

The third section discusses why most efforts should be devoted to examining and improving questionnaires, sampling designs, and interviewerm straining and supervision, instead of simply appealing to post-weighting. If everything necessary has been tried, weighting data may be the last resort to improve the estimates. But the justifications for the selection of auxiliary variables and the methods of calculating weight factors should be provided rather than doing it without any explicit considerations. It is also important to consider whether the consequence of weighting is positive or negative.

It is true that to weight data could be helpful, but this manipulation needs justifications. This paper therefore tries to argue that to weight data is no panacea and should not be taken for granted when considering the examples in Taiwan’s Election and Democratization Studies (TEDS) surveys. The first section discusses why weighted data is not necessarily representative of the population. As the TEDS surveys show, the turnout, the vote shares of parties, and marital status become more deviant from the population parameters after weighting the data.

If the focus is the relationships between variables, the correlations may be changed by weighting the data in bivariate or multivariate analysis. However, it is not clear whether we manufacture relationships which do not exist or if weighting the data actually helps us approximate the relationships that already exist in the population. Besides, it should be noted that to weight data set as a whole only deals with the problem of unit non-response, but does not solve the problem of item non-response.

The third section discusses why most efforts should be devoted to examining and improving questionnaires, sampling designs, and interviewerm straining and supervision, instead of simply appealing to post-weighting. If everything necessary has been tried, weighting data may be the last resort to improve the estimates. But the justifications for the selection of auxiliary variables and the methods of calculating weight factors should be provided rather than doing it without any explicit considerations. It is also important to consider whether the consequence of weighting is positive or negative.

On Minimum-Discrimination-Information (MDI) Method of Weighting: an Application to the 2001 Taiwan's Election and Democratization Study(TEDS) (in Chinese)

**Keyword**goodness-of-fit tests; raking; post-stratification weighting; minimum discrimination information weighting; TEDS; cross-entropy.

Show abstract

Full content

Goodness-of-fit tests allow us to examine if the sample at hand is representative enough of the population to ensure accurate statistical inferences of parameters. When the sample fails the tests, survey researchers often appeal to reweighting as a remedy. Post-stratification and raking are perhaps the two most popular weighting methods. However, post-stratification requires the knowledge of multivariate joint distribution of the population when more than one post-stratifying variable is considered. Without such detailed information, raking comes as a rescue since it requires only the knowledge of marginal distributions of selected variables. Popular as it may be, raking takes no account of associations among post-stratifying variables. Furthermore, it relies heavily on Chi-squared tests and a pre-selected p-value (usually 0.5) as the stopping rule of iteration, an ad hoc rule justified only by convenience.

This article proposes a third way of Weighting, which we call it the minimum-discrimination-information (MDI) method. MDI approach finds optimal (in terms of minimum cross-entropy) relative weights by treating sample joint distribution as prior and known population marginal distributions as constraints. We first explain the rationale behind this proposed MDI method and then use TEDS 2001 survey data to compare the estimates of raking and MDI weights. We find that nearly 70 percent of the latter indeed replicate the Census 2000 population joint distribution better than the former. We thus conclude that MDI method is an approach worth further theoretical investigation.

This article proposes a third way of Weighting, which we call it the minimum-discrimination-information (MDI) method. MDI approach finds optimal (in terms of minimum cross-entropy) relative weights by treating sample joint distribution as prior and known population marginal distributions as constraints. We first explain the rationale behind this proposed MDI method and then use TEDS 2001 survey data to compare the estimates of raking and MDI weights. We find that nearly 70 percent of the latter indeed replicate the Census 2000 population joint distribution better than the former. We thus conclude that MDI method is an approach worth further theoretical investigation.