SVREWEIGHT procedure

Modifies survey weights for particular observations, adjusting other weights in the sampling unit or stratum to ensure that the overall sum of the weights remains unchanged (S.D. Langton).


Options

PRINT = strings
Controls printed output (summary); default summ

METHOD = strings
What to reweight over (all, stratum, samplingunits, lowest); default lowe

WEIGHTS = variate
Initial weights

OUTWEIGHTS = variate
Final weights

STRATUMFACTOR = factor
Stratification factor; default * i.e. unstratified

OUTSTRATUMFACTOR = factor
Saves a modified stratification factor with the reweighted observations in their own stratum

SAMPLINGUNITS = factor
Factor indicating the primary sampling units; default *, i.e. single stage design

LABELS = variate, text or factor
Labels for each unit


Parameters

OBSERVATIONS = scalars, variates or texts
Observations to reweight

NEWWEIGHTS = scalars or variates
New weights (default inserts a missing value, indicating that the observation should be removed)


Description

Item non-response (i.e. a missing value for one question although valid responses are present for others) and outliers are two common problems in survey data. If the item response occurs entirely at random, one method of dealing with it is to analyse the question with a modified set of weights with the weight for the missing observation redistributed over the rest of the units in the stratum or sampling unit. This could be achieved by calculating the weights again from scratch, but it is often preferable to modify the existing weights variable. Similarly if the influence of outliers is reduced by giving them a reduced weight (see Lee 1995 for a discussion of this subject), the weights for the remaining observations must be adjusted to maintain the same sum of weights.

   The units whose weights are to be adjusted are specified by the OBSERVATIONS parameter. By default these are identified by the unit numbers of the observations, but, if the LABELS option is set to a variate, factor or text, the appropriate values are used instead. Multiple observations can be specified either as a list of scalars (or single-valued texts if appropriate), or by variates or texts with multiple values. By default the procedure assumes that the observations should have their weight set to missing so that they are excluded from analysis by TABULATE or SVTABULATE. Alternatively NEWWEIGHTS can be used to specify the required weights to insert. This can be set to a scalar if the same weight is to be used for every unit specified by the corresponding OBSERVATIONS variate, text or scalar. Alternatively, it can be set to a variate of the same length as the corresponding OBSERVATIONS setting.

   The METHOD option specifies the level at which the weights are redistributed, so that, for example, setting METHOD=stratum changes the other weights in the stratum containing the observation so that their total remains unchanged. If METHOD is unset the procedure works to the lowest specified level, i.e. sampling units if these are specified, or otherwise the strata. If the stratification factor is also unspecified the redistribution takes place over all other observations.

   Where reduced weights (typically 1.0) are allocated to outliers because they are genuine but not representative of the wider population, these units are often placed in their own stratum; the OUTSTRATUMFACTOR option can be used to create such a suitable stratification factor.

 

Options: PRINT, METHOD, WEIGHTS, OUTWEIGHTS, STRATUMFACTOR, OUTSTRATUMFACTOR, SAMPLINGUNITS, LABELS.

Parameters: OBSERVATIONS, NEWWEIGHTS.


Action with RESTRICT

Any restrictions are ignored.


References

Lee, H. (1995). Outliers in Business Surveys. Chapter 26 of Business Survey Methods (ed. Cox, Binder, Hinnappa, Christianson, Colledge & Kott). Wiley, New York.