National Panel Survey 2010-2011

The methodology described in this paper builds upon published documentation from established panel surveys, such as the Panel Study of Income Dynamics [PSID], conducted since 1968 by the Institute for Social Research at the University of Michigan; and the British Household Panel Survey [BHPS], whose first 13 waves were conducted between 1991 and 2003 by Institute for Social and Economic Research at the University of Essex. Both the PSID and the BHPS are nationally-representative panel surveys in the USA and the UK respectively.

The weights are developed following these steps:

1. Begin with the “base weights” or those calculated during the first round of the survey; The panel weight calculations are based on the 2008/2009 household weights. These weights are based on the inverse probability of selection, EA level non-response correction, trimming of outlier weights, and a post-stratification correction11. These probability weights form the first component of the 2010/2011 calculations.
   W1=W2008

2. incorporate fair-share weights for composition changes;
   Based on the tracking protocols, the tracking for split off rules for the TZNPS allow for the incorporation of people who now live with original sample members. For example a young adult living with his parents in 2008, may be 2010 have formed a new household, getting married and having a child. The wife and infant will be incorporated into the survey and thus require a probability of selection. Such corrections are routinely used to distribute weight to new sample members in panel surveys. See Rendtel and Harms (2009) for a discussion of several different methods of weight correction. Because split-off individuals are tracked and interviewed in their new households, there are multiple ways that a household can become part of the survey.

o Either by being selected initially for the first round of the TZNPS
o By receiving a member that came from a household that was selected for the first round of the TZNPS.

In an ideal world, it would be possible to know the probability of selection that each new member brought into the household, and adjust the household weight accordingly. This is necessary since households receiving members have higher probabilities of selection (and therefore lower weights) because the household could have been selected in multiple ways. Since we cannot know the probabilities of every member, we must make simplifying assumptions. The first simplifying assumption is that the arriving members arrived together from one other household. This would be the case if a man and woman get married and set up a new household, or in the case of an older relative moving in with adult children. In certain cases, however, arriving members come from more than one household. Assuming only two source households underestimates slightly the probability of selection (and therefore over-estimates the weights). Incidence of these cases is believed to be relatively rare, and any resulting bias should be negligible. The second simplifying assumption we make is that the arriving members have the same probability of selection, on average, as those members that are already there. This would not be true on a case-by-case basis but would be true in the aggregate. With these simplifying assumptions, we add a factor of ½ for all households, ‘split’ or ‘parent’ that have new members arriving from other households.
This takes into account the fact that they could have been selected in two ways, and assumes the probability of selection is equal.

A limitation of the panel methodology is that the represented population is not identical to the 2010 Tanzanian household population, as it does not include immigrants in new households. Inclusion of these groups would necessitate refreshing the sample with new households. However, the represented population is close enough to the 2010 Tanzanian population to permit the desired cross-sectional estimates.

3. derive attrition adjusted weights for all individuals, including split-off10 households, then aggregate these weights to the household level; All household panel surveys must tackle the problem of attrition, sample members selected for follow interview which cannot be located and/or interviewed. The methodology used to adjust weights for attrition in the UNPS follows Rosenbaum & Rubin (1984). We use predicted response probabilities from a logistic regression model based on the covariates to form the weighting classes or cells. This approach has also been adopted in the PSID; see for example, Gouskova (2008).

The total sample size for the second round of the NPS has a total sample size of 3924 households. This represents 3168 round one households, a re-interview rate of over 97 percent. In addition, of the 10,420 eligible adults (over age 15 in 2010), 9,338 were re-interviewed, a re-interview rate of approximately 90 percent. To obtain the attrition adjustment factor the probability that a sample household was successfully reinterviewed in the second round of surveys is modeled with the linear logistic model at the level of the
individual. A binary response variable is created by coding the response disposition for eligible households that do not respond in the second round as 0, and households that do respond as 1.

Then a logistic response propensity model is fitted, using 2005 UNHS household and individual characteristics measured in the first wave as covariates. In a few limited cases, values of unit level variables were missing from the 2008/2009 household dataset. These values were imputed using multivariate regression and logistic regression techniques. Imputations are done using the ‘impute’ command in Stata at the level of the UNPS strata (urban/rural and region). Overall, less than one percent of the variables required imputation to replace missing values. The estimated logistic model is used to obtain a predicted probability of response for each household member in the 2010/2011 survey. These response probabilities were then aggregated to the household level (by calculating the mean), the using the household-level predicted response probabilities as the ranking variable, all households are ranked into 10 equal groups (deciles). An attrition adjustment factor was then defined as the reciprocal of the empirical response rate for the household-level propensity score decile.

4. post-stratify the pooled weights to known population totals.

To reduce the overall standard errors, and weight the population totals up to the known population figures, a post-stratification correction is applied. Based on the projected number of households in the urban and rural segments of each region, adjustment factors are calculated. This correction also reduces overall standard errors (see Little et al, 1997).