Hello there.  I can see there have been a few posts about this issue already, but I am still struggling to understand when best to weight cases when doing multiple imputation on SPSS. 

I am a masters student in the midst of analysing survey data for my dissertation, and I am no statistician.  I have prepared a sample weight to deal with  suspected non-response bias and would like advice please on whether to apply these case weights prior to any imputation, as part of MI, or afterwards.  I ask about the middle option because of the "Analysis Weight" box that's on the Impute Missing Data Values variables page (see image), whose purpose I am unclear about.    Many thanks for any help/advice.

This is a complicate situation.  The MI procedure considers two kinds of weights - frequency weights and what it calls regression weights.  What you have constructed are probably the first kind as the second is mainly used to deal with unequal error variances in regression.  How did you construct your weights?  Raking?  Something else?

Frequency weights are expected to be integers for imputation purposes, but they should be normalized to add up to the actual sample size.  It probably does not matter if they are fractional.  How to proceed also depends on what sort of analysis you are planning to do.  Custom Tables also supports effective base weighting which allows for fractional weights.  Sampling weights are appropriate for the complex samples procedures.

In most cases, though, for the multiple imputation stage, I would recommend using the frequency weight for the imputation stage as it balances the sample to be more representative of the population.

Bear in mind, though, that when modeling, the underlying assumption is that the relationship expressed, say, in a regression equation, is that the same relationship applies to the whole sample (given any conditioning variables in the model), so if you run both weighted and unweighted alternatives, in expectation the coefficients should be the same.  The variances and sig levels will differ, but the general pattern should be similar.  If they are very different, that suggests a modeling problem.

Hello there.  I can see there have been a few posts about this issue already, but I am still struggling to understand when best to weight cases when doing multiple imputation on SPSS. 

I am a masters student in the midst of analysing survey data for my dissertation, and I am no statistician.  I have prepared a sample weight to deal with  suspected non-response bias and would like advice please on whether to apply these case weights prior to any imputation, as part of MI, or afterwards.  I ask about the middle option because of the "Analysis Weight" box that's on the Impute Missing Data Values variables page (see image), whose purpose I am unclear about.    Many thanks for any help/advice.

Many thanks for your response Jon.

Can I confirm that your recommendation was to apply the NRB weights during the MI process, i.e. to put the NRB weight variable into the "Analysis Weight" box as per my screen grab?

You ask about how I created the NRB weights. This was done via an online one-question 'micro-survey' with non-respondents answering a 5 point Likert question that exactly matched one in the main survey.  The aim was to help deal with suspected self-select bias – i.e. respondents were keener on the topic (pro-environmental  behaviours in homes) than the average in the sample frame population.  I normalised the dataset to give fractional weightings to apply to main survey respondents broken down by the level they chose to respond to the question, i.e. the weight to apply to those who responded level 5 is 1.12 to help them match the sample population, those who chose level 4 weighted at 0.908, etc. I guess this is what you term a "frequency weight" rather than a "sampling weight".

In terms of the analysis I plan, the topic of the survey is the influence of green primary schools on green behaviours in the family homes of their pupils. So  I intend to examine this relationship  using the main survey data from parents integrated with data from a separate survey of the 22 participating schools.

Does this info make any difference to your advice?

Thanks

This is a complicate situation.  The MI procedure considers two kinds of weights - frequency weights and what it calls regression weights.  What you have constructed are probably the first kind as the second is mainly used to deal with unequal error variances in regression.  How did you construct your weights?  Raking?  Something else?

Frequency weights are expected to be integers for imputation purposes, but they should be normalized to add up to the actual sample size.  It probably does not matter if they are fractional.  How to proceed also depends on what sort of analysis you are planning to do.  Custom Tables also supports effective base weighting which allows for fractional weights.  Sampling weights are appropriate for the complex samples procedures.

In most cases, though, for the multiple imputation stage, I would recommend using the frequency weight for the imputation stage as it balances the sample to be more representative of the population.

Bear in mind, though, that when modeling, the underlying assumption is that the relationship expressed, say, in a regression equation, is that the same relationship applies to the whole sample (given any conditioning variables in the model), so if you run both weighted and unweighted alternatives, in expectation the coefficients should be the same.  The variances and sig levels will differ, but the general pattern should be similar.  If they are very different, that suggests a modeling problem.

Hello there.  I can see there have been a few posts about this issue already, but I am still struggling to understand when best to weight cases when doing multiple imputation on SPSS. 

I am a masters student in the midst of analysing survey data for my dissertation, and I am no statistician.  I have prepared a sample weight to deal with  suspected non-response bias and would like advice please on whether to apply these case weights prior to any imputation, as part of MI, or afterwards.  I ask about the middle option because of the "Analysis Weight" box that's on the Impute Missing Data Values variables page (see image), whose purpose I am unclear about.    Many thanks for any help/advice.

Many thanks for your response Jon.

Can I confirm that your recommendation was to apply the NRB weights during the MI process, i.e. to put the NRB weight variable into the "Analysis Weight" box as per my screen grab?

You ask about how I created the NRB weights. This was done via an online one-question 'micro-survey' with non-respondents answering a 5 point Likert question that exactly matched one in the main survey.  The aim was to help deal with suspected self-select bias – i.e. respondents were keener on the topic (pro-environmental  behaviours in homes) than the average in the sample frame population.  I normalised the dataset to give fractional weightings to apply to main survey respondents broken down by the level they chose to respond to the question, i.e. the weight to apply to those who responded level 5 is 1.12 to help them match the sample population, those who chose level 4 weighted at 0.908, etc. I guess this is what you term a "frequency weight" rather than a "sampling weight".

In terms of the analysis I plan, the topic of the survey is the influence of green primary schools on green behaviours in the family homes of their pupils. So  I intend to examine this relationship  using the main survey data from parents integrated with data from a separate survey of the 22 participating schools.

Does this info make any difference to your advice?

Thanks

This is a complicate situation.  The MI procedure considers two kinds of weights - frequency weights and what it calls regression weights.  What you have constructed are probably the first kind as the second is mainly used to deal with unequal error variances in regression.  How did you construct your weights?  Raking?  Something else?

Frequency weights are expected to be integers for imputation purposes, but they should be normalized to add up to the actual sample size.  It probably does not matter if they are fractional.  How to proceed also depends on what sort of analysis you are planning to do.  Custom Tables also supports effective base weighting which allows for fractional weights.  Sampling weights are appropriate for the complex samples procedures.

In most cases, though, for the multiple imputation stage, I would recommend using the frequency weight for the imputation stage as it balances the sample to be more representative of the population.

Bear in mind, though, that when modeling, the underlying assumption is that the relationship expressed, say, in a regression equation, is that the same relationship applies to the whole sample (given any conditioning variables in the model), so if you run both weighted and unweighted alternatives, in expectation the coefficients should be the same.  The variances and sig levels will differ, but the general pattern should be similar.  If they are very different, that suggests a modeling problem.

Hello there.  I can see there have been a few posts about this issue already, but I am still struggling to understand when best to weight cases when doing multiple imputation on SPSS. 

I am a masters student in the midst of analysing survey data for my dissertation, and I am no statistician.  I have prepared a sample weight to deal with  suspected non-response bias and would like advice please on whether to apply these case weights prior to any imputation, as part of MI, or afterwards.  I ask about the middle option because of the "Analysis Weight" box that's on the Impute Missing Data Values variables page (see image), whose purpose I am unclear about.    Many thanks for any help/advice.

Hello there.  I can see there have been a few posts about this issue already, but I am still struggling to understand when best to weight cases when doing multiple imputation on SPSS. 

I am a masters student in the midst of analysing survey data for my dissertation, and I am no statistician.  I have prepared a sample weight to deal with  suspected non-response bias and would like advice please on whether to apply these case weights prior to any imputation, as part of MI, or afterwards.  I ask about the middle option because of the "Analysis Weight" box that's on the Impute Missing Data Values variables page (see image), whose purpose I am unclear about.    Many thanks for any help/advice.

When it comes to applying weights during multiple imputation, there are a few considerations to keep in mind. The choice of when to apply the weights depends on the specific goals of your analysis and the characteristics of your dataset. Here are three common approaches:

1. Apply weights before imputation:

   • If the weights are intended to adjust for non-response bias, you may consider applying them before performing the imputation. This approach allows the imputation process to take the weights into account, potentially improving the accuracy of the imputed values.
   • However, be cautious when using weights with imputation, as they can potentially affect the distribution of the imputed values and the analysis results.

2. Apply weights during imputation:

   • In SPSS, the "Analysis Weight" option in the Impute Missing Data Values dialog box is typically used to account for survey weights or other weights during imputation.
   • If you have a specific weight variable that you want to consider during imputation, you can specify it in this "Analysis Weight" box.
   • This approach allows the imputation process to incorporate the weights directly, potentially influencing the imputed values and the subsequent analyses.

3. Apply weights after imputation:

   • Alternatively, you can choose to apply the weights after completing the imputation process. Once you have imputed the missing values, you can then use the weighted data for subsequent analyses.
   • This approach treats the imputed data as complete cases and applies the weights as you would with any other complete dataset.

Ultimately, the decision on when to apply the weights depends on your research goals, the characteristics of your dataset, and the recommendations of your supervisor or dissertation committee. It may be helpful to consult with them to determine the best approach for your specific study.

------------------------------
Youssef Sbai Idrissi
Software Engineer

Original Message:
Sent: Mon July 03, 2023 04:55 AM
From: David Libbert
Subject: When and how to apply sample weights when doing multiple imputation.

Hello there.  I can see there have been a few posts about this issue already, but I am still struggling to understand when best to weight cases when doing multiple imputation on SPSS. 

I am a masters student in the midst of analysing survey data for my dissertation, and I am no statistician.  I have prepared a sample weight to deal with  suspected non-response bias and would like advice please on whether to apply these case weights prior to any imputation, as part of MI, or afterwards.  I ask about the middle option because of the "Analysis Weight" box that's on the Impute Missing Data Values variables page (see image), whose purpose I am unclear about.    Many thanks for any help/advice.

------------------------------
David Libbert
------------------------------

Hello there.  I can see there have been a few posts about this issue already, but I am still struggling to understand when best to weight cases when doing multiple imputation on SPSS. 

I am a masters student in the midst of analysing survey data for my dissertation, and I am no statistician.  I have prepared a sample weight to deal with  suspected non-response bias and would like advice please on whether to apply these case weights prior to any imputation, as part of MI, or afterwards.  I ask about the middle option because of the "Analysis Weight" box that's on the Impute Missing Data Values variables page (see image), whose purpose I am unclear about.    Many thanks for any help/advice.