HBAI Datasets Guidance for the Production and Checking of Analysis

Transcription

1 UK Data Archive Study Number Households Below Average Income, 1994/ /17 HBAI Datasets Guidance for the Production and Checking of Analysis This user documentation has been designed for SAS users. However, much of this material is relevant to other users. End User (or Anonymised) Datasets 1. These datasets contain no names, addresses, telephone numbers, bank account details, NINOs or anything that can be considered as being disclosive under the terms of the ONS Disclosure Control. The changes to the datasets are as follows; All ages above 80 are coded to 80 years of age. The variable for the amount of Council tax liability for the household has been dropped. All amount variables have been rounded to the nearest 1. Large households (with 10 or more individuals) have been removed. 2. All of the data is available without these restrictions on the Special Licence versions of the datasets. Before starting the query 3. Perform an initial assessment of the complexity of the analysis. This might feed into the level of checking which must be performed on the analysis once it is completed. This might include assessing whether the values come straight from the publication 1, whether the publication can be used to check totals, whether analysis is performed on the same basis as the publication (e.g. it asks for number of individuals living in a certain family type, as opposed to individual receipt of a benefit). 4. In addition to this, think of what will the data be used for. This will enable you to provide the most appropriate figures and advice. 5. Are there are old queries that can be used or amended for the current query? 1 1

2 6. Do you want Before Housing Costs (BHC), After Housing Costs (AHC) or both? Should the income be equivalised or unequivalised? Net or gross? If unsure, think if what the data will be used for, and assess what would be best for your purposes. 7. If, after proportion in low income values, are you interested in absolute or contemporary low income? 8. Are you interested in data for individuals, families or households for example? 9. If the query is complex, requires a lot of tables, and you are unsure as to what you actually want, it might be worth considering creating some empty tables first which you can then agree on or amend as necessary. Note: datasets HBAI 02/03 to HBAI 11/12 have been revised to include grossing factors that incorporate 2011 Census based mid-year population estimates. Data-steps in SAS 10. Which years or year does the analysis relate to? If you want the most recent, are you using the most recent dataset? 11. Do you require a time series? If so, do you want the output in publication prices, or want the values uprated/deflated to a common value. Have you used the correct inflation/deflation factors? HBAI CPI Deflators HBAI year Inflation index BHC AHC 1994/ / / / / / / / / / /

3 HBAI year Inflation index 2005/ / / / / / / / / / / / The deflators need to be used as follows. Income values need to be multiplied by a factor x/y, where x=the year you are converting into y=the year you are converting from So for example, if you want to transform a value of 100 from the 1996/7 dataset into average 2002/03 prices, you need to do the following: Uprated value = 100 * (2002/03 factor / 1996/97 factor) Note: from 2014/15 CPI, rather than RPI, deflators are used in HBAI to in-year deflate income variables and to uprate/deflate to a common year. See the HBAI Quality and Methodology Information Report 2 for further details. 12. Are you merging datasets of different levels? If so, are your merged datasets the correct level, e.g. individual, benefit unit, household, and have you used the right level variables for your analysis? One way of checking that you are merging to the correct levels is to examine the number of observations in your input datasets, and compare this to the merged dataset. You can do this by looking in the log

4 13. If you have merged datasets, have you used the correct by variables? When merging two datasets that are on the same level, you must use by variables down to the level of the datasets. So for example, if you are merging two datasets that are both on a benefit unit level, you must merge by sernum benunit. When merging two datasets that are on different levels, you must use by variables that go down to the level of the highest dataset. So for example, if you were to merge an adult with a benefit unit level dataset (and so create an adult level dataset), the higher level dataset is the benefit unit level dataset, so you would use by sernum benunit. You wouldn t be able to use person for the adult dataset, as this information isn t on the benefit unit level dataset, and so SAS won t be able to compute a match. An easier way of thinking about this is to match by those variables that are common to both datasets. 14. When merging HBAI datasets with FRS datasets, you must only merge those cases that appear on both datasets, and therefore not include FRS cases that we exclude due to missing spouses. To do this, after the HBAI dataset name in the set statement, write (in=a). Then, when merging, after the by statement, add a new line that says if a; This does the following: it creates a temporary variable a that doesn't stay in the final dataset. It is 1 if there was a record so the "if a;" statement acts as a gatekeeper only allowing records where a=1 through the rest of the datastep. 15. If you have used the retain statement, have you used the correct by values? You only need to go down to the level of x that your if first.x and if last.x statement goes down to. (Whatever x is, that is the level that your new dataset will be on). So for example, if you are using a benefits level table, and summing up to a benefit unit level, you would need to use by sernum benunit as well as using if first.benunit and if last.benunit. 16. Some variables are useful only at certain levels of dataset. For example, if you had a household level dataset, you would need to use depchldh as the number of children, not depchldb. Further to this, see the point above on the use of the retain statement for looping through records to produce a dataset at a higher level. 17. Have you used the right grossing factor? This is very important, and depends on what level the dataset is at, and also what you are grossing to. You can use your gross total to check that you are multiplying up to the correct amount (using the population totals in the publication). 4

5 TOTAL THAT YOU WANT TO GROSS TO 18. The table below gives a matrix of the level your dataset is on, and what you want to gross up to, as a means of finding the correct grossing factor. LEVEL YOUR DATASET IS AT: Individual Benefit Unit Household Number of children gs_newbu gs_newch can t be done Number of adults gs_newbu gs_newad can t be done Number of Working age adults Number of pensioners gs_newbu gs_newwa can t be done gs_newbu gs_newpn can t be done Total population gs_newbu gs_newpp can t be done Number of families can t be done gs_newbu can t be done Number of households can t be done can t be done gs_newhh or gs_newbu* * Dataset is restricted to where benunit = 1. Some extra, interesting notes/observations on grossing factors: a) If presenting gross income averages, you need to restrict to household level where benunit=1 and use gs_newbu. b) If using equivalised net income, use the benunit level file and use gs_newpp. Therefore, the total average net income will be the same as in the HBAI publication. c) Gs_newbu and gs_newhh are the same as each other, and are the equivalent of dividing gs_newpp by the sum of depchldb and adultb within each benefit unit. In fact, we use gs_newbu to construct gs_newpp, and we do this by multiplying gs_newbu by the sum of depchldb and adultb. d) If, for example, you are calculating any average of a variable that is on a household basis, you need to use gs_newhh. 19. Have you used the correct client group, e.g. if you are only interested in working-age adults, have you restricted your figure to this group? 20. Have you restricted it to the correct subset, e.g. only families with children as opposed to all families? 21. If you are making amendments to the income distribution, for example by stripping out disability benefits in order to see the result of this, as you are 5

6 looking for the effects of this in the real world, you should recalculate the mean and median income after deduction. 22. When using any FRS monetary values or HBAI monetary values other than S_OE_BHC and S_OE_AHC for use with HBAI, these values need to be equivalised and multiplied by the within-year deflator in order to get them into an average year price. 23. In general, good SAS coding means use of comments throughout that can help a reader understanding what is happening in the code at each step. Good code is also efficient code, which means keeping only those variables in a program that you actually need, and dropping all the others. This can especially make a difference when you have a large or unwieldy program. Robustness/Quality assurance checks 24. If you are looking at quite detailed breakdowns of the population, or a small group, are sample sizes sufficient for analysis? Percentages: The sample size relating to the denominator (i.e. the row/column total for row/column percentages) must be at least 100. If not suppress the entire row/column (..). Averages: The cell sample size must be at least 50. If not suppress the cell (..). 25. In addition to the validation section below, a further way of checking your figures is to take out any restrictions you have placed on your dataset, so that you now have the whole population. This way, you can check that your totals gross up to known tables. So for example, if you are looking at risk of low income for children in England, if you remove the regional restraint, does the total match the total number of children? Outputting the query to Excel 26. It is worth setting up links to a final summary table (or however you are going to be presenting the data) as opposed to copying and pasting values into a final table. This way, if the table needs tweaking slightly, any new output will automatically feed into the summary table. 27. If you are presenting a time series, breakdowns such as by region or ethnicity should not be presented as a time series without giving consideration to the volatility due to small sample sizes. To account for this, produce a three-year average estimates for ethnicity and region figures. For example, average for yr1-yr3 = (estimate in y1 + estimate in yr2 + estimate in yr3)/3. 6

7 28. If you have used a proc freq and a pivot table, and have then re-run code, have you refreshed the contents of the pivot table? If you haven t you might be using old output. 29. One thing that needs consideration is whether categories from the FRS/HBAI need switching round or re-ordering so that they are presented in a different order. An example of this is the presentation of Scotland and Wales in the publication, which is different to how these two categories are ordered in our datasets. It is easier to switch them in the code, as opposed to in Excel, as it is less messy, and has less potential for confusion when linking between different spreadsheets. Displaying the query in Excel 30. The following rounding conventions are used: Percentages and percentage changes: rounded to the nearest whole percentage point. Population totals: in millions, rounded to 1 decimal place. Weekly income values: rounded to the nearest pound. Annual income values: rounded to the nearest 100. Validation of figures/quality assurance 31. Are there any figures (either for population or risk etc) that can be validated using any figure in the publication? 32. Are there any old queries whose output can be matched against your output? 33. Finally, have a look at the figures and check whether they look sensible. For example, are BHC and AHC figures different (generally they will be, so if they are exactly the same, check the links and check that the final sheet is linking to the correct output). A handy hint is to view the spreadsheet as its underlying formulae (go to tools, options, and check the formulas tickbox under the View tab). This provides a quick way of checking that output has been sent where it really should go. If you are presenting a time trend, do the figures look sensible, and do they tell a story that appears sensible? If you are presenting compositional figures, do they sum to 100? 7