Exercise 3. Attitude on Social Security Spending and the Presidential Vote

Step A. Create and interpret Table 3A

Studies of voting behavior often attempt to explain voting behavior on the basis of certain policy issues. The hypothesis is that voters choose between candidates based on their issue positions. One public policy issue that has received attention in recent years concerns federal entitlement programs, such as Social Security or Medicare. Democrats have long been viewed as more supportive of these programs.

Social Security was not a major issue in the 2016 campaign, but there did appear to be some differences between the candidates on this issue. Clinton opposed any benefit cuts to Social Security, and she supported increased funding for the program through a tax increase; the Democratic Party Platform took similar positions. Trump was ambiguous about his proposals for Social Security: he said that he would not cut benefits and that he would save Social Security, but he also opposed increased funding through tax increases, a position that the Republican Party platform endorsed. Thus, there did seem to be a clear difference between the candidates on whether increased Social Security taxes were desirable, particularly for higher income individuals.

Did attitudes on Social Security influence voters in 2016? Perhaps the lack of attention given to this issue and the ambiguous positions of Trump resulted in little effect. On the other hand, it seems plausible to hypothesize that voters who wanted more spending on Social Security would have been more likely to vote for Clinton than would voters who did not favor increasing spending on Social Security. To examine whether attitudes on this issue affected how people voted, we can create and look at a table (Create Table 3A) that relates attitude on Social Security spending (J18) to the presidential vote. For the reasons suggested in exercise 1, you should use the recoded version of A02 that you created for that exercise, so that you examine only the major-party vote (i.e., only the Clinton and Trump voters).

If you ran Table 3A as suggested, you should have a table with five columns and two rows. Attitude toward Social Security spending (the independent variable) should be on the top of the table (the column variable), and the two-party presidential vote (the dependent variable) should be on the side of the table (the row variable). Percentages should be calculated by column (i.e., they should sum to 100 percent for each column). In reading your table, take care to properly interpret the percentages, remembering that they are column percentages, not row percentages.

You should attempt to answer these questions to see if you are able to correctly read the table and interpret the data:

  1. What percentage of the voters who favored increased spending on Social Security cast a ballot for Clinton? What percentage of the voters who favored less spending cast a ballot for Clinton? How did those who wanted the same level of spending vote?Because we are using the two-party vote, it makes no difference whether we look at the percentage voting for Clinton or for Trump. If those who favored more spending were more likely to vote for Clinton, then those who opposed more spending were more likely to vote for Trump.
  2. Overall, is there a clear relationship between these two variables? How strong is the relationship?

After examining Table 3A, you should conclude that those who favored increasing Social Security spending had a greater propensity to for Clinton than did those who wanted to hold spending constant or even reduce it. Does this prove that voters cast their ballots in part on the basis of this issue, or is there something else that might cause this relationship? In social science research, we have to be careful about attributing causality. Often, two variables are associated not because one affects the other but because both are influenced by a confounding variable.

Step B. Create a three-variable table

One possible confounding variable is party identification. The relationship in Table 3A could be due to both variables being affected by party identification. To examine this possibility, you need to create Table 3B, a three-variable table that shows the relationship between attitude on Social Security spending, presidential vote, and party identification. To keep the table simple and to have a sufficient N for each column, you should recode J18 so that it has just two categories (increased spending versus same or decreased spending), as there are too few voters who favored decreased spending to keep that as a separate category. Also, use the recoded version of party identification that you created for exercise one (Democrats, independents, and Republicans) to keep the table simple and have a sufficient N in each column.

J18 has three response categories, running from increased spending to decreased spending. While these three categories are fine for the two-variable table that you generated to start this exercise, keeping all the categories of J18 will create some problems when you generate a three-variable table in which party identification is a control variable.

There are not that many voters who favored decreased spending, so the result would be a three-variable table in which there are too few respondents in some table columns, making the percentages in those columns unreliable estimates of population patterns. Recoding J18 so that it has only two categories—more spending versus same or less spending—will solve that problems.

Thus, you should recode J18 so that those who favor more spending are in one category and those who favor either the same level of spending or less spending are in a second category.

To create a three-variable table, you must specify a row variable, a column variable, and a control variable. You normally would set up the table by putting your independent variable as the column variable (this will put it on the top of each subtable), your dependent variable as the row variable (this will put it on the side of each subtable), and your control variable so that it specifies each subtable (there will be one subtable for each category of the control variable). When you are in the SDA crosstabulation program, enter your variables in the row, column, and control dialog boxes. You also will have to enter your recoding instructions for each variable in the appropriate dialog box.

In this example, your control variable should be party identification. You want to recreate Table 3A for each category of party identification—Democrats, independents, and Republicans. You want to use your recoded version of party identification, which has only three groups, because if you use the full seven-category version of party identification, you will have too many subtables, and the result will be many columns with small Ns and a fairly complex table that will be difficult to interpret.

If the table is set up in the above fashion, you normally would want percentages by columns. You should click this option under table options.

You should be sure to have the weight on and that you have selected the weighted Ns to appear in the table. The unweighted data are not a representative sample of the electorate, so be careful not to mistakenly use them.

Because the data are weighted, which means that individual respondents may count as more or less than one person (e.g., as .75 or 1.35 persons), the number of respondents in each cell (the Ns) probably will not be whole numbers. If you prefer to have Ns that are whole numbers, you can revise the output to do that by using the “revise the display” options that appears to the left of the table that you generated.

If statistics are desired, that option should be checked under table options. For a discussion of the statistics that are commonly used for contingency tables, see the section on data analysis. In these exercises, we have not asked you to generate statistics, but your instructor may suggest doing so.

The SDA crosstabulation program will produce both a table and a chart, but the chart is not necessary, as all of the information that you need will be contained in the table that you generate. You can revise the output to drop the chart if you like by using the “revise the display” option.

Step C. Interpret Table 3B

If you ran Table 3B as suggested, you should have a table that consists of three subtables, one for each category of party identification. Each subtable should have two columns and two rows. Each subtable should have the recoded version of attitude toward Social Security spending (the independent variable) on the top of the table (the column variable), and the two-party presidential vote (the dependent variable) on the side of the table (the row variable). Percentages should be calculated by column (i.e., they should sum to 100% for each column). There should be one subtable for Democrats, one for independents, and one for Republicans. In reading your subtables, take care to properly interpret the percentages, remembering that they are column percentages, not row percentages.

You should attempt to answer these questions to see if you are able to correctly read the table and interpret the data:

  1. What is the relationship between attitude toward Social Security spending and presidential vote among Democrats? How about among independents? Among Republicans? Overall, what relationship exists between this attitude and voting?
  2. How does the relationship between attitude toward Social Security spending and presidential vote in each subtable compare to the original relationship in Table 3A? Is it stronger, weaker, or about the same? What conclusion do you draw from this?

Your interpretation of Table 3B should have concluded that there was very little association between attitude toward Social Security spending and presidential vote once party identification is controlled for, which is clear when we examine each subtable:

  • Among Democrats, the percentage voting for Clinton is just as high among those who do not favor more spending as it is among those who want more.
  • Among independents, those who favor more spending are more likely to vote for Trump, which is the opposite of what we found in the original two-variable table.
  • Among Republicans, there is virtually no difference in the vote between those who want more spending and those who do not. The two percentage point difference between these two groups is too small to treat as a true difference in the vote.

This is an excellent example of a spurious association because the original relationship in Table 3A disappears in all three subtables.

In this example, the relationship between attitude toward Social Security spending and presidential vote is extremely weak once we control for party identification. This tells us that the original relationship between the attitude and the vote is basically spurious.