Ch 6

Multiple indicators - several q on survey addressing same concept, also interviewers asking 'essential q' and then 'extra q' that asks that same info slightly differently.

Composite measures are used freq in quantitative res: 1. no single indicator covers meaning, 2. want to use ordinal measure of var with a range of variation, 3. data analysis.

Indexes and scales (esp scales) are efficient data reduction devices: assign scores, not loosing details of response.

Index v Scale:
Both:
ordinal
rank-order units of analysis in terms of vars
score gives indication of relative 'religiosity'
use composite measures: measurements based on more than one data item. SHOULD ONLY MEASURE ONE DIMENSION. (unidimensionality)
Diffs:
Index: accumulate score assigned to individual attributes (1 point for each q).
Scale: assign score to patterns of resp; takes advantage of diff in intensity among attribs of same var to ID distinct patterns of response.
'idealized action patterns'

Index Construction:

1. In selecting items for index, does the item have face validity?

2. Are you measuring the concept in a general way or a specific aspect of the concept? (bal measure of religiosity or a measure only of ritual participation?)

3. Select items differing in variance (1 item ID as conservative, another might pick up a few more)

*I don't like the 'general or specific' and the 'exam of empirical rel' step on p150. Help?

4. Examine empirical rel among items included in proto-index. (Empirical Rel - when answers to one q let predict answers to other qs).

5. Find bivariate relationships among items and drop items w/o relationships to other items on index, unlikely that they really measure concept. Bivariate Rel - rel bet 2 vars, responses on 2 vars likely to get same responses. Also, drop items that VERY strongly correlate as they're prob the same q.

ASIDE:

Indicators should be rel if they are 'effects' of same var. However, not case when indicators are 'cause' rather than 'effect' of variable.

Social interaction - time spent w/ fam, friends, coworkers. 3 indicators 'cause' degree of social interaction.

Self-esteem - 'good person', 'like self'. Person w/ high self esteem should Y both.

Decide if indicators are causes or effects of var before using inter correlations to assess validity.

Here's another place I fall apart -- do we need to know the percentage tables he used to analyze his physician example? I don't remember Nigem covering it in class or making a big deal out of it - do we not need to know it? Can you summarize? (?!! P 154 - 155?!!)

Index Scoring:
Assign scores for particular responses.
Decide desirable range of index scores (how many index 'points' is conservative?).
How far into extremes does index extend (consider variance and the tails of the normal curve). Your goal is to have an adequate # of cases at each point on the index, generally index scoring is equally weighted.

How do you handle missing data?

• If few cases, simply exclude from index construction. (Will exclusion result in biased sample?)

• Treat missing data as one of available responses (you might decide that failure to answer meant no, if respondent answered yes and left some blank)

• Analysis may yield meaning - respondents failed to answer a Q were generally consistently conservative on other items -- you may decide to score accordingly.

• Assign 'middle' value

• Use proportions of what observed (if 4/4 answered strongly conservative, may score '6', if 2/4, may score '3')

Best method is to construct index through multiple methods and compare results.

Index Validation - does the index measure what it says it measures? (Does your index rank-order people in their degree of conservatism?)

1. Item Analysis - internal validation, examine extent to which composite index is related to or predicts responses to individual items it comprises. If item adds nothing, trash it.

2. External Validation - people who scored as politically conservative on your index should score as conservative by other methods as well. (most conservative index scorers should be most conservative on all other q on survey)

3. Bad Index vs Bad Valida tors - This can be a problem, check carefully.

Scale Construction:

Scales offer more assurance of ordinality by taking into consideration intensity structures among indicators. (Is the senator who voted for 7 moderately conservative bills more conservative than the senator who voted for 4 strongly conservative bills (rejecting the others cause they were too moderate)?

Bogardus Social Distance Scale -- teq for determining willingness of people to socially relate to certain other people. If person allows contact next door, they'd allow person to live in country... etc. Logical structure of intensity.

1. live in country?

2. Live in community?

3. Live in neighborhood?

4. Next door?

5. Marry child?

Thurstone Scale - format for generating groups of indicators of var that have empirical structure to them. Judges given list of indicators of a var and rated on intensity. Disagreement among Judges gets indicators tossed as ambiguous. Then items selected to represent each scale score, which then used in survey. Respondents who hit a strength of 5 would be expected to 'hit' the lower indicators too, but not hit indicators above 5.
Incredibly resource intensive, would have to be updated periodically.

Likert Scaling - goes one step beyond regular index construction, calcs avg index score for those agreeing with individual statements making up 'index'. As result of item analysis, respondents could be rescored using the avg index score for each item.
Too complex to be used frequently.

Semantic Differential - determine dimensions you want subjects to judge something and then find 2 polar opposite words along each dimension. (dimension of music: enjoyability, use enjoyable, unenjoyable. Dimension of music: complexity, use complex and simple. Etc). Allow individuals to check box along those continuums.

Guttman Scaling - based on notion that anyone giving strong indicator of some variable will also give weaker indicators.

Scale types - patterns of response that form a scale. See Table 6-2 for example.

I'm iffy on the ex given in book - I understand the example but can't extract a def from it. Skipped rest of Guttman Scale.

Typologies: summary of intersection of 2 or more variables, creating set of categories or types. Typologies MUST be used as the independent variable.

Also iffy on this, need def from somewhere else??