This study employed a multi-stage sampling procedure to achieve a random, representative sample of the driving public age 16 or older. The design employed was a stratified Random-Digit-Dial (RDD) sample design, which resulted in a sample that was consistent with earlier rounds of the study. It was very important to maintain a consistent sampling structure with earlier execution in order for the samples to be comparable across time.

First, the universe of residential telephone listings was identified within each of the geographic U.S. Census regions. A systematic sample of telephone number banks within each region was then drawn. A telephone number bank consists of the blocks of 100 sequential telephone numbers where only the last two digits of the 10-digit telephone number vary. For example, within the area code 202, exchange 366, one bank would be 202-366-12XX. This procedure provides for an equal probability of selection for each working residential telephone number in the U.S. (both listed and unlisted residential telephone households). A random-digit-dial (RDD) procedure was used to generate the last two numbers for a full ten-digit phone number within each selected 100-number bank

The second stage of selection occurred at the household level. Once a telephone number had been selected for inclusion, one person age 16 or older living in that household was randomly selected to participate. The household-level selection was made using the most recent birthday method, which represents a true random selection of household members, and is considered much less intrusive than the purely random selection method or grid selection that require enumeration of all household member in order to make a respondent selection. Once a person was selected for inclusion in the study, that person could not be replaced by another person in the household. If the selected person refused to participate, refusal conversion attempts were made to obtain the responses from the selected person. If refusal conversion attempts failed, the entire household was substituted to maintain the representativeness of the sample.

Telephone interviews were completed with persons age 16 or older living in the United States during each execution of the study. Interviewing took place generally over a two- to three-month period in the Fall of the study year (1993, 1995 and 1997). Some interviews in each wave were completed in Spanish using a Spanish-language version of the questionnaire.

All sample management, interview scheduling, conducting and monitoring of interviews, and progress reporting of data collection was handled by Gallup’s state-of-the-art computer assisted telephone interviewing (CATI) system. A comprehensive data collection plan was maintained to ensure that high response rates, high data quality and low respondent burden were achieved. The plan involved a call design scheme to optimize telephone coverage and contact with respondents, and to minimize no contacts and refusals. Gallup’s internal interviewer recruitment, training and monitoring procedures are also designed to support these aims for this and all studies conducted by Gallup.

The CATI programming process included identification of data locations, keying in question text, responses and corresponding codes, as well as acceptable response ranges, consistency checks, interviewer instructions, skip patterns, and help screens. Two kinds of range and consistency checks were programmed: hard and soft checks. Responses initially entered by interviewers that were outside the hard range were not accepted by CATI. These required the interviewer to clarify with the respondent their initial response (e.g., if the question asked how many days of the past 30 they consumed alcohol, a response of 31 would not be accepted by the CATI system). Soft range checks prompt the interviewer to verify the response. The questionnaire design and layout pass through a strict internal hard copy "proofing review" before it reaching the programming stage. The CATI program was once again proofed before interviewing began. Separate questionnaires were programmed in both English- and Spanish-language.

Weighting occurred in multiple stages in each study wave and was designed to equalize selection probabilities at both the household and individual levels as well as adjust for non-response bias by demographics. Each step was made using data weighted from the previous step.

All data was post-stratify weighted to correct for the imposed disproportional sample and to adjust for any disproportionality by age, gender, and race due to selecting just one person within a household (clustering effect) and unequal participation rates. The weighting was conducted in a three-stage procedure. In the first stage weighting, Gallup adjusted for the unequal probability of selection for households. In the RDD procedure, households with more than one telephone line had a higher probability of selection in our sample. This disproportionality was corrected by applying an inverse weight to each respondent based on the number of residential telephone lines in his/her household.

The second stage weighting adjusted for any unequal probability of selection within a household. While the study is based on the total non-institutionalized residential population of the U.S., the actual sample units are households. Persons living in households with only one person of driving age have a higher probability of selection than those in households with several eligible persons. In the second stage weight, Gallup applied a weight to each respondent in the inverse to the number of persons of driving age in the household.

The third stage weight corrected for any unintentional disproportionality due to unequal participation rates among respondents by key demographic characteristics. At this stage, Gallup weighted the actual respondent database (weighted in the first two stages) to match the known demographic characteristics of the U.S. population by age, race, and gender based on the most recent Census Population Projections. This demographic weighting is done in several stages.

The proportion of Hispanics/non-Hispanics were first adjusted to reflect the most recent Census Bureau estimates by census region. White/nonwhite distributions were then reviewed by census region and adjusted if necessary. The data were then examined and the distribution of gender by age (using three age categories (16-34;35-54;55+) corrected if more than +/- 3% variation from the population projections by census region.

In interpreting survey results, it should be borne in mind that all sample surveys are subject to various types of potential errors. Errors may occur due to non-response (where selected respondents are never reached or refuse to participate), interviewer administration error (where a response can be mis-keyed or misinterpreted by the interviewer), incomplete or inaccurate answers from the respondent or sampling less than the total population, among others.

The sampling design employed in this study was used to produce an unbiased estimate of the stated target population. An unbiased sample will have the same characteristics and behaviors as those of the total population from which it was drawn. In other words, with a properly drawn sample, we can make statements about the target population within a specific range of certainty. Sampling errors can be estimated and their measure used to help interpret the final data results. The size of such sampling errors depends largely on the number of interviews and the complexity of the sampling design.

The confidence interval for sample estimates of population proportions, assuming a simple random sample without replacement is calculated using the following formula:

The sample of telephone households in this study was drawn as a simple random sample within each region. However, the stratification by region and the disproportionate sampling of persons within household introduces a design effect that could possibly suggest that the sample reflect other than a simple random sample.

To test the belief that the resultant sample approximated one of a simple random sample, the sampling errors were calculated under a stratified design and were compared to the sampling errors for the same measures and sample sizes under the assumption of a simple random sample. These sample error comparisons were made for 23 of the key measures in the study. The net impact over the 23 measures was found to be about a 10% wider band of confidence around the estimates gathered in the study. That is, if an estimate from a study conducted under a pure a simple random sample method had an error range of + 1.5. The more precise error range due to the more complex sampling method would be about + 1.6%.

Since the data presented in this report are rounded to whole numbers, the incremental increase in the sampling error range generally did not translate into a wider band around the estimate. Given the relatively small average design effect, the table of expected sampling error ranges based on a simple random sample is a useful approximation of the precision of the sample estimates.

The following tables may be used in estimating the sampling error in any percentage in this report. They may be interpreted as indicating the approximate range (plus or minus the figure shown) within which the results of repeated sampling in the same time period could be expected to vary 95% of the time, assuming the same sampling procedures, the same interviewers, and the same questionnaire.

Table A shows how much allowance should be made for the sampling error around a single percentage estimate in the study.

* The chances are 95 in 100 that the sampling error is not larger than the figures shown.

The table would be used in the following manner: Let us say a reported percentage is 30 for a group that includes about 300 respondents. Then we go to the column labeled "Percentages near 30/70%" in the table and go down to the row labeled "300." The number at this point is 5.2, which means that the 27% obtained in the sample is subject to a sampling error or ±5 points. Another way of saying this is that 95 times out of 100 the true figure in the population would be somewhere between 25% and 35%.

In comparing survey results in two samples – for example, Whites and Blacks– the question arises as to how large a difference between them must exist before one can be reasonably sure that it reflects a real difference. In Table B, the number of points, which must be allowed for in such comparisons, is shown.

Here is an example of how the table would be used: Let us say that 53% of Non-Hispanic Whites who drove after drinking in the past year report a particular behavior, while in 47% of Non-Hispanic Black drinker-drivers report the same behavior, for a difference of six percentage points between them. Can we say with any assurance that the six-percentage point difference reflects a real difference between Whites and Blacks? The sample contains approximately 2277 Whites and 160 Blacks. We consult Table B, we look at the column headed 2,000 and the row labeled 200: we see the number 7.2 here. This means that the allowance for error should be 7.2 percentage points. Since the 6% difference found in the data is less than the 7.2% in the table, the six -point difference is inconclusive.

*The changes are 95 in 100 that the sampling error is not larger than the figures shown.

The table provided is for percentages near 50. For percentages higher or lower than 50%, the error to be allowed for is somewhat smaller than those shown in the table.