Sampling is the most efficient way to conduct a survey. My favorite is stratified random sampling which is the most cost-effective and efficient sampling method, and also the most complex. Sampling and post-stratification weighting is one of the best ways to treat non-response bias. Below are a few examples:
Sample of public school teachers in 2, 5 and 8 grades
In order to construct a sample of grade 2, 5 and 8 teachers in the U.S, A two stage sampling process was designed. First 175 public schools were selected then two teachers per school were selected from each grade. Final sample size was 1050 teachers (350 x 3).
A sampling frame of elementary and secondary public schools was downloaded from IPED’s Common Core Data (CCD) was consisting of about 100,000 schools. This sampling frame was cleaned and divided into 3 sections, grade 2, 5, and 8. From each section 175 schools were selected stratified by size (number of teachers), region, and NCES code for location. Strict sample replacement was ensued where sample replacements were selected from the same strata.
In the next stage these 525 schools were contacted and names of two grades 2, 5 and 8 teachers taken chosen according to a random letter. Then these teachers were administered the survey.
Sampling error was ±3% for a sample of 1050 teachers at 95% confident level.
Sample of TV and daily newspaper journalists
The Bacon’s Media Source Premium Research Module database was used to construct four lists newspaper organizations, television organizations, newspaper journalists, and television journalists.
Independent two-stage sampling processes were conducted for television and newspaper journalists. In the first stage a list of 140 newspaper organizations was randomly selected out of 1562 while 263 television organizations were selected out of 916. Television and newspaper organizations were divided into five stratums based on circulation and number of working journalists.
In the next stage two lists of journalists from selected organizations were constructed. Size of daily newspaper sampling frame was about 10000 journalists while the television sampling frame was about 8000 journalists.
The sample was designed to obtain 300 completed interviews among journalists, with 180 from newspapers and 120 from television. A strict sample replacement was ensued where sample replacements were taken from the same organization.
Sampling error is ±5% adjusted for a sample size of 300 at 95% confidence level.
Sample of full-time sworn municipal police officers
The sample requirement was 700 police officers from the top 10 cities in the U.S. and 2500 police officers from rest of the nation.
We identified top 10 cities in the United States by using Law Enforcement Management Administrative Statistics (LEMAS) compiled by Bureau of Justice Statistics (BJS). Every four years BJS conducts comprehensive census of all State and local law enforcement agencies in the United States. The data is stored in the National Archive of Criminal Justice Date (NACJD) at Inter-University Consortium for Political and Social Research (ICPSR) at the University of Michigan. Registration with ICPSR enabled us to download a .wk1 data file containing year 2000 information about all State and Local law enforcement agencies in the U.S. The variables included State, County, City, Name of the Agency, Type of Agency, Population, Address, Telephone number, Web site, contact e-mail, number of full-time, sworn officers. The .wk1 data file was converted to SPSS for practicality reasons. Original file contained 17784 police departments in the United States. Out of that 12409 municipal police departments were selected as our cluster. It included the top 10 cities which are New York (NY), Los Angeles (CA), Chicago (IL), Houston (TX), Philadelphia (PA), Phoenix (AZ), San Diego (CA), Dallas (TX), San Antonio (TX) and Las Vegas (NV).
Two stage sampling process was designed. A sampling frame of all the State and Local law enforcement agencies was constructed by using BJS’s NACJ data. The sampling frame was about 12,500 agencies consisting of about 450,000 police officers.
First a sample of law enforcement agencies 100 from top ten cities and 500 from the rest of the country was selected. The sample was stratified by number of full-time sworn police officers. Questionnaires were mailed to these agencies based on the assumed response rate. The officer in charge of the morning briefing was given instructions to hand over the questionnaire to randomly selected police officers.
Sampling error is ±2% adjusted for a sample size of 3200 at 95% confidence level.
Random Digit Dialing (RDD) phone sampling
RDD sampling was utilized almost all of the surveys we conducted at CSRA. The proliferation of cell-only populations, decline on land phones, and increasingly on-line and text surveys had introduced challenges to the RDD surveys. Still, it is highly utilized. We bought RDD samples and cleaned them up using a private vendor. The Computer Aided Telephone Interviewing (CATI) software we used, is highly capable of launching RDD samples.
Stratified Random Sample (SRR) of classrooms
I designed a stratified sample for the Strategic Initiatives survey at the University of New Haven. The survey received a very good response. This survey methodology helped to reduce non-response bias and conduct the survey in a shorter period of time, still, producing a highly scientific data set. The survey was also multi-mode paper and web. The web survey was designed by the university IT department using Lime Survey.