MCHB/EPI Miami Conference — December 7 - 9, 2005

Methods Can Help — Transcript

GENET BURKA: Hi, my name is Genet Burka. I'm an NCTP (inaudible) Fellow--used to be Louisiana Office of Public Health. My presentation topic is characteristic of respondent and nonrespondent in Louisiana PRAMS Survey, and is there any potential nonresponse bias. In any survey there are always individuals, you know, who cannot or will not cooperate or cannot be reached or cannot be contacted. Even if it is voluntary, nonresponse is going to be a problem for any kind of survey. We can get--even if we know that how important survey, if we survey this year hundred people, I'm sure nobody hundred percent is not going to get response rate, and nonresponse always will be a problem as long as we have a survey.

And it's one of the most serious--nonresponse in general (inaudible) data from sampled individual and this can be item nonresponse or unit nonresponse. And item responses are when respondent answers some questions but not all, which means when we have incomplete survey. And unit respondent is when we don't have anything that whole respond from our sample population. That's one of the most serious types of sampling error, which usually we cannot directly measure. It's bad because it reduce our effective, you know, our sample size and introduce bias and jeopardize the general reliability of our collected data. And always as nonresponse increase, the bias also increase, whatever survey we're doing. And that degree of bias depends on two factors: the percentage of sample not responding and also the extent to which nonrespondents differ systematically from the study population.

Saying that but there is no cut point where we can say that. I have heard so many people that they say that. This is a response rate good. That's nonresponse rate good. And you can have high response rate, 95 percent, let's assume. But if we have five percent of the population who are not responding and distinctively different from your respondent, then your data is not going represent the whole study population, still is going to be biased.

And to help reduce nonresponse bias, there are different methods adopted, and one of the methods which usually we use is adjustment. We weight the data to adjust for nonrespondent. The assumption behind this when you are adjusting nonrespondents, the respondents--our respondent would have provided similar answer to our nonrespondent. But that's not always true. If that assumption holds, which is good; we're reducing our bias, but is that always true? Is our respondent always represent the nonrespondent? And there are always possibility even if we adjust for the nonrespondent that there's going to be bias exist after adjustment.

And it is very difficult to quantify nonresponse bias. But there are different methods where they evaluate nonresponse bias. One of the simple and forward method is if you have the data from both respondent and nonrespondent, to compare the nonrespondent and the respondent group if they are systematically significant between each other. And it's a very straightforward analysis. It's not complicated. And for this analysis, we examine the characteristic of respondent and nonrespondent for Louisiana PRAMS 2000/2002. We use weighted sample. And we link the data with vital record; birth and death certificate we have more information what we have in vital record. And we measure the association of the respondent and nonrespondent with the response data with (inaudible) square test, and we try to use realistic regression to identify predictors of nonresponse bias.

Most of you know what PRAMS is. I don't think it's going to be but just to give you fast and general idea. PRAMS is just of state specific population births survey of women who deliver live-born infant and collect information for maternal behavior and experience that occur during and after pregnancy and specifically designed to generate state-specific data by weighting the sample just for sampling probability and nonresponse and (inaudible). And Louisiana begins the PRAMS in 1997 and the state draw around 150 to 200 sample and we use random stratify system sampling. And the sample size are usually present about 3.5 to five percent of the total (inaudible) of the state. And the states stratify the sample in four strata: by birth weight, by geographic--by region, (inaudible). The reason birth weight is used is chosen for the state is it's the highest grouping. We want more people--more data to get from that specific group and the states over-sample the new birth weight strata. And data collected to six months after birth by using mail (inaudible) and telephone (inaudible).

This is done with the response rates for--by race. As you see, the overall response rates for Louisiana 2000/2002 PRAMS Survey was around 72 percent. And the rate-specific response rates for blacks and whites were about 78 for black and 65 for--I mean 78 for white and 65 percent for black. But we don't see--there was no significant change in response rate by year except, you know, but we can see the significant difference response rate between black and white.

These figures show the cumulative response rate from the mailed and telephone-phase of the survey. Most white mothers you see that--I think the color is quite clear. And most white mothers respond in the first mail. If you see 48 percent almost white women are responding at the first mail comparing to black, which is only 25 percent. The complete mail phase generated 61 percent and 75 percent of the response for black and white respectively. It's quite a difference.

And further the telephone-phase survey generated around 18 percent and 30 percent of the response rate, which right in the cumulative response rate for blacks--for whites by 28 percent and for blacks by 83 percent, which is 65 percent. That's a big difference. And the question is who are the other 27 percent of nonrespondent. Are they different from those who do respond? Is that there a quotation for nonresponse (inaudible)? That's the question we're trying to answer by doing simple analysis, straight forward.

If you see here, when you compare the respondent and nonrespondent by the two column by different characteristics, the rates--chosen rates are the significant one. Women who did not respond in the survey were blacks around 55 percent and had less than high school education, less than 12 years--younger than 20 years and were not married and used tobacco. Also they are urban resident and multiparas women. They also enter prenatal care first trimester and had preterm infants and they have also low-birth infants. They are all high risk group women are not responding, if you see by social demographic characteristic.

By obstetric history, if you see, I already say that the multipara women--the majority were not responding and those who did not have prenatal care or who enter prenatal care after first trimester are the most likely not to respond.

This is by (inaudible). The same thing which I say the gestational age, the preterms, the birth weight, the low-birth weight moms, and the abnormal conditions, those who had most probably, you know, with babies whose abnormal condition are more likely not to respond. And since we already linked our data with the infant death file and we found out that women whose infants died at neonatal period are less likely to respond for the PRAMS survey.

And we tried to compare also the early and late respondents, those who responded in mail and those who responded by telephone. The telephone respondents--if you compare by telephone, the phone respondents are the same thing. They're black, less than 12 years education, less than 20 years old, and single, and those who use tobacco.

If you compare--I did--I put one column of nonrespondent, the one which you saw on the previous slide here just to compare how they are similar, the nonrespondent and the late respondent. If you see almost the same character of people who responded later for an interview are the nonrespondent. They have more similarities than with the respondent.

This is the same thing by obstetric history. If you compare the nonrespondent and the phone respondent, the multiparas are the more likely not--the least likely to respond. The previous women who have history of previous preterm births or small gestational age babies are also less likely to respond. Also the prenatal entry--it's the same thing. And you see the similarity of the nonrespondent also the phone respondent population.

This is by (inaudible) found the same thing. If you compare the mail and phone respondent is the same thing: the preterm birth, the low birth-weight moms, those whose abnormal conditions are less likely to respond to PRAMS questionnaire.

We did our analysis by race also, but I didn't have the table just (inaudible). But most likely when we compare the black and white, the average age significantly lower for whites which comparing 24 versus for white 26. And nonresponse rates are higher among blacks with less than 12 years education; comparing 75 percent of blacks versus 62 percent of whites are nonrespondent. Teenagers 24--almost 25 percent versus 16 percent. And those who use tobacco triple in whites. If you say that's the most--you know, the whites is predominantly least likely to respond are those who use tobacco. Almost triple in whites: 22 percent of whites nonrespondent are smokers comparing to only 8 percent of blacks. And single black women very significantly different. Around almost 80 percent of black women, and are single, will not respond to the survey comparing to only 76 percent white. And the blacks who did not have prenatal care--all enter prenatal care after the first trimester are also the least likely to respond. And black women who are preterm and also urban residents than rural are more likely not to respond. If you see the percentage, 69 percent of urban residents are respondent, nonrespondent versus 33 percent.

We did race-specific (inaudible) analysis just to identify characteristic which can predict a nonresponse. As you see, for all race, maternal education, marital status, tobacco use, parity, abnormal condition, and urban rural--place of residence, or whether they have, you know, neonatal days in the history or not, they are more likely not to respond. When we did it by race specific analysis, what we found out is marital status is no more significant for blacks comparing to whites. And for blacks, those who are neonatal, in the history when residents are significantly, which is not significant for whites.

In conclusion, what we can say from this study is the study characteristics of nonrespondent to PRAMS are similar to those mothers who are at higher risk for outcome. And we found the difference between nonrespondent and respondent current weight adjustment for nonresponse does not account for some of the body of the thing. We have to be caution when we are making inference based on PRAMS analysis. And what we can do is to capture the nonresponse group is we might need to develop a new response instrument technique. As they say, it's always better to prevent nonresponse than cure it, assuming that we can weight it. We just don't have to take whatever response rate we have. We have to try. And by minimizing--especially for Louisiana when you see most of the nonrespondent are high-risk group. Louisiana almost 50/50 percent is black and white population. And those are the more--and as you see on the first graph, the majority of blacks are responding by telephone, which is--if we maybe increase the phone call for those people who are not responding, maybe we might gain more respondent among those group which nonrespondent. And also to minimize refusing by training the interviewer not to give, you know, just to be sensitive--an incentive. We tried, I guess, in Louisiana to do it using incentive. We don't know how much it's going to work. We still are working on it. But that's another thing we have to think. That's it. If you have any questions--