Association between screen time and depression: a population-based study using the 2011-2012 national health and nutrition examination sample

The goal is to replicate (and extend) some of the analysis results in the research question that has been studied by K.C. Madhava et al. (2017) in their paper Association between screen time and depression among US adults which used the 2011-2012 National Health and Nutrition Examination survey (NHANES) sample data where it concluded that there was an association between people who spent more than 4 hours per day on TV/computer with experiencing moderate or severe depression. Various survey sampling techniques are implemented throughout the analysis.

The most important (and difficult) part is to obtain the variance estimates for the interested quantities (population proportions).

Four-Stage Sampling Design

The study of this report is conducted using the sample from the 2011-2012 National Health and Nutrition Examination survey (NHANES), a large cross-sectional U.S. survey conducted by the Centers for Disease Control (CDC) and National Center for Health Statistics (NCHS). The sampling design consists of four stages which is quite complex and is summarized as follows:

• Stage 1: Selecting PSUs from a frame of all U.S. counties
• Stage 2: Selecting area segments from each PSU
• Stage 3: Selecting households from each area segment
• Stage 4: Selecting persons within each household

• Each individual has a sample weight assigned which represents the number of people in the study population in NHANES. The design uses unequal probability of selection (which means the sample weights are different for each sampled individual) in order to produce an unbiased estimate for the nation
• Oversampling is a common approach in large survey sampling designs which is used to adjust the distribution of subsamples of a sample dataset, so that the bias in estimations can be corrected. Here, NHANES uses oversampling to sample more individuals from certain subgroups for particular health interest so that the precision of estimates about the health information for these subgroups can be improved
• Finally, because the survey only covers a short period of time, longitudinal techniques are not involved

Estimation of population proportions using the sample data

• Analysis 1: Weighted analysis
• Analysis 2. Unweighted analysis
• For each analysis, descriptive statistics of the study population are computed include estimated population proportions, their standard errors, and Rao-Scott chi-square test with second order corrections (which corrects both mean and variance) to test the association between depression and the study variables.

It is noted that the estimates and their standard errors are significantly different between the weighted and unweighted analyses. This makes sense as the NHANES sampling design uses unequal probability selection where the sample weights are different to reflect the number of people in the population that they represent. In this case, using an unweighted analysis will be inappropriate as equal sample probability increases the bias of the estimates as well as decreases the precision of the estimates. This is particularly true for categories that involves a lot of oversampling.

Statistical analysis

• Univariate logistic regression - Unadjusted odds ratios and confidence intervals
• Multivariate logistic regression - Adjusted odds ratios and confidence intervals

Last updated on Nov 1, 2019