Transition into Adulthood Supplement - Revision history

Talmujah: /* TAS Sampling Weights */

2021-04-03T16:40:30Z

TAS Sampling Weights

← Older revision		Revision as of 12:40, 3 April 2021
Line 384:		Line 384:

	The calibration adjustment was performed using a raking ratio (or iterative proportional fitting) method (Deming and Stephan, 1940<ref>Deming, W. E., & Stephan, F. F. (1940). <u>On a least squares adjustment of a sampled frequency table when the expected marginal totals are known.</u> The Annals of Mathematical Statistics, 11(4), 427-444.</ref> through a SAS macro developed by Izrael, Battaglia and Frankel, 2009.<ref> Izrael, D., Battaglia, M. P., & Frankel, M. R. (2009). <u>Extreme survey weight adjustment as a component of sample balancing (aka raking)</u>. In Proceedings from the Thirty-Fourth Annual SAS Users Group International Conference.</ref> An advantage of this SAS implementation is that apart from running the raking procedure to adjust the weights to enforce the weighted sample distribution to match the population margins in the selected calibration dimensions, it also simultaneously trim the weights according to trimming parameters, in order to mitigate the increase of the sampling variance due to the weight variability. The final cross-sectional weight for TAS-2017 respondents was derived from the output weights of this calibration adjustment with trimming. Table 5 reports key summary statistics for the final TAS-2017 cross-sectional weight.		The calibration adjustment was performed using a raking ratio (or iterative proportional fitting) method (Deming and Stephan, 1940<ref>Deming, W. E., & Stephan, F. F. (1940). <u>On a least squares adjustment of a sampled frequency table when the expected marginal totals are known.</u> The Annals of Mathematical Statistics, 11(4), 427-444.</ref> through a SAS macro developed by Izrael, Battaglia and Frankel, 2009.<ref> Izrael, D., Battaglia, M. P., & Frankel, M. R. (2009). <u>Extreme survey weight adjustment as a component of sample balancing (aka raking)</u>. In Proceedings from the Thirty-Fourth Annual SAS Users Group International Conference.</ref> An advantage of this SAS implementation is that apart from running the raking procedure to adjust the weights to enforce the weighted sample distribution to match the population margins in the selected calibration dimensions, it also simultaneously trim the weights according to trimming parameters, in order to mitigate the increase of the sampling variance due to the weight variability. The final cross-sectional weight for TAS-2017 respondents was derived from the output weights of this calibration adjustment with trimming. Table 5 reports key summary statistics for the final TAS-2017 cross-sectional weight.


	{\|class="wikitable"		{\|class="wikitable"
Line 475:		Line 474:

	As the final step in weight development, the newly constructed TAS-2017 longitudinal weight was trimmed to reduce the influence of extreme weight values on the variances of sample estimates of population statistics. The cases with the weight values in the top one percent and in the bottom one percent of the weight distribution were assigned values corresponding to the 99th and 1st percentiles of the weight distribution, respectively. Table 8 reports key summary statistics for the final TAS-2017 longitudinal weight for original CDS-I participants.		As the final step in weight development, the newly constructed TAS-2017 longitudinal weight was trimmed to reduce the influence of extreme weight values on the variances of sample estimates of population statistics. The cases with the weight values in the top one percent and in the bottom one percent of the weight distribution were assigned values corresponding to the 99th and 1st percentiles of the weight distribution, respectively. Table 8 reports key summary statistics for the final TAS-2017 longitudinal weight for original CDS-I participants.


	{\|class="wikitable"		{\|class="wikitable"

Talmujah: /* TAS Sampling Weights */

2021-04-03T16:39:44Z

TAS Sampling Weights

← Older revision		Revision as of 12:39, 3 April 2021
Line 389:		Line 389:
	!colspan="2"\|Table 5: Summary Statistics for the TAS-2017 Sample Weights		!colspan="2"\|Table 5: Summary Statistics for the TAS-2017 Sample Weights
	\|-		\|-
	\|colspan\| <center>Description</center>		!colspan\| <center>Description</center>
	\|<center>Value</center>		!<center>Value</center>
	\|-		\|-
	\|<center>N</center>		\|<center>N</center>

Talmujah: /* TAS Sampling Weights */

2021-04-03T16:39:02Z

TAS Sampling Weights

← Older revision		Revision as of 12:39, 3 April 2021
Line 480:		Line 480:
	!colspan="2"\|Table 8: Summary Statistics for the TAS-2017 Sample Weights		!colspan="2"\|Table 8: Summary Statistics for the TAS-2017 Sample Weights
	\|-		\|-
	\|colspan\| <center>Description</center>		!colspan\| <center>Description</center>
	\|<center>Value</center>		!<center>Value</center>
	\|-		\|-
	\|<center>N</center>		\|<center>N</center>
Line 499:		Line 499:
	\|-		\|-
	\|}		\|}

			To examine the properties of the TAS-2017 longitudinal weight, we compared weighted estimates for selected demographic, geographic, and socio-economic variables in the CDS-I data computed in two ways. The first set of estimates is based on the full CDS-I sub-sample that remained eligible for TAS-2017. The CDS-I weight was used to create these estimates for the full TAS-2017 sample. The second set of estimates is based only on the TAS-2017 respondent cases and employs the TAS-2017 longitudinal weight that adjusts for longitudinal nonresponse among the eligible cases in the TAS-2017 wave of data collection. The results are provided in Table 9 and show that the distributions of the selected characteristics are similar in the appropriately-weighted TAS-2017 eligible sub-sample of CDS-I original respondents and in the TAS-2017 interview sample, suggesting that the attrition adjustment for the TAS-2017 weight compensates for potential attrition bias for variables included in the analysis. It is important to note, however, that this comparison does not necessarily rule out the possibility of selection bias associated with other characteristics of the respondents.

	==References==		==References==

Talmujah: /* TAS Sampling Weights */

2021-04-03T16:37:17Z

TAS Sampling Weights

@@ Line 420: / Line 420: @@
 Among the projected eligible sample, which excludes deceased (n=7) and non-sample individuals (n=71), a total of 1,410 interviews were completed, resulting in a cumulative unweighted response rate of 64.4% (i.e., 1,410/(1,410+780) =0.644)<ref>The cumulative response rate is defined as a ratio of the number of cases that were successfully interviewed in TAS-2017 to the number of cases that were projected to be eligible for TAS-2017 in 1997, excluding deceased and non-sample individuals.</ref>. See Chapter 5 for a description of the TAS-2017 wave-specific response rate (87%) and data collection procedures and outcomes.
 {|class="wikitable"
@@ Line 469: / Line 468: @@
 <u>'''''TAS-2017 Individual Longitudinal Weight for Original CDS-I participants'''''</u>
-The TAS-2017 individual longitudinal weight for original CDS-I participants was designed to account for the differential attrition between the baseline CDS-I (in 1997) and TAS-2017, i.e. t=2017 and t-1=1997.  Thus, the TAS-2017 longitudinal weight is a product of the CDS-I weight, i.e., the individual-level primary caregiver/child weight (stored in the weight variable named ‘CH97PRWT’), and the attrition adjustment factor<ref>For a description of the 1997 CDS-I weights, see [https://psidonline.isr.umich.edu/CDS/weightsdoc.pdf].</u>.
+The TAS-2017 individual longitudinal weight for original CDS-I participants was designed to account for the differential attrition between the baseline CDS-I (in 1997) and TAS-2017, i.e. t=2017 and t-1=1997.  Thus, the TAS-2017 longitudinal weight is a product of the CDS-I weight, i.e., the individual-level primary caregiver/child weight (stored in the weight variable named ‘CH97PRWT’), and the attrition adjustment factor<ref>For a description of the 1997 CDS-I weights, see [https://psidonline.isr.umich.edu/CDS/weightsdoc.pdf].</ref>.
 To obtain the attrition adjustment classes, the probability that a sample person was nonresponse in TAS-2017 was estimated using a logistic regression model.  The dependent variable for this nonresponse propensity model is Y=1 if the eligible sample person was a nonrespondent in 2017 and Y=0 if they were a respondent.  The estimated parameters and standard errors for the logistic model of nonresponse attrition are reported in Table 7.  For example, the results indicate that the odds of attrition between 1997 and 2017 were significantly higher among males as compared to females, white respondents as compared to non-white respondents, those in Northeast and South regions as compared to the West region, and significantly lower among older respondents, SRC sample as compared to non-SRC sample, and those whose reference person was male as compared to female reference persons, holding all else equal.
 ==References==

Talmujah: /* TAS Sampling Weights */

2021-04-03T16:33:05Z

TAS Sampling Weights

@@ Line 419: / Line 419: @@
 Most of the TAS-2017 respondents were all originally selected for CDS-I in 1997 when they were 0-8 years of age. CDS-I selected sample children between the ages of 0-12 in 1997, and those who were aged 9-12 in 1997 aged out of the TAS age eligible range by 2017. Of the 3,563 children who participated in the original 1997 CDS-I interview, 2,268 were projected to be eligible for participation in TAS-2017, based on their participation in the CDS, TAS, and Core PSID studies. For these 2,268 cases, Table 6 summarizes the final contact and interview dispositions in TAS-2017.
-Among the projected eligible sample, which excludes deceased (n=7) and non-sample individuals (n=71), a total of 1,410 interviews were completed, resulting in a cumulative unweighted response rate of 64.4% (i.e., 1,410/(1,410+780) =0.644).3 See Chapter 5 for a description of the TAS-2017 wave-specific response rate (87%) and data collection procedures and outcomes.
+Among the projected eligible sample, which excludes deceased (n=7) and non-sample individuals (n=71), a total of 1,410 interviews were completed, resulting in a cumulative unweighted response rate of 64.4% (i.e., 1,410/(1,410+780) =0.644)<ref>The cumulative response rate is defined as a ratio of the number of cases that were successfully interviewed in TAS-2017 to the number of cases that were projected to be eligible for TAS-2017 in 1997, excluding deceased and non-sample individuals.</ref>. See Chapter 5 for a description of the TAS-2017 wave-specific response rate (87%) and data collection procedures and outcomes.
@@ Line 455: / Line 455: @@
 |<center>Deceased</center>
 |}
 ==References==

Talmujah: /* TAS Sampling Weights */

2021-04-03T16:22:51Z

TAS Sampling Weights

← Older revision		Revision as of 12:22, 3 April 2021
Line 420:		Line 420:

	Among the projected eligible sample, which excludes deceased (n=7) and non-sample individuals (n=71), a total of 1,410 interviews were completed, resulting in a cumulative unweighted response rate of 64.4% (i.e., 1,410/(1,410+780) =0.644).3 See Chapter 5 for a description of the TAS-2017 wave-specific response rate (87%) and data collection procedures and outcomes.		Among the projected eligible sample, which excludes deceased (n=7) and non-sample individuals (n=71), a total of 1,410 interviews were completed, resulting in a cumulative unweighted response rate of 64.4% (i.e., 1,410/(1,410+780) =0.644).3 See Chapter 5 for a description of the TAS-2017 wave-specific response rate (87%) and data collection procedures and outcomes.


			{\|class="wikitable"
			!colspan="2"\|Table 6: TAS-2017 Sample Disposition
			\|-
			!<center>Sample Count</center>
			!<center>Description</center>
			\|-
			\|<center>2,268</center>
			\|<center>Total projected eligible</center>
			\|-
			\|<center>1,410 </center>
			\|<center>Completed TAS-2017 interview</center>
			\|-
			\|<center>780</center>
			\|<center>Non-response</center>
			\|-
			\|<center>577</center>
			\|<center>Non-response before the 2017 interview</center>
			\|-
			\|<center>172</center>
			\|<center>Non-response in 2017</center>
			\|-
			\|<center>31</center>
			\|<center>Difficult to access/outside of the U.S.</center>
			\|-
			\|<center>78</center>
			\|<center>No longer eligible</center>
			\|-
			\|<center>71</center>
			\|<center>Not a sample person</center>
			\|-
			\|<center>7</center>
			\|<center>Deceased</center>
			\|}

	==References==		==References==

Talmujah: /* TAS Sampling Weights */

2021-04-03T16:05:37Z

TAS Sampling Weights

@@ Line 361: / Line 361: @@
 ''
-<u>'''Calibration'''''</u>
+<u>'''''Calibration'''''</u>
 As the final step in weight development, the nonresponse-adjusted weights are used as input in a calibration adjustment , in which the TAS-2017 sample weighted distribution are matched to population totals estimated from the ACS 2017 1-year PUMS data for individuals between 18 and 28 years by December 31st, 2017 on selected auxiliary variables. Similar to the nonresponse adjustment, if such variables are predictive of the survey outcomes, this calibration adjustment can reduce non-sampling biases (such as coverage and nonresponse) and improve the precision of the survey estimates. For this reason, we tested the main and interaction effects on the seven selected survey outcomes of the following variables:
@@ Line 384: / Line 384: @@
 The calibration adjustment was performed using a raking ratio (or iterative proportional fitting) method (Deming and Stephan, 1940<ref>Deming, W. E., & Stephan, F. F. (1940). <u>On a least squares adjustment of a sampled frequency table when the expected marginal totals are known.</u> The Annals of Mathematical Statistics, 11(4), 427-444.</ref> through a SAS macro developed by Izrael, Battaglia and Frankel, 2009.<ref> Izrael, D., Battaglia, M. P., & Frankel, M. R. (2009). <u>Extreme survey weight adjustment as a component of sample balancing (aka raking)</u>. In Proceedings from the Thirty-Fourth Annual SAS Users Group International Conference.</ref> An advantage of this SAS implementation is that apart from running the raking procedure to adjust the weights to enforce the weighted sample distribution to match the population margins in the selected calibration dimensions, it also simultaneously trim the weights according to trimming parameters, in order to mitigate the increase of the sampling variance due to the weight variability. The final cross-sectional weight for TAS-2017 respondents was derived from the output weights of this calibration adjustment with trimming. Table 5 reports key summary statistics for the final TAS-2017 cross-sectional weight.
 ==References==

Talmujah: /* TAS Sampling Weights */

2021-04-03T15:48:14Z

TAS Sampling Weights

← Older revision		Revision as of 11:48, 3 April 2021
Line 319:		Line 319:

	===TAS Sampling Weights===		===TAS Sampling Weights===

			This chapter provides an overview of the sample weights for TAS-2017. There will be three weights for TAS-2017, a cross-sectional weight and two longitudinal weights. The cross-sectional weight accounts for unequal selection probabilities due to the PSID sample design, while the longitudinal weights are panel weights for any TAS Respondent who was interviewed in (1) the Original CDS cohort or (2) the Original or Ongoing CDS. The second longitudinal weight will be included in Release 2 of the TAS-2017 data.

			<u>'''TAS-2017 Cross-Sectional Weight'''</u>

			The TAS-2017 data are provided with a cross-sectional weight for every survey respondent to be used in data analysis and obtain unbiased estimates for population parameters. We describe in this chapter the construction of the TAS-2017 cross-sectional weight.

			<u>'''''Weighting Methodology'''''</u>

			The TAS-2017 cross-sectional weight was designed to account for the unequal selection probabilities, due to the original PSID sample design, and for differential eligibility and nonresponse. These weights were also calibrated to selected demographic variables of the target population to further mitigate any coverage and nonresponse error, and to improve the precision of survey estimates. We describe below the three main components of the TAS-2017 cross-sectional weights.

			<u>'''''Base Weights'''''</u>

			In order to account for differential selection probability and nonresponse during the PSID recruitment, we used as base weights for TAS-2017 the 2017 PSID Individual Longitudinal Weights. However, due to various eligibility criteria (including, for example, having to respond to both PSID 2013 and 2015), not every sample person at PSID 2017 that belongs to the TAS target population (individuals between 18 and 28 years old by December 31st, 2017) was eligible for TAS-2017. In order to account for this difference, the base weights of such ineligible cases were re-distributed across the eligible sampled individuals, so that the sum of the weights reflects, on average, the size of the target population.

			<u>'''''Nonresponse adjustment'''''</u>

			Unit nonresponse poses a threat to the quality of survey estimates as respondents and nonrespondents might differ in terms of the study outcomes, which can ultimately cause nonresponse bias in such estimates. If nonresponse follows a missing at random (MAR) mechanism (Little and Rubin, 20024), this nonresponse bias can be attenuated through certain statistical adjustments. To that end, a nonresponse weighting adjustment was performed over the TAS-2017 data using a response propensity procedure. In this approach, the weights are inversely proportional to estimate of the probability of response to the survey. These estimated probabilities of responding the survey, also referred to response propensities, are computed using a logistic regression model of the survey response indicator over a set of covariates available for both respondents and nonrespondents. In order to reduce nonresponse bias while not increasing sampling variance of the survey estimates, the covariates used in this adjustment should be correlated with both the survey response and the study outcomes (Little and Vartivarian, 2003<ref>Little, R.J.A., and Vartivarian, S. (2003). <u>On weighting the rates in nonresponse weights.</u> Statistics in Medicine, 22, 1589-1599.</ref> ). For this reason, the following TAS-2017 outcomes were selected to assist in this adjustment:

			*Body Mass Index (BMI),
			*Weeks of employment in previous year (WKSEMPPY),
			*Weeks of employment in the past two years (WKSEMPPPY),
			*Completed education of mother (MOCED),
			*Completed education of father (FACED),
			*Marital/cohabitation status (TAMS), and
			*School enrollment status (EDSTAT)

			Ideally, we would like to use these survey outcomes as covariates in the response propensity modelling. However, we only observe them for the survey respondents. Instead, we computed predictions for both nonrespondents and respondents for each of these survey outcomes using regression models over covariates available for every sampled individual, including:

			*'''Census variables''': block group or tract level variables from the Census Planning Database that can be appended to the sampling frame, such as percentage and medians of population, households and housing units by socioeconomic characteristics.

			*'''Paradata''': variables generated as a byproduct of the data collection itself, such as number of call attempts by survey mode (face-to-face, telephone, e-mail), indicator of refusal in previous waves.


			These sets of covariates are typically available for every sampled element. There are a few cases in the TAS-2017 sample with missing values in some of these variables though. Therefore, as a first step in this nonresponse adjustment, we used regression-based single imputation to fill in the missing values in those variables. For the Census variables, the data were aggregated and imputed at the tract-level, such that elements within the same block group or tract received the same imputed values. All other variables were imputed at the element-level.

			Next, we fitted a regression model for each of the seven selected outcome variables over the respondents’ data using all the covariates mentioned above. Given the large number of covariates in the models, we used the Lasso (Least absolute shrinkage and selection operator<ref>Friedman, J., Hastie, T., & Tibshirani, R. (2001). <u>The elements of statistical learning</u>. New York: Springer series in statistics.</ref>) for both variable selection and estimation in each of these models. Using these fitted regression models and the observed/imputed covariates for every sampled individual, we predicted the survey outcomes for both respondents and nonrespondents. These predictions can be seen as proxy summaries of the covariates correlated to the selected survey variables, thus satisfying one of the conditions for a successful nonresponse adjustment (Little and Vartivarian, 20036).
			The probability that a sample person was a respondent in TAS-2017 was estimated using a logistic regression model. The dependent variable for this response propensity model is Y=1 if the eligible sample person was a respondent in 2017 and Y=0 otherwise. The independent variables were the predicted values of the selected survey outcomes for numeric variables (BMI, WKSEMPPY, WKSEMPPPY, MOCED and FACED) and predicted probabilities for each category for the categorical variables (TAMS and EDSTAT). The estimated parameters and standard errors for this logistic model are reported in Table 1. For example, the results indicate that the odds of response were significant higher for those with higher chance of were never been married, not cohabiting.

			To reduce variation in response propensity weights and lower the reliance on correct model specification of the logistic regression, 10 nonresponse adjustment classes were created based on deciles of the predicted response probability (propensity score stratification; Little and Rubin, 2002<ref>Deville, J. C., & Särndal, C. E. (1992). <u>Calibration estimators in survey sampling</u>. Journal of the American Statistical Association, 87(418), 376-382.)</ref> estimated using the logistic model in Table 4. The inverse of the mean response probability for TAS-2017 eligible sample cases in each decile was assigned as the nonresponse adjustment factor for that weighting class. The final nonresponse-adjusted weight for TAS-2017 respondents was computed as the product of their base weights (2017 PSID Individual Longitudinal Weights adjusted by eligibility) and their weighting class nonresponse adjustment factor.
			''

			<u>'''Calibration'''''</u>

			As the final step in weight development, the nonresponse-adjusted weights are used as input in a calibration adjustment , in which the TAS-2017 sample weighted distribution are matched to population totals estimated from the ACS 2017 1-year PUMS data for individuals between 18 and 28 years by December 31st, 2017 on selected auxiliary variables. Similar to the nonresponse adjustment, if such variables are predictive of the survey outcomes, this calibration adjustment can reduce non-sampling biases (such as coverage and nonresponse) and improve the precision of the survey estimates. For this reason, we tested the main and interaction effects on the seven selected survey outcomes of the following variables:

			*Sex (Male, Female)
			*Race/Ethnicity (Hispanic, Non-Hispanic While Alone, Non-Hispanic Black or African American, Non-Hispanic Asian alone/AIAN/NHPI/Some other race alone, Non-Hispanic Two or more races)
			*Family type and employment status (Married-couple family: Husband and wife in Labor Force; Married-couple family: Husband in labor force, wife not in Labor Force; Married-couple family: Husband not in Labor Force, wife in Labor Force; Married-couple family: Neither husband nor wife in Labor Force; Male householder, no wife present, in Labor Force; Male householder, no wife present, not in Labor Force; Female householder, no husband present, in Labor Force; Female householder, no husband present, not in Labor Force;)
			*Region (Northeast, Midwest, South, West)
			*Age (17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28)
			*Household size (1, 2, 3, 4 or more)

			Due to the high dimensionality that would have been created if we accounted for all the interaction terms, we decided to test only the two-way interactions. We then kept all the main effects (regardless of their level of explanation on the survey outcomes) and only the two-way interactions that were significant, at a 5% level, to at least four of the survey outcomes, which on this case were:

			*Sex by Race/Ethnicity
			*Sex by Family type and employment status
			*Sex by Region

			In order to avoid undue increase in the variability of the weights, the following calibration cells with small sample sizes were collapsed for the calibration procedure:

			*Gender by Married-couple family: Neither husband nor wife in Labor Force (Family type and employment status)
			*Gender by Male householder, no wife present, not in Labor Force (Family type and employment status)

			The calibration adjustment was performed using a raking ratio (or iterative proportional fitting) method (Deming and Stephan, 1940<ref>Deming, W. E., & Stephan, F. F. (1940). <u>On a least squares adjustment of a sampled frequency table when the expected marginal totals are known.</u> The Annals of Mathematical Statistics, 11(4), 427-444.</ref> through a SAS macro developed by Izrael, Battaglia and Frankel, 2009.<ref> Izrael, D., Battaglia, M. P., & Frankel, M. R. (2009). <u>Extreme survey weight adjustment as a component of sample balancing (aka raking)</u>. In Proceedings from the Thirty-Fourth Annual SAS Users Group International Conference.</ref> An advantage of this SAS implementation is that apart from running the raking procedure to adjust the weights to enforce the weighted sample distribution to match the population margins in the selected calibration dimensions, it also simultaneously trim the weights according to trimming parameters, in order to mitigate the increase of the sampling variance due to the weight variability. The final cross-sectional weight for TAS-2017 respondents was derived from the output weights of this calibration adjustment with trimming. Table 5 reports key summary statistics for the final TAS-2017 cross-sectional weight.

	==References==		==References==

Talmujah: /* TAS Sample by Survey Year, Age in Years, and CDS Affiliation */

2021-04-03T15:29:19Z

TAS Sample by Survey Year, Age in Years, and CDS Affiliation

← Older revision		Revision as of 11:29, 3 April 2021
Line 317:		Line 317:
	\|-		\|-
	\|}		\|}

			===TAS Sampling Weights===

	==References==		==References==

Pfomby at 13:10, 16 February 2021

2021-02-16T13:10:40Z

← Older revision		Revision as of 09:10, 16 February 2021
Line 122:		Line 122:
	<u>''Age eligibility''</u>. Age eligibility at each wave is determined by birth year. Birth cohorts for each wave appear below. Participants are at least 18 years old and no older than 28 years in in the survey year.		<u>''Age eligibility''</u>. Age eligibility at each wave is determined by birth year. Birth cohorts for each wave appear below. Participants are at least 18 years old and no older than 28 years in in the survey year.

	TAS 2005: 1984-1987 TAS 2007: 1984-1989		TAS 2005: 1984-1987

			TAS 2007: 1984-1989

	TAS 2009: 1984-1991		TAS 2009: 1984-1991