Background/Context: Employment and training programs in the United States, which often provide postsecondary education to non-traditional students, aim to improve earnings for program participants. Effectively estimating the impact of a job training program on earnings requires thoughtful research design work well in advance of a multiyear study. The first goal of this paper is to be a resource for thoughtful design by supporting design-stage sample size calculations, also known as "statistical power analyses." The paper uses a large dataset of administrative earnings records for nearly 70,000 participants from evaluations of employment and training programs to report empirical estimates of the standard deviation of earnings, share of earnings variance explained by covariates (R2), and the intra-class correlation (ICC). The second goal of this paper is to explore recent evidence of the impact of job training programs on earnings. Recent evaluations of Project Quest, Year Up, and Per Scholas demonstrate strong evidence of sizable and sustained impacts on earnings (Fein & Dastrup, 2022; Roder & Elliott, 2021; Kanengiser & Schaberg, 2022); but many other seemingly strong programs have failed to produce evidence of sustained earnings gains. These include eight programs (outside Year Up) that participated in the Pathways for Advancing Careers and Education (PACE) study; the Health Profession Opportunity Grants (HPOG) Program, of which there were two rounds of funding (HPOG 1.0 and HPOG 2.0); and WorkAdvance programs at St. Nicks Alliance, Madison Strategies Group, and Towards Employment (Juras et al., 2022; Kanengiser & Schaberg, 2022; Klerman et al., 2022; Peck, Litwok, & Walton, 2022). One plausible explanation for this pattern of findings is that the true impact of these programs is too small to be reliably detectable in these studies given their sample sizes. To explore this hypothesis, we calculate the minimum sample size needed to detect an earnings impact of reasonable magnitude ($500) and compare that sample size to those from recently published studies in this literature. Setting: For over a decade, the Administration for Children & Families (ACF) has funded, administered, and evaluated nearly 100 employment and training programs through the Pathways for Advancing Careers and Education (PACE) project and two rounds of the Health Professions Opportunity Grants Programs (known as HPOG 1.0 and HPOG 2.0). Across all three projects, the programs provide education, occupational training, and support and employment services to TANF recipients and other adults with low incomes with an emphasis on career pathways in the healthcare sector. Data: Through their evaluation efforts, the Office of Planning, Research, and Evaluation at ACF has compiled administrative quarterly earnings records from the National Directory of New Hires for a total of nearly 70,000 study participants from evaluations of 89 job training programs. Specific details vary by study, but records are available from as early as 2010 through 2022. The analysis focuses on 24 quarters of study follow-up available for the entire PACE and HPOG 1.0 samples and on 5 quarters of follow-up available for the entire HPOG 2.0 sample. Analysis: To support future studies that could have many potential designs, our analysis reports empirical estimates for inputs that commonly appear in expressions for variance of the impact estimate when the outcome measure is earnings: the size of total variation, the ICC, and the proportion of variability explained by covariates. We estimate individual-level variability, program-level variability, and the ICC using two separate multi-level models in which individuals are clustered within programs. The first does not include any baseline covariates, such that our estimates of variance are unconditional. The second model includes fixed coefficients for baseline covariates at both the individual and program levels. We estimate the proportion of variation explained by covariates by fitting the multi-level model both with and without covariates, saving the estimates of variance, and dividing the change in estimated variance by the unconditional variance. We also explore the degree to which our estimates of inputs vary with baseline characteristics of study participants, including race/ethnicity, sex, and baseline levels of educational attainment. As the field focuses on designing research and evaluation that centers underserved communities, understanding how design parameters vary for key populations of interest is critical. Discussion: We conduct a simple hypothetical power analysis in which we set the minimum detectable effect to $500 in quarterly earnings and calculate the appropriate total sample for detecting that impact with high probability. This exercise demonstrates that a substantially larger sample is required to detect an impact of the same magnitude at later follow-up quarters. It also demonstrates that differences in input values for subgroups can imply substantial variation in required sample for subgroup analyses. We also compare our estimates of appropriate total sample to actual sample sizes from a recent meta-analysis of employment and training programs. Of 15 impact evaluations that used an experimental design and reported on earnings as an outcome measure, only 2 have a total sample size exceeding our estimate of appropriate total sample for detecting an impact of $500 per quarter. This suggest that researchers estimating the causal effects of employment and training programs should carefully consider both the program's anticipated effect size as well as the sample needed to adequately power their statistical analysis.