You can do it. Read then that all the assumptions are correct. Read SPSS survival manual (Pallant) then you know what needs to be correct.

Independent bivariate variables = female and male or urban and rural need not to be normally distributed.

In regression analysis it is more about continuous and not continuous variables and VIF and tolerance than normality. Your table of correlations is the important table in regression. The correlations between variable cannot be too high = .90

The normality assumption in linear regression concerns the residuals (error terms), not the independent or dependent variables.

There are plenty of examples of skewed data with ratio models/linear regression in the following, and multiple linear regression, even quadratic linear regression, regardless of data distributions:

"Application of Efficient Sampling with Prediction for Skewed Data," JSM 2022:

https://www.researchgate.net/publication/362370770_Application_of_Efficient_Sampling_with_Prediction_for_Skewed_Data

Knaub, J.R., Jr.(2021), "When Would Heteroscedasticity in Regression Occur?" Pak. J. Statist., Vol. 37(4), 315-367, https://www.pakjs.com/,

http://www.pakjs.com/wp-content/uploads/2021/07/37401-1.pdf,

https://www.researchgate.net/publication/354854317_WHEN_WOULD_HETEROSCEDASTICITY_IN_REGRESSION_OCCUR

And

https://www.researchgate.net/publication/319914742_Quasi-Cutoff_Sampling_and_the_Classical_Ratio_Estimator_-_Application_to_Establishment_Surveys_for_Official_Statistics_at_the_US_Energy_Information_Administration_-_Historical_Development

As to whether 84 is a large enough sample, that depends upon (1) have you covered different strata or subpopulations well enough, i.e., bias, and (2) is variance low enough for your needs.

To address the theoretical question, the answer is no. There is no finite sample size large enough for all cases. First, the central limit theorem doesn't apply in all cases (e.g., if your data have something like a Cauchy distribution). Second, the sample size to get away with violations of normality depends on the degree of violation - sufficiently extreme skew or kurtosis could always cause problems.

In practice you may well work with data where sample sizes of 30 or 100 or 1000 might suffice, but that depends on context.

There are two parts to the variance of the prediction error for a given predicted-yi, by the way. One part does not get smaller with increased sample size, but the other does. That is because the standard errors for the b-coefficients are made smaller with larger sample sizes.

However, as I noted you have to watch for bias in your sample selection.

As has been noted many times on ResearchGate, by many of us, normality of the data is not an issue. It is desirable for residuals.

I would add to some of the other replies that normality of the errors is a sufficient, but not a necessary condition--see slides 9-11 in the attached PDF. And thank goodness for that, at least if you believe what George Box famously said about normal distributions in the real world.